About this Author
DBL%20Hendrix%20small.png College chemistry, 1983

Derek Lowe The 2002 Model

Dbl%20new%20portrait%20B%26W.png After 10 years of blogging. . .

Derek Lowe, an Arkansan by birth, got his BA from Hendrix College and his PhD in organic chemistry from Duke before spending time in Germany on a Humboldt Fellowship on his post-doc. He's worked for several major pharmaceutical companies since 1989 on drug discovery projects against schizophrenia, Alzheimer's, diabetes, osteoporosis and other diseases. To contact Derek email him directly: Twitter: Dereklowe

Chemistry and Drug Data: Drugbank
Chempedia Lab
Synthetic Pages
Organic Chemistry Portal
Not Voodoo

Chemistry and Pharma Blogs:
Org Prep Daily
The Haystack
A New Merck, Reviewed
Liberal Arts Chemistry
Electron Pusher
All Things Metathesis
C&E News Blogs
Chemiotics II
Chemical Space
Noel O'Blog
In Vivo Blog
Terra Sigilatta
BBSRC/Douglas Kell
Realizations in Biostatistics
ChemSpider Blog
Organic Chem - Education & Industry
Pharma Strategy Blog
No Name No Slogan
Practical Fragments
The Curious Wavefunction
Natural Product Man
Fragment Literature
Chemistry World Blog
Synthetic Nature
Chemistry Blog
Synthesizing Ideas
Eye on FDA
Chemical Forums
Symyx Blog
Sceptical Chymist
Lamentations on Chemistry
Computational Organic Chemistry
Mining Drugs
Henry Rzepa

Science Blogs and News:
Bad Science
The Loom
Uncertain Principles
Fierce Biotech
Blogs for Industry
Omics! Omics!
Young Female Scientist
Notional Slurry
Nobel Intent
SciTech Daily
Science Blog
Gene Expression (I)
Gene Expression (II)
Adventures in Ethics and Science
Transterrestrial Musings
Slashdot Science
Cosmic Variance
Biology News Net

Medical Blogs
DB's Medical Rants
Science-Based Medicine
Respectful Insolence
Diabetes Mine

Economics and Business
Marginal Revolution
The Volokh Conspiracy
Knowledge Problem

Politics / Current Events
Virginia Postrel
Belmont Club
Mickey Kaus

Belles Lettres
Uncouth Reflections
Arts and Letters Daily

In the Pipeline

Category Archives

June 5, 2015

Artificial Intelligence For Biology?

Email This Entry

Posted by Derek

A new paper in PLoS Computational Biology is getting a lot of attention (which event, while not trying to be snarky about it, is not something that happens every day). Here's the press release, which I can guarantee that most of the articles written about this work will be based on. That's because the paper itself becomes heavy going after a bit - the authors (from Tufts) have applied machine learning to the various biochemical pathways involved in flatworm regeneration.

That in itself sounds somewhat interesting, but not likely to attract the attention of the newspapers. But here's the claim being made for it:

An artificial intelligence system has for the first time reverse-engineered the regeneration mechanism of planaria--the small worms whose extraordinary power to regrow body parts has made them a research model in human regenerative medicine.

The discovery by Tufts University biologists presents the first model of regeneration discovered by a non-human intelligence and the first comprehensive model of planarian regeneration, which had eluded human scientists for over 100 years.

The "100 years" part is hyperbole, because it's not like people have been doing a detailed mechanistic search for that amount of time. Biology wasn't up to the job, as the earlier biologists well knew. But is the artificial intelligence part hyperbole, or not? As the many enzymes and other proteins involved in planarians have been worked out, it has definitely been a challenge to figure out what's doing what to what else for which reasons, and when. (That's the shortest description of pathway elucidation that I can come up with!) The questions about this work are (1) is the model proposed correct (or at least plausibly correct)? (2) Was it truly worked out by a computational process? And (3) does this process rise to the level of "artificial intelligence"?

We'll take those in order. I'm actually willing to stipulate the first point, pending the planarian people. There are a lot of researchers in the regeneration field who will be able to render a more meaningful opinion than mine, and I'll wait for them to weigh in. I can look at the proposed pathways and say things like "Yeah, beta-catenin would probably have to be involved, damn thing is everywhere. . .yeah, don't see how you can leave Wnt out of it. . ." and other such useful comments, but that doesn't help us much.

What about the second point? What the authors have done is apply evolutionary algorithms to a modeled version of the various pathways involved, and let it rip, rearranging and tweaking the orders and relationships until it recapitulates the experimental data. It is interesting that this process didn't spit out a wooly Ptolemaic scheme full of epicycles and special pleading, but rather a reasonably streamlined account of what could be going on. The former is always what you have to guard against with machine-learning systems - overfitting. You can make any model work if you're willing to accept sufficient wheels within wheels, but at some point you have to wonder if you're optimizing towards reality.

How close is the proposed scheme to what people already might have been thinking (or might have already proposed themselves?) In other words, did we need a ghost come from the grave to tell us this? I am not up on the planarian stem-cell literature, but my impression is that this new model really is more comprehensive than anything that's been proposed before. It provides testable hypotheses. For example, it interprets the results of some experiments as inferring the existence of (yet unknown) regulatory molecules and genes. (The authors present candidates for two of these, and I would guess that experimental evidence in this area will be coming soon).

It's also important to note, as the authors do, that this model is not comprehensive. It only takes into account 2-D morphology, and has nothing to say about (for example) the arrangement of planarian internal organs. This, though, seems to be a matter of degree, only - if you're willing to collect more data, code it up, and run the model for longer after doing some more coding on it, its successor should presumably be able to deal with this sort of thing.

And that brings us to point three: is this a discovery made via artificial intelligence? Here we get into the sticky swamp of defining intelligence, there to recognize the artificial variety. The arguments here have not ceased, and probably won't cease until an AI hosts its own late-night talk show. Is the Siri software artificial intelligence? Are the directions you get from Google Maps? A search done through the chemical literature on SciFinder or the like? An earlier age would have probably answered "yes" (and an even earlier age would have fled in terror) but we've become more used to this sort of thing.

I think that one big problem in this area is that the word "intelligence" is often taken (consciously or not) to mean "human intelligence". That doesn't have to be true, but it does move the argument to whether border collies or African grey parrots demonstrate intelligence. (Personally, I think they do, just at a lower level and in different ways than humans). Is Google Maps as smart, in its own field, as a border collie? As a hamster? As a fire ant, or a planarian? Tough question, and part of the toughness is that we expect intelligence to be able to handle more than one particular problem. Ants are very good at what they do, but they seem to me clearly to be bundles of algorithms, and is a computer program any different, fundamentally? (Is a border collie merely a larger bundle of more complex algorithms? Are we? I will defer discussion of this disturbing question, because I see no way to answer it).

One of the hardest parts of the work in this current paper, I think, was the formalization step, where the existing phenomena from the experimental literature were coded into a computable framework. Now that took intelligence. Designing all the experiments (decades worth) that went into this hopper took quite a bit of it, too. Banging through it all, though, to come up with a model that fit the data, tweaking and prodding and adjusting and starting all over when it didn't work - which is what the evolutionary algorithms did - takes something else: inhuman patience and focus. That's what computers are really good at, relentless grinding. I can't call it intelligence, and I can call it artificial intelligence only in the sense that an inflatable palm is an artificial tree. I realize that we do have to call it something, though, but the term "artificial intelligence" probably confuses more than it illuminates.

Comments (30) + TrackBacks (0) | Category: Biological News | In Silico | Who Discovers and Why

April 15, 2015

Using the Same FEP Ruler

Email This Entry

Posted by Derek

With free energy perturbation having its time in the calculational spotlight, thanks to Schrödinger and others, it seems worthwhile to link to this new paper. It's a proposal for a common framework to analyze the results of such work. That's needed, because (as far as I can tell, as a definite outsider) every group seems to have its own idea of how to do that. This situation makes it difficult-to-impossible to compare various approaches, so even if this isn't the best possible set of benchmarking tools (I'm not qualified to say), just getting everyone to use the same ones would be a step up. (Thanks to Ash/Wavefunction on Twitter for pointing this one out).

Comments (14) + TrackBacks (0) | Category: In Silico

April 6, 2015

Levels of Data

Email This Entry

Posted by Derek

Here's a brief article in Science that a lot of us should keep a copy of. Plenty of journalists and investors should do the same. It's a summary of what sort of questions get asked of data sets, and the differences between them. There are six broad data analysis categories:

1. Descriptive. This is the simplest case, where you're just summarizing a data set and describing the totals in it.

2. Exploratory. The next step - you search through the descriptive analysis looking for trends or relationships, with which to develop new hypotheses. No guarantees, of course - you'll have to confirm these with more work.

3. Inferential. This one looks at an exploratory treatment and tried to determine whether those trends are likely to hold up. As the authors say, this is probably the most common statistical workup in the literature - better than randome chance, or not? But it can't tell you why something is happening, of course.

4. Predictive. An inferential study is necessarily done on a large sample (well, it had better be, at any rate, if you're going to infer with much confidence). A predictive analysis uses some subset of the data to predict how individual cases will go. The example from drug development would be the use of biomarkers to predict whether a given patient in a trial will respond to some new investigational drug.

5. Causal. At this level, you're trying to see what the magnitude of changes are across the system when you start changing things - what often gets called the "tone" of the system. What are the most important variables, and what has little effect on the outcome?

6. Mechanistic. With the information at the causal level available, now you can really get down to the nuts and bolts. Change A causes effect B, through this detailed mechanism. We don't see this as much with anything involving biology - there always seem to be exceptions. This is more the realm of engineering and physics, although a lot of time and money is going into trying to change that.

It's only at the causal and mechanistic levels that you can start doing detailed modeling with confidence. That's where everyone would like to be with computational binding predictions, but we don't understand them well enough yet. And think how far we have to go to get predictive toxicology to those levels! We can do that sort of thing on a small scale - for example, saying that a compound that (say) inhibits angiotensin-converting enzyme, to this degree, and with that average half-life in vivo, will be expected to lower X% of a random population's members blood pressure by at least Y%. That's after decades of experience and data-gathering, keep in mind.

But that's not aeronautical engineering. Those folks don't tell you that wing design A will provide at least so much lift on a certain percentage of the airframes it gets bolted on to. Nope, those folks get to build their airframes to the same exact specifications, not just take whatever shows up at the factory needing wings, and those airframe/wing combinations had better perform within some very tight tolerances or something has gone seriously wrong. This is just another way of stating the "built by humans" difference I was talking about the other day.

So some of that data analysis hierarchy above is, well, aspirational for those of us doing drug research. The authors of the Science article are well aware of this themselves, saying that "Outside of engineering, mechanistic data analysis is extremely challenging and rarely achievable.". But that level is where many people expect science to be, most of the time, which leads to a lot of frustration: "Look, is this pill going to help me or not?" We should remember where we are on the scale and try to work our way up.

Comments (12) + TrackBacks (0) | Category: Clinical Trials | Drug Development | General Scientific News | In Silico

April 3, 2015

Sanofi Bets on Schrödinger

Email This Entry

Posted by Derek

Sanofi has signed an expanded deal with Schrödinger, the computational chemistry folks. Here's something from the press release:

Schrödinger has made a number of key scientific breakthroughs in recent years in the areas of protein and ligand structure determination and potency prediction that promise to have a transformative impact on the discovery of drugs. The collaboration with Sanofi aims to deploy this and related technologies at a level that is unprecedented in the pharmaceutical industry.

I'd guess that this is going to involve the FEP calcuations that they've been talking about (blogged here). Schrödinger is also very strong in doing molecular dynamics simulations (for similar reasons), although, as with everything in this field, there's room to argue about what that can do for you. So this will be very interesting to watch. I'm glad that Schrödinger's technology is being given such a thorough real-world test, because that's the only way to see what it can do.

Comments (37) + TrackBacks (0) | Category: In Silico

March 24, 2015

A Couple of Ycombinator's Startups

Email This Entry

Posted by Derek

Last year I mentioned reports that the startup incubator Ycombinator was thinking of getting into the biopharma field. Here's a look at the current crop of potential companies.

One thing that stands out is that most of these seem to be focused on patient care or some sort of diagnostic. One exception is 20n, which is looking to engineer microbes to produce known pharmaceuticals or intermediates. That's not at all a crazy idea, but the example given on the site (acetaminophen) is not a particularly compelling example by itself, since it's extremely easy to make, from cheap precursors, on an industrial scale. And I'm not sure what to make of that "map of every chemical that can be made biologically". It's a nice graphic for the middle of the page, but there's no telling what it means. I like a lot of the ideas kicking around in the synthetic biology field, but I can't really say what 20n is up to yet.

The other one on the list I noticed is Atomwise, and I'll let them speak for themselves:

Medicines are getting more expensive to develop. These days, it takes about $1.8 billion and 15 years for a single new drug. Atomwise aims to change that by using supercomputers to predict, in advance, which potential medicines will work, and which won't. Our tools can tell the difference between great drug candidates and toxic ones, and discover new uses for old medicines.

Actually, what will speak for themselves are the results. If Atomwise can do this, at all, even poorly, then there are billions of dollars waiting out there for them to scoop up. But just the act of saying that you can do things like this makes me suspicious that they really can't do things like this. Here's a bit more:

Previous attempts haven't always met expectations. The techniques of the day were limited by the knowledge and computers available. Today, things are different. We have invented cutting-edge machine learning algorithms that are built specifically for the world's most powerful computers. We use one of the world's top supercomputers to analyze databases 1000 times larger than those used in the past. This lets us deliver what many others can't: precise and reliable medicinal predictions.

I could go on about this for a while, but in the end, these arguments are settled by data. Come on down and try it, guys. There's plenty of room, and plenty to work on, so let's see what you can get done. I'll be watching with interest, and so will others.

Update, from the comments: "Hello Everyone, I'm the CEO of Atomwise and a long-time semi-lurker here. (I'm the anonymous who keeps asking what it would take to convince people that in silico methods work.) I agree that the proof will come from data; that's one reason why we're doing a large-scale evaluation with Merck. Send me an email ( and I'd be happy to present our data from previous prospective validation projects to you. Or, if you have some minimally proprietary data against which you'd like to evaluate our predictive capability, let's run the experiment. Best, Abraham Heifets"

Comments (33) + TrackBacks (0) | Category: Biological News | Business and Markets | In Silico

March 6, 2015

Not Even Wrong

Email This Entry

Posted by Derek

This paper is not going to make a lot of computational chemists very happy at all. It's from Dan Singleton and Erik Plata at Texas A&M, and it's on the Morita-Bayliss-Hillman reaction. More specifically, though, it's on the many computational attempts to decide on the mechanism of the MBH reaction, and taken together, they're not a pretty sight. The authors do some good old physical organic chemistry to help establish the real mechanism (which had already been proposed some years ago), and let's just say that things don't always match up very well.

Computational methods are simply scientific models. Any model makes some inaccurate predictions but models can retain utility despite significant propensities for inaccuracy. Inaccurate predictions aid the choice of models for future predictions. Because of this, the central scientific problem in the computational study of the MBH mechanism is not the inaccuracy of the predictions. Rather, it is the absence of any particular prediction at all. Fully-defined computational methods (including the choice of basis set, entropy calculation, and solvent model) of course make quite specific predictions. However, there is neither a consensus best-choice method nor a common view on the right way to choose a method. When evaluable, the most accurate choice varies with the system at hand. In the MBH reaction, defensible and expectantly publishable choices of computational approaches lead to predictions of the facility of the proton-shuttle process that vary by 35 orders of magnitude in the stability of 19, while also diverging in the geometry and preferred stereochemistry of transition state 13. This variance is in practical terms indistinguishable from making no prediction. In addition, studies of the MBH mechanism have not been considered falsified by extreme inaccuracies in predictions. In the terminology of Pauli, computational mechanistic chemistry is “not even wrong” about the MBH mechanism.

Here's a C&E News article if you don't have access to JACS. It's true that predicting reaction mechanisms is a challenge for computational methods, because you are, out of necessity, looking at high-energy molecular states and trying to distinguish between them. It's especially tough with a polar reaction mechanism, because solvation effects (which we still don't have as good a handle on as we need) become very important in stabilizing transition states, assisting proton transfers, and so on. But at the same time, this sort of problem is just the sort of thing that many such groups work on: the MBH mechanism has been the subject of 11 separate computational papers.

The authors here try to figure out what has gone wrong. The errors mostly seem to be in the enthalpy term, which would suggest trouble with those polar interactions. A good number of the earlier studies predicted a proton-shuttle mechanism, which turns out not to be operating at all. The problem is that current programs have a much easier time handling proton-shuttle mechanisms, while full-scale proton transfer to and from a solvent molecule is much harder to model. So there's a constant danger of arriving at a mechanism because it's computationally tractable, not because it's real. Digging into the individual equilibria, it appears that some approaches did very well on particular reaction steps, but blew up completely on others: 14, 20, or 35 orders of magnitude off for the equilibrium constants, I would say, is enough to warrant that description. And it's very hard to see what factors led to the failures or successes - in fact, it's quite possible that some of the best individual predictions were themselves fortuitous. Overall, though, no computational approach got things anywhere near correct.

The problems in the computational study of mechanisms encountered in the MBH reaction certainly cannot be used to paint all computational mechanistic studies. Many, either by simplicity or carefully designed use of the computations, would not be susceptible to the difficulties encountered here. At least, however, it would seem that studies of complex multimolecular polar reactions in solution should be undertaken and interpreted only with extreme care.

That's for sure. And while this is a harder problem, in many ways, than docking a ligand into a protein, we should keep in mind that polar interactions and the treatment of solvation are very important parts of those calculations, too, and looking under this particular hood tells us that we have a long way to go on those.

Comments (26) + TrackBacks (0) | Category: Chemical News | In Silico

March 5, 2015

Diversity and Similarity Scoring: Does One Size Ever Fit All?

Email This Entry

Posted by Derek

We spend a lot of time talking about compound similarities in this business. All those schemes for virtual screening, to find new activities for old compounds, to predict side effects and general toxicity, and many others besides rely on some sort of measurement of how similar various compounds (and collections of compounds) are to each other.

But how do you determine similarities? Some might answer "By comparing Tanimoto coefficients, of course", but that's an example of a little knowledge being a dangerous thing. Tanimoto distance calculations are a meat grinder, which will grind whatever you shove into the hopper. How, then, are you characterizing the compounds themselves? That's where things get tricky. There are all sorts of "fingerprint" descriptors for molecules, and different ones will give you different measures of similarity (or of chemical diversity, depending on which end of the tube you're looking through).

And the problem is, all of them have a tendency to look funny. Many medicinal chemists have experienced this, looking over a list of compounds ranked by similarities. You come across one with a high score, but it doesn't look that similar to you, because the algorithm liked what look to you like unimportant details. The next list has two compounds that are supposed to be quite different, but they're a methyl ester versus a t-butyl ester, or something of the sort, and just how different is that when the rest of the molecule is the same? And these are just two-dimensional comparisons. If you want to talk similarity in conformational space, and our drug targets generally want to talk that way, then you're in for an even bigger universe of choices and tradeoffs. (Here's a good recent overview from J. Med. Chem.)

The Matsy algorithm is supposed to generate results that look a little less alien, but I haven't used it myself. I'd be interested in hearing what people have found to be the most useful in their own hands for such measurements. Is this always a case-by-case thing, or are there methods that have enough of a Swiss-army-knife character to them to stick with? Any favorites out there?

Comments (4) + TrackBacks (0) | Category: In Silico

March 4, 2015

Neural Networks for Drug Discovery: A Work in Progress

Email This Entry

Posted by Derek

There have been many attempts over the years to bring together large amounts of biological and drug activity data, winnow them computationally, and come up with insights that would not be obvious to human eyes. It's a natural thing to imagine - there are so many drug targets in the body, doing so many things, and there's an awful lot of information out there about thousands and thousands of compounds. There's no way that a human observer could pick up on all the things that are going on; you need tireless software to sift through the piles.

The success record of this sort of thing has been mixed, though. Early attempts can now be set aside as underpowered, but what to make of current attempts at "virtual clinical trials" and the like? (We're probably still underpowered for that sort of thing). Less ambitiously, people have tried to mine for new targets and new drug activities by rooting through the Big Data. But this sort of thing is not without controversy: many of us, chemists and biologists alike, don't have the mathematical background to say if the methods being used are appropriate, or what their weaknesses and blind spots might be.

A new paper has gotten me thinking about all this again. It's a collaboration between several researchers at Stanford and Google (press release here) on machine learning for drug discovery. What that means is that they're trying to improve virtual screening techniques, using a very Google-centric approach that might be summed up as "MOAR DATA!" (That phrase does not appear in the paper, sadly).

In collaboration with the Pande Lab at Stanford University, we’ve released a paper titled "Massively Multitask Networks for Drug Discovery", investigating how data from a variety of sources can be used to improve the accuracy of determining which chemical compounds would be effective drug treatments for a variety of diseases. In particular, we carefully quantified how the amount and diversity of screening data from a variety of diseases with very different biological processes can be used to improve the virtual drug screening predictions.

Using our large-scale neural network training system, we trained at a scale 18x larger than previous work with a total of 37.8M data points across more than 200 distinct biological processes. Because of our large scale, we were able to carefully probe the sensitivity of these models to a variety of changes in model structure and input data. In the paper, we examine not just the performance of the model but why it performs well and what we can expect for similar models in the future. The data in the paper represents more than 50M total CPU hours.

I end up with several trains of thought about this kind of thing. On track one, I appreciate that if virtual screening is going to work well, it needs to draw from the largest data sets possible, since there are so many factors at work. But on track two, I wonder how good the numbers going into this hopper really are, since I (like anyone else in the business) have seen some pretty garbagey screening numbers, both in person and in the literature. Piling more noise into the computations cannot improve them, even if your hardware is capable of dealing with landfills of the stuff. (The authors do note that they didn't do any preprocessing of the data sets to remove potential artifacts. The data come from four main sources (see the paper, which is open access, for more), and only one of these has probably been curated to that level.) And that brings us to track three: my innate (and doubtless somewhat unfair) suspicions go up when I see a lot of talk about just how Incredibly Large the data sets are, and how Wildly Intense all the computations were.

Not to be too subtle about it, asking for a virtual screen against some target is like asking for a ditch to be dug from Point A to Point B. Can you dig the ditch, or not? Does it get to where it's supposed to go, and do what a ditch is supposed to do? If so, then to a good approximation, I don't care how many trained badgers you herded in for the job, or (alternatively) about the horsepower and fuel requirements of the earth-moving equipment you rented. If someone spends a lot of time telling me about these things (those engines! those badgers!) then I wonder if they're trying to distract me from what really matters to me, which is the final product.
Well, I'm willing to accept that that's not a completely fair criticism, but it's something that always crosses my mind, and I may not be alone in this. Let's take a look at the ditch - uh, the virtual screening - and see how well it came out.

In this work, we investigate several aspects of the multitask learning paradigm as applied to virtual screening. We gather a large collection of datasets containing nearly 40 million experimental measurements for over 200 targets. We demonstrate that multitask networks trained on this collection achieve significant improvements over baseline machine learning methods. We show that adding more tasks and more data yields better performance. This effect diminishes as more data and tasks are added, but does not appear to plateau within our collection. Interestingly, we find that the total amount of data and the total number of tasks both have significant roles in this improvement. Furthermore, the features extracted by the multitask networks demonstrate some transferability to tasks not contained in the training set. Finally, we find that the presence of shared active compounds is moderately correlated with multitask improvement, but the biological class of the target is not.

As the paper notes, this is similar to Merck's Kaggle challenge of a couple of years back (and I just noticed this morning that they cite that blog post, and its comments, as an example of the skepticism that it attracted from some parts of the med-chem community). In this case, the object isn't (yet) to deliver up a bunch of virtual screening hits, so much as it is to see what the most appropriate architecture for such a search might be.

One of the biggest problems with these papers (as this one explicitly states) is that the criteria used to evaluate the performance of these systems are not standardized. So it's basically impossible to compare one analysis with another, because they're scoring by different systems. But that graphic gives some idea of how things worked on different target classes. The Y axis is the difference between using multitask models (as in this paper) and single-task neural network models, and it shows that in most cases, most of the time, multitask modeling was better. But I note that almost every class has some cases where this doesn't hold, and that (for reasons unknown) the GPCR targets seem to show the least improvement.

But what I don't know is how well these virtual screening techniques compared to the actual screening data. The comparisons in the paper are all multi-task versus single-task (which, to the fair, is the whole focus of the work), but I'd be interested in an absolute-scale measurement. That shows up, though, in Table B2 in the appendix, where they use Jain and Nicholls' "enrichment" calculation. Assuming that I'm reading these correctly, which may or may not be warranted, the predictions look to be anywhere from about 5% to about 25% better than random, depending on what false-positive rate you're looking at, with occasional hops up to the 40% better range. Looking at the enrichment figures, though, I don't see this model performing much better than the Random Forest method, which has already been applied to med-chem work and activity prediction many times. Am I missing something in that comparison? Or does this all have quite a ways to go yet?

Comments (25) + TrackBacks (0) | Category: In Silico | In Silico

Neural Networks for Drug Discovery: A Work in Progress

Email This Entry

Posted by Derek

There have been many attempts over the years to bring together large amounts of biological and drug activity data, winnow them computationally, and come up with insights that would not be obvious to human eyes. It's a natural thing to imagine - there are so many drug targets in the body, doing so many things, and there's an awful lot of information out there about thousands and thousands of compounds. There's no way that a human observer could pick up on all the things that are going on; you need tireless software to sift through the piles.

The success record of this sort of thing has been mixed, though. Early attempts can now be set aside as underpowered, but what to make of current attempts at "virtual clinical trials" and the like? (We're probably still underpowered for that sort of thing). Less ambitiously, people have tried to mine for new targets and new drug activities by rooting through the Big Data. But this sort of thing is not without controversy: many of us, chemists and biologists alike, don't have the mathematical background to say if the methods being used are appropriate, or what their weaknesses and blind spots might be.

A new paper has gotten me thinking about all this again. It's a collaboration between several researchers at Stanford and Google (press release here) on machine learning for drug discovery. What that means is that they're trying to improve virtual screening techniques, using a very Google-centric approach that might be summed up as "MOAR DATA!" (That phrase does not appear in the paper, sadly).

In collaboration with the Pande Lab at Stanford University, we’ve released a paper titled "Massively Multitask Networks for Drug Discovery", investigating how data from a variety of sources can be used to improve the accuracy of determining which chemical compounds would be effective drug treatments for a variety of diseases. In particular, we carefully quantified how the amount and diversity of screening data from a variety of diseases with very different biological processes can be used to improve the virtual drug screening predictions.

Using our large-scale neural network training system, we trained at a scale 18x larger than previous work with a total of 37.8M data points across more than 200 distinct biological processes. Because of our large scale, we were able to carefully probe the sensitivity of these models to a variety of changes in model structure and input data. In the paper, we examine not just the performance of the model but why it performs well and what we can expect for similar models in the future. The data in the paper represents more than 50M total CPU hours.

I end up with several trains of thought about this kind of thing. On track one, I appreciate that if virtual screening is going to work well, it needs to draw from the largest data sets possible, since there are so many factors at work. But on track two, I wonder how good the numbers going into this hopper really are, since I (like anyone else in the business) have seen some pretty garbagey screening numbers, both in person and in the literature. Piling more noise into the computations cannot improve them, even if your hardware is capable of dealing with landfills of the stuff. (The authors do note that they didn't do any preprocessing of the data sets to remove potential artifacts. The data come from four main sources (see the paper, which is open access, for more), and only one of these has probably been curated to that level.) And that brings us to track three: my innate (and doubtless somewhat unfair) suspicions go up when I see a lot of talk about just how Incredibly Large the data sets are, and how Wildly Intense all the computations were.

Not to be too subtle about it, asking for a virtual screen against some target is like asking for a ditch to be dug from Point A to Point B. Can you dig the ditch, or not? Does it get to where it's supposed to go, and do what a ditch is supposed to do? If so, then to a good approximation, I don't care how many trained badgers you herded in for the job, or (alternatively) about the horsepower and fuel requirements of the earth-moving equipment you rented. If someone spends a lot of time telling me about these things (those engines! those badgers!) then I wonder if they're trying to distract me from what really matters to me, which is the final product.
Well, I'm willing to accept that that's not a completely fair criticism, but it's something that always crosses my mind, and I may not be alone in this. Let's take a look at the ditch - uh, the virtual screening - and see how well it came out.

In this work, we investigate several aspects of the multitask learning paradigm as applied to virtual screening. We gather a large collection of datasets containing nearly 40 million experimental measurements for over 200 targets. We demonstrate that multitask networks trained on this collection achieve significant improvements over baseline machine learning methods. We show that adding more tasks and more data yields better performance. This effect diminishes as more data and tasks are added, but does not appear to plateau within our collection. Interestingly, we find that the total amount of data and the total number of tasks both have significant roles in this improvement. Furthermore, the features extracted by the multitask networks demonstrate some transferability to tasks not contained in the training set. Finally, we find that the presence of shared active compounds is moderately correlated with multitask improvement, but the biological class of the target is not.

As the paper notes, this is similar to Merck's Kaggle challenge of a couple of years back (and I just noticed this morning that they cite that blog post, and its comments, as an example of the skepticism that it attracted from some parts of the med-chem community). In this case, the object isn't (yet) to deliver up a bunch of virtual screening hits, so much as it is to see what the most appropriate architecture for such a search might be.

One of the biggest problems with these papers (as this one explicitly states) is that the criteria used to evaluate the performance of these systems are not standardized. So it's basically impossible to compare one analysis with another, because they're scoring by different systems. But that graphic gives some idea of how things worked on different target classes. The Y axis is the difference between using multitask models (as in this paper) and single-task neural network models, and it shows that in most cases, most of the time, multitask modeling was better. But I note that almost every class has some cases where this doesn't hold, and that (for reasons unknown) the GPCR targets seem to show the least improvement.

But what I don't know is how well these virtual screening techniques compared to the actual screening data. The comparisons in the paper are all multi-task versus single-task (which, to the fair, is the whole focus of the work), but I'd be interested in an absolute-scale measurement. That shows up, though, in Table B2 in the appendix, where they use Jain and Nicholls' "enrichment" calculation. Assuming that I'm reading these correctly, which may or may not be warranted, the predictions look to be anywhere from about 5% to about 25% better than random, depending on what false-positive rate you're looking at, with occasional hops up to the 40% better range. Looking at the enrichment figures, though, I don't see this model performing much better than the Random Forest method, which has already been applied to med-chem work and activity prediction many times. Am I missing something in that comparison? Or does this all have quite a ways to go yet?

Comments (25) + TrackBacks (0) | Category: In Silico | In Silico

February 23, 2015

Is FEP Ready For the World?

Email This Entry

Posted by Derek

Here's a paper that basically throws down the computational gauntlet. A large group of authors from Schrödinger, Nimbus, Columbia, Yale, and UC-Irvine say that their implementation of free energy perturbation (FEP) calculations really does lead to a significant number of more active compounds being predicted. That's as compared to other computational methods, or to straight med-chem intuition and synthesis.

Here, we report an FEP protocol that enables highly accurate affinity predictions across a broad range of ligands and target classes (over 200 ligands and 10 targets). The ligand perturbations include a wide range of chemical modifications that are typically seen in medicinal chemistry efforts, with modifications of up to 10 heavy atoms routinely included. Critically, we have applied the method in eight prospective discovery projects to date, with the results from two of those projects disclosed in this work. The high level of accuracy obtained in the prospective studies demonstrates the ability of this approach to drive decisions in lead optimization.

They say that these improvements are due to a better force field, better sampling algorithms, increased computing power, and automated work flow to get through things in an organized fashion. The paper shows some results against BACE, CDK2, JNK1, MCL1, p38, PTP1b, and thrombin, which seems like a reasonably diverse real-world set of targets. Checking the predicted binding energies versus experiment, most of them are within 1 kcal/mol, and only about 5% are 2 kcal/mol or worse. (To put these into med-chem terms, the rule of thumb is that a 10x difference in Ki represents 1.36 kcal/mol). These calculation should, in theory, be capturing the lot: hydrogen bonding, hydrophobic interaction, displacement of bound waters, pi-pi interactions, what have you. The two prospective projects mentioned are IRAK4 and TYK2. In both of these, the average error between theory and experiment was about 1 kcal/mol.

But this is not yet the Rise of the Machines:

The preceding notwithstanding, a highly accurate and robust FEP methodology is not, in any way, a replacement for a creative and technically strong medicinal chemistry team; it is necessary to generate the ideas for optimization of the lead compound that are synthetically tractable and have acceptable values for a wide range of druglike properties (e.g., solubility, membrane permeability, metabolism, etc.). Rather, the computational approach described here can be viewed as a tool to enable medicinal chemists to pursue modifications and new synthetic directions that would have been considered too risky without computational validation or to eliminate compounds that would be unlikely to meet the desired target affinity. This is particularly significant when considering whether to make an otherwise highly attractive molecule that may be synthetically challenging. If such a molecule is predicted to achieve the project potency targets by reliable FEP calculations, this substantially reduces the risk of taking on such synthetic challenges.

There's no reason, a priori why this shouldn't work; it's all down to limits in how well the algorithms at the heart of the process deal with the energies involved, and how much computing force can be thrown at the problem. To that point, these calculations were done by running on graphics processing units (GPUs), which really do have a lot more oomph for the buck (although it's still not just as trivial as plugging in some graphics processor cards and standing back). GPUs are getting more capable all the time themselves, and are a mass-market item, which bodes well for their application in drug design. Have we reached the tipping point here? Or is this another in a very long series of false dawns? I look forward to seeing how this plays out.

Comments (59) + TrackBacks (0) | Category: In Silico

February 4, 2015

The Old Binding Mode Switcheroo

Email This Entry

Posted by Derek

When a medicinal chemist gets a promising lead compound, the first thing that's done to it is to start making modifications to the structure. Add methyls or fluorines, take some of the substituents back off, switch stereochemistry, switch carbons and nitrogens around, that sort of thing. And there you start to build up a structure-activity relationship.

Maybe. Sometimes, though, the SAR doesn't seem to make a whole lot of sense. And in these cases, it may be that the underlying assumption of SAR is the problem. That is, we assume that the starting scaffold binds in pretty much the same way no matter what we do to it. There's a single binding mode, and some things make it better, and some things make it worse. But that's not always true. Fragment-based drug discovery illustrates that quite often, with minor changes to fragment-sized compounds leading to new (and often rather hard to predict) orientations in the binding site.

This new paper in Angewandte Chemie is a great illustration of just how weird things can get. The authors, from Marburg, were investigating a series o endothiapepsin inhibitors, with variations as shown at left. But when they did that ester-amide switch, X-ray crystallography showed a new binding mode. In fact, out of 8 compounds, they had four different binding, with the starting material the only singleton in the bunch. More potent compounds could be found in more than one series; it's not that any one particular mode was noticeably better.

The current example highlights the complexity of binding events and their strong dependence on seemingly minor effects of scaffold decoration and modifications. That 1, 7, 8, and 9 do not adopt the same binding geometry as observed in the green cluster can be rationalized retrospectively, but without the crystal structure determinations the adopted binding modes would have been difficult to predict, and, if so, would have hardly been believed or attracted sustained attention without our experimental evidence

The odds of getting so many different modes in so few compounds are not high, but they sure aren't zero, either. That's the question: just how often does this sort of thing go on, but pass unnoticed, written off as "just one of those things"? You can't always get X-ray structures of everything, but the authors here note that some discrepancies between affinity in the assay and those found by a thermal-shift assay seemed (in retrospect) to indicate some of the binding mode changes. So that might be a way to spot them. Not everyone will believe you if you keep using this an excuse, but it could be more valid than people realize. . .

Comments (10) + TrackBacks (0) | Category: Drug Assays | In Silico

January 13, 2015

Odd Structures, Subjected to Powerful Computations

Email This Entry

Posted by Derek

Here's a paper in Nature Chemistry on computational simulation of GPCR activation, using the beta-2 receptor as a model. I'm writing as someone who's worked on GPCRs, who is interested in such mechanisms, but who is not a computational chemist. And as such, I have some real reservations about the paper. Are my misgivings well-founded or not?

What this team at Stanford has done is a massive amount of molecular dynamics work, attempting to capture details of the the conformational changes that have to be taking place during receptor activation. They present plots illustrating the movements of key residues versus time for agonists, antagonists, and inverse agonists, at a very fine level of detail. I am not competent to judge the fitness of their MD software, their use of Markov state models, the methods by which they reduce 3,000 of those MSMs down to a ten-state model of the receptor dynamics, and so on. I'd be glad to hear from people who are - for now, I'll assume that all this has been done to a high standard.

But the paper does get into some areas that I feel able to question. Fellow chem-blogger Wavefunction's Twitter account first alerted me to the feature that I find most disturbing. Take a look at these structures. They're supposed to be from the GPCR Ligand and Decoy Database, published here, which this paper used as a source of more agonists, antagonists, and comparison compounds. But when I search the beta-2 files from that server (both agonists and antagonists, ligands and decoys), I can't find any compounds with the left-hand sides of compounds 2 or 3. They're not in the ZINC database, either, as far as I can tell. (A minor point is that the current paper refers to them as catecholamines, but that's not right, either: a catechol is a dihydroxybenzene, and there's an extra methylene in these. Such structures can be found in beta-adrenergic ligands, but they're not catecholamines).

Then there's the more serious matter of that hemiaminal. That's not the usual pharmacophore for an adrenergic receptor, which has amine and OH on adjacent carbons, and it's not even really a stable group under most conditions. That was the first thing that struck me when I saw the structures - how are these things GPCR ligands, when I wouldn't even be sure that they're stable in buffer?
So I don't know where these compounds come from, or if some mistake has been made along the way, but that's how it looks to me after a bit of digging around. I took a look through the literature for structures like these, and I did find some in an old Theravance patent, WO03042164 (see compounds 16 through 26). Looking over the patent procedures, though, I think that those structures are a mistake. The claimed chemical matter (and the synthetic procedures used in the rest of the patent) are all directed to making traditional beta-hydroxy-amino structures. The patent has a run of these hemiaminal things, supposedly made by the same coupling procedures that make the real beta-receptor ligands in the rest of the patent, and that's not going to work. My guess at the moment is that some beta-receptor ligand database has been corrupted by inclusion of these structures, which may well have propagated from this application or others in the series. At any rate, I don't see, at the moment, how these things are beta-receptor agonists, and I would have to say that running molecular dynamics simulations on them is not the best use of computing cycles.

And that brings up one last problem I have with this paper, which may be a minor one (or maybe not). The title is "Cloud-based simulations on Google Exacycle reveal ligand modulation of GPCR activation pathways". I can't help but thinking of something Bill James (the baseball statistics guy) wrote back in the 1980s. He was doing one of his exercises to try to predict performance, and said that he'd done a "computer projection". But then he backed up, and said that he figured that this phrase was probably going to disappear from use in the coming years, because the only way to do these things was with a computer, and you didn't say that you'd done a "pencil projection" or something. The tool would become so common that it would disappear into the background. To a large extent, that's just what's happened - but it was a sufficiently novel thought back then that I found it striking.

So I have to wonder, perhaps unfairly, if the "Google Exacycle" part is there to bring in some more attention. It's true that the cloud-computing aspect of this work did allow the authors to do a lot more than the usual MD simulation, and it may well be a difference of kind rather than just a difference of degree (again, I'd be glad to hear from computational folks on this point). But it can't help sounding cutting-edge, can it?

Update: Both Stuart Cantrill of Nature Chemistry and lead author Vijay Pande have shown up in the comments section, and I appreciate both of them coming by. These issues are being looked at - more later as details become clear.

Second Update: Prof. Pande says in the comments that this is indeed a drawing error. The correct structures are more reasonable beta-receptor ligands, and those are the ones that were docked. The paper is being corrected.

Comments (44) + TrackBacks (0) | Category: In Silico

November 21, 2014

Virtual Covalent Screening

Email This Entry

Posted by Derek

Covalent drugs have been a big item in R&D over the last few years, and I wrote here about covalent fragments. The whole topic of reactive groups in small molecules and their interaction with living systems and biomolecules is a complicated one, with many interesting twists.

Now the Shoichet group at UCSF has what could be a useful computational approach to the field. They're reporting "DOCKovalent", a virtual screening platform for covalent inhibitors, and illustrate it with examples of cyanoacrylamides and boronic acids across several different enzymes. The calculations are based on the angles needed for an electrophile to react with a residue like cysteine - these reactions, as organic chemists know, can be rather constrained in what approaches the reacting partners have to take. In solution you can see stereoelectronic effects that arise from the structures of the small molecules, but most reactions can find a way. When this process is taking place in the clefts of a protein, though, the number of feasible approaches can get cut down considerably. I'm generally pretty hard to convince when it comes to virtual screening, but the number of constraints needed here gives me more hope than usual for meaningful results.

Running their program retrospectively, with known covalent inhibitors of various enzymes versus decoy molecules, showed that the virtual screening did (in most cases) give hit sets that were well enriched in the real binders. Like any such technique, it's going to fail sometimes, both on individual compounds and on whole runs, but the Shoichet group has made the tool available on the web for anyone to try for free. Here's the link, and I applaud them for putting it out there. The appeal of computational approaches has always been the low barrier to entry, and this lowers it even more. That's the hazard of computational approaches, too, of course, but the results from a virtual screen like this should at least be easy to subject to a real-world test.

Comments (15) + TrackBacks (0) | Category: In Silico

October 20, 2014

Compound Properties: Starting a Renunciation

Email This Entry

Posted by Derek

I've been thinking a lot recently about compound properties, and what we use them for. My own opinions on this subject have been changing over the years, and I'm interested to see if I have any company on this.

First off, why do we measure things like cLogP, polar surface area, aromatic ring count, and all the others? A quick (and not totally inaccurate) answer is "because we can", but what are we trying to accomplish? Well, we're trying to read the future a bit and decrease the horrendous failure rates for drug candidates, of course. And the two aspects that compound properties are supposed to help with are PK and tox.

Of the two, pharmacokinetics is the one with the better shot at relevance. But how fine-grained can we be with our measurements? I don't think it's controversial to say that compounds with really high cLogP values are going to have, on average, more difficult PK, for various reasons. Compounds with lots of aromatic rings in them are, on average, going to have more difficult PK, too. But how much is "lots" or "really high"? That's the problem, because I don't think that you can draw a useful line and say that things on one side of it are mostly fine, and things on the other are mostly not. There's too much overlap, and too many exceptions. The best you can hope for, if you're into line-drawing, is to draw one up pretty far into the possible range and say that things below it may or may not be OK, but things above it have a greater chance of being bad. (This, to my mind, is all that we mean by all the "Rule of 5" stuff). But what good does that do? Everyone doing drug discovery already knows that much, or should. Where we get into trouble is when we treat these lines as if they were made of electrified barbed wire.

That's because of a larger problem with metrics aimed at PK: PK is relatively easy data to get. When in doubt, you should just dose the compound and find out. This makes predicting PK problems a lower-value proposition - the real killer application would be predicting toxicology problems. I fear that over the years many rule-of-five zealots have confused these two fields, out of a natural hope that something can be done about the latter (or perhaps out of thinking that the two are more related than they really are). That's unfortunate, because to my mind, this is where compound property metrics get even less useful. That recent AstraZeneca paper has had me thinking, the one where they state that they can't reproduce the trends reported by Pfizer's group on the influences of compound properties. If you really can take two reasonably-sized sets of drug discovery data and come to opposite conclusions about this issue, what hope does this approach have?

Toxicology is just too complicated, I think, for us to expect that any simple property metrics can tell us enough to be useful. That's really annoying, because we could all really use something like that. But increasingly, I think we're still on our own, where we've always been, and that we're just trying to make ourselves feel better when we think otherwise. That problem is particularly acute as you go up the management ladder. Avoiding painful tox-driven failures is such a desirable goal that people are tempted to reach for just about anything reasonable-sounding that holds out hope for it. And this one (compound property space policing) has many other tempting advantages - it's cheap to implement, easy to measure, and produces piles of numbers that make for data-rich presentations. Even the managers who don't really know much chemistry can grasp the ideas behind it. How can it not be a good thing?

Especially when the alternative is so, so. . .empirical. So case-by-case. So disappointingly back-to-where-we-started. I mean, getting up in front of the higher-ups and telling them that no, we're not doing ourselves much good by whacking people about aromatic ring counts and nitrogen atom counts and PSA counts, etc., that we're just going to have to take the compounds forward and wait and see like we always have. . .that doesn't sound like much fun, does it? This isn't what anyone is wanting to hear. You're going to do a lot better if you can tell people that you've Identified The Problem, and How to Address It, and that this strategy is being implemented right now, and here are the numbers to prove it. Saying, in effect, that we can't do anything about it runs the risk of finding yourself replaced by someone who will say that we can.

But all that said, I really am losing faith in property-space metrics as a way to address toxicology. The only thing I'm holding on to are some of the structure-based criteria. I really do, for example, think that quinones are bad news. I think if you advance a hundred quinones into the clinic, that a far higher percentage of them will fail due to tox and side effects than a hundred broadly similar non-quinones. Same goes for rhodanines, and a few other classes, those "aces in the PAINS deck" I referred to the other day. I'm still less doctrinaire about functional groups than I used to be, but I still have a few that I balk at.

And yes, I know that there are drugs with all these groups in them. But if you look at the quinones, for example, you find mostly cytotoxics and anti-infectives which are cytotoxins with some selectivity for non-mammalian cells. If you're aiming at a particularly nasty target (resistant malaria, pancreatic cancer), go ahead and pull out all the stops. But I don't think anyone should cheerfully plow ahead with such structures unless there are such mitigating circumstances, or at least not without realizing the risks that they're taking on.

But this doesn't do us much good, either - most medicinal chemists don't want to advance such compounds anyway. In fact, rather than being too permissive about things like quinones, most of us are probably too conservative about the sorts of structures we're willing to deal with. There are a lot of funny-looking drugs out there, as it never hurts to remind oneself. Peeling off the outer fringe of these (and quinones are indeed the outer fringe) isn't going to increase anyone's success rate much. So what to do?

I don't have a good answer for that one. I wish I did. It's a rare case when we can say, just by looking at its structure, that a particular compound just won't work. I've been hoping that the percentages would allow us to say more than that about more compounds. But I'm really not sure that they do, at least not to the extent that we need them to, and I worry that we're kidding ourselves when we pretend otherwise.

Comments (32) + TrackBacks (0) | Category: Drug Assays | Drug Development | In Silico

October 17, 2014

More on "Metabolite Likeness" as a Predictor

Email This Entry

Posted by Derek

A recent computational paper that suggested that similarity to known metabolites could help predict successful drug candidates brought in a lot of comments around here. Now the folks at Cambridge MedChem Consulting have another look at it here.

The big concern (as was expressed by some commenters here as well) is the Tanimoto similarity cutoff of 0.5. Does that make everything look too similar, or not? CMC has some numbers across different data sets, and suggests that this cutoff is, in fact, too permissive to allow for much discrimination. People with access to good comparison sets of compounds that made it and compounds that didn't - basically, computational chemists inside large industrial drug discovery organizations - will have a better chance to see how all this holds up.

Comments (6) + TrackBacks (0) | Category: Drug Development | In Silico

October 6, 2014

A New Way to Estimate a Compound's Chances?

Email This Entry

Posted by Derek

Just a few days ago we were talking about whether anything could be predicted about a molecule's toxicity by looking over its biophysical properties. Some have said yes, this is possible (that less polar compounds tend to be more toxic), but a recent paper has said no, that no such correlation exists. This is part of the larger "Rule of 5" discussion, about whether clinical success in general can be (partially) predicted by such measurements (lack of unexpected toxicity is a big factor in that success). And that discussion shows no sign of resolving any time soon, either.

Now comes a new paper that lands right in the middle of this argument. Douglas Kell's group at Manchester has analyzed a large data set of known human metabolites (the Recon2 database, more here) and looked at how similar marketed drugs are to the structures in it. Using MACCS structural fingerprints, they find that 90% of marketed drugs have a Tanimoto similarity of more than 0.5 to at least one compound in the database, and suggest that this could be a useful forecasting tool for new structures.

Now, that's an interesting idea, and not an implausible one, either. But the next things to ask are "Is it valid?" and "What could be wrong with it?" That's the way we learn how to approach pretty much anything new that gets reported in science, of course, although people do tend to take it the wrong way around the dinner table. Applying that in this case, here's what I can think of that could be off:

1. Maybe the reason that everything looks like one of the metabolites in the database is that the database contains a bunch of drug metabolites to start with, perhaps even the exact ones from the drugs under discussion? This isn't the case, though: Recon2 contains endogenous metabolites only, and the Manchester group went through the list removing compounds that are listed as drugs but are also known metabolites (nutritional supplements, for the most part).

2. Maybe Tanimoto similarities aren't the best measurement to use, and overestimate things? Molecular similarity can be a slippery concept, and different people often mean different things by it. The Tanimoto coefficient is the ratio of shared features of two molecules to their unique features, so a Tanimoto of 1 means that the two are identical. What does a coefficient of 0.5 tell us? That depends on how those "features" are counted, as one could well imagine, and the various ones are usually referred to as compound "fingerprints". The Manchester group tried several of these, and settled on the 166 descriptors of the MACCS set. And that brings up the next potential problem. . .

3. Maybe MACCS descriptors aren't the best ones to use? I'm not enough of an informatics person to say, although this point did occur to the authors. They don't seem to know the answer, either, however:

However, the cumulative plots of the (nearest metabolite Tanimoto similarity) for each drug using different fingerprints do differ quite significantly depending on which fingerprint is used, and clearly the well-established MACCS fingerprints lead to a substantially greater degree of ‘metabolite-likeness’ than do almost all the other encodings (we do not pursue this here).

So this one is an open question - it's not for sure if there's something uniquely useful about the MACCS fingerprint set here, or if there's something about the MACCS fingerprint set that makes it just appear to be uniquely useful. The authors do note in the paper that they tried to establish that the patterns they saw were ". . .not a strange artefact of the MACCS encoding itself." And there's another possibility. . .

4. Maybe the universe of things that make this cutoff is too large to be informative? That's another way of asking "What does a Tanimoto coefficient of 0.5 or greater tell you?" The authors reference a paper (Baldi and Nasr) on that very topic, which says:

Examples of fundamental questions one would like to address include: What threshold should one use to assess significance in a typical search? For instance, is a Tanimoto score of 0.5 significant or not? And how many molecules with a similarity score above 0.5 should one expect to find? How do the answers to these questions depend on the size of the database being queried, or the type of queries used? Clear answers to these questions are important for developing better standards in chemoinformatics and unifying existing search methods for assessing the significance of a similarity score, and ultimately for better understanding the nature of chemical space.

The Manchester authors say that applying the methods of that paper to their values show that they're highly significant. I'll take their word for that, since I'm not in a position to run the numbers, but I do note that the earlier paper emphasizes that a particular Tanimoto score's significance is highly dependent on the size of the database, the variety of molecules in it, and the representations used. The current paper doesn't (as far as I can see) go into the details of applying the Baldi and Nasr calculations to their own data set, though.

The authors have done a number of other checks, to make sure that they're not being biased by molecular weights, etc. They looked for trends that could be ascribed to molecular properties like cLogP, but found none. And they tested their hypothesis by running 2000 random compounds from Maybridge through, which did indeed generate much different-looking numbers than the marketed drugs.

As for whether their overall method is useful, here's the Manchester paper's case:

. . .we have shown that when encoded using the public MDL MACCS keys, more than 90 % of individual marketed drugs obey a ‘rule of 0.5’ mnemonic, elaborated here, to the effect that a successful drug is likely to lie within a Tanimoto distance of 0.5 of a known human metabolite. While this does not mean, of course, that a molecule obeying the rule is likely to become a marketed drug for humans, it does mean that a molecule that fails to obey the rule is statistically most unlikely to do so.

That would indeed be a useful thing to know. I would guess that people inside various large drug organizations are going to run this method over their own internal database of compounds to see how it performs on their own failures and successes - and that is going to be the real test. How well it performs, though, we may not hear for a while. But I'll keep my ears open, and report on anything useful.

Comments (35) + TrackBacks (0) | Category: Drug Assays | In Silico

October 2, 2014

We Can't Calculate Our Way Out of This One

Email This Entry

Posted by Derek

Clinical trial failure rates are killing us in this industry. I don't think there's much disagreement on that - between the drugs that just didn't work (wrong target, wrong idea) and the ones that turn out to have unexpected safety problems, we incinerate a lot of money. An earlier, cheaper read on either of those would transform drug research, and people are willing to try all sorts of things to those ends.

One theory on drug safety is that there are particular molecular properties that are more likely to lead to trouble. There have been several correlations proposed between high logP (greasiness) and tox liabilities, multiple aromatic rings and tox, and so on. One rule proposed in 2008 by a group at Pfizer is that clogP >3 and total polar surface area less than 75 square angstroms is a good cutoff - compounds on the other side of it are about 2.5 times more likely to run into trouble. But here's a paper in MedChemComm that asks if any of this has any validity:

What is the likelihood of real success in avoiding attrition due to toxicity/safety from using such simple metrics? As mentioned in the beginning, toxicity can arise from a wide variety of reasons and through a plethora of complex mechanisms similar to some of the DMPK endpoints that we are still struggling to avoid. In addition to the issue of understanding and predicting actual toxicity, there are other hurdles to overcome when doing this type of historical analysis that are seldom discussed.

The first of these is making sure that you're looking at the right set of failed projects - that is, ones that really did fail because of unexpected compound-associated tox, and not some other reason (such as unexpected mechanism-based toxicity, which is another issue). Or perhaps a compound could have been good enough to make it on its own under other circumstances, but the competitive situation made it untenable (something else came up with a cleaner profile at about the same time). Then there's the problem of different safety cutoffs for different therapeutic areas - acceptable tox for a pancreatic cancer drug will not cut it for type II diabetes, for example.

The authors did a thorough study of 130 AstraZeneca development compounds, with enough data to work out all these complications. (This is the sort of thing that can only be done from inside a company's research effort - you're never going to have enough information, working from outside). What they found, right off, was that for this set of compounds the Pfizer rule was completely inverted. The compounds on the too-greasy side actually had shown fewer problems (!) The authors looked at the data sets from several different angles, and conclude that the most likely explanation is that the rule is just not universally valid, and depends on the dataset you start with.

The same thing happens when you look at the fraction of sp3 carbons, which is a characteristic (the "Escape From Flatland" paper) that's also been proposed to correlate with tox liabilities. The AZ set shows no such correlation at all. Their best hypothesis is that this is a likely correlation with pharmacokinetics that has gotten mixed in with a spurious correlation with toxicity (and indeed, the first paper on this trend was only talking about PK). And finally, they go back to an earlier properties-based model published by other workers at AstraZeneca, and find that it, too, doesn't seem to hold up on the larger, more curated data set. Their-take home message: ". . .it is unlikely that a model of simple physico-chemical descriptors would be predictive in a practical setting."

Even more worrisome is what happens when you take a look at the last few years of approved drugs and apply such filters to them (emphasis added):

To investigate the potential impact of following simple metric guidelines, a set of recently approved drugs was classified using the 3/75 rule (Table 3). The set included all small molecule drugs approved during 2009–2012 as listed on the ChEMBL website. No significant biases in the distribution of these compounds can be seen from the data presented in Table 3. This pattern was unaffected if we considered only oral drugs (45) or all of the drugs (63). The highest number of drugs ends up in the high ClogP/high TPSA class and the class with the lowest number of drugs is the low ClogP/low TPSA. One could draw the conclusion that using these simplistic approaches as rules will discard the development of many interesting and relevant drugs.

One could indeed. I hadn't seen this paper myself until the other day - a colleague down the hall brought it to my attention - and I think it deserves wider attention. A lot of drug discovery organizations, particularly the larger ones, use (or are tempted to use) such criteria to rank compounds and candidates, and many of us are personally carrying such things around in our heads. But if these rules aren't valid - and this work certainly makes it look as if they aren't - then we should stop pretending as if they are. That throws us back into a world where we have trouble distinguishing troublesome compounds from the good ones, but that, it seems, is the world we've been living in all along. We'd be better off if we just admitted it.

Comments (25) + TrackBacks (0) | Category: Drug Assays | Drug Development | In Silico | Toxicology

September 22, 2014

Chemical Space

Email This Entry

Posted by Derek

I'm listening to Jean-Louis Reymond of Bern talking about the GDB data set, the massive enumerated set of possible molecules. That's the set of chemically feasible molecules at or below a certain heavy atom count - the first iteration was GDB11 (blogged about here), and it's since been extended to GDB13, which has nearly one billion compounds with up to 13 C, N, O, S and Cl atoms. (Note, as always, that huge vast heaps of poly-small-ring compounds, especially concatenations of 3-membered rings, are pre-filtered out of these sets, because otherwise they would overwhelm them completely). They're working now on GDB17, which is a truly huge mound of data.

I was particularly taken with the image shown (from this paper), an artificial set of compounds (up to heavy atoms counts of 500) from several main classes of real molecules. It's a 3-D principle components analysis plot, which tunes things up to emphasize the differences, of course, and there's what chemical space looks like from this angle. There go the proteins and nucleic acids, off into their own zones, and similarly the linear alkanes and diamond-like lattices, beaming off in separate directions. In the middle are drug-like compounds - and don't imagine for a minute that any substantial number of those have actually been prepared, either. This is where we live, all of us organic chemists.

Comments (4) + TrackBacks (0) | Category: In Silico

June 23, 2014

The Virtual Clinical Trial: Not Quite Around the Corner

Email This Entry

Posted by Derek

Here's one of those "Drug Discovery of. . .the. . .Future-ure-ure-ure" articles in the popular press. (I need a reverb chamber to make that work property). At The Atlantic, they're talking with "medical futurists" and coming up with this:

The idea is to combine big data and computer simulations—the kind an engineer might use to make a virtual prototype of a new kind of airplane—to figure out not just what's wrong with you but to predict which course of treatment is best for you. That's the focus of Dassault Systèmes, a French software company that's using broad datasets to create cell-level simulations for all different kinds of patients. In other words, by modeling what has happened to patients like you in previous cases, doctors can better understand what might happen if they try certain treatments for you—taking into consideration your age, your weight, your gender, your blood type, your race, your symptom, any number of other biomarkers. And we're talking about a level of precision that goes way beyond text books and case studies.

I'm very much of two minds about this sort of thing. On the one hand, the people at Dassault are not fools. They see an opportunity here, and they think that they might have a realistic chance at selling something useful. And it's absolutely true that this is, broadly, the direction in which medicine is heading. As we learn more about biomarkers and individual biochemistry, we will indeed be trying to zero in on single-patient variations.

But on that ever-present other hand, I don't think that you want to make anyone think that this is just around the corner, because it's not. It's wildly difficult to do this sort of thing, as many have discovered at great expense, and our level of ignorance about human biochemistry is a constant problem. And while tailoring individual patient's therapies with known drugs is hard enough, it gets really tricky when you talk about evaluating new drugs in the first place:

Charlès and his colleagues believe that a shift to virtual clinical trials—that is, testing new medicines and devices using computer models before or instead of trials in human patients—could make new treatments available more quickly and cheaply. "A new drug, a succesful drug, takes 10 to 12 years to develop and over $1 billion in expenses," said Max Carnecchia, president of the software company Accelrys, which Dassault Systèmes recently acquired. "But when it is approved by FDA or other government bodies, typically less than 50 percent of patients respond to that therapy or drug." No treatment is one-size-fits-all, so why spend all that money on a single approach?

Carnecchia calls the shift toward algorithmic clinical trials a "revolution in drug discovery" that will allow for many quick and low-cost simulations based on an endless number of individual cellular models. "Those models start to inform and direct and focus the kinds of clinical trials that have historically been the basis for drug discovery," Carnecchia told me. "There's the benefit to drug companies from reduction of cost, but more importantly being able to get these therapies out into the market—whether that's saving lives or just improving human health—in such a way where you start to know ahead of time whether that patient will actually respond to that therapy."

Speed the day. The cost of clinical trials, coupled with their low success rate, is eating us alive in this business (and it's getting worse every year). This is just the sort of thing that could rescue us from the walls that are closing in more tightly all the time. But this talk of shifts and revolutions makes it sound as if this sort of thing is happening right now, which it isn't. No such simulated clinical trial, one that could serve as the basis for a drug approval, is anywhere near even being proposed. How long before one is, then? If things go really swimmingly, I'd say 20 to 25 years from now, personally, but I'd be glad to hear other estimates.

To be fiar, the article does go on to mentions something like this, but it just says that "it may be a while" before said revolution happens. And you get the impression that what's most needed is some sort of "cultural shift in medicine toward openness and resource sharing". I don't know. . .I find that when people call for big cultural shifts, they're sometimes diverting attention (even their own attention) from the harder parts of the problem under discussion. Gosh, we'd have this going in no time if people would just open up and change their old-fashioned ways! But in this case, I still don't see that as the rate-limiting step at all. Pouring on the openness and sharing probably wouldn't hurt a bit in the quest for understanding human drug responses and individual toxicology, but it's not going to suddenly open up any blocked-up floodgates, either. We don't know enough. Pooling our current ignorance can only take us so far.

Remember there are hundreds of billions of dollars waiting to be picked up off the ground by anyone who can do these things. It's not like there are no incentives to find ways to make clinical trials faster and cheaper. Anything that gives the impression that there's this one factor (lack of cooperation, too much regulation, Evil Pharma Executives, what have you) holding us back from the new era, well. . .that just might be an oversimplified view of the situation.

Comments (15) + TrackBacks (0) | Category: Clinical Trials | In Silico | Regulatory Affairs | Toxicology

June 9, 2014

Hosed-Up X-Ray Structures: A Big Problem

Email This Entry

Posted by Derek

X-ray crystallography is great stuff, no doubt about it. But it's not magic. It takes substantial human input to give a useful structure of a ligand bound to a protein - there are decisions to be made and differences to be split. It's important to emphasize, for those of us who are not crystallographers, that unless you have resolution down below 1Å - and I'll bet you don't - then your X-ray structures are not quite "structures"; they're models. A paper several years ago emphasized these factors for chemists outside the field.

About ten years ago, I wrote about this paper, which suggested that many ligand-bound structures seemed to have strain energy in them that wouldn't have been predicted. One interpretation is that there's more to ligand (and binding site) reorganization than people tend to realize, and that ligands don't always bind in their lowest-energy conformations. And while I still think that's true, the situation is complicated by another problem that's become more apparent over the years: many reported X-ray structures for ligand-bound proteins are just messed up.
Here's an editorial in ACS Medicinal Chemistry Letters that shows how bad the problem may well be. Reviews of the crystallographic databases have suggested that there are plenty of poorly refined structures hiding in there. But I didn't realize that they were as poorly refined as some of these. Take a look at the phosphate in 1xqd, and note how squashed-out those oxygens are around the first phosphorus. Or try the olefin in 4g93, which has been yanked 90 degrees out of plane. It's bad that there are such ridiculous structures in the literature, but the larger number of semi-plausible (but still wrong) structures is even worse.

Those structures at the left illustrate what's going on. The top one is an old PDB structure, 3qad, for an IKK inhibitor. It's a mess. Note that there's a tetrahedralish aromatic carbon (not happening), and a piperazine in a boat conformation (only slightly less unlikely). The structure was revised after this was pointed out to the middle version (3rzf), but that one still has some odd features - those two aromatic groups are flat-on in the same plane, and the amine between them and the next aryl is rather odd, too. Might be right, might be wrong - who's to know?

The most recent comprehensive look (from 2012) suggests that about 25% of the reported ligand-bound structures are mangled to the point of being misleading. This new editorial goes on to mention some computational tools that could help to keep this from happening, such as this one. If we're all going to draw conclusions from these things (and that's what they're there for, right?) we'd be better off using the best ones we can.

Comments (20) + TrackBacks (0) | Category: Drug Assays | In Silico

June 4, 2014

Predicting New Targets - Another Approach

Email This Entry

Posted by Derek

So you make a new chemical structure as part of a drug research program. What's it going to hit when it goes into an animal?

That question is a good indicator of the divide between the general public and actual chemists and pharmacologists. People without any med-chem background tend to think that we can predict these things, and people with it know that we can't predict much at all. Even just predicting activity at the actual desired target is no joke, and guessing what other targets a given compound might hit is, well, usually just guessing. We get surprised all the time.

That hasn't been for lack of trying, of course. Here's an effort from a few years ago on this exact question, and a team from Novartis has just published another approach. It builds on some earlier work of theirs (HTS fingerprints, HTSFP) that tries to classify compounds according to similar fingerprints of biological activity in suites of assays, rather than by their structures, and this latest one is called HTSFP-TID (target ID, and I think the acronym is getting a bit overloaded at that point).

We apply HTSFP-TID to make predictions for 1,357 natural products (NPs) and 1,416 experimental small molecules and marketed drugs (hereafter generally referred to as drugs). Our large-scale target prediction enables us to detect differences in the protein classes predicted for the two data sets, reveal target classes that so far have been underrepresented in target elucidation efforts, and devise strategies for a more effective targeting of the druggable genome. Our results show that even for highly investigated compounds such as marketed drugs, HTSFP-TID provides fresh hypotheses that were previously not pursued because they were not obvious based on the chemical structure of a molecule or against human intuition.

They have up to 230 or so assays to pick from, although it's for sure that none of the compounds have been through all of them. They required that any given compound have at least 50 different assays to its name, though (and these were dealt with as standard deviations off the mean, to keep things comparable). And what they found shows some interesting (and believable) discrepancies between the two sets of compounds. The natural product set gave mostly predictions for enzyme targets (70%), half of them being kinases. Proteases were about 15% of the target predictions, and only 4% were predicted GPCR targets. The drug-like set also predicted a lot of kinase interactions (44%), and this from a set where only 20% of the compounds were known to hit any kinases before. But it had only 5% protease target predictions, as opposed to 23% GPCR target predictions.

The group took a subset of compounds and ran them through new assays to see how the predictions came out, and the results weren't bad - overall, about 73% of the predictions were borne out by experiment. The kinase predictions, especially, seemed fairly accurate, although the GPCR calls were less so. They identified several new modes of action for existing compounds (a few of which they later discovered buried in the literature). They also tried a set of predictions based on chemical descriptor (the other standard approach), but found a lower hit rate. Interestingly, though, the two methods tended to give orthogonal predictions, which suggests that you might want to run things both ways if you care enough. Such efforts would seem particularly useful as you push into weirdo chemical or biological space, where we'll take whatever guidance we can get.

Novartis has 1.8 million compounds to work with, and plenty of assay data. It would be worth knowing what some other large collections would yield with the same algorithms: if you used (say) Merck's in-house data as a training set, and then applied it to all the compounds in the CHEMBL database, how similar would the set of predictions for them be? I'd very much like for someone to do something like this (and publish the results), but we'll see if that happens or not.

Comments (11) + TrackBacks (0) | Category: Drug Assays | In Silico

April 4, 2014

Ancient Modeling

Email This Entry

Posted by Derek

I really got a kick out of this picture that Wavefunction put up on Twitter last night. It's from a 1981 article in Fortune, and you'll just have to see the quality of the computer graphics to really appreciate it.

That sort of thing has hurt computer-aided drug design a vast amount over the years. It's safe to say that in 1981, Merck scientists did not (as the article asserts) "design drugs and check out their properties without leaving their consoles". It's 2014 and we can't do it like that yet. Whoever wrote that article, though, picked those ideas up from the people at Merck, with their fuzzy black-and-white monitor shots of DNA from three angles. (An old Evans and Sutherland terminal?) And who knows, some of the Merck folks may have even believed that they were close to doing it.

But computational power, for the most part, only helps you out when you already know how to calculate something. Then it does it for you faster. And when people are impressed (as they should be) with all that processing power can do for us now, from smart phones on up, they should still realize that these things are examples of fast, smooth, well-optimized versions of things that we know how to calculate. You could write down everything that's going on inside a smart phone with pencil and paper, and show exactly what it's working out when it display this pixel here, that pixel there, this call to that subroutine, which calculates the value for that parameter over there as the screen responds to the presence of your finger, and so on. It would be wildly tedious, but you could do it, given time. Someone, after all, had to program all that stuff, and programming steps can be written down.

The programs that drove those old DNA pictures could be written down, too, of course, and in a lot less space. But while the values for which pixels to light up on the CRT display were calculated exactly, the calculations behind those were (and are) a different matter. A very precise-looking picture can be drawn and animated of an animal that does not exist, and there are a lot of ways to draw animals that do not exist. The horse on your screen might look exact in every detail, except with a paisley hide and purple hooves (my daughter would gladly pay to ride one). Or it might have a platypus bill instead of a muzzle. Or look just like a horse from outside, but actually be filled with helium, because your program doesn't know how to handle horse innards. You get the idea.

The same for DNA, or a protein target. In 1981, figuring out exactly what happened as a transcription factor approached a section of DNA was not possible. Not to the degree that a drug designer would need. The changing conformation of the protein as it approaches the electrostatic field of the charged phosphate residues, what to do with the water molecules between the two as they come closer, the first binding event (what is it?) between the transcription factor and the double helix, leading to a cascade of tradeoffs between entropy and enthalpy as the two biomolecules adjust to each other in an intricate tandem dance down to a lower energy state. . .that stuff is hard. It's still hard. We don't know how to model some of those things well enough, and the (as yet unavoidable) errors and uncertainties in each step accumulate the further you go along. We're much better at it than we used to be, and getting better all the time, but there's a good way to go yet.

But while all that's true, I'm almost certainly reading too much into that old picture. The folks at Merck probably just put one of their more impressive-looking things up on the screen for the Fortune reporter, and hey, everyone's heard of DNA. I really don't think that anyone at Merck was targeting protein-DNA interactions 33 years ago (and if they were, they splintered their lance against that one, big-time). But the reporter came away with the impression that the age of computer-designed drugs was at hand, and in the years since, plenty of other people have seen progressively snazzier graphics and thought the same thing. And it's hurt the cause of modeling for them to think that, because the higher the expectations get, the harder it is to come back to reality.

Update: I had this originally as coming from a Forbes article; it was actually in Fortune.

Comments (22) + TrackBacks (0) | Category: Drug Industry History | In Silico

March 17, 2014

Predicting What Group to Put On Next

Email This Entry

Posted by Derek

Here's a new paper in J. Med. Chem. on software that tries to implement matched-molecular-pair type analysis. The goal is a recommendation - what R group should I put on next?

Now, any such approach is going to have to deal with this paper from Abbott in 2008. In that one, an analysis of 84,000 compounds across 30 targets strongly suggested that most R-group replacements had, on average, very little effect on potency. That's not to say that they don't or can't affect binding, far from it - just that over a large series, those effects are pretty much a normal distribution centered on zero. There are also analyses that claim the same thing for adding methyl groups - to be sure, there are many dramatic "magic methyl" enhancement examples, but are they balanced out, on the whole, by a similar number of dramatic drop-offs, along with a larger cohort of examples where not much happened at all?

To their credit, the authors of this new paper reference these others right up front. The answer to these earlier papers, most likely, is that when you average across all sorts of binding sites, you're going to see all sorts of effects. For this to work, you've got a far better chance of getting something useful if you're working inside the same target or assay. Here we get to the nuts and bolts:

The predictive method proposed, Matsy, relies on the hypothesis that a particular matched series tends to have a preferred activity order, for example, that not all six possible orders of [Br, Cl, F] are equally frequent. . .Although a rather straightforward idea, we have been unable to find any quantitative analysis of this question in the literature.

So they go on to provide one, with halogen substituents. There's not much to be found comparing pairs of halogen compounds head to head, but when you go to the longer series, you find that the order Br > Cl > F > H is by far the most common (and that appears to be just a good old grease effect). The next most common order just swaps the bromine and chlorine, but the third most common is the original order, in reverse. The other end of the distribution is interesting, too - for example, the least most common order is Br > H > F > Cl, which is believable, since it doesn't make much sense along any property axis.

They go on to do the same sorts of analyses for other matched series, and the question then becomes, if you have such a matched series in your own SAR, what does that order tell you about what to make next? The idea of "SAR transfer" has been explored, and older readers will remember the Topliss tree for picking aromatic substituents (do younger ones?)

The Matsy algorithm may be considered a formalism of aspects of how a medicinal chemist works in practice. Observing a particular trend, a chemist considers what to make next on the basis of chemical intuition, experience with related compounds or targets, and ease of synthesis. The structures suggested by Matsy preserve the core features of molecules while recommending small modifications, a process very much in line with the type of functional group replacement that is common in lead optimization projects. This is in contrast to recommendations from fingerprint-based similarity comparisons where the structural similarity is not always straightforward to rationalize and near-neighbors may look unnatural to a medicinal chemist.

And there's a key point: prediction and recommendation programs walk a fine line, between "There's no way I'm going out of my way to make that" and "I didn't need this program to tell me this". Sometimes there's hardly any space between those two territories at all. Where do this program's recommendations fall? As companies try this out in-house, some people will be finding out. . .

Comments (13) + TrackBacks (0) | Category: Drug Development | In Silico

February 28, 2014

Computational Nirvana

Email This Entry

Posted by Derek

Wavefunction has a post about this paper from J. Med. Chem. on a series of possible antitrypanosomals from the Broad Institute's compound collection. It's a good illustration of the power of internal hydrogen bonds - in this case, one series of isomers can make the bond, but that ties up their polar groups, making them less soluble but more cell-permeable. The isomer that doesn't form the internal H-bond is more polar and more soluble, but less able to get into cells. Edit - fixed this part.

So if your compound has too many polar functionalities, an internal hydrogen bond can be just the thing to bring on better activity, because it tones things down a bit. And there are always the conformational effects to keep in mind. Tying a molecule up like that is the same as any other ring-forming gambit in medicinal chemistry: death or glory. Rarely is a strong conformational restriction silent in the SAR - usually, you either hit the magic conformer, or you move it forever out of reach.

I particularly noticed Wavefunction's line near the close of his post: "If nothing else they provide a few more valuable data points on the way to prediction nirvana.". I know what he's talking about, and I think he's far from the only computational chemist with eschatological leanings. Eventually, you'd think, we'd understand enough about all the things we're trying to model for the models to, well, work. And yes, I know that there are models that work right now, but you don't know that they're going to work until you've messed with them a while, and there are other models that don't work but look equally plausible at first, etc., and very much etc. "Prediction nirvana" would be the state where you have an idea for a new structure, you enter it into your computational model, and it immediately tells you the right answer, every single time. In theory, I think this is a reachable state of affairs. In practice, it is not yet implemented.

And remember, people have spotted glows on that horizon before and proclaimed the imminent dawn. The late 1980s were such a time, but experiences like those tend to make people more reluctant to immanentize the eschaton, or at least not where anyone can hear. But we are learning more about enthalpic and entropic interactions, conformations, hydrogen bonds, nonpolar interactions, all those things that go into computational prediction of structure and binding interactions. And if we continue to learn more, as seems likely, won't there come a point when we've learned what we need to know? If not true computational nirvana, then surely (shrink those epsilons and deltas) as arbitrarily close an approach as we like?

Comments (8) + TrackBacks (0) | Category: In Silico

February 19, 2014

Ligand Efficiency: A Response to Shultz

Email This Entry

Posted by Derek

I'd like to throw a few more logs on the ligand efficiency fire. Chuck Reynolds of J&J (author of several papers on the subject, as aficionados know) left a comment to an earlier post that I think needs some wider exposure. I've added links to the references:

An article by Shultz was highlighted earlier in this blog and is mentioned again in this post on a recent review of Ligand Efficiency. Shultz’s criticism of LE, and indeed drug discovery “metrics” in general hinges on: (1) a discussion about the psychology of various metrics on scientists' thinking, (2) an assertion that the original definition of ligand efficiency, DeltaG/HA, is somehow flawed mathematically, and (3) counter examples where large ligands have been successfully brought to the clinic.

I will abstain from addressing the first point. With regard to the second, the argument that there is some mathematical rule that precludes dividing a logarithmic quantity by an integer is wrong. LE is simply a ratio of potency per atom. The fact that a log is involved in computing DeltaG, pKi, etc. is immaterial. He makes a more credible point that LE itself is on average non-linear with respect to large differences in HA count. But this is hardly a new observation, since exactly this trend has been discussed in detail by previous published studies (here, here, here, and here). It is, of course, true that if one goes to very low numbers of heavy atoms the classical definition of LE gets large, but as a practical matter medicinal chemists have little interest in extremely small fragments, and the mathematical catastrophe he warns us against only occurs when the number of heavy atoms goes to zero (with a zero in the denominator it makes no difference if there is a log in the numerator). Why would HA=0 ever be relevant to a med. chem. program? In any case a figure essentially equivalent to the prominently featured Figure 1a in the Shultz manuscript appears in all of the four papers listed above. You just need to know they exist.

With regard to the third argument, yes of course there are examples of drugs that defy one or more of the common guidelines (e.g MW). This seems to be a general problem of the community taking metrics and somehow turning them into “rules.” They are just helpful, hopefully, guideposts to be used as the situation and an organization’s appetite for risk dictate. One can only throw the concept of ligand efficiency out the window completely if you disagree with the general principle that it is better to design ligands where the atoms all, as much as possible, contribute to that molecule being a drug (e.g. potency, solubility, transport, tox, etc.). The fact that there are multiple LE schemes in the literature is just a natural consequence of ongoing efforts to refine, improve, and better apply a concept that most would agree is fundamental to successful drug discovery.

Well, as far as the math goes, dividing a log by an integer is not any sort of invalid operation. I believe that [log(x)]/y is the same as saying log(x to the one over y). That is, log(16) divided by 2 is the same as the log of 16 to the one-half power, or log(4). They both come out to about 0.602. Taking a BEI calculation as real chemistry example, a one-micromolar compound that weighs 250 would, by the usual definition, -log(Ki)/(MW/1000), have a BEI of 6/0.25, or 24. By the above rule, if you want to keep everything inside the log function, then say -log(0.0000001) divided by 0.25, that one-micromolar figure should be raised to the fourth power, then you take the log of the result (and flip the sign). One-millionth to the fourth power is one times ten to the minus twenty-fourth, so that gives you. . .24. No problem.

Shultz's objection that LE is not linear per heavy atom, though, is certainly valid, as Reynolds notes above as well. You have to realize this and bear it in mind while you're thinking about the topic. I think that one of the biggest problems with these metrics - and here's a point that both Reynolds and Shultz can agree on, I'll bet - is that they're tossed around too freely by people who would like to use them as a substitute for thought in the first place.

Comments (19) + TrackBacks (0) | Category: Drug Assays | Drug Development | In Silico

February 14, 2014

"It Is Not Hard to Peddle Incoherent Math to Biologists"

Email This Entry

Posted by Derek

Here's a nasty fight going on in molecular biology/bioinformatics. Lior Pachter of Berkeley describes some severe objections he has to published work from the lab of Manolis Kellis at MIT. (His two previous posts on these issues are here and here). I'm going to use a phrase that Pachter hears too often and say that I don't have the math to address those two earlier posts. But the latest one wraps things up in a form that everyone can understand. After describing what does look like a severe error in one of the Manolis group's conference presentations, which Pachter included in a review of the work, he says that:

. . .(they) spun the bad news they had received as “resulting from combinatorial connectivity patterns prevalent in larger network structures.” They then added that “…this combinatorial clustering effect brings into question the current definition of network motif” and proposed that “additional statistics…might well be suited to identify larger meaningful networks.” This is a lot like someone claiming to discover a bacteria whose DNA is arsenic-based and upon being told by others that the “discovery” is incorrect – in fact, that very bacteria seeks out phosphorous – responding that this is “really helpful” and that it “raises lots of new interesting open questions” about how arsenate gets into cells. Chutzpah. When you discover your work is flawed, the correct response is to retract it.

I don’t think people read papers very carefully. . .

He goes on to say:

I have to admit that after the Grochow-Kellis paper I was a bit skeptical of Kellis’ work. Not because of the paper itself (everyone makes mistakes), but because of the way he responded to my review. So a year and a half ago, when Manolis Kellis published a paper in an area I care about and am involved in, I may have had a negative prior. The paper was Luke Ward and Manolis Kellis “Evidence for Abundant and Purifying Selection in Humans for Recently Acquired Regulatory Functions”, Science 337 (2012) . Having been involved with the ENCODE pilot, where I contributed to the multiple alignment sub-project, I was curious what comparative genomics insights the full-scale $130 million dollar project revealed. The press releases accompanying the Ward-Kellis paper (e.g. The Nature of Man, The Economist) were suggesting that Ward and Kellis had figured out what makes a human a human; my curiosity was understandably piqued.

But a closer look at the paper, Pachter says, especially a dig into the supplementary material (always a recommended move) shows that the conclusions of the paper were based on what he terms "blatant statistically invalid cherry picking". See, I told you this was a fight. He also accuses Kellis of several other totally unacceptable actions in his published work, the sorts of things that cannot be brushed off as differences in interpretations or methods. He's talking fraud. And he has a larger point about how something like this might persist in the computational biology field (emphasis added):

Manolis Kellis’ behavior is part of a systemic problem in computational biology. The cross-fertilization of ideas between mathematics, statistics, computer science and biology is both an opportunity and a danger. It is not hard to peddle incoherent math to biologists, many of whom are literally math phobic. For example, a number of responses I’ve received to the Feizi et al. blog post have started with comments such as

“I don’t have the expertise to judge the math, …”

Similarly, it isn’t hard to fool mathematicians into believing biological fables. Many mathematicians throughout the country were recently convinced by Jonathan Rothberg to donate samples of their DNA so that they might find out “what makes them a genius”. Such mathematicians, and their colleagues in computer science and statistics, take at face value statements such as “we have figured out what makes a human human”. In the midst of such confusion, it is easy for an enterprising “computational person” to take advantage of the situation, and Kellis has.

You can peddle incoherent math to medicinal chemists, too, if you feel the urge. We don't use much of it day-to-day, although we've internalized more than we tend to realize. But if someone really wants to sell me on some bogus graph theory or topology, they'll almost certainly be able to manage it. I'd at least give them the benefit of the doubt, because I don't have the expertise to call them on it. Were I so minded, I could probably sell them some pretty shaky organic chemistry and pharmacokinetics.

But I am not so minded. Science is large, and we have to be able to trust each other. I could sit down and get myself up to speed on topology (say), if I had to, but the effort required would probably be better spent doing something else. (I'm not ruling out doing math recreationally, just for work). None of us can simultaneously be experts across all our specialities. So if this really is a case of publishing junk because, hey, who'll catch on, right, then it really needs to be dealt with.

If Pachter is off base, though, then he's in for a rough ride of his own. Looking over his posts, my money's on him and not Kellis, but we'll all have a chance to find out. After this very public calling out, there's no other outcome.

Comments (32) + TrackBacks (0) | Category: Biological News | In Silico | The Dark Side | The Scientific Literature

January 14, 2014

A New Metabolism Predictor

Email This Entry

Posted by Derek

Drug metabolism is a perennial topic for us small-molecule people. Watching your lovingly optimized molecules go through the shredding-machine of the liver is an instructive experience, not least when you consider how hard it would be for you to do some of the chemistry that it does. (For reference and getting up to speed on the details, the comments section here has had reader recommendations for the Drug Metabolism and Pharmacokinetics Quick Guide).

Here's a review of a new sites-of-metabolism predictor, FAME, a decision-tree type program that's been trained on data from 20,000 known compounds. It handles both Phase I and Phase II metabolism (a "Pharma 101" entry on that topic is here, for those who'd like to know more), and it looks like it's well worth considering if you're in need for something like this.

Here's my question for the med-chem and PK types: have you made use of predictive metabolism software? Did it save you time, or did you either go down the wrong alleys or not see anything you wouldn't have predicted yourself? I'm interested in real-world experiences, since I haven't had too many myself in this area.

Comments (10) + TrackBacks (0) | Category: In Silico | Pharmacokinetics

December 20, 2013

Picking Diverse Compounds

Email This Entry

Posted by Derek

Diversity deck, diversity set, diversity collection: most chemical screening efforts try to have some bunch of compounds that are selected for being as unlike each other as possible. Fragment-based collections, being smaller by design, are particularly combed through for this property, in order to cover the most chemical space possible. But how, exactly, do you evaluate chemical diversity?

There are a lot of algorithmic approaches, and a new paper helpfully tries to sort them out for everyone. Here's the take-home:

We assessed both the similar behavior of the descriptors in assessing the diversity of chemical libraries, and their ability to select compounds from libraries that are diverse in bioactivity space, which is a property of much practical relevance in screening library design. This is particularly evident, given that many future targets to be screened are not known in advance, but that the library should still maximize the likelihood of containing bioactive matter also for future screening campaigns. Overall, our results showed that descriptors based on atom topology (i.e., fingerprint-based descriptors and pharmacophore-based descriptors) correlate well in rank-ordering compounds, both within and between descriptor types. On the other hand, shape-based descriptors such as ROCS and PMI showed weak correlation with the other descriptors utilized in this study, demonstrating significantly different behavior.

One of the best-performing methods was Bayes Activity Fingerprints, a technique proposed a few years ago by a group at Novartis. That (at least to my non-computational eyes) doesn't seem too surprising, since this new paper is trying to see how well diversity measure perform when compared to bioactivity space, and that earlier one was specifically adding in a measure to account for bioactivity space as well.

On the other hand, shape-based descriptors were problematic. One that turns up a lot is Principle Moments of Inertia (PMI), the scheme that separates compounds into rod-like, disk-like, and sphere-like shape families, but it and ROCS (based on overlaying molecular volumes) were definitely off in their own world when compared to the other descriptors. In fact, the authors found that there seemed to be no correlation at all between PMI diversity and diverse bioactivity, which should be worth thinking about. You'd apparently do better just picking things randomly than using PMI.

Comments (13) + TrackBacks (0) | Category: In Silico

December 10, 2013

Standards of Proof

Email This Entry

Posted by Derek

Here are some slides from Anthony Nicholls of OpenEye, from his recent presentation here in Cambridge on his problems with molecular dynamics calcuations. Here's his cri du coeur (note: fixed a French typo from the original post there):

. . .as a technique MD has many attractive attributes that have nothing to do with its actual predictive capabilities (it makes great movies, it’s “Physics”, calculations take a long time, it takes skill to do right, “important” people develop it, etc). As I repeatedly mentioned in the talk, I would love MD to be a reliable tool - many of the things modelers try to do would become much easier. I just see little objective, scientific evidence for this as yet. In particular, it bothers me that MD is not held to the same standards of proof that many simpler, empirical approaches are - and this can’t be good for the field or MD.

I suspect he'd agree with the general principle that while most things that are worthwhile are hard, not everything that's hard is worthwhile. His slides are definitely fun to read, and worthwhile even if you don't give a hoot about molecular dynamics. The errors he's warning about apply to all fields of science. For example,he starts off with the definition of cognitive dissonance from Wikipedia, and proposes that a lot of the behavior you see in the molecular dynamics field fits the definitions of how people deal with this. He also maintains that the field seems to spend too much of its time justifying data retrospectively, and that this isn't a good sign.

I especially enjoyed his section on the "Tanimoto of Truth". That's comparing reality to experimental results. You have the cases where there should have been a result and the experiment showed it, and there shouldn't have been one, and the experiment reproduced that, too : great! But there are many more cases where only that first part applies, or gets published (heads I win, tails just didn't happen). And you have the inverse of that, where there was nothing, in reality, but your experiment told you that there was something. These false positives get stuck in the drawer, and no one hears about them at all. The next case, the false negatives, often end up in the "parameterize until publishable" category (as Nicholls puts it), or they get buried as well. The last category (should have been negative, experiment says they're negative) are considered so routine and boring that no one talks about them at all, although logically they're quite important.

All this can impart a heavy, heavy publication bias: you only hear about the stuff that worked, even if some of the examples you hear about really didn't. And unless you do a lot of runs yourself, you don't usually have a chance to see how robust the system really is, because the data you'd need aren't available. The organic synthesis equivalent is when you read one of those papers that do, in fact, work on the compounds in Table 1, but hardly any others. And you have to play close attention to Table 1 to realize that you know, there aren't any basic amines on that list (or esters, or amides, or what have you), are there?

The rest of the slides get into the details of molecular dynamic simulations, but he has some interesting comments on the paper I blogged about here, on modeling of allosteric muscarinic ligands. Nicholls says that "There are things to admire about this paper- chiefly that a prospective test seems to have been done, although not by the Shaw group." That caught my eye as well; it's quite unusual to see that, although it shouldn't be. But he goes on to say that ". . .if you are a little more skeptical it is easy to ask what has really been done here. In their (vast) supplementary material they admit that GLIDE docking results agree with mutagenesis as well (only, “not quite as well’, whatever that means- no quantification, of course). There’s no sense, with this data, of whether there are mutagenesis results NOT concordant with the simulations." And that gets back to his Tanimoto of Truth argument, which is a valid one.

He also points out that the predictions ended up being used to make one compound, which is not a very robust standard of proof. The reason, says Nicholls, is that molecular dynamics papers are held to a lower standard, and that's doing the field no good.

Comments (9) + TrackBacks (0) | Category: In Silico | Who Discovers and Why

December 9, 2013

Low Energy Records

Email This Entry

Posted by Derek

Pick an empirical formula. Now, what's the most stable compound that fits it? Not an easy question, for sure, and it's the topic of this paper in Angewandte Chemie. Most chemists will immediately realize that the first problem is the sheer number of possibilities, and the second one is figuring out their energies. A nonscientist might think that this is the sort of thing that would have been worked out a long time ago, but that definitely isn't the case. Why think about these things?

What is this “Guinness” molecule isomer search good for? Some astrochemists think in such terms when they look for molecules in interstellar space. A rule with exceptions says that the most stable isomers have a higher abundance (Astrophys. J.­ 2009, 696, L133), although kinetic control undoubtedly has a say in this. Pyrolysis or biotechnology processes, for example, in anaerobic biomass-to-fuel conversions, may be classified on the energy scale of their products. The fate of organic aerosols upon excitation with highly energetic radiation appears to be strongly influenced by such sequences because of ion-catalyzed chain reactions (Phys. Chem. Chem. Phys.­ 2013, 15, 940). The magic of protein folding is tied to the most stable atomic arrangement, although one must keep in mind that this is a minimum-energy search with hardly any chemical-bond rearrangement. We should rather not think about what happens to our proteins in a global search for their minimum-energy structure, although the peptide bond is not so bad in globally minimizing interatomic energy. Regularity can help and ab initio crystal structure prediction for organic compounds is slowly coming into reach. Again, the integrity of the underlying molecule is usually preserved in such searches.

Things get even trickier when you don't restrict yourself to single compounds. It's pointed out that the low-energy form of the hexose empirical formula (C6H12O6) might well be a mixture of methane and carbon dioxide (which sounds like the inside of a comet to me). That brings up another reason this sort of thinking is useful: if you want to sequester carbon dioxide, what's the best way to do it? What molecular assemblies are most energetically favorable, and at what temperatures do they exist, and what level of complexity? At larger scales, we'll also need to think about such things in the making of supramolecular assemblies for nanotechnology.

The author, Martin Suhm of Göttingen, calls for a database of the lowest-energy species for each given formula as an invitation for people to break the records. I'd like to see someone give it a try. It would provide challenges for synthesis, spectroscopy and (especially) modeling and computational chemistry.

Comments (13) + TrackBacks (0) | Category: Chemical News | In Silico

November 18, 2013

When the "c" in cLogP Stands For "Crazy" (Update: Fixed!)

Email This Entry

Posted by Derek

I don't know how many readers out there use the cLogP function in ChemDraw, but you might want to take a look at the illustration here before you use it again. A reader alerted me to this glitch: drawing in explicit hydrogens sends it into an even stranger world of fantasy than most calculated logP values inhabit. There seems to be a limit, though, as that cyclopropane series illustrates. (Note that the "logP" function resists temptation).

PerkinElmer says that there are several logP calculating functions built in, but that cLogP is from BioByte. I don't think that they got their money's worth from them. Now, this isn't the first glitch like this I've seen - a lot of chemical drawing and searching programs choke on details here and there. But this is the sort of thing you'd have thought would have been fixed by now.

Update: it has indeed been fixed, according to the fast-acting Phillip Skinner at PerkinElmer. But apparently many of us out in user-land aren't using updated versions. For what it's worth, I have ChemBioDraw Ultra 13.02 v. 3020 - are others getting more reliable numbers with newer software?

Note: here's where the patch can be downloaded.

Comments (32) + TrackBacks (0) | Category: In Silico

October 31, 2013

Models and Reality

Email This Entry

Posted by Derek

I have to admit that I enjoyed seeing this question asked: how come we still use wind tunnels in aerodynamic engineering? Why don't we just model everything in software? The answers, from people who've actually done some of the work, are what you might expect: the models all involve degrees of approximation, gloss over some effects that are sometimes important, can help you but have to be given a reality check, etc. (Turbulent flow is no joke, theoretically or computationally, as any physicist can tell you).

The exact same sorts of answers, with a few nouns swapped out, could be given for the similar question of why we don't just design drugs using computer simulations. I'm asked that often by people outside the field, and have run into many people over the years who assume that it's just the way that drug discovery is done. But no, the continued existence of med-chem departments, such as they are, is testimony to the empirical nature of the business. It's sometimes maddening, but reality can be that way.

Note: every good modeler I've worked with has made it very clear that they know that they're working with approximations of reality - in fact, I think that's a prerequisite for someone to be a good modeler.

Comments (18) + TrackBacks (0) | Category: In Silico

October 29, 2013

Unraveling An Off-Rate

Email This Entry

Posted by Derek

Medicinal chemists talk a lot more about residence time and off rate than they used to. It's become clear that (at least in some cases) a key part of a drug's action is its kinetic behavior, specifically how quickly it leaves its binding site. You'd think that this would correlate well with its potency, but that's not necessarily so. Binding constants are a mix of on- and off-rates, and you can get to the same number by a variety of different means. Only if you're looking at very similar compounds with the same binding modes can you expect the correlation your intuition is telling you about, and even then you don't always get it.

There's a new paper in J. Med. Chem. from a team at Boehringer Ingelheim that takes a detailed look at this effect. The authors are working out the binding to the muscarinic receptor ligand tiotropium, which has been around a long time. (Boehringer's efforts in the muscarinic field have been around a long time, too, come to think of it). Tiotropium binds to the m2 subtype with a Ki of 0.2 nM, and to the m3 subtype with a Ki of 0.1 nM. But the compound has a much slower off rate on the m3 subtype, enough to make it physiologically distinct as an m3 ligand. Tiotropium is better known by its brand name Spiriva, and if its functional selectivity at the m3 receptors in the lungs wasn't pretty tight, it wouldn't be a drug. By carefully modifying its structure and introducing mutations into the receptor, this group hoped to figure out just why it's able to work the way it does.
The static details of tiotropium binding are well worked out - in fact, there's a recent X-ray structure, adding to the list of GPCRs that have been investigated by X-ray crystallography. There are plenty of interactions, as those binding constants would suggest:

The orthosteric binding sites of hM3R and hM2R are virtually identical. The positively charged headgroup of the antimuscarinic agent binds to (in the class of amine receptors highly conserved) Asp3.32 (D1483.32) and is surrounded by an aromatic cage consisting of Y1493.33, W5046.48, Y5076.51, Y5307.39, and Y5347.43. In addition to that, the aromatic substructures of the ligands dig into a hydrophobic region close to W2004.57 and the hydroxy groups, together with the ester groups, are bidentally interacting with N5086.52, forming close to optimal double hydrogen bonds. . .

The similarity of these binding sites was brought home to me personally when I was working on making selective antagonists of these myself. (If you want a real challenge, try differentiating m2 and m4). The authors point out, though, and crucially, that if you want to understand how different compounds bind to these receptors, the static pictures you get from X-ray structures are not enough. Homology modeling helps a good deal, but only if you take its results as indicators of dynamic processes, and not just swapping out residues in a framework.

Doing point-by-point single changes in both the tiotropium structure and the the receptor residues lets you use the kinetic data to your advantage. Such similar compounds should have similar modes of dissociation from the binding site. You can then compare off-rates to the binding constants, looking for the ones that deviate from the expected linear relationship. What they find is that the first event when tiotropium leaves the binding site is the opening of the aromatic cage mentioned above. Mutating any of these residues led to a big effect on the off-rate compared to the effect on the binding constant. Mutations further up along the tunnel leading to the binding site behaved in the same way: pretty much identical Ki values, but enhanced off-rates.

These observations, the paper says with commendable honesty, don't help the medicinal chemists all that much in designing compounds with better kinetics. You can imagine finding a compound that takes better advantage of this binding (maybe), but you can also imagine spending a lot of time trying to do that. The interaction with the asapragine at residue 508 is more useful from a drug design standpoint:

Our data provide evidence that the double hydrogen interaction of N5086.52 with tiotropium has a crucial influence on the off-rates beyond its influence on Ki. Mutation of N5086.52 to alanine accelerates the dissociation of tiotropium more than 1 order of magnitude than suggested by the Ki. Consequently, tiotropium derivatives devoid of the interacting hydroxy group show overproportionally short half-lives. Microsecond MD simulations show that this double hydrogen bonded interaction hinders tiotropium from moving into the exit channel by reducing the frequency of tyrosine-lid opening movements. Taken together, our data show that the interaction with N5086.52 is indeed an essential prerequisite for the development of slowly dissociating muscarinic receptor inverse agonists. This hypothesis is corroborated by the a posteriori observation that the only highly conserved substructure of all long-acting antimuscarinic agents currently in clinical development or already on the market is the hydroxy group.

But the extracellular loops also get into the act. The m2 subtype's nearby loop seems to be more flexible than the one in m3, and there's a lysine in the m3 that probably contributes some electrostatic repulsion to the charged tiotropium as it tries to back out of the protein. That's another effect that's hard to take advantage of, since the charged region of the molecule is a key for binding down in the active site, and messing with it would probably not pay dividends.

But there are some good take-aways from this paper. The authors note that the X-ray structure, while valuable, seems to have large confirmed the data generated by mutagenesis (as well it should). So if you're willing to do lots of point mutations, on both your ligand and your protein, you can (in theory) work some of these fine details out. Molecular dynamics simulations would seem to be of help here, too, also in theory. I'd be interested to hear if people can corroborate that with real-world experience.

Comments (20) + TrackBacks (0) | Category: Drug Assays | In Silico | Pharmacokinetics | The Central Nervous System

October 28, 2013

Molecular Dynamics, Pro and Con

Email This Entry

Posted by Derek

Readers will remember the spirited discussion about the usefulness of molecular dynamics in the comments thread to this post. Now Anthony Nicholls of OpenEye Software, who ignited that particular powder train, is going to defend his position in person this Wednesday. There will be a lunchtime seminar and discussion at the OpenEye offices (222 3rd Street, Suite 3210 in Cambridge). If you're in the area and going, you'll need to contact them to make sure that there's room. There will surely be a report from Ash at Curious Wavefunction if you can't make it, though.

Comments (1) + TrackBacks (0) | Category: In Silico

October 23, 2013

Allosteric Binding Illuminated?

Email This Entry

Posted by Derek

G-protein coupled receptors are one of those areas that I used to think I understood, until I understood them better. These things are very far from being on/off light switches mounted in drywall - they have a lot of different signaling mechanisms, and none of them are simple, either.

One of those that's been known for a long time, but remains quite murky, is allosteric modulation. There are many compounds known that clearly are not binding at the actual ligand site in some types of GPCR, but (equally clearly) can affect their signaling by binding to them somewhere else. So receptors have allosteric sites - but what do they do? And what ligands naturally bind to them (if any)? And by what mechanism does that binding modulate the downstream signaling, and are there effects that we can take advantage of as medicinal chemists? Open questions, all of them.

There's a new paper in Nature that tries to make sense of this, and trying by what might be the most difficult way possible: through computational modeling. Not all that long ago, this might well have been a fool's errand. But we're learning a lot about the details of GPCR structure from the recent X-ray work, and we're also able to handle a lot more computational load than we used to. That's particularly true if we are David Shaw and the D. E. Shaw company, part of the not-all-that-roomy Venn diagram intersection of quantitative Wall Street traders and computational chemists. Shaw has the resources to put together some serious hardware and software, and a team of people to make sure that the processing units get frequent exercise.
They're looking at the muscarinic M2 receptor, an old friend of mine for which I produced I-know-not-how-many antagonist candidates about twenty years ago. The allosteric region is up near the surface of the receptor, about 15A from the acetylcholine binding site, and it looks like all the compounds that bind up there do so via cation/pi interactions with aromatic residues in the protein. (That holds true for compounds as diverse as gallamine, alcuronium, and strychnine), and the one shown in the figure. This is very much in line with SAR and mutagenesis results over the years, but there are some key differences. Many people had thought that the aromatic groups of the ligands the receptors must have been interacting, but this doesn't seem to be the case. There also don't seem to be any interactions between the positively charged parts of the ligands and anionic residues on nearby loops of the protein (which is a rationale I remember from my days in the muscarinic field).

The simulations suggest that the two sites are very much in communication with each other. The width and conformation of the extracellular vestibule space can change according to what allosteric ligand occupies it, and this affects whether the effect on regular ligand binding is positive or negative, and to what degree. There can also, in some cases, be direct electrostatic interactions between the two ligands, for the larger allosteric compounds. I was very glad to see that the Shaw group's simulations suggested some experiments: one set with modified ligands, which would be predicted to affect the receptor in defined ways, and another set with point mutations in the receptor, which would be predicted to change the activities of the known ligands. These experiments were carried out by co-authors at Monash University in Australia, and (gratifyingly) seem to confirm the model. Too many computational papers (and to be fair, too many non-computational papers) don't get quite to the "We made some predictions and put our ideas to the test" stage, and I'm glad this one does.

Comments (14) + TrackBacks (0) | Category: Biological News | In Silico | The Central Nervous System

September 4, 2013

More Thoughts on Compound Metrics

Email This Entry

Posted by Derek

Over at Practical Fragments, Dan Erlanson has comments on the Michael Shultz paper that I wrote about here. He goes into details on some of the problems that turn up when you try to apply various compound metrics across a broad range of molecular weights and/or lipophilicities. In the most obvious example, the indices that are based on Heavy Atom Count (HAC) will jump around much more in the low-molecular-weight range, and none of the proposed refinements can quite fix this. And with the alternative LELP measure, you have to watch out when you're at very low LogP values.

Shultz's preferred LipE/LLE metric avoids that problem, and it's size-independent as well. That part can be either a bug or a feature, depending on your perspective. For the most part, I think that's useful, but in the early stages of fragment optimization, I think that a size-independent measurement is not what you want. The whole point in that stage is to pick the starting points with the most binding for their size, and a well-designed fragment library shouldn't have too many big problems with lipophilicity (those will come along later). So I take Shultz's point about the validity of LLE in general, but I think that I'll be using it and either LE or BEI (HAC-driven and molecular-weight driven) measure of binding efficiency when I'm working in the fragment end of things. How to weight those will be a judgment call, naturally, but judgment calls are, in theory, what we're being paid for, right?

Comments (3) + TrackBacks (0) | Category: Drug Assays | In Silico

August 22, 2013

Too Many Metrics

Email This Entry

Posted by Derek

Here's a new paper from Michael Shultz of Novartis, who is trying to cut through the mass of metrics for new compounds. I cannot resist quoting his opening paragraph, but I do not have a spare two hours to add all the links:

Approximately 15 years ago Lipinski et al. published their seminal work linking molecular properties with oral absorption.1 Since this ‘Big Bang’ of physical property analysis, the universe of parameters, rules and optimization metrics has been expanding at an ever increasing rate (Figure 1).2 Relationships with molecular weight (MW), lipophilicity,3 and 4 ionization state,5 pKa, molecular volume and total polar surface area have been examined.6 Aromatic rings,7 and 8 oxygen atoms, nitrogen atoms, sp3 carbon atoms,9 chiral atoms,9 non-hydrogen atoms, aromatic versus non-hydrogen atoms,10 aromatic atoms minus sp3 carbon atoms,6 and 11 hydrogen bond donors, hydrogen bond acceptors and rotatable bonds12 have been counted and correlated.13 In addition to the rules of five came the rules of 4/40014 and 3/75.15 Medicinal chemists can choose from composite parameters (or efficiency indices) such as ligand efficiency (LE),16 group efficiency (GE), lipophilic efficiency/lipophilic ligand efficiency (LipE17/LLE),18 ligand lipophilicity index (LLEAT),19 ligand efficiency dependent lipophilicity (LELP), fit quality scaled ligand efficiency (LE_scale),20 percentage efficiency index (PEI),21 size independent ligand efficiency (SILE), binding efficiency index (BEI) or surface binding efficiency index (SEI)22 and composite parameters are even now being used in combination.23 Efficiency of binding kinetics has recently been introduced.24 A new trend of anthropomorphizing molecular optimization has occurred as molecular ‘addictions’ and ‘obesity’ have been identified.25 To help medicinal chemists there are guideposts,21 rules of thumb,14 and 26 a property forecast index,27 graphical representations of properties28 such as efficiency maps, atlases,29 ChemGPS,30 traffic lights,31 radar plots,32 Craig plots,33 flower plots,34 egg plots,35 time series plots,36 oral bioavailability graphs,37 face diagrams,28 spider diagrams,38 the golden triangle39 and the golden ratio.40

He must have enjoyed writing that one, if not tracking down all the references. This paper is valuable right from the start just for having gathered all this into one place! But as you read on, you find that he's not too happy with many of these metrics - and since there's no way that they can all be equally correct, or equally useful, he sets himself the task of figuring out which ones we can discard. The last reference in the quoted section below is to the famous "Can a biologist fix a radio?" paper:

While individual composite parameters have been developed to address specific relationships between properties and structural features (e.g. solubility and aromatic ring count) the benefit may be outweighed by the contradictions that arise from utilizing several indices at once or the complexity of adopting and abandoning various metrics depending on the stage of molecular optimization. The average medicinal chemist can be overwhelmed by the ‘analysis fatigue’ that this plethora of new and contradictory tools, rules and visualizations now provide, especially when combined with the increasing number of safety, off-target, physicochemical property and ADME data acquired during optimization efforts. Decision making is impeded when evaluating information that is wrong or excessive and thus should be limited to the absolute minimum and most relevant available.

As Lazebnik described, sometimes the more facts we learn, the less we understand.

And he discards quite a few. All the equations that involve taking the log of potency and dividing by the heavy atom count (HAC), etc., are playing rather loose with the math:

To be valid, LE must remain constant for each heavy atom that changes potency 10-fold. This is not the case as a 15 HAC compound with a pIC50 of 3 does not have the same LE as a 16 HAC compound with a pIC50 of 4 (ΔpIC50 = 1, ΔHAC = 1, ΔLE = 0.07). A 10-fold change in potency per heavy atom does not result in constant LE as defined by Hopkins, nor will it result in a constant SILE, FQ or LLEAT values. These metrics do not mathematically normalize size or potency because they violate the quotient rule of logarithms. To obey this rule and be a valid mathematical function HAC would subtracted from pIC50 and rendered independent of size and reference potency.

Note that he's not recommending that last operation as a guideline, either. Another conceptual problem with plain heavy atom counting is that it treats all atoms the same, but that's clearly an oversimplification. But dividing by some form of molecular weight is an oversimplification, too: a nitrogen differs from an oxygen by a lot more than that 1 mass unit. (This topic came up here a little while back). But oversimplified or not - heck, mathematically valid or not - the question is whether these things help out enough when used as metrics in the real world. And Shultz would argue that they don't. Keeping LE the same (or even raising it) is supposed to be the sign of a successful optimization, but in practice, LE usually degrades. His take on this is that "Since lower ligand efficiency is indicative of both higher and lower probabilities of success (two mutually exclusive states) LE can be invalidated by not correlating with successful optimization."

I think that's too much of a leap - because successful drug programs have had their LE go down during the process, that doesn't mean that this was a necessary condition, or that they should have been aiming for that. Perhaps things would have been even better if they hadn't gone down (although I realize that arguing from things that didn't happen doesn't have much logical force). Try looking at it this way: a large number of successful drug programs have had someone high up in management trying to kill them along the way, as have (obviously) most of the unsuccessful ones. That would mean that upper management decisions to kill a program are also indicative of both higher and lower probabilities of success, and can thus be invalidated, too. Actually, he might be on to something there.

Shultz, though, finds that he's not able to invalidate LipE (or LLE), variously known as ligand-lipophilicity efficiency or lipophilic ligand efficiency. That's p(IC50) - logP, which at least follows the way that logarithms of quotients are supposed to work. And it also has been shown to improve during known drug optimization campaigns. The paper has a thought experiment, on some hypothetical compounds, as well as some data from a tankyrase inhibitor series that seem to show the LipE behave more rationally than other metrics (which sometimes start pointing in opposite directions).

I found the chart below to be quite interesting. It uses the cLogP data from Paul Leeson and Brian Springthorpe's original LLE paper (linked in the above paragraph) to show what change in potency you would expect when you change a hydrogen in your molecule to one of the groups shown if you're going to maintain a constant LipE value. So while hydrophobic groups tend to make things more potent, this puts a number on it. A t-butyl, for example, should make things about 50-fold more potent if it's going to pull its weight as a ball of grease. (Note that we're not talking about effects on PK and tox here, just sheer potency - if you play this game, though, you'd better be prepared to keep an eye on things downstream).
On the other end of the scale, a methoxy should, in theory, cut your potency roughly in half. If it doesn't, that's a good sign. A morpholine should be three or four times worse, and if it isn't, then it's found something at least marginally useful to do in your compound's binding site. What we're measuring here is the partitioning between your compound wanting to be in solution, and wanting to be in the binding site. More specifically, since logP is in the equation, we're looking at the difference in the partitioning of your compound between octanol and water, versus its partitioning between the target protein and water. I think we can all agree that we'd rather have compounds that bind because they like something about the active site, rather than just fleeing the solution phase.

So in light of this paper, I'm rethinking my ligand-efficiency metrics. I'm still grappling with how LipE performs down at the fragment end of the molecular weight scale, and would be glad to hear thoughts on that. But Shultz's paper, if it can get us to toss out a lot of the proposed metrics already in the literature, will have done us all a service.

Comments (38) + TrackBacks (0) | Category: Drug Assays | Drug Development | In Silico | Pharmacokinetics

August 16, 2013

An HIV Structure Breakthrough? Or "Complete Rubbish"?

Email This Entry

Posted by Derek

Structural biology needs no introduction for people doing drug discovery. This wasn't always so. Drugs were discovered back in the days when people used to argue about whether those "receptor" thingies were real objects (as opposed to useful conceptual shorthand), and before anyone had any idea of what an enzyme's active site might look like. And even today, there are targets, and whole classes of targets, for which we can't get enough structural information to help us out much.

But when you can get it, structure can be a wonderful thing. X-ray crystallography of proteins, and protein-ligand complexes has revealed so much useful information that it's hard to know where to start. It's not the magic wand - you can't look at an empty binding site and just design something right at your desk that'll be a potent ligand right off the bat. And you can't look at a series of ligand-bound structures and say which one is the most potent, not in most situations, anyway. But you still learn things from X-ray structures that you could never have known otherwise.

It's not the only game in town, either. NMR structures are very useful, although the X-ray ones can be easier to get, especially in these days of automated synchroton beamlines and powerful number-crunching. But what if your protein doesn't crystallize? And what if there are things happening in solution that you'd never pick up on from the crystallized form? You're not going to watch your protein rearrange into a new ligand-bound conformation with X-ray crystallography, that's for sure. No, even though NMR structures can be a pain to get, and have to be carefully interpreted, they'll also show you things you'd never had seen.

And there are more exotic methods. Earlier this summer, there was a startling report of a structure of the HIV surface proteins gp120 and gp41 obtained through cryogenic electron microscopy. This is a very important and very challenging field to work in. What you've got there is a membrane-bound protein-protein interaction, which is just the sort of thing that the other major structure-determination techniques can't handle well. At the same time, though, the number of important proteins involved in this sort of thing is almost beyond listing. Cryo-EM, since it observes the native proteins in their natural environment, without tags or stains, has a lot of potential, but it's been extremely hard to get the sort of resolution with it that's needed on such targets.

Joseph Sodroski's group at Harvard, longtime workers in this area, published their 6-angstrom-resolution structure of the protein complex in PNAS. But according to this new article in Science, the work has been an absolute lightning rod ever since it appeared. Many other structural biologists think that the paper is so flawed that it never should have seen print. No, I'm not exaggerating:

Several respected HIV/AIDS researchers are wowed by the work. But others—structural biologists in particular—assert that the paper is too good to be true and is more likely fantasy than fantastic. "That paper is complete rubbish," charges Richard Henderson, an electron microscopy pioneer at the MRC Laboratory of Molecular Biology in Cambridge, U.K. "It has no redeeming features whatsoever."

. . .Most of the structural biologists and HIV/AIDS researchers Science spoke with, including several reviewers, did not want to speak on the record because of their close relations with Sodroski or fear that they'd be seen as competitors griping—and some indeed are competitors. Two main criticisms emerged. Structural biologists are convinced that Sodroski's group, for technical reasons, could not have obtained a 6-Å resolution structure with the type of microscope they used. The second concern is even more disturbing: They solved the structure of a phantom molecule, not the trimer.

Cryo-EM is an art form. You have to freeze your samples in an aqueous system, but without making ice. The crystals of normal ice formation will do unsightly things to biological samples, on both the macro and micro levels, so you have to form "vitreous ice", a glassy amorphous form of frozen water, which is odd enough that until the 1980s many people considered it impossible. Once you've got your protein particles in this matrix, though, you can't just blast away at full power with your electron beam, because that will also tear things up. You have to take a huge number of runs at lower power, and analyze them through statistical techniques. The Sodolski HIV structure, for example, is the product of 670,000 single-particle images.

But its critics say that it's also the product of wishful thinking.:

The essential problem, they contend, is that Sodroski and Mao "aligned" their trimers to lower-resolution images published before, aiming to refine what was known. This is a popular cryo-EM technique but requires convincing evidence that the particles are there in the first place and rigorous tests to ensure that any improvements are real and not the result of simply finding a spurious agreement with random noise. "They should have done lots of controls that they didn't do," (Sriram) Subramaniam asserts. In an oft-cited experiment that aligns 1000 computer-generated images of white noise to a picture of Albert Einstein sticking out his tongue, the resulting image still clearly shows the famous physicist. "You get a beautiful picture of Albert Einstein out of nothing," Henderson says. "That's exactly what Sodroski and Mao have done. They've taken a previously published structure and put atoms in and gone down into a hole." Sodroski and Mao declined to address specific criticisms about their studies.

Well, they decline to answer them in response to a news item in Science. They've indicated a willingness to take on all comers in the peer-reviewed literature, but otherwise, in print, they're doing the we-stand-by-our-results-no-comment thing. Sodroski himself, with his level of experience in the field, seems ready to defend this paper vigorously, but there seem to be plenty of others willing to attack. We'll have to see how this plays out in the coming months - I'll update as things develop.

Comments (34) + TrackBacks (0) | Category: Analytical Chemistry | Biological News | In Silico | Infectious Diseases

August 13, 2013

Druggability: A Philosophical Investigation

Email This Entry

Posted by Derek

I had a very interesting email the other day, and my reply to it started getting so long that I thought I'd just turn it into a blog post. Here's the question:

How long can we expect to keep finding new drugs?

By way of analogy, consider software development. In general, it's pretty hard to think of a computer-based task that you couldn't write a program to do, at least in principle. It may be expensive, or may be unreasonably slow, but physical possibility implies that a program exists to accomplish it.

Engineering is similar. If it's physically possible to do something, I can, in principle, build a machine to do it.

But it doesn't seem obvious that the same holds true for drug development. Something being physically possible (removing plaque from arteries, killing all cancerous cells, etc.) doesn't seem like it would guarantee that a drug will exist to accomplish it. No matter how much we'd like a drug for Alzheimer's, it's possible that there simply isn't one.

Is this accurate? Or is the language of chemistry expressive enough that if you can imagine a chemical solution to something, it (in principle) exists. (I don't really have a hard and fast definition of 'drug' here. Obviously all bets are off if your 'drug' is complicated enough to act like a living thing.)

And if it is accurate, what does that say about the long-term prospects for the drug industry? Is there any risk of "running out" of new drugs? Is drug discovery destined to be a stepping-stone until more advanced medical techniques are available?

That's an interesting philosophical point, and one that had never occurred to me in quite that way. I think that's because programming is much more of a branch of mathematics. If you've got a Universal Turing Machine and enough tape to run through it, then you can, in theory, run any program that ever could be run. And any process that can be broken down into handling ones and zeros can be the subject of a program, so the Church-Turing thesis would say that yes, you can calculate it.

But biochemistry is most definitely a different thing, and this is where a lot of people who come into it from the math/CS/engineering side run into trouble. There's a famous (infamous) essay called "Can A Biologist Fix A Radio" that illustrates the point well. The author actually has some good arguments, and some legitimate complaints about the way biochemistry/molecular biology has been approached. But I think that his thesis breaks down eventually, and I've been thinking on and off for years about just where that happens and how to explain what makes things go haywire. My best guess is algorithmic complexity. It's very hard to reduce the behavior of biochemical systems to mathematical formalism. The whole point of formal notation is to express things in the most compact and information-rich way possible, but trying to compress biochemistry in this manner doesn't give you much of an advantage, at least not in the ways we've tried to do it so far.

To get back to the question at hand, let's get philosophical. I'd say that at the most macro level, there are solutions to all the medical problems. After all, we have the example of people who don't have multiple sclerosis, who don't have malaria, who don't have diabetes or pancreatic cancer or what have you. We know that there are biochemical states where these things do not exist; the problem is then to get an individual patient's state back to that situation. Note that this argument does not apply to things like life extension, limb regeneration, and so on: we don't know if humans are capable of these things or not yet, even if there may be some good arguments to be made in their favor. But we know that there are human brains without Alzheimer's.

To move down a level from this, though, the next question is whether there are ways to put a patient's cells and organs back into a disease-free state. In some cases, I think that the answer has to be, for all practical purposes, "No". I tend to think that the later stages of Alzheimer's (for example) are in fact incurable. Neurons are dead and damaged, what was contained in them and in their arrangement is gone, and any repair system can only go so far. Too much information has been lost and too much entropy has been let in. I would like to be wrong about this, but I don't think I am.

But for less severe states and diseases, you can imagine various interventions - chemical, surgical, genetic - that could restore things. So the question here becomes whether there are drug-like solutions. The answer is tricky. If you look at a biochemical mechanism and can see that there's a particular pathway involving small molecules, then certainly, you can say that there could be a molecule to be found as a treatment, even if we haven't found it yet. But the first part of that last sentence has to be unpacked.

Take diabetes. Type I diabetes is proximately caused by lack of insulin, so the solution is to take insulin. And that works, although it's certainly not a cure, since you have to take insuin for the rest of your life, and it's impossible to take it in a way that perfectly mimics the way your body would adminster it, etc. A cure would be to have working beta-cells again that respond just the way they're supposed to, and that's less likely to be achieved through a drug therapy. (Although you could imagine some small molecule that affects a certain class of stem cell, causing it to start the program to differentiate into a fully-formed beta cell, and so on). You'd also want to know why the original population of cells died in the first place, and how to keep that from happening again, which might also take you to some immunological and cell-cycle pathways that could be modulated by drug molecules. But all of these avenues might just as easily take you into genetically modified cloned cell lines and surgical implantation, too, rather than anything involving small-molecule chemistry.

Here's another level of complexity, then: insulin is certainly a drug, but it's not a small molecule of the kind I'd be making. Is there a small molecular that can replace it? You'd do very well with that indeed, but the answer (I think) is "probably not". If you look at the receptor proteins that insulin binds to, the recognition surfaces that are used are probably larger than small molecules can mimic. No one's ever found a small molecule insulin mimetic, and I don't think anyone is likely to. (On the other hand, if you're trying to disrupt a protein-protein interaction, you have more hope, although that's still an extremely difficult target. We can disrupt things a lot more easily than we can make them work). Even if you found a small-molecule-insulin, you'd be faced with the problem of dosing it appropriately, which is no small challenge for a tightly and continuously regulated system like that one. (It's no small challenge for administering insulin itself, either).

And even for mechanisms that do involve small-molecule signaling, like the G-protein coupled receptors, there are still things to worry about. Take schizophrenia. You can definitely see problems with neural systems in the brain when you study that disease, and these neurons respond to, among other things, small-molecue neurotransmitters that the body makes and uses itself - dopamine, serotonin, acetylcholine and others. There are a certain number of receptors for each of those, and although we don't have all the combinations yet, I could imagine, on a philosophical level, that we could eventually have selective drugs that are agonists, antagonists, partial agonists, inverse agonists, what have you at all the subtypes. We have quite a few of them now, for some of the families. And I can even imagine that we could eventually have most or all of the combinations: a molecule that's a dopamine D2 agonist and a muscarinic M4 antagonist, all in one, and so on and so on. That's a lot more of a stretch, to be honest, but I'll stipulate that it's possible.

So you have them all. Now, which ones do you give to help a schizophrenic? We don't know. We have guesses and theories, but most of them are surely wrong. Every biochemical theory about schizophrenia is either wrong or incomplete. We don't know what goes wrong, or why, or how, or what might be done to bend things back in the right direction. It might be that we're in the same area as Alzheimer's: perhaps once a person's brain has developed in such a way that it slips into schizophrenia, that there is no way at all to rewire things, in the same way that we can't ungrow a tree in order to change the shape of its canopy. I've no idea, and we're going to know a lot more about the brain by the time we can answer that one.

So one problem with answering this question is that it's bounded not so much by chemistry as by biology. Lots and lots of biology, most of it unknown. But thinking in terms of sheer chemistry is interesting, too. Consider "The Library of Babel", the famous story by Jorge Luis Borges. It takes place in some sort of universe that is no more (and no less) than a vast library containing every possible book that can be be produced with a 25-character set of letters and punctuation marks. This is, as a bit of reflection will show, a very, very large number, one large enough to contain everything that can possibly be written down. And all the slight variations. And all the misprints. And all the scrambled coded versions of everything, and so on and so on. (W. v. O. Quine extended this idea to binary coding, which brings you back to computability).

Now think about the universe of drug-like molecules. It is also very large, although it is absolutely insignificant compared to the terrifying Library of Babel. (It's worth noting that the Library contains all of the molecules that can ever exist, coded in SMILES strings - that thought just occurred to me at this very moment, and gives me the shivers). The universe of proteins works that way, too - an alphabet of twenty-odd letters for amino acids gives you the exact same situation as the Library, and if you imagine some hideous notation for coding in all the folding variants and post-translational modifications, all the proteins are written down as well.

These, then, encompass everything chemical compound up to some arbitrary size, and the original question is, is this enough? Are there questions for which none of these words are the answer? That takes you into even colder and deeper philosophical waters. Wittgenstein (among many others) wondered the same thing about our own human languages, and seems to have decided that there are indeed things that cannot be expressed, and that this marks the boundary of philosophy itself. Famously, his Tractacus ends with the line "Wovon man nicht sprechen kann, darüber muss man schweigen": whereof we cannot speak, we must pass over in silence.

We're not at that point in the language of chemistry and pharmacology yet, and it's going to be a long, long time before we ever might be. Just the fact, though, that computability seems like such a more reasonable proposition in computer science than druggability does in biochemistry tells you a great deal about how different the two fields are.

Update: On the subject of computabiity, I'm not sure how I missed the chance to bring Gödel's Incompleteness Theorem into this, just to make it a complete stewpot of math and philosophy. But the comments to this post point out that even if you can write a program, you cannot be sure whether it will ever finish the calculation. This Halting Problem is one of the first things ever to be proved formally undecidable, and the issues it raises are very close to those explored by Gödel. But as I understand it, this is decidable for a machine with a finite amount of memory, running a deterministic program. The problem is, though, that it still might take longer than the expected lifetime of the universe to "halt", which leaves you, for, uh, practical purposes, in pretty much the same place as before. This is getting pretty far afield from questions of druggability, though. I think.

Comments (40) + TrackBacks (0) | Category: Drug Development | Drug Industry History | In Silico

August 8, 2013

The 3D Fragment Consortium

Email This Entry

Posted by Derek

Fragment-based screening comes up here fairly often (and if you're interested in the field, you should also have Practical Fragments on your reading list). One of the complaints both inside and outside the fragment world is that there are a lot of primary hits that fall into flat/aromatic chemical space (I know that those two don't overlap perfectly, but you know the sort of things I mean). The early fragment libraries were heavy in that sort of chemical matter, and the sort of collections you can buy still tend to be.

So people have talked about bringing in natural-product-like structures, and diversity-oriented-synthesis structures and other chemistries that make more three-dimensional systems. The commercial suppliers have been catching up with this trend, too, although some definitions of "three-dimensional" may not match yours. (Does a biphenyl derivative count, or is that what you're trying to get away from?)

The UK-based 3D Fragment Consortium has a paper out now in Drug Discovery Today that brings together a lot of references to work in this field. Even if you don't do fragment-based work, I think you'll find it interesting, because many of the same issues apply to larger molecules as well. How much return do you get for putting chiral centers into your molecules, on average? What about molecules with lots of saturated atoms that are still rather squashed and shapeless, versus ones full of aromatic carbons that carve out 3D space surprisingly well? Do different collections of these various molecular types really have differences in screening hit rates, and do these vary by the target class you're screening against? How much are properties (solubility, in particular) shifting these numbers around? And so on.

The consortium's site is worth checking out as well for more on their activities. One interesting bit of information is that the teams ended up crossing off over 90% of the commercially available fragments due to flat structures, which sounds about right. And that takes them where you'd expect it to:

We have concluded that bespoke synthesis, rather than expansion through acquisition of currently available commercial fragment-sized compounds is the most appropriate way to develop the library to attain the desired profile. . .The need to synthesise novel molecules that expand biologically relevant chemical space demonstrates the significant role that academic synthetic chemistry can have in facilitating target evaluation and generating the most appropriate start points for drug discovery programs. Several groups are devising new and innovative methodologies (i.e. methyl activation, cascade reactions and enzymatic functionalisation) and techniques (e.g. flow and photochemistry) that can be harnessed to facilitate expansion of drug discovery-relevant chemical space.

And as long as they stay away from the frequent hitters/PAINS, they should end up with a good collection. I look forward to future publications from the group to see how things work out!

Comments (3) + TrackBacks (0) | Category: Analytical Chemistry | Chemical News | Drug Assays | In Silico

July 31, 2013

Evolving Enzymes: Let 'Em Rip

Email This Entry

Posted by Derek

Evolutionary and genetic processes fascinate many organic chemists, and with good reason. They've provided us with the greatest set of chemical catalysts we know of: enzymes, which are a working example of molecular-level nanotechnology, right in front of us. A billion years of random tinkering have accomplished a great deal, but (being human) we look at the results and wonder if we couldn't do things a bit differently, with other aims in mind than "survive or die".

This has been a big field over the years, and it's getting bigger all the time. There are companies out there that will try to evolve enzymes for you (here's one of the most famous examples), and many academic labs have tried their hands at it as well. The two main routes are random mutations and structure-based directed changes - and at this point, I think it's safe to say that any successful directed-enzyme project has to take advantage of both. There can be just too many possible changes to let random mutations do all the work for you (20 to the Xth power gets out of hand pretty quickly, and that's just the natural amino acids), and we're usually not smart enough to step in and purposefully tweak things for the better every time.

Here's a new paper that illustrates why the field is so interesting, and so tricky. The team (a collaboration between the University of Washington and the ETH in Zürich) has been trying to design a better retro-aldolase enzyme, with earlier results reported here. That was already quite an advance (15,000x rate enhancement over background), but that's still nowhere near natural enzymes of this class. So they took that species as a starting point and did more random mutations around the active site, with rounds of screening in between, which is how we mere humans have to exert selection pressure. This gave a new variant with another lysine in the active site, which some aldolases have already. Further mutational rounds (error-prone PCR and DNA shuffling) and screening let to a further variant that was over 4000x faster than the original enzyme.

But when the team obtained X-ray structures of this enzyme in complex with an inhibitor, they got a surprise. The active site, which had already changed around quite a bit with the addition of that extra lysine, was now a completely different place. A new substrate-binding pocket had formed, and the new lysine was now the catalytic residue all by itself. The paper proposes that the mechanistic competition between the possible active-site residues was a key factor, and they theorize that many natural enzymes may have evolved through similar paths. But given this, there are other questions:

The dramatic changes observed during RA95 evolution naturally prompt the question of whether generation of a highly active retro-aldolase required a computational design step. Whereas productive evolutionary trajectories might have been initiated from random libraries, recent experiments with the same scaffold dem- onstrate that chemical instruction conferred by computation greatly increases the probability of identifying catalysts. Although the programmed mechanisms of other computationally designed enzymes have been generally reinforced and refined by directed evolution, the molecular acrobatics observed with RA95 attest to the functional leaps that unanticipated, innovative mutations—here, replacement of Thr83 by lysine—can initiate.

So they're not ready to turn off the software just yet. But you have to wonder - if there were some way to run the random-mutation process more quickly, and reduce the time and effort of the mutation/screening/selection loop, computational design might well end up playing a much smaller role. (See here for more thoughts on this). Enzymes are capable of things that we would never think of ourselves, and we should always give them the chance to surprise us when we can.

Comments (14) + TrackBacks (0) | Category: Chemical Biology | In Silico

July 29, 2013

More Whitesides on Ligand Binding

Email This Entry

Posted by Derek

George Whitesides and his lab have another paper out on the details of how ligands bind to proteins. They're still using the favorite model enzyme of all time (carbonic anhydrase), the fruit fly and nematode of the protein world. Last time around, using a series of ligands and their analogs with an extra phenyl in their structure. The benzo-ligands had increased affinity, and this seemed to be mostly an enthalpic effect. After a good deal of calorimetry, etc., they concluded that the balancing act between enthalpy and entropy they saw over the group was different for ligand binding than it was for logP partitioning, and that means that it doesn't really match up with the accepted definition of a "hydrophobic effect".

In this study, they're looking at fluorinated analogs of the same compounds to see what that might do to the binding process. That makes the whole thing interesting for a medicinal chemist, because we make an awful lot of fluorinated analogs. You can start some interesting discussions about whether these are more hydrophobic than their non-F analogs, though, and this paper lands right in the middle of that issue.

The first result was that the fluorinated analogs bound to the enzyme (in their X-ray structures) with almost identical geometry. That makes the rest of the discussion easier to draw conclusions from (and more relevant). It's worth remembering, though, that very small changes can still add up. There was a bit of a shift in the binding pocket, actually, which they attribute to an unfavorable interaction between the fluorines and the carbonyl of a threonine residue. But the carbonic anhydrase pocket is pretty accomodating - the overall affinity of the compounds did not really change. That led to this conclusion:

Values of DG8bind, combined with an overall conserved binding geometry of each set of benzo- and fluorobenzo-extended ligands suggest that binding depends on a fine balance of interactions between HCA, the ligand, and molecules of water filling the pocket and surrounding the ligand, and that a simple analysis of interactions between the protein and ligand (Figure1E) is insufficient to understand (or more importantly, predict) the free energy of binding.

But although the overall free energy didn't change, the enthalpic and entropic components did (but arrived at the same place, another example to add to the long list of systems that do this). The differences seem to be in the Coulombic interaction with the binding pocket (worse enthalpy term - is that what shifted the structure over a bit in the X-ray?) and changes in energy of solvation as the ligand binds (better entropy term). Matched pairs of compounds didn't really show a difference in how many waters they displaced from the binding site.

So the take-home is that the hydrophobic effect is not all about releasing waters from protein binding surfaces, as has been proposed by some. It's a mixture of stuff, and especially depends on the structure of the water in the binding pocket and around the ligands, and the changes in these as the compounds leave bulk solvent and find their way into the binding site.

That makes things tricky for many compounds. Hydrophobic effects seem to be a big part of the binding energy of a lot of drug molecules (despite various efforts to cut back on this), and these Whitesides studies would say that modeling and predicting these energetic changes are going to be hard. Computationally, we'd have an easier time figuring out direct interactions between the protein and the ligand, the way we do with enthalpic interactions like hydrogen bonds. Keeping track of all those water molecules is more painful - but necessary.

Comments (10) + TrackBacks (0) | Category: Drug Assays | In Silico

July 17, 2013

MedChemica: When One Compound Collection Isn't Enough

Email This Entry

Posted by Derek

According to SciBx, here's another crack at computational solutions for drug discovery: MedChemica, a venture started by several ex-AstraZeneca scientists. They're going to be working with data from both AZ and Roche, using what sounds like a "matched molecular pair" approach:

Although other algorithms try to relate structure to biological function, most of the analyses look at modifications across a wide array of diverse structures. MedChemica's approach is to look at modifications in a set of similar structures and see how minor differences affect the compounds' biological activity.
Al Dossetter, managing director of MedChemica, said the advantage of the company's platform is the WizePairZ algorithm that looks at pairs of fragments that are similar in structure but differ by a chemical group, such as a change from chlorine to fluorine or the addition of a methyl group.
This platform, he told SciBX, captures the chemical environment of the fragment change. For example, it incorporates the fact that the effect of changing chlorine to fluorine on a molecule will depend on the surrounding structure. The result is a rule that is context dependent.
The MedChemica approach applies to small molecules and uses only partial chemical structures, thus keeping compound identities out of the picture.
Because the platform does not reveal compound identities, AstraZeneca and Roche can share knowledge without disclosing proprietary information.

The belief is that neither company's database on its own gives quite enough statistical power for this approach to work, so they're trying it on the pooled data:

smaller databases only allow researchers to extract one to five matched pairs, which have a low fidelity of prediction. Ten matched pairs are sufficient to draw a prediction, but reliability increases significantly with 20 matched pairs.
The MedChemica database contains 1.2 million datapoints, each of which represents a single molecule fragment in a single assay. It includes 31 different assays, although more are likely to be added in the future, and not all molecules have been tested in all assays.

The article says that AZ and Roche are in discussions with other companies about joining the collaboration. Everyone who joins will get a copy of the pooled database, in addition to being able to share in whatever insights MedChemica comes up with. A limitation is mentioned as well: this is all in vitro data, and its translation to animals or to the clinic provides room to argue.

That's a real concern, I'd say, although I can certainly see why they're doing things the way that they are. It's probably hard enough coming up with in vitro assays across the two companies that are run under similar enough conditions to be usefully paired. In vivo protocols are more varied still, and are notoriously tricky to compare across projects even inside the same company. Just off the top of my head, you have the dosing method (i.v., p.o., etc.), the level of compound given, the vehicle and formulation (a vast source of variability all in itself), the species and strain of animal, the presence of any underlying disease model (versus control animals), what time of day they were dosed and whether they were fed or fasted, whether they were male or female, how old the animals were, and so on and so on. And these factors would be needed just to compare things like PK data, blood levels and so on. If you're talking about toxicology or other effects, there's yet another list of stuff to consider. So yes, the earlier assays will be enough to handle for now.

But will they be enough to provide useful information? Here's where the arguing starts. Limitations of working with only in vitro data aside, you could also say that any trends that are subtle enough to need multi-company-sized pools of data might be too subtle to affect drug discovery very much. The counterargument to that is that some of these rules might still be quite real, but lost in the wilds of chemical diversity space due to lack of effective comparisons. (And the counterargument to that is that if you don't have very many example, how are you so sure that it's a rule?) I'm not sure which side of that one I come down on - "skeptical but willing to listen to data" probably describes me here - but this is the key question that MedChemica will presumably answer, one way or another.

Even so, that in vitro focus is going to be a long-term concern. One of the founders is quoted in the article as saying that the goal is to learn how to predict which compounds shouldn't be made. Fine, but "shouldn't have been made" is a characteristic that's often assigned only after a compound has been dosed in vivo. In the nastier cases, the ones you want to avoid the most, it's only realized after a compound has been in hundreds or thousands of humans in the clinic. The kinds of rules that MedChemica will come up with won't have any bearing on efficacy failures (nor are they meant to), but efficacy failures - failures of biological understanding - are depressingly common. Perhaps they've got a better chance at cutting down the number of "unexplained tox" failures, but that's still a very tall order as well as a very worthy goal.

Falling short of that, I worry, will mean that the MedChemica approach might end up - even if it works - by only optimizing a bit the shortest and cheapest part of the whole drug discovery process, preclinical med-chem. I sympathize - most of my own big ideas, when I get them, bear only on that part of the business, too. But is it the part that needs to be fixed the most? The hope is that there's a connection, but it takes quite a while to prove if one exists.

Comments (8) + TrackBacks (0) | Category: In Silico

June 17, 2013

GPCRs Are As Crazy As You Thought

Email This Entry

Posted by Derek

That's my take-away from this paper, which takes a deep look at a reconstituted beta-adrenergic receptor via fluorine NMR. There are at least four distinct states (two inactive ones, the active one, and an intermediate), and the relationships between them are different with every type of ligand that comes in. Even the ones that look similar turn out to have very different thermodynamics on their way to the active state. If you're into receptor signaling, you'll want to read this one closely - and if you're not, or not up for it, just take away the idea that the landscape is not a simple one. As you'd probably already guessed.

Note: this is a multi-institution list of authors, but it did catch my eye that David Shaw of Wall Street's D. E. Shaw does make an appearance. Good to see him keeping his hand in!

Comments (6) + TrackBacks (0) | Category: Analytical Chemistry | Biological News | In Silico

May 22, 2013

How Many Binding Pockets Are There?

Email This Entry

Posted by Derek

Just how many different small-molecule binding sites are there? That's the subject of this new paper in PNAS, from Jeffrey Skolnick and Mu Gao at Georgia Tech, which several people have sent along to me in the last couple of days.

This question has a lot of bearing on questions of protein evolution. The paper's intro brings up two competing hypotheses of how protein function evolved. One, the "inherent functionality model", assumes that primitive binding pockets are a necessary consequence of protein folding, and that the effects of small molecules on these (probably quite nonspecific) motifs has been honed by evolutionary pressures since then. (The wellspring of this idea is this paper from 1976, by Jensen, and this paper will give you an overview of the field). The other way it might have worked, the "acquired functionality model", would be the case if proteins tend, in their "unevolved" states, to be more spherical, in which case binding events must have been much more rare, but also much more significant. In that system, the very existence of binding pockets themselves is what's under the most evolutionary pressure.

The Skolnick paper references this work from the Hecht group at Princeton, which already provides evidence for the first model. In that paper, a set of near-random 4-helical-bundle proteins was produced in E. coli - the only patterning was a rough polar/nonpolar alternation in amino acid residues. Nonetheless, many members of this unplanned family showed real levels of binding to things like heme, and many even showed above-background levels of several types of enzymatic activity.

In this new work, Skolnick and Gao produce a computational set of artificial proteins (called the ART library in the text), made up of nothing but poly-leucine. These were modeled to the secondary structure of known proteins in the PDB, to produce natural-ish proteins (from a broad structural point of view) that have no functional side chain residues themselves. Nonetheless, they found that the small-molecule-sized pockets of the ART set actually match up quite well with those found in real proteins. But here's where my technical competence begins to run out, because I'm not sure that I understand what "match up quite well" really means here. (If you can read through this earlier paper of theirs at speed, you're doing better than I can). The current work says that "Given two input pockets, a template and a target, (our algorithm) evaluates their PS-score, which measures the similarity in their backbone geometries, side-chain orientations, and the chemical similarities between the aligned pocket-lining residues." And that's fine, but what I don't know is how well it does that. I can see poly-Leu giving you pretty standard backbone geometries and side-chain orientations (although isn't leucine a little more likely than average to form alpha-helices?), but when we start talking chemical similarities between the pocket-lining residues, well, how can that be?

But I'm even willing to go along with the main point of the paper, which is that there are not-so-many types of small-molecule binding pockets, even if I'm not so sure about their estimate of how many there are. For the record, they're guessing not many more than about 500. And while that seems low to me, it all depends on what we mean by "similar". I'm a medicinal chemist, someone who's used to seeing "magic methyl effects" where very small changes in ligand structure can make big differences in binding to a protein. And that makes me think that I could probably take a set of binding pockets that Skolnick's people would call so similar as to be basically identical, and still find small molecules that would differentiate them. In fact, that's a big part of my job.

But in general, I see the point they're making, but it's one that I've already internalized. There are a finite number of proteins in the human body. Fifty thousand? A couple of hundred thousand? Probably not a million. Not all of these have small-molecule binding sites, for sure, so there's a smaller set to deal with right there. Even if those binding sites were completely different from one another, we'd be looking at a set of binding pockets in the thousands/tens of thousands range, most likely. But they're not completely different, as any medicinal chemist knows: try to make a selective muscarinic agonist, or a really targeted serine hydrolase inhibitor, and you'll learn that lesson quickly. And anyone who's run their drug lead through a big selectivity panel has seen the sorts of off-target activities that come up: you hit someof the other members of your target's family to greater or lesser degree. You hit the flippin' sigma receptor, not that anyone knows what that means. You hit the hERG channel, and good luck to you then. Your compound is a substrate for one of the CYP enzymes, or it binds tightly to serum albumin. Who has even seen a compound that binds only to its putative target? And this is only with the counterscreens we have, which is a small subset of the things that are really out there in cells.

And that takes me to my main objection to this paper. As I say, I'm willing to stipulate, gladly, that there are only so many types of binding pockets in this world (although I think that it's more than 500). But this sort of thing is what I have a problem with:

". . .we conclude that ligand-binding promiscuity is likely an inherent feature resulting from the geometric and physical–chemical properties of proteins. This promiscuity implies that the notion of one molecule–one protein target that underlies many aspects of drug discovery is likely incorrect, a conclusion consistent with recent studies. Moreover, within a cell, a given endogenous ligand likely interacts at low levels with multiple proteins that may have different global structures.

"Many aspects of drug discovery" assume that we're only hitting one target? Come on down and try that line out in a drug company, and be prepared for rude comments. Believe me, we all know that our compounds hit other things, and we all know that we don't even know the tenth of it. This is a straw man; I don't know of anyone doing drug discovery that has ever believed anything else. Besides, there are whole fields (CNS) where polypharmacy is assumed, and even encouraged. But even when we're targeting single proteins, believe me, no one is naive enough to think that we're hitting those alone.

Other aspects of this paper, though, are fine by me. As the authors point out, this sort of thing has implications for drawing evolutionary family trees of proteins - we should not assume too much when we see similar binding pockets, since these may well have a better chance of being coincidence than we think. And there are also implications for origin-of-life studies: this work (and the other work in the field, cited above) imply that a random collection of proteins could still display a variety of functions. Whether these are good enough to start assembling a primitive living system is another question, but it may be that proteinaceous life has an easier time bootstrapping itself than we might imagine.

Comments (16) + TrackBacks (0) | Category: Biological News | In Silico | Life As We (Don't) Know It

February 28, 2013

IBM's Watson Does Drug Discovery?

Email This Entry

Posted by Derek

I saw this story this morning, about IBM looking for more markets for its Watson information-sifting system (the one that performed so publicly on "Jeopardy". And this caught my eye for sure:

John Baldoni, senior vice president for technology and science at GlaxoSmithKline, got in touch with I.B.M. shortly after watching Watson’s “Jeopardy” triumph. He was struck that Watson frequently had the right answer, he said, “but what really impressed me was that it so quickly sifted out so many wrong answers.”

That is a huge challenge in drug discovery, which amounts to making a high-stakes bet, over years of testing, on the success of a chemical compound. The failure rate is high. Improving the odds, Mr. Baldoni said, could have a huge payoff economically and medically.

Glaxo and I.B.M. researchers put Watson through a test run. They fed it all the literature on malaria, known anti-malarial drugs and other chemical compounds. Watson correctly identified known anti-malarial drugs, and suggested 15 other compounds as potential drugs to combat malaria. The two companies are now discussing other projects.

“It doesn’t just answer questions, it encourages you to think more widely,” said Catherine E. Peishoff, vice president for computational and structural chemistry at Glaxo. “It essentially says, ‘Look over here, think about this.’ That’s one of the exciting things about this technology.”

Now, without seeing some structures and naming some names, it's completely impossible to say how valuable the Watson suggestions were. But I would very much like to know on what basis these other compounds were suggested: structural similarity? Mechanisms in common? Mechanisms that are in the same pathway, but hadn't been specifically looked at for malaria? Something else entirely? Unfortunately, we're probably not going to be able to find out, unless GSK is forthcoming with more details.

Eventually, there's coing to be another, somewhat more disturbing answer to that "what basis?" question. As this Slate article says, we could well get to the point where such systems make discoveries or correlations that are correct, but beyond our ability to figure out. Watson is most certainly not there yet. I don't think anything is, or is really all that close. But that doesn't mean it won't happen.

For a look at what this might be like, see Ted Chiang's story "Catching Crumbs From the Table", which appeared first in Nature, and then in his collection Stories of Your Life and Others, which I highly recommend, as "The Evolution of Human Science".

Comments (32) + TrackBacks (0) | Category: In Silico | Infectious Diseases

February 8, 2013

All Those Drug-Likeness Papers: A Bit Too Neat to be True?

Email This Entry

Posted by Derek

There's a fascinating paper out on the concept of "drug-likeness" that I think every medicinal chemist should have a look at. It would be hard to count the number of publications on this topic over the last ten years or so, but what if we've been kidding ourselves about some of the main points?

The big concept in this area is, of course, Lipinski criteria, or Rule of Five. Here's what the authors, Peter Kenny and Carlos Montanari of the University of São Paulo, have to say:

No discussion of drug-likeness would be complete without reference to the influential Rule of 5 (Ro5) which is essentially a statement of property distributions for compounds taken into Phase II clinical trials. The focus of Ro5 is oral absorption and the rule neither quantifies the risks of failure associated with non-compliance nor provides guidance as to how sub-optimal characteristics of compliant compounds might be improved. It also raises a number of questions. What is the physicochemical basis of Ro50s asymmetry with respect to hydrogen bond donors and acceptors? Why is calculated octanol/water partition coefficient (ClogP) used to specify Ro50s low polarity limit when the high polarity cut off is defined in terms of numbers of hydrogen bond donors and acceptors? It is possible that these characteristics reflect the relative inability of the octanol/water partitioning system to ‘see’ donors (Fig. 1) and the likelihood that acceptors (especially as defined for Ro5) are more common than donors in pharmaceutically-relevant compounds. The importance of Ro5 is that it raised awareness across the pharmaceutical industry about the relevance of physico- chemical properties. The wide acceptance of Ro5 provided other researchers with an incentive to publish analyses of their own data and those who have followed the drug discovery literature over the last decade or so will have become aware of a publication genre that can be described as ‘retrospective data analysis of large proprietary data sets’ or, more succinctly, as ‘Ro5 envy’.

There, fellow med-chemists, doesn't this already sound like something you want to read? Thought so. Here, have some more:

Despite widespread belief that control of fundamental physicochemical properties is important in pharmaceutical design, the correlations between these and ADMET properties may not actually be as strong as is often assumed. The mere existence of a trend is of no interest in drug discovery and strengths of trends must be known if decisions are to be accurately described as data-driven. Although data analysts frequently tout the statistical significance of the trends that their analysis has revealed, weak trends can be statistically significant without being remotely interesting. We might be confident that the coin that lands heads up for 51 % of a billion throws is biased but this knowledge provides little comfort for the person charged with predicting the result of the next throw. Weak trends can be beaten and when powered by enough data, even the feeblest of trends acquires statistical significance.

So, where are the authors going with all this entertaining invective? (Not that there's anything wrong with that; I'm the last person to complain). They're worried that the transformations that primary drug property data have undergone in the literature have tended to exaggerate the correlations between these properties and the endpoints that we care about. The end result is pernicious:

Correlation inflation becomes an issue when the results of data analysis are used to make real decisions. To restrict values of properties such as lipophilicity more stringently than is justified by trends in the data is to deny one’s own drug-hunting teams room to maneuver while yielding the initiative to hungrier, more agile competitors.

They illustrate this by reference to synthetic data sets, showing how one can get rather different impressions depending on how the numbers are handled along the way. Representing sets of empirical points by using their average values, for example, can cause the final correlations to appear more robust than they really are. That, the authors say, is just what happened in this study from 2006 ("Can we rationally design promiscuous drugs?) and in this one from 2007 ("The influence of drug-like concepts on decision-making in medicinal chemistry"). The complaint is that showing a correlation between cLogP and median compound promiscuity does not imply that there is one between cLogP and compound promiscuity per se. And the authors note that the two papers manage to come to opposite conclusions about the effect of molecular weight, which does make one wonder. The "Escape from flatland" paper from 2009 and the "ADMET rules of thumb" paper from 2008 (mentioned here) also come in for criticism on this point - binning averaged data from a large continuous set and then treated those as real objects for statistic analysis. Ones conclusions depend strongly on how many bins one uses. Here's a specific take on that last paper:

The end point of the G2008 analysis is ‘‘a set of simple interpretable ADMET rules of thumb’’ and it is instructive to examine these more closely. Two classifications (ClogP<4 and MW<400 Da; ClogP>4 or MW>400 Da) were created and these were combined with the four ionization state classifications to define eight classes of compound. Each combination of ADMET property and compound class was labeled according to whether the mean value of the ADMET property was lower than, higher than or not significantly different from the average for all compounds. Although the rules of thumb are indeed simple, it is not clear how useful they are in drug discovery. Firstly, the rules only say whether or not differences are significant and not how large they are. Secondly, the rules are irrelevant if the compounds of interest are all in the same class. Thirdly, the rules predict abrupt changes in ADMET properties going from one class to another. For example, the rules predict significantly different aqueous solubility for two neutral compounds with MW of 399 and 401 Da, provided that their ClogP values do not exceed 4. It is instructive to consider how the rules might have differed had values of logP and MW of 5 and 500 Da (or 3 and 300 Da) had been used to define them instead of 4 and 400 Da.

These problems also occur in graphical representations of all these data, as you'd imagine, and the authors show several of these that they object to. A particular example is this paper from 2010 ("Getting physical in drug discovery"). Three data sets, whose correlations in their primary data do not vary significantly, generate very different looking bar charts. And that leads to this comment:

Both the MR2009 and HY2010 studies note the simplicity of the relationships that the analysis has revealed. Given that drug discovery would appear to be anything but simple, the simplicity of a drug-likeness model could actually be taken as evidence for its irrelevance to drug discovery. The number of aromatic rings in a molecule can be reduced by eliminating rings or by eliminating aromaticity and the two cases appear to be treated as equivalent in both the MR2009 and HY2010 studies. Using the mnemonic suggested in MR2009 one might expect to make a compound more developable by replacing a benzene ring with cyclohexadiene or benzoquinone.

The authors wind up by emphasizing that they're not saying that things like lipophilicity, aromaticity, molecular weight and so on are unimportant - far from it. What they're saying, though, is that we need to be aware of how strong these correlations really are so that we don't fool ourselves into thinking that we're addressing our problems, when we really aren't. We might want to stop looking for huge, universally applicable sets of rules and take what we can get in smaller, local data sets within a given series of compounds. The paper ends with a set of recommendations for authors and editors - among them, always making primary data sets part of the supplementary material, not relying on purely graphical representations to make statistical points, and a number of more stringent criteria for evaluating data that have been partitioned into bins. They say that they hope that their paper "stimulates debate", and I think it should do just that. It's certainly given me a lot of things to think about!

Comments (14) + TrackBacks (0) | Category: Drug Assays | Drug Development | In Silico | The Scientific Literature

January 30, 2013

Farewell to Bioinformatics

Email This Entry

Posted by Derek

Here are some angry views that I don't necessarily endorse, but I can't say that they're completely wrong, either. A programmer bids an angry farewell to the bioinformatics world:

Bioinformatics is an attempt to make molecular biology relevant to reality. All the molecular biologists, devoid of skills beyond those of a laboratory technician, cried out for the mathematicians and programmers to magically extract science from their mountain of shitty results.

And so the programmers descended and built giant databases where huge numbers of shitty results could be searched quickly. They wrote algorithms to organize shitty results into trees and make pretty graphs of them, and the molecular biologists carefully avoided telling the programmers the actual quality of the results. When it became obvious to everyone involved that a class of results was worthless, such as microarray data, there was a rush of handwaving about “not really quantitative, but we can draw qualitative conclusions” followed by a hasty switch to a new technique that had not yet been proved worthless.

And the databases grew, and everyone annotated their data by searching the databases, then submitted in turn. No one seems to have pointed out that this makes your database a reflection of your database, not a reflection of reality. Pull out an annotation in GenBank today and it’s not very long odds that it’s completely wrong.

That's unfair to molecular biologists, but is it unfair to the state of bioinformatic databases? Comments welcome. . .

Update: more comments on this at Ycombinator.

Comments (62) + TrackBacks (0) | Category: Biological News | In Silico

January 28, 2013

The Hydrophobic Effect: I Don't Understand It, Either

Email This Entry

Posted by Derek

We medicinal chemists talk a good game when it comes to the the hydrophobic effect. It's the way that non-water-soluble molecules (or parts of molecules) like to associate with each other, right? Sure thing. And it works because of. . .well, van der Waals forces. Or displacement of water molecules from protein surfaces. Or entropic effects. Or all of those, plus some other stuff that, um, complicated to explain. Something like that.

Here's a paper in Angewandte Chemie that really bears down on the topic. The authors study the binding of simple ligands to thermolysin, a well-worked-out system for which very high-resolution X-ray structures are available. And what they find is, well, that things really are complicated to explain:

In summary, there are no universally valid reasons why the hydrophobic effect should be predominantly “entropic” or “enthalpic”; small structural changes in the binding features of water molecules on the molecular level determine whether hydrophobic binding is enthalpically or entropically driven.

Admittedly, this study reaches the limits of experimental accuracy accomplishable in contemporary protein–ligand structural work. . .Surprising pairwise systematic changes in the thermodynamic data are experienced for complexes of related ligands, and they are convincingly well reflected by the structural properties. The present study unravels small but important details. Computational methods simulate molecular properties at the atomic level, and are usually determined by the summation of many small details. However, details such as those observed here are usually not regarded by these computational methods as relevant, simply because we are not fully aware of their importance for protein–ligand binding, structure–activity relationships, and rational drug design in general. . .

I think that there are a lot of things in this area of which we're not fully aware. There are many others that we treat as unified phenomena, because we've given them names that make us imagine that they are. The hydrophobic effect is definitely one of these - George Whitesides is right when he says that there are many of them. But when all of these effects, on closer inspection, break down into tiny, shifting, tricky arrays of conflicting components, can you blame us for simplifying?

Comments (14) + TrackBacks (0) | Category: "Me Too" Drugs | Chemical News | In Silico

January 17, 2013

Halogen Bonds

Email This Entry

Posted by Derek

Here's a recent paper in J. Med. Chem. on halogen bonding in medicinal chemistry. I find the topic interesting, because it's an effect that certainly appears to be real, but is rarely (if ever) exploited in any kind of systematic way.

Halogens, especially the lighter fluorine and chlorine, are widely used substituents in medicinal chemistry. Until recently, they were merely perceived as hydrophobic moieties and Lewis bases in accordance with their electronegativities. Much in contrast to this perception, compounds containing chlorine, bromine, or iodine can also form directed close contacts of the type R–X···Y–R′, where the halogen X acts as a Lewis acid and Y can be any electron donor moiety. . .

What seems to be happening is that the electron density around the halogen atom is not as smooth as most of us picture it. You'd imagine a solid cloud of electrons around the bromine atom of a bromoaromatic, but in reality, there seems to be a region of slight positivecharge (the "sigma hole") out on the far end. (As a side effect, this give you more of a circular stripe of negative charge as well). Both these effects have been observed experimentally.

Now, you're not going to see this with fluorine; that one is more like most of us picture it (and to be honest, fluorine's weird enough already). But as you get heavier, things become more pronounced. That gives me (and probably a lot of you) an uneasy feeling, because traditionally we've been leery of putting the heavier halogens into our molecules. "Too much weight and too much hydrophobicity for too little payback" has been the usual thinking, and often that's true. But it seems that these substituents can actually earn out their advance in some cases, and we should be ready to exploit those, because we need all the help we can get.

Interestingly, you can increase the effect by adding more fluorines to the haloaromatic, which emphasizes the sigma hole. So you have that option, or you can take a deep breath, close your eyes, and consider. . .iodos:

Interestingly, the introduction of two fluorines into a chlorobenzene scaffold makes the halogen bond strength comparable to that of unsubstituted bromobenzene, and 1,3-difluoro-5-bromobenzene and unsubstituted iodobenzene also have a comparable halogen bond strength. While bromo and chloro groups are widely employed substituents in current medicinal chemistry, iodo groups are often perceived as problematic. Substituting an iodoarene core by a substituted bromoarene scaffold might therefore be a feasible strategy to retain affinity by tuning the Br···LB (Lewis base) halogen bond to similar levels as the original I···LB halogen bond.

As someone who values ligand efficiency, the idea of putting in an iodine gives me the shivers. A fluoro-bromo combo doesn't seem much more attractive, although almost anything looks good compared to a single atom that adds 127 mass units at a single whack. But I might have to learn to love one someday.

The paper includes a number of examples of groups that seem to be capable of interacting with halogens, and some specific success stories from recent literature. It's probably worth thinking about these things similarly to the way we think about hydrogen bonds - valuable, but hard to obtain on purpose. They're both directional, and trying to pick up either one can cause more harm than good if you miss. But keep an eye out for something in your binding site that might like a bit of positive charge poking at it. Because I can bet that you never thought to address it with a bromine atom!

Update: in the spirit of scientific inquiry, I've just sent in an iodo intermediate from my current work for testing in the primary assay. It's not something I would have considered doing otherwise, but if anyone gives me any grief, I'll tell them that it's 2013 already and I'm following the latest trends in medicinal chemistry.

Comments (16) + TrackBacks (0) | Category: Chemical Biology | Chemical News | In Silico

January 14, 2013

Virtual Screening, The Good Parts and the Bad Ones

Email This Entry

Posted by Derek

Virtual screening is what many people outside the field are thinking of when they talk about the use of computational models in drug discovery. There are many other places where modeling can pitch in, but one of the dreams has always been to take a given protein target and a long list of chemical structures, hit the button, and come back to a sorted list of which ones are going to bind well. That list could be as long as "every compound in our screening deck", or "every available compound in the commercial catalogs", or "everything that our chemists can think of and draw on a whiteboard, whether it's ever been made or not". So these virtual collections can get rather large, but that's what computer power is for, right?

Despite what some people might think, we're not exactly there yet. But we're not exactly not there, either, if you know what I mean. Like much of drug discovery, it's in that awkward age. Virtual screening is certainly real, and it can be useful, but it can also waste your time if you're not careful. And that's where this paper comes in - it's a fine overview of the issue that you need to think about if you're interested in trying this technique.

For one thing, you need to decide if you're going to be taking a drug target whose structure you know pretty well and modeling a bunch of small compounds into it, or if you're taking a bunch of small molecules whose activities you know pretty well and trying to find more compounds like them. These two approaches call for some different methods, and have different potential problems. The second one, especially in the older literature, often goes under the name of QSAR, for quantitative structure-activity relationship. But as the authors point out, "virtual screening" as a name has some advantages, because many people have been burned by things labeled "QSAR" over the years. They're also being used for different purposes, which is probably a good thing:

A fundamental assumption inherent in QSAR and pharmacophore-based VS is the “similar property principle”, that is, the general observation that molecules with similar structure are likely to have similar properties. While this assumption holds true in many cases, there are many counter-examples in the field of QSAR which lead to erroneous predictions and can shake the confidence of the experimental community in the prospective utility of QSAR modeling. Interestingly, this has not yet (or not to the same extent) been the case with VS. The difference is that QSAR is typically employed to evaluate a limited number of synthetic candidates, where errors are more noticeable and costly. However, when these techniques are applied on a massive scale to screen large chemical libraries, errors are much more easily tolerated as the objective is to increase the number and diversity of hits over what would have been otherwise a random selection.

The authors extensively cover the previous literature on computational screening - successful examples, warnings of trouble, theoretical predictions both optimistic and pessimistic. It would take you quite a while to assemble this list on your own, so that by itself recommends this paper to anyone interested in the area. But they go on to codify the various pitfalls to look out for.

"Such as expecting it to work", the cynics in the audience will remark. I say that sort of thing under my breath for time to time myself - or audibly, as the case may be. But this is the sort of paper that I can really endorse, because it's a completely realistic view of what you can expect with current technology. And that comes down to "Less than you want", but still "More than you might think". You're not going to able to feed the software the complete pile of all the chemical supplier catalogs and come back to find the nanomolar leads printing out. But you can get pointers toward parts of chemical space that you wouldn't have thought about (or wouldn't have been able to physically screen).

One tricky part is that when a virtual screening effort is successful (for whatever value you assign to "success"), it can be hard to tell why, and likewise for failures. There are so many places where things can disconnect - proteins are mobile, and small molecules even more so, and accounting for these conformational ensembles is not trivial. Binding interactions are not always well understood, or well modeled. Water molecules are pesky, but can be vitally important. You might have picked inappropriate controls (positive or negative), or be weighting the various computed factors in the wrong way. Either of those will send your calculation further and further off the rails.

And so on. The paper goes into detail on these possibilities and more; I highly recommend it for anyone getting into virtual screening (or for anyone already doing it, to keep the troubleshooting guide in one handy place).

Comments (2) + TrackBacks (0) | Category: In Silico

January 10, 2013

Automated Ligand Design?

Email This Entry

Posted by Derek

There's a paper out in Nature with the provocative title of "Automated Design of Ligands to Polypharmcological Profiles". Admittedly, to someone outside my own field of medicinal chemistry, that probably sounds about as dry as the Atacama desert, but it got my attention.

It's a large multi-center contribution, but the principal authors are Andrew Hopkins at Dundee and Bryan Roth at UNC-Chapel Hill. Using James Black's principle that the best place to find a new drug is to start with an old drug, what they're doing here is taking known ligands and running through a machine-learning process to see if they can introduce new activities into them. Now, those of us who spend time trying to take out other activities might wonder what good this is, but there are a some good reasons: for one thing, many CNS agents are polypharmacological to start with. And there certainly are situations where you want dual-acting compounds, CNS or not, which can be a major challenge. And read on - you can run things to get selectivity, too.

So how well does their technique work? The example they give starts with the cholinesterase inhibitor donepezil (sold as Aricept), which has a perfectly reasonable med-chem look to its structure. The groups' prediction, using their current models, was the it had a reasonable chance of having D4 dopaminergic activity, but probably not D2 (which numbers were borne out by experiment, and might have something to do with whatever activity Aricept has for Alzheimer's). I'll let them describe the process:

We tested our method by evolving the structure of donepezil with the dual objectives of improving D2 activity and achieving blood–brain barrier penetration. In our approach the desired multi-objective profile is defined a priori and then expressed as a point in multi-dimensional space termed ‘the ideal achievement point’. In this first example the objectives were simply defined as two target properties and therefore the space has two dimensions. Each dimension is defined by a Bayesian score for the predicted activity and a combined score that describes the absorption, distribution, metabolism and excretion (ADME) properties suitable for blood–brain barrier penetration (D2 score = 100, ADME score = 50). We then generated alternative chemical structures by a set of structural transformations using donepezil as the starting structure. The population was subsequently enumerated by applying a set of transformations to the parent compound(s) of each generation. In contrast to rules-based or synthetic-reaction-based approaches for generating chemical structures, we used a knowledge-based approach by mining the medicinal chemistry literature. By deriving structural transformations from medicinal chemistry, we attempted to mimic the creative design process.

Hmm. They rank these compounds in multi-dimensional space, according to distance from the ideal end point, filter them for chemical novelty, Lipinski criteria, etc., and then use the best structures as starting points for another round. This continues until you reach close enough to the desired point, or until you dead-end on improvement. In this case, they ended up with fairly active D2 compounds, by going to a lactam in the five-membered ring, lengthening the chain a bit, and going to an arylpiperazine on the end. They also predicted, though, that these compounds would hit a number of other targets, which they indeed did on testing.

How about something a bit more. . .targeted? They tried taking these new compounds through another design loop, this time trying to get rid of all the alpha-adrenergic activity they'd picked up, while maintaining the 5-HT1A and dopamine receptor activity they now had. They tried it both ways - running the algorithms with filtration of the alpha-active compounds at each stage, and without. Interestingly, both optimizations came up with very similar compounds, differing only out on the arylpiperazine end. The alpha-active series wanted ortho-methoxyphenyl on the piperazine, while the alpha-inactive series wanted 2-pyridyl. These preferences were confirmed by experiment as well. Some of you who've worked on adrenergics might be saying "Well, yeah, that's what the receptors are already known to prefer, so what's the news here?" But keep in mind, what the receptors are known to prefer is what's been programmed into this process, so of course, that's what it's going to recapitulate. The idea is for the program to keep track of all the known activities - the huge potential SAR spreadsheet - so you don't have to try to do it yourself, with you own grey matter.

The last example asks whether, starting from donezepil, potent and selective D4 compounds could be evolved. I'm going to reproduce the figure from the paper here, to give an idea of the synthetic transformations involved:
So, donezepil (compound 1) is 614 nM against D4, and after a few rounds of optimization, you get structure 13, which is 9 nM. Not bad! Then if you take 13 as a starting point, and select for structural novelty along the way, you get 18 (five micromolar against D4), 20, 21, and (S)-27 (which is 90 nM at D4). All of these compounds picked up a great deal more selectivity for D4 compared to the earlier donezepil-derived scaffolds as well.

Well, then, are we all out of what jobs we have left? Not just yet. You'll note that the group picked GPCRs as a field to work in, partly because there's a tremendous amount known about their SAR preferences and cross-functional selectivities. And even so, of the 800 predictions made in the course of this work, the authors claim about a 75% success rate - pretty impressive, but not the All-Seeing Eye, quite yet. I'd be quite interested in seeing these algorithms tried out on kinase inhibitors, another area with a wealth of such data. But if you're dwelling among the untrodden ways, like Wordsworth's Lucy, then you're pretty much on your own, I'd say, unless you 're looking to add in some activity in one of the more well-worked-out classes.

But knowledge piles up, doesn't it? This approach is the sort of thing that will not be going away, and should be getting more powerful and useful as time goes on. I have no trouble picturing an eventual future where such algorithms do a lot of the grunt work of drug discovery, but I don't foresee that happened for a while yet. Unless, of course, you do GPCR ligand drug discovery. In that case, I'd be contacting the authors of this paper as soon as possible, because this looks like something you need to be aware of.

Comments (12) + TrackBacks (0) | Category: Drug Assays | In Silico | The Central Nervous System

December 11, 2012

Did Kaggle Predict Drug Candidate Activities? Or Not?

Email This Entry

Posted by Derek

I notied this piece on Slate (originally published in New Scientist) about Kaggle, a company that's working on data-prediction algorithms. Actually, it might be more accurate to say that they're asking other people to work on data-prediction algorithems, since they structure their tasks as a series of open challenges, inviting all comers to submit their best shots via whatever computational technique they think appropriate.

PA: How exactly do these competitions work?
JH: They rely on techniques like data mining and machine learning to predict future trends from current data. Companies, governments, and researchers present data sets and problems, and offer prize money for the best solutions. Anyone can enter: We have nearly 64,000 registered users. We've discovered that creative-data scientists can solve problems in every field better than experts in those fields can.

PA: These competitions deal with very specialized subjects. Do experts enter?
JH: Oh yes. Every time a new competition comes out, the experts say: "We've built a whole industry around this. We know the answers." And after a couple of weeks, they get blown out of the water.

I have a real approach-avoidance conflict with this sort of thing. I tend to root for outsiders and underdogs, but naturally enough, when they're coming to blow up what I feel is my own field of expertise, that's a different story, right? And that's just what this looks like: the Merck Molecular Activity Challenge, which took place earlier this fall. Merck seems to have offered up a list of compounds of known activity in a given assay, and asked people to see if they could recapitulate the data through simulation.

Looking at the data that were made available, I see that there's a training set and a test set. They're furnished as a long run of molecular descriptors, but the descriptors themselves are opaque, no doubt deliberately (Merck was not interested in causing themselves any future IP problems with this exercise). The winning team was a group of machine-learning specialists from the University of Toronto and the University of Washington. If you'd like to know a bit more about how they did it, here you go. No doubt some of you will be able to make more of their description than I did.

But I would be very interested in hearing some more details on the other end of things. How did the folks at Merck feel about the results, with the doors closed and the speaker phone turned off? Was it better or worse than what they could have come up with themselves? Are they interested enough in the winning techniques that they've approached the high-ranking groups with offers to work on virtual screening techniques? Because that's what this is all about: running a (comparatively small) test set of real molecules past a target, and then switching to simulations and screening as much of small molecule chemical space as you can computationally stand. Virtual screening is always promising, always cost-attractive, and sometimes quite useful. But you never quite know when that utility is going to manifest itself, and when it's going to be another goose hunt. It's a longstanding goal of computational drug design, for good reason.

So, how good was this one? That also depends on the data set that was used, of course. All of these algorithm-hunting methods can face a crucial dependence on the training sets used, and their relations to the real data. Never was "Garbage In, Garbage Out" more appropriate. If you feed in numbers that are intrinsically too well-behaved, you can emerge with a set of rules that look rock-solid, but will take ou completely off into the weeds when faced with a more real-world situation. And if you go to the other extreme, starting with wooly multi-binding-mode SAR with a lot of outliers and singletons in it, you can end up fitting equations to noise and fantasies. That does no one any good, either.

Back last year, I talked about the types of journal article titles that make me keep on scrolling past them, and invited more. One of the comments suggested "New and Original strategies for Predictive Chemistry: Why use knowledge when fifty cross-correlated molecular descriptors and a consensus of over-fit models will tell you the same thing?". What I'd like to know is, was this the right title for this work, or not?

Comments (28) + TrackBacks (0) | Category: In Silico

November 26, 2012

Chemistry Software Questions Here

Email This Entry

Posted by Derek

As mentioned the other day, this will be a post for people to ask questions directly to Philip Skinner (SDBioBrit) of Perkin-Elmer/Cambridgesoft. He's doing technical support for ChemDraw, ChemDraw4Excel, E-Notebook, Inventory, Registration, Spotfire, Chem3D, etc., and will be monitoring the comments and posting there. Hope it helps some people out!

Note - he's out on the West Coast of the US, so allow the poor guy time to get up and get some coffee in him!

Comments (76) + TrackBacks (0) | Category: Chemical News | In Silico

August 22, 2012

Watch that Little Letter "c"

Email This Entry

Posted by Derek

Hang around a bunch of medicinal chemists (no, really, it's more fun than you'd think) and you're bound to hear discussion of cLogP. For the chemists in the crowd, I should warn you that I'm about to say nasty things about it.

For the nonchemists in the crowd, logP is a measure of how greasy (or how polar) a compound is. It's based on a partition experiment: shake up a measured amount of a compound with defined volumes of water and n-octanol, a rather greasy solvent which I've never seen referred to in any other experimental technique. Then measure how much of the compound ends up in each layer, and take the log of the octanol/water ratio. So if a thousand times as much compound goes into the octanol as goes into the water (which for drug substances is quite common, in fact, pretty good), then the logP is 3. The reason we care about this is that really greasy compounds (and one can go up to 4, 5, 6, and possibly beyond), have problems. They tend to dissolve poorly in the gut, have problems crossing membranes in living systems, get metabolized extensively in the liver, and stick to a lot of proteins that you'd rather they didn't stick to. Fewer high-logP compounds are capable of making it as drugs.

So far, so good. But there are complications. For one thing, that description above ignores the pH of the water solution, and for charged compounds that's a big factor. logD is the term for the distribution of all species (ionized or not), and logD at pH 7.4 (physiological) is a valuable measurement if you've got the possibility of a charged species (and plenty of drug molecules do, thanks to basic amines, carboxylic acids, etc.) But there are bigger problems.

You'll notice that the experiment outlined in the second paragraph could fairly be described as tedious. In fact, I have never seen it performed. Not once, and I'll bet that the majority of medicinal chemists never have, either. And it's not like it's just being done out of my sight; there's no roomful of automated octanol/water extraction machines clanking away in the basement. I should note that there are other higher-throughput experimental techniques (such as HPLC retention times) that also correlate with logP and have been used to generate real numbers, but even those don't account for the great majority of the numbers that we talk about all the time. So how do we manage to do that?

It has to do with a sleight of hand I've performed while writing the above sections, which some of you have probably already noticed. Most of the time, when we talk about logP values in early drug discovery, we're talking about cLogp. That "c" stands for calculated. There are several programs that estimate logP based on known values for different rings and functional groups, and with different algorithms for combining and interpolating them. In my experience, almost all logP numbers that get thrown around are from these tools; no octanol is involved.

And sometimes that worries me a bit. Not all of these programs will tell you how solid those estimates are. And even if they will, not all chemists will bother to check. If your structure is quite close to something that's been measured, then fine, the estimate is bound to be pretty good. But what if you feed in a heterocycle that's not in the lookup table? The program will spit out a number, that's what. But it may not be a very good number, even if it goes out to two decimal places. I can't even remember when I might have last seen a cLogP value with a range on it, or any other suggestion that it might be a bit fuzzy.

There are more subtle problems, too - I've seen some oddities with substitutions on saturated heterocyclic rings (morpholine, etc.) that didn't quite seem to make sense. Many chemists get these numbers, look at them quizzically, and say "Hmm, I didn't know that those things sorted out like that. Live and learn!" In other words, they take the calculated values as reality. I've even had people defend these numbers by explaining to me patiently that these are, after all, calculated logP values, and the calculated log P values rank-order like so, and what exactly is my problem? And while it's hard to argue with that, we are not putting our compounds into the simulated stomachs of rationalized rodents. Real-world decisions can be made based on numbers that do not come from the real world.

Comments (39) + TrackBacks (0) | Category: Drug Assays | In Silico | Life in the Drug Labs

June 12, 2012

Predicting Toxicology

Email This Entry

Posted by Derek

One of the major worries during a clinical trial is toxicity, naturally. There are thousands of reasons a compound might cause problem, and you can be sure that we don't have a good handle on most of them. We screen for what we know about (such as hERG channels for cardiovascular trouble), and we watch closely for signs of everything else. But when slow-building low-incidence toxicity takes your compound out late in the clinic, it's always very painful indeed.

Anything that helps to clarify that part of the business is big news, and potentially worth a lot. But advanced in clinical toxicology come on very slowly, because the only thing worse than not knowing what you'll find is thinking that you know and being wrong. A new paper in Nature highlights just this problem. The authors have a structural-similarity algorithm to try to test new compounds against known toxicities in previously tested compounds, which (as you can imagine) is an approach that's been tried in many different forms over the years. So how does this one fare?

To test their computational approach, Lounkine et al. used it to estimate the binding affinities of a comprehensive set of 656 approved drugs for 73 biological targets. They identified 1,644 possible drug–target interactions, of which 403 were already recorded in ChEMBL, a publicly available database of biologically active compounds. However, because the authors had used this database as a training set for their model, these predictions were not really indicative of the model's effectiveness, and so were not considered further.

A further 348 of the remaining 1,241 predictions were found in other databases (which the authors hadn't used as training sets), leaving 893 predictions, 694 of which were then tested experimentally. The authors found that 151 of these predicted drug–target interactions were genuine. So, of the 1,241 predictions not in ChEMBL, 499 were true; 543 were false; and 199 remain to be tested. Many of the newly discovered drug–target interactions would not have been predicted using conventional computational methods that calculate the strength of drug–target binding interactions based on the structures of the ligand and of the target's binding site.

Now, some of their predictions have turned out to be surprising and accurate. Their technique identified chlorotrianisene, for example, as a COX-1 inhibitor, and tests show that it seems to be, which wasn't known at all. The classic antihistamine diphenhydramine turns out to be active at the serotonin transporter. It's also interesting to see what known drugs light up the side effect assays the worst. Looking at their figures, it would seem that the topical antiseptic chlorhexidine (a membrane disruptor) is active all over the place. Another guanidine-containing compound, tegaserod, is also high up the list. Other promiscuous compounds are the old antipsychotic fluspirilene and the semisynthetic antibiotic rifaximin. (That last one illustrates one of the problems with this approach, which the authors take care to point out: toxicity depends on exposure. The dose makes the poison, and all that. Rifaximin is very poorly absorbed, and it would take very unusual dosing, like with a power drill, to get it to hit targets in a place like the central nervous system, even if this technique flags them).

The biggest problem with this whole approach is also highlighted by the authors, to their credit. You can see from those figures above that about half of the potentially toxic interactions it finds aren't real, and you can be sure that there are plenty of false negatives, too. So this is nowhere near ready to replace real-world testing; nothing is. But where it could be useful is in pointing out things to test with real-world assays, activities that you probably hadn't considered at all.

But the downside of that is that you could end up chasing meaningless stuff that does nothing but put the fear into you and delays your compound's development, too. That split, "stupid delay versus crucial red flag", is at the heart of clinical toxicology, and is the reason it's so hard to make solid progress in this area. So much is riding on these decisions: you could walk away from a compound, never developing one that would go on to clear billions of dollars and help untold numbers of patients. Or you could green-light something that would go on to chew up hundreds of millions of dollars of development costs (and even more in opportunity costs, considering what you could have been working on instead), or even worse, one that makes it onto the market and has to be withdrawn in a blizzard of lawsuits. It brings on a cautious attitude.

Comments (21) + TrackBacks (0) | Category: Drug Development | In Silico | Toxicology

April 4, 2012

The Artificial Intelligence Economy?

Email This Entry

Posted by Derek

Now here's something that might be about to remake the economy, or (on the other robotic hand) it might not be ready to just yet. And it might be able to help us out in drug R&D, or it might turn out to be mostly beside the point. What the heck am I talking about, you ask? The so-called "Artificial Intelligence Economy". As Adam Ozimek says, things are looking a little more futuristic lately.

He's talking about things like driverless cars and quadrotors, and Tyler Cowen adds the examples of things like Apple's Siri and IBM's Watson, as part of a wider point about American exports:

First, artificial intelligence and computing power are the future, or even the present, for much of manufacturing. It’s not just the robots; look at the hundreds of computers and software-driven devices embedded in a new car. Factory floors these days are nearly empty of people because software-driven machines are doing most of the work. The factory has been reinvented as a quiet place. There is now a joke that “a modern textile mill employs only a man and a dog—the man to feed the dog, and the dog to keep the man away from the machines.”

The next steps in the artificial intelligence revolution, as manifested most publicly through systems like Deep Blue, Watson and Siri, will revolutionize production in one sector after another. Computing power solves more problems each year, including manufacturing problems.

Two MIT professors have written a book called Race Against the Machine about all this, and it appears to be sort of a response to Cowen's earlier book The Great Stagnation. (Here's an article of theirs in The Atlantic making their case).

One of the export-economy factors that it (and Cowen) bring up is that automation makes a country's wages (and labor costs in general) less of a factor in exports, once you get past the capital expenditure. And as the size of that expenditure comes down, it becomes easier to make that leap. One thing that means, of course, is that less-skilled workers find it harder to fit in. Here's another Atlantic article, from the print magazine, which looked at an auto-parts manufacturer with a factory in South Carolina (the whole thing is well worth reading):

Before the rise of computer-run machines, factories needed people at every step of production, from the most routine to the most complex. The Gildemeister (machine), for example, automatically performs a series of operations that previously would have required several machines—each with its own operator. It’s relatively easy to train a newcomer to run a simple, single-step machine. Newcomers with no training could start out working the simplest and then gradually learn others. Eventually, with that on-the-job training, some workers could become higher-paid supervisors, overseeing the entire operation. This kind of knowledge could be acquired only on the job; few people went to school to learn how to work in a factory.
Today, the Gildemeisters and their ilk eliminate the need for many of those machines and, therefore, the workers who ran them. Skilled workers now are required only to do what computers can’t do (at least not yet): use their human judgment.

But as that article shows, more than half the workers in that particular factory are, in fact, rather unskilled, and they make a lot more than their Chinese counterparts do. What keeps them employed? That calculation on what it would take to replace them with a machine. The article focuses on one of those workers in particular, named Maddie:

It feels cruel to point out all the Level-2 concepts Maddie doesn’t know, although Maddie is quite open about these shortcomings. She doesn’t know the computer-programming language that runs the machines she operates; in fact, she was surprised to learn they are run by a specialized computer language. She doesn’t know trigonometry or calculus, and she’s never studied the properties of cutting tools or metals. She doesn’t know how to maintain a tolerance of 0.25 microns, or what tolerance means in this context, or what a micron is.

Tony explains that Maddie has a job for two reasons. First, when it comes to making fuel injectors, the company saves money and minimizes product damage by having both the precision and non-precision work done in the same place. Even if Mexican or Chinese workers could do Maddie’s job more cheaply, shipping fragile, half-finished parts to another country for processing would make no sense. Second, Maddie is cheaper than a machine. It would be easy to buy a robotic arm that could take injector bodies and caps from a tray and place them precisely in a laser welder. Yet Standard would have to invest about $100,000 on the arm and a conveyance machine to bring parts to the welder and send them on to the next station. As is common in factories, Standard invests only in machinery that will earn back its cost within two years. For Tony, it’s simple: Maddie makes less in two years than the machine would cost, so her job is safe—for now. If the robotic machines become a little cheaper, or if demand for fuel injectors goes up and Standard starts running three shifts, then investing in those robots might make sense.

At this point, some similarities to the drug discovery business will be occurring to readers of this blog, along with some differences. The automation angle isn't as important, or not yet. While pharma most definitely has a manufacturing component (and how), the research end of the business doesn't resemble it very much, despite numerous attempts by earnest consultants and managers to make it so. From an auto-parts standpoint, there's little or no standardization at all in drug R&D. Every new drug is like a completely new part that no one's ever built before; we're not turning out fuel injectors or alternators. Everyone knows how a car works. Making a fundamental change in that plan is a monumental challenge, so the auto-parts business is mostly about making small variations on known components to the standards of a given customer. But in pharma - discovery pharma, not the generic companies - we're wrenching new stuff right out of thin air, or trying to.

So you'd think that we wouldn't be feeling the low-wage competitive pressure so much, but as the last ten years have shown, we certainly are. Outsourcing has come up many a time around here, and the very fact that it exists shows that not all of drug research is quite as bespoke as we might think. (Remember, the first wave of outsourcing, which is still very much a part of the business, was the move to send the routine methyl-ethyl-butyl-futile analoging out somewhere cheaper). And this takes us, eventually, to the Pfizer-style split between drug designers (high-wage folks over here) and the drug synthesizers (low-wage folks over there). Unfortunately, I think that you have to go the full reducio ad absurdum route to get that far, but Pfizer's going to find out for us if that's an accurate reading.

What these economists are also talking about is, I'd say, the next step beyond Moore's Law: once we have all this processing power, how do we use it? The first wave of computation-driven change happened because of the easy answers to that question: we had a lot of number-crunching that was being done by hand, or very slowly by some route, and we now had machines that could do what we wanted to do more quickly. This newer wave, if wave it is, will be driven more by software taking advantage of the hardware power that we've been able to produce.

The first wave didn't revolutionize drug discovery in the way that some people were hoping for. Sheer brute force computational ability is of limited use in drug discovery, unfortunately, but that's not always going to be the case, especially as we slowly learn how to apply it. If we really are starting to get better at computational pattern recognition and decision-making algorithms, where could that have an impact?

It's important to avoid what I've termed the "Andy Grove fallacy" in thinking about all this. I think that it is a result of applying first-computational-wave thinking too indiscriminately to drug discovery, which means treating it too much like a well-worked-out human-designed engineering process. Which it certainly isn't. But this second-wave stuff might be more useful.

I can think of a few areas: in early drug discovery, we could use help teasing patterns out of large piles of structure-activity relationship data. I know that there are (and have been) several attempts at doing this, but it's going to be interesting to see if we can do it better. I would love to be able to dump a big pile of structures and assay data points into a program and have it say the equivalent of "Hey, it looks like an electron-withdrawing group in the piperidine series might be really good, because of its conformational similarity to the initial lead series, but no one's ever gotten back around to making one of those because everyone got side-tracked by the potency of the chiral amides".

Software that chews through stacks of PK and metabolic stability data would be worth having, too, because there sure is a lot of it. There are correlations in there that we really need to know about, that could have direct relevance to clinical trials, but I worry that we're still missing some of them. And clinical trial data itself is the most obvious place for software that can dig through huge piles of numbers, because those are the biggest we've got. From my perspective, though, it's almost too late for insights at that point; you've already been spending the big money just to get the numbers themselves. But insights into human toxicology from all that clinical data, that stuff could be gold. I worry that it's been like the concentration of gold in seawater, though: really there, but not practical to extract. Could we change that?

All this makes me actually a bit hopeful about experiments like this one that I described here recently. Our ignorance about medicine and human biochemistry is truly spectacular, and we need all the help we can get in understanding it. There have to be a lot of important things out there that we just don't understand, or haven't even realized the existence of. That lack of knowledge is what gives me hope, actually. If we'd already learned what there is to know about discovering drugs, and were already doing the best job that could be done, well, we'd be in a hell of a fix, wouldn't we? But we don't know much, we're not doing it as well as we could, and that provides us with a possible way out of the fix we're in.

So I want to see as much progress as possible in the current pattern-recognition and data-correlation driven artificial intelligence field. We discovery scientists are not going to automate ourselves out of business so quickly as factory workers, because our work is still so hypothesis-driven and hard to define. (For a dissenting view, with relevance to this whole discussion, see here). It's the expense of applying the scientific method to human health that's squeezing us all, instead, and if there's some help available in that department, then let's have it as soon as possible.

Comments (32) + TrackBacks (0) | Category: Drug Assays | Drug Development | Drug Industry History | In Silico | Pharmacokinetics | Toxicology

February 21, 2012

Rational Drug Design. Hmm.

Email This Entry

Posted by Derek

Here's a huge review that goes over most everything you may have wanted to know about what's called "rational drug design". The authors are especially addressing selectivity, but that's a broad enough topic to cover all the important features. (If you can't access the paper, here's a key graphic from it).

"Rational", it should be understood, generally tends to mean "computationally modeled" in the world of drug discovery. And that's certainly how this review is pitched. I'm of two minds - at least - about the whole area (a personal bias that has made for some lively discussions over the years). Some of those discussions have taken place between my own ears as well, because I'm still not sure that all my opinions about computational drug design are self-consistent.

On the one hand, drug potency is a physical act which is mediated by physical laws. Computing the change in free energy during such a process should be feasible. But it turns out to be rather difficult - proteins flex and bonds rotate, water molecules assist and interfere, electrostatic charges help and hinder, hydrogen bonds are vital (and hard to model), and a dozen other sorts of interactions between clouds of electrons weigh in as well. Never forget, too, that free energy changes have an entropy component, and that's not trivial to model, either. I keep wondering if the error bars of the various assumptions and approximations don't end up swamping out the small changes that we're interested in predicting.

But, on that other hand, there are certainly cases where modeling has helped out a great deal. A cynic would say that we've been sure to hear about those, while the cases where it had no impact at all (or did actual harm) don't make the journals very often. It can't be denied, though, that modeling really has been (at times) the tool for the job. It would be interesting to know if the frequency of that happening has been increasing over time, as our tools get better.

Because on the third hand, it's been a poor bet to go against the relentless computational tide over the last few decades. You'd have to think that sheer computing power will end up making molecular modeling ever more capable and useful, as we learn more about what we're doing. Mind you, there were people back in the mid-1980s who thought we'd already reached that point. I'm not saying that they were the best-informed people at that time, but they certainly did exist. I wonder sometimes what it would have been like, to show people in 1985 what the state of rational drug design would be like in 2012. Would they be excited, or vaguely disappointed?

And then there's that word "rational". I think that its adoption might have been the best advertising that the field's ever achieved, because it makes everything else seem irrational (or at least arational) by default. I mean, do you just wanna make compounds, or do you want to think about what you're doing? I also wonder what might have changed if that phrase had never been adopted - perhaps expectations wouldn't have gotten out of hand in the computational field's early days, but it might not have received the attention (and money) that it did, either. . .

Comments (35) + TrackBacks (0) | Category: In Silico

January 26, 2012

Putting a Number on Chemical Beauty

Email This Entry

Posted by Derek

There's a new paper out in Nature Chemistry called "Quantifying the Chemical Beauty of Drugs". The authors are proposing a new "desirability score" for chemical structures in drug discovery, one that's an amalgam of physical and structural scores. To their credit, they didn't decide up front which of these things should be the miost important. Rather, they took eight properties over 770 well-known oral drugs, and set about figuring how much to weight each of them. (This was done, for the info-geeks among the crowd, by calculating the Shannon entropy for each possibility to maximize the information contained in the final model). Interestingly, this approach tended to give zero weight to the number of hydrogen-bond acceptors and to the polar surface area, which suggests that those two measurements are already subsumed in the other factors.

And that's all fine, but what does the result give us? Or, more accurately, what does it give us that we haven't had before? After all, there have been a number of such compound-rating schemes proposed before (and the authors, again to their credit, compare their new proposal with the others head-to-head). But I don't see any great advantage. The Lipinski "Rule of 5" is a pretty simple metric - too simple for many tastes - and what this gives you is a Rule of 5 with both categories smeared out towards each other to give some continuous overlap. (See the figure below, which is taken from the paper). That's certainly more in line with the real world, but in that real world, will people be willing to make decisions based on this method, or not?
The authors go for a bigger splash with the title of the paper, which refers to an experiment they tried. They had chemists across AstraZeneca's organization assess some 17,000 compounds (200 or so for each) with a "Yes/No" answer to "Would you undertake chemistry on this compound if it were a hit?" Only about 30% of the list got a "Yes" vote, and the reasons for rejecting the others were mostly "Too complex", followed closely by "Too simple". (That last one really makes me wonder - doesn't AZ have a big fragment-based drug design effort?) Note also that this sort of experiment has been done before.

Applying their model, the mean score for the "Yes" compounds was 0.67 (s.d.0.16), and the mean score for the "No" compounds was 0.49 (s.d. 0.23, which they say was statistically significant, although that must have been a close call. Overall, I wouldn't say that this test has an especially strong correlation with medicinal chemists' ideas of structural attractiveness, but then, I'm not so sure of the usefulness of those ideas to start with. I think that the two ends of the scale are hard to argue with, but there's a great mass of compounds in the middle that people decide that they like or don't like, without being able to back up those statements with much data. (I'm as guilty as anyone here).

The last part of the paper tries to extend the model from hit compounds to the targets that they bind to - a druggability assessment. The authors looked through the ChEMBL database, and ranked the various target by the scores of the ligands that are associated with them. They found that their mean ligand score for all the targets in there is 0.478. For the targets of approved drugs, it's 0.492, and for the orally active ones it's 0.539 - so there seems to be a trend, although if those differences reached statistical significance, it isn't stated in the paper.

So overall, I find nothing really wrong with this paper, but nothing spectacularly right with it, either. I'd be interested in hearing other calls on it as it gets out into the community. . .

Comments (22) + TrackBacks (0) | Category: Drug Development | Drug Industry History | In Silico | Life in the Drug Labs

January 9, 2012

A Look Into the Future?

Email This Entry

Posted by Derek

For a look into a possible drug-discovery future (from the computational optimist viewpoint), you might want to check out a brief bit of science fiction, "Alpha Shock", in the Journal of Computer-Aided Molecular Design. Some excerpts to give you the general idea:

". . .Of course, the compounds were of little value if they couldn’t be formulated. Sanjay was pressed for time, and nanobot development still took several weeks, so he had to go “old school.” Sanjay accessed World Crystallography Repository’s (WCR) formulation suite and entered the 2D structures of his compounds. The system linked to the Amazon Hyper-Cloud and initiated a series of quantum chemical calculations to develop a custom force field for the solid phase simulations. Unfortunately the preliminary results were disappointing, even after more than 100 million combinations of excipients, particle sizes, focusing tails, and polymorphs had been analyzed in detail. He would run a more complete search overnight, but the chances were that the 10-min simulation was telling him what he needed to know: don’t expect these exact compounds to be quite right. . .

. . .“In fact,” Dmitri continued, “I think the best tactic is to turn down the interaction of this transcription factor”—a protein popped out of one node on the map—“with that protein”—another protein materialized—“and this stretch of DNA.” A 3D model of the complex assembled in front of him, slowly rotating, with the most likely binding sites and points of intervention highlighted. “Of course, you only want to disrupt this interaction in the hippocampus, and only when D7 receptor functioning is high.” The relevant pathway maps showed the effects of the blockage on downstream signaling. “Oh, and naturally you also want to turn down oxphos in the mitochondria. So we need either a single molecule that can do both things, or a two-drug combo.”

The overall impression is a bit like Charles Stross, in its deliberate you-haven't-extrapolated-wildly-enough approach. But Stross doesn't put in as many computational chemistry inside jokes, which is probably better for his sales. My first impulse is the same one I have to, say, Ray Kurzweil, that all this stuff may (in fact probably is) on its way, but not by the dates stated. That position allows me to take flak from both sides, which must be some sort of feature that I value.

Comments (16) + TrackBacks (0) | Category: In Silico

November 7, 2011

Where's the Best Place to Apply Modeling to Drug Discovery?

Email This Entry

Posted by Derek

An e-mail correspondent and I were discussing this question, and I thought it would be an interesting one for everyone. He's a computational guy, and he's been wondering where the best use of computation/modeling effort in drug research might be. The obvious place to apply it is in lead generation and SAR development - but is that the best place? Is it the rate-limiting step enough of the time?

Problem is, the things that are often limiting steps are not as amenable to modeling. These are things like toxicology, target selection, and the like, and I'm not sure what they're susceptible to, except that simulation is probably not the answer. Or not yet, anyway. So what's the sweet spot, the place that maximizes importance and feasibility?

Update: an early vote for clinical trial design, which is a strong contender. Can't say that that doesn't get right to the hard part. . .

Comments (38) + TrackBacks (0) | Category: In Silico

September 20, 2011

Foldit Notches a Protein Structure Success (And Some Failures)

Email This Entry

Posted by Derek

I wrote last year about Foldit, a collaborative effort to work on protein structure problems that's been structured as an open-access game. Now the team is back with another report on how the project is going, and it's interesting stuff. The headlines have generally taken the "Computer Gamers Solve Incredible Protein Problem That Baffled Scientists!" line, but that's not exactly the full story.

The Foldit collaboration participated in the latest iteration of a regular protein-structure prediction challenge, CASP9. And their results varied - in the category of proteins with known structural homologs, for example, they didn't perform all that well. The players, it turned out, sort of over-worked the structures, and made a lot of unnecessary changes to the peripheral parts of the proteins. Another category took on proteins that have no identified structural homologs, a much harder problem. But that had its problems, too, which illustrate both the difficulties of the Foldit approach and protein modeling in general:

For prediction problems for which there were no identifiable homologous protein structures—the CASP9 Free Modeling category—Foldit players were given the five Rosetta Server CASP9 submissions (which were publicly available to other prediction groups) as starting points, along with the Alignment Tool. . .In this Free Modeling category, some of the shortcomings of the Foldit predictions became clear. The main problem was a lack of diversity in the conformational space explored by Foldit players because the starting models were already minimized with the same Rosetta energy function used by Foldit. This made it very difficult for Foldit players to get out of these local minima, and the only way for the players to improve their Foldit scores was to make very small changes ('tunneling' to the nearest local minimum) to the starting structures. However, this tunneling did lead to one of the most spectacular successes in the CASP9 experiment.

. . .the Rosetta Server, which carried out a large-scale search for the lowest-energy structure using computing power from Rosetta@home volunteers, produced a remarkably accurate model . . . However, the server ranked this model fourth out of the five submissions. The Foldit Void Crushers team correctly selected this near-native model and further improved it by accurately moving the terminal helix, producing the best model for this target of any group and one of the best overall predictions at CASP9 . . . Thus, in a situation where one model out of several is in a near-native conformation, Foldit players can recognize it and improve it to become the best model. Unfortunately for the other Free Modeling targets, there were no similarly outstanding Rosetta Server starting models, so Foldit players simply tunneled to the nearest incorrect local minima.

In the Refinement challenge, where participants take a minimized structure and try to improve its accuracy, the Foldit players had similar problems with starting from structures that had already been minimized by the same tools that they were using. Every change tended to make things look worse. The team improved their performance by reposting one of the structures as a new challenge, this time keeping the parts that were known with confidence to be near-native, while more or less randomizing the other parts to give a greater diversity to the starting points.

And those really are some of the key problems in this work. There are an awful lot of energy minima out there, and which ones you can get to depend crucially on where you start looking. In order to get to a completely different manifold of protein structures, even ones with much better energies, you may well have to go through a zone where you look like you're ruining everything. (And most of the time, you probably are ruining everything - there's no way to know if there's a safe haven on the other side or not).

But this paper also reports the results that are getting the headlines, a structure for the Mason-Pfizer monkey retroviral protease. This is an interesting protein, because although it crystallizes readily (in several different forms), and although the structures of other retroviral proteases are known, no one has been able to solve this one from the available X-ray data. The Foldit players, however, came up with several proposals that fit the data well enough for the structure to finally fall out of the diffraction data. It does have some odd features in its protein loops, different enough from the other proteases for no one to have hit on it before.

And that really is an accomplishment, and the way it was solved (with different players building on the results of others, competing to get the best optimization scores) really is the way the Foldit is supposed to work. Their less impressive performance on the CASP9 problems, though, shows that the same protein prediction difficulties apply to Foldit players as apply to the rest of the modeling field. This isn't a magic technique, and Foldit gamers are not going to rampage through the structural biology world solving all the extant problems any time soon. But it's nothing to sneeze at, either.

Comments (16) + TrackBacks (0) | Category: In Silico | Press Coverage

August 26, 2011

Design a Molecule, Win an IPad (Which is More Than You Usually Get)

Email This Entry

Posted by Derek

For those of you who are (or have always wanting to try being) molecular modelers, Cresset Design is holding a contest you might enjoy. They're putting up a molecule and giving out temporary licenses to their modeling software, and inviting people to come up with the closest bioisosteric match. The winner gets a free IPad2.

Of course, you're not going to be able to win by suggesting a para-fluoro group or by making a tetrazole-for-carboxylate switch. In their words:

We will use the Field alignment score for your molecule to the reference molecule as the primary judgment in designing the winner. However, molecules with high 2D similarity or high calculated logP with receive a penalty and are unlikely to win. Also entries with reasonable chemistry and good synthetic feasibility will be favoured. Feedback showing the score for your molecule and describing which properties of the molecule are being penalised will be provided on request. The winner will be the molecule that, in the opinion of the judges, represents the best design chosen from the top scoring results.

Fair enough, I'd say. I look forward to a follow-up from them at the end of the contest; I'd like to see what sort of stuff comes in.

Comments (18) + TrackBacks (0) | Category: In Silico

March 29, 2011

Modeling and Structure

Email This Entry

Posted by Derek

Man, am I getting all kinds of comments (here and by e-mail) about my views on modeling, QSAR, and the like. I thought it might be helpful for me to clarify my position on these things.

First off, structure. It's a valuable thing to have. My comments on the recent Nature Reviews Drug Discovery article were not meant to suggest otherwise, just to point out that the set of examples the authors picked to make this point was (in my view) flawed. It's actually surprisingly hard to come up with good comparison sets that isolate the effect of having structural information on the success of drug discovery projects. There are too many variables, and too many of them aren't independent. But just because a question (does having structural information help, overall?) is hard to answer doesn't mean that the answer is "no".

As an aside, since I've talked here about my admiration for fragment-based approaches, my own opinion should have been pretty clear already. Doing fragment-based drug discovery without good structural information looks to be very hard indeed.

Now, that said, there's structure and there's structure. Like every other tool in our kit, this one can be used well or used poorly. I think that fragment projects (to pick one example) get a lot of bang-for-the-buck out of structural data, and at the opposite end of the scale are those projects that only get good X-ray data after they've sent their compound to the clinic. No, wait, let me take that back. In those cases, the structure did no good, but it also did no harm. At the true opposite end of the scale are the projects where having structural data actually slowed things down. That's not frequent, but it does happen. Sometimes you have solid data, but for one reason or another the X-ray isn't corresponding to what's happening in real life. And sometimes this kicks in when medicinal chemists try to make too much out of less compelling structural data, just because it's all they have.

Now for in silico techniques. I have a similar attitude towards modeling of all kinds, but at one further remove than physical structure data. That is, I think it can be used well or used poorly, but I think that (for various reasons) the chances of using it poorly are somewhat increased. One reason is that modeling can be very hard to do well, naturally. And at the same time, tools with which to model conformations, docking, and so on are pretty widely available, which leads to a fair amount of work from people who really don't know what they're doing. Another reason is that the validity of any given model is of limited scope, as is the case with any mental construct that we have about what our molecules are doing, whether we used a software package or waved our hands around in the air. The software-package version of some binding model is more likely to have a wider range of usefulness than the hand-waving one, but they'll both break down at some point as you explore a range of compounds.

The key then is to figure out as quickly as possible if the project you're working on would be enhanced by modeling, or if such modeling would be merely ornamental, or even harmful. And that's not always easy to do. Any reasonable model is going to need a few iterations to get up to speed, generally requiring some specific compounds to be made by the chemists, and if you're running a project, you have to decide how much effort is worth spending to do that. You don't want to end up endlessly trying to refine the model, but at the same time, that model could turn out to be very useful after a few more turns of the crank. Which way to go? The same decisions apply, naturally, to the folks standing in front of the hoods, even without any modeling. How many more compounds are worth making in a given series? Would that effort be better used somewhere else? These calls are why we're paid the approximation of the big bucks.

So, while I don't think that modeling is an invariable boon to a project, neither do I think it's a waste of time. Sometimes it's one, and sometimes it's the other, and most of the time it's a mix of each - just like ideas at the bench. When modeling works, it can be a real help in sending the chemists down a productive path. On the other hand, you can certainly run a whole project with no modeling at all, just good old-fashioned analoging from the labs. It's the job of modelers to make the first possibility more likely and more attractive, and the job of the chemists and project managers to be open to that (and to be ready to emphasize or de-emphasize things as they develop).

This point of view seems reasonable to me (which is why I hold it!) But it also exposes me to complaints from people at both ends of the spectrum. I'm a lot more skeptical of in silico approaches than are many true believers, but I don't want to make the mistake of dismissing them outright.

Comments (15) + TrackBacks (0) | Category: In Silico

March 28, 2011

Value in Structure?

Email This Entry

Posted by Derek

A friend on the computational/structural side of the business sent along this article from Nature Reviews Drug Discovery. The authors are looking through the Thomson database at drug targets that are the subject of active research in the industry, and comparing the ones that have structural information available to the ones that don't: enzyme targets (with high-resolution structures) and and GPCRs without it. They're trying to to see if structural data is worth enough to show up in the success rates (and thus the valuations) of the resulting projects.

Overall, the Thomson database has over a thousand projects in it from these two groups, a bit over 600 from the structure-enabled enzymes and just under 500 GPCR projects. What they found was that 70% of the projects in the GPCR category were listed as "suspended" or "discontinued", but only 44% of the enzyme projects were so listed. In order to correct for probability of success across different targets, the authors picked ten targets from each group that have led, overall, to similar numbers of launched drugs. Looking at the progress of the two groups, the structure-enabled projects are again lower in the "stopped" categories, with corresponding increases in discovery and the various clinical phases.

You have to go to the supplementary info for the targets themselves, but here they are: for the enzymes, it's DPP-IV, BCR-ABL, HER2 kinase, renin, Factor Xa, HDAC, HIV integrase, JAK2, Hep C protease, and cathepsin K. For the receptor projects, the list is endothelin A receptor, P2Y12, CXCR4, angiogensin II receptor, sphingosine-1-phosphate receptor, NK1, muscarinic M1, vasopressin V2, melatonin receptor, and adenosine A2A.

Looking over these, though, I think that the situation is more complicated than the authors have presented. For example, DPP-IV has good structural information now, but that's not how people got into the area. The cyanopyrrolidine class of inhibitors, which really jump-started the field, were made by analogy to a reported class of prolyl endopeptidase inhibitors (BOMCL 1996, p. 1163). Three years later, the most well-characterized Novartis compound in the series was being studied by classic enzymology techniques, because it still wasn't possible to say just how it was binding. But even more to the point, this is a well-trodden area now. Any DPP-IV project that's going on now is piggybacking not only on structural information, but on an awful lot of known SAR and toxicology.

And look at renin. That's been a target forever, structure or not. And it's safe to say that it wasn't lack of structural information that was holding the area back, nor was it the presence of it that got a compound finally through the clinic. You can say the same things about Factor Xa. The target was validated by naturally occurring peptides, and developed in various series by classical SAR. The X-ray structure of one of the first solid drug candidates in the area (rivaroxaban) bound to its target, came after the compound had been identified and the SAR had been optimized. Factor Xa efforts going on now also are standing on the shoulders of an awful lot of work.

In the case of histone deacetylase, the first launched drug in that category (SAHA, vorinostat) has already been identified before any sort of X-ray structure was available. Overall, that target is an interesting addition to the list, since there are actually a whole series of them, some of which have structural information and some of which don't. The big difficulty in that area is that we don't really know what the various roles of the different isoforms are, and thus how the profiles of different compounds might translate to the clinic, so I wouldn't say that structural data is helping with the rate-determining steps in the field.

On the receptor side, I also wouldn't say that it's lack of structural information that's necessarily holding things back in all of those cases, either. Take muscarinic M1 - muscarinic ligands have been known for a zillion years. That encompasses fairly selective antagonists, and hardly-selective-at-all agonists, so I'm not sure which class the authors intended. If they're talking about antagonists, then there are plenty already known. And if they're talking about agonists, I doubt that even detailed structural information would help, given the size of the native ligand (acetylcholine).

And the vasopressin V2 case is similar to some of the enzyme ones, in that there's already an approved drug in this category (tolvaptan), with several others in the same structural class chasing it. Then you have the adenosine A2A field, where long lists of agonists and antagonists have been found over the years, structure or not. The problem there has been finding a clinical use for them; all sorts of indications have been chased over the years, a problem that structural information would have not helped with in the least.

Now, it's true that there are projects in these categories where structure has helped out quite a bit, and it's also true that detailed GPCR structures would be welcome (and are slowly coming along, for that matter). I'm not denying either of those. But what does strike me is that there are so many confounding variables in this particular comparison, especially among the specific targets that are the subject of the article's featured graphic, that I just don't think that its conclusions follow.

Comments (32) + TrackBacks (0) | Category: Drug Development | Drug Industry History | In Silico

August 9, 2010

Maybe We Should Make It More of a Game

Email This Entry

Posted by Derek

David Baker's lab at the University of Washington has been working on several approaches to protein structure problems. I mentioned Rosetta@home here, and now the team has published an interesting paper on another one of their efforts, FoldIt.

That one, instead of being a large-scale passive computation effort, is more of an active process - in fact, it's active enough that it's designed as a game:

We hypothesized that human spatial reasoning could improve both the sampling of conformational space and the determination of when to pursue suboptimal conformations if the stochastic elements of the search were replaced with human decision making while retaining the deterministic Rosetta algorithms as user tools. We developed a multiplayer online game, Foldit, with the goal of producing accurate protein structure models through gameplay. Improperly folded protein conformations are posted online as puzzles for a fixed amount of time, during which players interactively reshape them in the direction they believe will lead to the highest score (the negative of the Rosetta energy). The player’s current status is shown, along with a leader board of other players, and groups of players working together, competing in the same puzzle.

So how's it working out? Pretty well, actually. It turns out that human players are willing to do more extensive rearrangements to the protein chains in the quest for lower energies than computational algorithms are. They're also better at evaluating which positions to start from. Both of these remind me of the differences between human chess play and machine play, as I understand them, and probably for quite similar reasons. Baker's team is trying to adapt the automated software to use some of the human-style approaches, when feasible.

There are several dozen participants who clearly seem to have done better in finding low-energy structures than the rest of the crowd. Interestingly, they're mostly not employed in the field, with "Business/Financial/Legal" making up the largest self-declared group in a wide range of fairly even-distributed categories. Compared to the "everyone who's played" set, the biggest difference is that there are far fewer students in the high-end group, proportionally. That group of better problem solvers also tends to be slightly more female (although both groups are still mostly men), definitely older (that loss of students again), and less well-stocked with college graduates and PhDs. Make of that what you will.

Their conclusion is worth thinking about, too:

The solution of challenging structure prediction problems by Foldit players demonstrates the considerable potential of a hybrid human–computer optimization framework in the form of a massively multiplayer game. The approach should be readily extendable to related problems, such as protein design and other scientific domains where human three-dimensional structural problem solving can be used. Our results indicate that scientific advancement is possible if even a small fraction of the energy that goes into playing computer games can be channelled into scientific discovery.

That's crossed my mind, too. In my more pessimistic moments, I've imagined the human race gradually entertaining itself to death, or at least to stasis, as our options for doing so become more and more compelling. (Reading Infinite Jest a few years ago probably exacerbated such thinking). Perhaps this is one way out of that problem. I'm not sure that it's possible to make a game compelling enough when it's hooked up to some sort of useful gear train, but it's worth a try.

Comments (16) + TrackBacks (0) | Category: Biological News | In Silico | Who Discovers and Why

June 22, 2010

Free Software

Email This Entry

Posted by Derek

The folks at Cresset sent me a note about a free download of some software that they've developed for molecular fields (an approach you can read more about here). Fieldview is a free tool for trying this out yourself, and can be had here. Worth a look for the computationally curious, especially at the price. . .

Comments (11) + TrackBacks (0) | Category: In Silico

June 18, 2010

What Has Bioinformatics Ever Done For Us?

Email This Entry

Posted by Derek

A reader points me to this discussion, which is trying to figure out what the most useful discovery made via bioinfomatics is so far. There's a $100 prize for the winning suggestion, just to keep the discussion moving (and no, I don't anticipate offering cash bounties around here any time soon!) The early going seems to have ended up in the "Hold it, that's not bioinformatics, is it?" ditch, but that's not a useless discussion, either.

So if you have some suggestions, hop over there and add them to the fray, or vote for the ones that you like so far. I'm racking my brain a bit myself.

Comments (11) + TrackBacks (0) | Category: In Silico

May 17, 2010

Modeling in Drug Discovery: Questions?

Email This Entry

Posted by Derek

I'll have the opportunity to sit in on a few talks during a conference on free energy calculations in drug design. Since I'm not a computational guy myself, I'll be picking my sessions carefully, but I am interested in hearing what the state of the art is.

If we could just walk right up and calculate the free energies of binding events reliably, that would mean that the era of high throughput screening would begin to come to its end - well, in the physical world, anyway. Depending on how lengthy the computations needed to be, we could (in theory) just sit back and let the hardware hum while it ran through all the compounds we could think up - then we'd come back in on Monday and see who the winners were. Despite what some of you outside the field of medicinal chemistry might have read, we are not exactly to this point yet. That phrase "in theory" covers an awful lot of ground. But progress is apparently being made (here's a recent paper (PDF) with background).

So here's a question for the readership: what would you most want such calculations to be able to do for you? What would convince you that they're actually believable? And how close to you think that we actually are to that? Your comments will go directly to the ears of a roomful of high-powered modelers, so feel free to unload.

That thought of a roomful of computational chemists, though, reminds me inexorably of a story about Robert Oppenheimer that Freeman Dyson retells here. At a theoretical physics conference in Vancouver, the attendees were on a boat ride among the islands when the weather turned impenetrably foggy. Someone asked what the consequences for physics would be if the boat sank, and Oppenheimer instantly said "It wouldn't do any permanent good". There, that should ensure me a warm welcome at the meeting!

Comments (62) + TrackBacks (0) | Category: In Silico

May 10, 2010

Unlovely Polyphenols

Email This Entry

Posted by Derek

Here's a new paper from the folks at the Burnham Institute and UCSD on a new target for vaccinia virus. They're going after a virulence factor (N1L) through computational screening, which is a challenge, since this is a protein-protein interaction.

They pulled out a number of structures, which have some modest activity in cell infection assays. In addition, they showed through calorimetry that the compounds do appear to be affecting the target protein, specifically its equilibrium between monomeric and oligomeric forms. But the structures of their best hits. . .well, here's the table. You can ignore compounds 6 and 8; they show up as cytotoxic. But the whole list is pretty ghastly, at least to my eyes.

These sorts of highly aromatic polyphenol structures have two long traditions in medicinal chemistry: showing activity in assays, for the first part, and not being realizable as actual drugs, for the second. There's no doubt that they can do a lot of things; it's just that getting them to do them in a real-world situation is not trivial. Part of the problem is specificity (and associated toxicity) and part of it is pharmacokinetics. As you'd imagine, these compounds can have rather funky clearance behavior, what with all those phenols.

So I'd regard these as proof-of-concept compounds that validate N1L as a target. I think that we'll need to wait for someone to format up an assay for high-throughput (non-virtual) screening to see if something more tractable comes up. Either that, or rework the virtual screens on the basis that we've seen enough polyphenols come up on this target already. . .

Note: readers of the paper will note that our old friend resveratrol turns up as an active compound as well. It's very much in the polyphenol tradition; make of that what you will.

Comments (25) + TrackBacks (0) | Category: In Silico | Infectious Diseases | Pharmacokinetics

Bill Gates Put Some Money On Schrödinger

Email This Entry

Posted by Derek

My take on the recent news that Bill Gates has invested ten million dollars in the computational drug design company Schrödinger is here at Nature News. (They've recently made all their stories open-access, by the way, so you don't need a subscription to get the full stories).

In short, I think that patient billionaire money is just the sort of thing the field needs, because anyone with a short timeline and a need for a good return is going to have a rough time of it. . .

Comments (7) + TrackBacks (0) | Category: In Silico

May 4, 2010

Another Proposal For the Scientific Literature

Email This Entry

Posted by Derek

I was talking with a colleague yesterday, and I suddenly had an insight into an opportunity in scientific publishing. We were discussing the various computational/modeling papers that you see out in the literature. Some of them are quite interesting, many are worth looking at if it's your particular field - but many others are, well, not so great. I should mention up front that the same objections apply - and how - to the non-computational literature, of course. But there are a number of second-tier (and lower) journals to soak up those sorts of papers in the other disciplines.

What surprises me is that there's no Computational Chemistry Letters or some such. Communications in Computational Chemistry? CADD Comm? This would be the dumping ground for the piles of unconvincing computer-driven stuff that gets sent around by people who have paid a bit too much attention to the sales brochures that came with their software packages.

The barriers for entry to such things have been getting lower and lower, while the real state of the art has been getting more and more complicated. That's created a gap into which too much stuff falls. Who will speak for the bottom-dwelling "We modeled it, therefore it's real" constituency? The advent of systems biology has created more opportunities than ever for these folks. Isn't it time that there was an expensive, low-impact, completely disregardable journal for them, too?

Comments (13) + TrackBacks (0) | Category: In Silico | The Scientific Literature

March 29, 2010

Compounds and Proteins

Email This Entry

Posted by Derek

For the medicinal chemists in the audience, I wanted to strongly recommend a new paper from a group at Roche. It's a tour through the various sorts of interactions between proteins and ligands, with copious examples, and it's a very sensible look at the subject. It covers a number of topics that have been discussed here (and throughout the literature in recent years), and looks to be an excellent one-stop reference.

In fact, read the right way, it's a testament to how tricky medicinal chemistry is. Some of the topics are hydrogen bonds (and why they can be excellent keys to binding or, alternatively, of no use whatsoever), water molecules bound to proteins (and why disturbing them can account for large amounts of binding energy, or, alternatively, kill your compound's chances of ever binding at all), halogen bonds (which really do exist, although not everyone realizes that), interactions with aryl rings (some of which can be just as beneficial coming in 90 degrees to where you might imagine), and so on.

And this is just to get compounds to bind to their targets, which is the absolute first step on the road to a drug. Then you can start worrying about how to have your compounds not bind to things you don't want (many of which you probably don't even realize even are out there). And about how to get it to decent blood levels, for a decent amount of time, and into the right compartments of the body. And at that point, it's nearly time to see if it does any good for the disease you're trying to target!

Comments (5) + TrackBacks (0) | Category: Drug Assays | In Silico | Life in the Drug Labs

March 23, 2010

We Don't Know Beans About Biotin

Email This Entry

Posted by Derek

You know, you'd think that we'd understand the way things bind to proteins well enough to be able to explain why biotin sticks so very, very tightly to avidins. That's one of the most impressive binding events in all of biology, short of pushing electrons and forming a solid chemical bond - biotin's stuck in there at femtomolar levels. It's so strong and so reliable that this interaction is the basis for untold numbers of laboratory and commercial assays - just hang a biotin off one thing, expose it to something else that has an avidin (most often streptavidin) coated on it, and it'll stick, or else something is Very Wrong. So we have that all figured out.

Wrong. Turns out that there's a substantial literature given to arguing about just why this binding is so tight. One group holds out for hydrophobic interactions (which seems rather weird to me, considering that biotin's rather polar by most standards). Another group has a hydrogen-bonding explanation, which (on the surface) seems more feasible to me. Now a new paper says that the computational methods applied so far can't handle electrostatic factors well, and that those are the real story.

I'm not going to take a strong position on any of these; I'll keep my head down while the computational catapults launch at each other. But it's definitely worth noting that we apparently can't explain the strongest binding site interaction that we know of. It's the sort of thing that we'd all like to be able to generate at will in our med-chem programs, but how can we do that when we don't even know what's causing it?

Comments (12) + TrackBacks (0) | Category: Drug Assays | In Silico

March 22, 2010

Benford's Law, Revisited

Email This Entry

Posted by Derek

I mentioned Benford's Law in passing in this post (while speculating on how long people report their reactions to have run when publishing their results). That's the rather odd result that many data sets don't show a random distribution of leading digits - rather, 1 is the first digit around 30% of the time, 2 leads off about 18% of the time, and so on down.

For data that come from some underlying power-law distribution, this actually makes some sense. In that case, the data points spend more time being collected in the "lag phase" when they're more likely to start with a 1, and proportionally less and less time out in the higher-number-leading areas. The law only holds up when looking at distributions that cover several orders of magnitude - but all the same, it also seems to apply to data sets where there's no obvious exponential growth driving the numbers.

Lack of adherence to Benford's Law can be acceptable as corroborative evidence of financial fraud. Now a group from Astellas reports that several data sets used in drug discovery (such as databases of water solubility values) obey the expected distribution. What's more, they're suggesting that modelers and QSAR people check their training data sets to make sure that those follow Benford's Law as well, as a way to make sure that the data have been randomly selected.

Is anyone willing to try this out on a bunch of raw clinical data to see what happens? Could this be a way to check the integrity of reported data from multiple trial centers? You'd have to pick your study set carefully - a lot of the things we look for don't cover a broad range - but it's worth thinking about. . .

Comments (9) + TrackBacks (0) | Category: Clinical Trials | In Silico | The Dark Side

December 10, 2009

Selective Scaffolds

Email This Entry

Posted by Derek

We spend a lot of time in this business talking about molecular scaffolds - separate chemical cores that we elaborate into more advanced compounds. And there's no doubt that such things exist, but is part of the reason they exist just an outcome of the way chemical research is done? Some analysis in the past has suggested that chemical types get explored in a success-breeds-success fashion, so that the (over)representation of some scaffold might not mean that it has unique properties. It's just that it's done what's been asked of it, so people have stuck with it.

A new paper in J. Med. Chem. from a group in Bonn takes another look at this question. They're trying to see if the so-called "privileged substructures" really exist: chemotypes that have special selectivity for certain target classes. Digging through a public-domain database (BindingDB), they found about six thousand compounds with activity toward some 259 targets. Many of these compounds hit more than one target, as you'd expect, so there were about 18,000 interactions to work with.

Isolating structural scaffolds from the compound set and analyzing them for their selectivity showed some interesting trends. They divide the targets up into communities (kinases, serine proteases, and so on), and they definitely find community-selective scaffolds, which is certainly the experience of medicinal chemists. Inside these sets, various scaffolds also show tendencies for selectivity against individual members of the community. Digging through their supporting information, though, it appears that a good number of the most-selective scaffolds tend to come from the serine protease community (their number 3), with another big chunk coming from kinases (their number 1a). Strip those (and some adenosine receptor ligands and DPP inhibitors, numbers 11 and 8) out, and you've taken out all the really eye-catching selectivity numbers out of their supplementary table S5. So I'm not sure that they've identified as many hot structures as one might think.

Another problem I have, when I look at these structures, is that a great number of them look too large for any useful further development. That's just a function of the data this team had to start with, but this gets back to the question of "drug-like" versus "lead-like" structures. I have a feeling that too many of the compounds in the BindingDB set are in the former category, or even beyond, which skews things a bit. Looking at a publication on it from 2007, I get the impression that a majority of compounds in it have a molecular weight greater than 400, with a definite long tail toward the higher weights. What medicinal chemists would like, of course, is a set of smaller scaffolds that will give them a greater chance of landing in a selective chemical space that can be developed. Some of the structures in this paper qualify, but definitely not all of them. . .

Comments (6) + TrackBacks (0) | Category: Drug Assays | Drug Development | In Silico

December 7, 2009

Why Don't We Have More Protein-Protein Drug Molecules?

Email This Entry

Posted by Derek

Almost all of the drugs on the market target one or more small-molecule binding sites on proteins. But there's a lot more to the world than small-molecule binding sites. Proteins spend a vast amount of time interacting with other proteins, in vital ways that we'd like to be able to affect. But those binding events tend to be across broader surfaces, rather than in well-defined binding pockets, and we medicinal chemists haven't had great success in targeting them.

There are some successful examples, with a trend towards more of them in the recent literature. Inhibitors of interactions of the oncolocy target Bcl are probably the best known, with Abbott's ABT-737 being the poster child of the whole group.

But even though things seem to be picking up in this area, there's still a very long way to go, considering the number of possible useful interactions we could be targeting. And for every successful molecule that gets published, there are surely an iceberg of failed attempts that never make the literature. What's holding us back?

A new article in Drug Discovery Today suggests, as others have, that our compound libraries aren't optimized for finding hits in such assays. Given that the molecular weights of the compounds that are known to work tend toward the high side, that may well be true - but, of course, since the amount of chemical diversity up in those weight ranges is ridiculously huge, we're not going to be able to fix the situation through brute-force expansion of our screening libraries. (We'll table, for now, the topic of the later success rate of such whopper molecules).

Some recent work has suggested that there might be overall molecular shapes that are found more often in protein-protein inhibitors, but I'm not sure if everyone buys into this theory or not. This latest paper does a similar analysis, using 66 structurally diverse protein-protein inhibitors (PPIs) from the literature compared to a larger set (557 compounds) of traditional drug molecules. The PPIs tend to be larger and greasier, as feared>. They tried some decision-tree analysis to see what discriminated the two data sets, and found a shape description and another one that correlated more with aromatic ring/multiple-bond count. Overall, the decision tree stuff didn't shake things down as well as it does with data sets for more traditional target classes, which doesn't come as a surprise, either.

So the big questions are still out there: can we go after protein-protein targets with reasonably-sized molecules, or are they going to have to be big and ugly? And in either case, are there structures that have a better chance of giving us a lead series? If that's true, is part of the problem that we don't tend to have such things around already? If I knew the answers to these questions, I'd be out there making the drugs, to be honest. . .

Comments (14) + TrackBacks (0) | Category: Drug Assays | Drug Industry History | In Silico

November 17, 2009

Side Effects, Predicted?

Email This Entry

Posted by Derek

There's a new paper out in Nature that presents an intriguing way to look for off-target effects of drug candidates. The authors (a large multi-center team) looked at a large number of known drugs (or well-characterized clinical candidates) and their activity profiles. They then characterized the protein targets by the similarities of the molecules that were known to bind to them.

That gave a large number of possible combinations - nearly a million, actually, and in most cases, no correlations showed up. But in about 7,000 examples, a drug matched some other ligand set to an interesting degree. On closer inspection, some of these off-target effects turned out to be already known (but had not been picked up during their initial searching using the MDDR database). Many others turned out to be trivial variations on other known structures.

But what was left over was a set of 3,832 predictions of meaningful off-target binding events. The authors took 184 of these out to review them carefully and see how well they held up. 42 of these turned out to be already confirmed in the primary literature, although not reported in any of the databases they'd used to construct the system - that result alone is enough to make one think that they might be on the right track here.

Of the remaining 142 correlations, 30 were experimentally feasible to check directly. Of these, 23 came back with inhibition constants less than 15 micromolar - not incredibly potent, but something to think about, and a lot better hit rate than one would expect by chance. Some of the hits were quite striking - for example, an old alpha-blocker, indoramin, showed a strong association for dopamine receptors, and turned out to be an 18 nM ligand for D4, which is better than it does on the alpha receptors themselves. In general, they uncovered a lot of new GPCR activities for older CNS drugs, which doesn't surprise me, given the polypharmacy that's often seen in that area.

But they found four examples of compounds that jumped into completely new target categories. Rescriptor (delavirdine), a reverse transcriptase inhibitor used against HIV, showed a strong score against histamine subtypes, and turned out to bind H4 at about five micromolar. That may not sound like much, but the drug's blood levels make that a realistic level to think about, and its side effects include a skin rash that's just what you might expect from such off-target binding.

There are some limitations. To their credit, the authors mention in detail a number of false positives that their method generated - equally compelling predictions of activities that just aren't there. This doesn't surprise me much - compounds can look quite similar to existing classes and not share their activity. I'm actually a bit surprised that their methods works as well as it does, and look forward to seeing refined versions of it.

To my mind, this would be an effort well worth some collaborative support by all the large drug companies. A better off-target prediction tool would be worth a great deal to the whole industry, and we might be able to provide a lot more useful data to refine the models used. Anyone want to step up?

Update: be sure to check out the comments section for other examples in this field, and a lively debate about which methods might work best. . .

Comments (20) + TrackBacks (0) | Category: Drug Assays | In Silico | Toxicology

Warren DeLano

Email This Entry

Posted by Derek

I've been remiss in not mentioning this, but I just found out recently that Warren DeLano (the man behind the excellent open-source PyMOL program) passed away suddenly earlier this month. He was 37 - another unfortunate loss of a scientist who had done a lot of fine work and was clearly on the way to doing much more.

I notice that as I write this I have a PyMOL window open on my desktop; I use the program regularly to look at protein structures. Si monumentum requiris, circumspice.

Comments (8) + TrackBacks (0) | Category: Current Events | In Silico

July 16, 2009

The Further In You Go, The Bigger It Gets

Email This Entry

Posted by Derek

I had a printout of the structure of maitotoxin on my desk the other day, mostly as a joke to alarm anyone who came into my office. "Yep, here's the best hit from the latest screen. . .I hear that you're on the list to run the chemistry end. . .what's that you say?"
This is, needless to say, one of the largest and scariest marine natural product structures ever determined (and that determination has been no stroll past the dessert table, either).

But that' hasn't stopped people from messing around with it. And there's much speculation that other people are strongly considering messing around with it, too - you synthetic chemists can guess the sorts of people that this might be, and their names, and what it might be like to sit through the seminars that result, and so on.

I fear that a total synthesis of maitotoxin would be largely a waste of time, but I'm willing to hear arguments against that position. Just looking at it, though, inspires thought. This eldrich beastie has 98 chiral centers. So let's do some math. If you're interested in the SAR of such molecules, you have your choice of (two to the 98th) possible isomers, which comes out to a bit over (3 times ten to the 29th) compounds. This is. . .a pretty large number. If you're looking for 10mg of each isomer to add to your screening collection (no sense in going back and making them again), then you're looking at a good bit over half the mass of the entire Earth. And that's just in sheer compounds; we're not counting the weight of vials, which will, I'd say, safely move you up toward the planetary weight of a low-end gas giant. We will ignore shelving considerations in the interest of time.

Recall that yesterday's post gave a number of about 27 million compounds below 11 heavy atoms. You could toss 27 million compounds into a collection of ten to the 29th and never see them again, of course. But that brings up two points: one, that the small-compound estimate ignores stereochemistry, and we've been getting those insane maitotoxin numbers by considering nothing but. The thing is, with only 11 non-hydrogen atoms, there aren't quite as many chances for things to get out of control. The GDB compound set goes up only to 110 million or so if you consider stereoisomers, which actually isn't nearly as much as I'd thought.

But the second point is that this shows you why the Berne group stopped at 11 heavy atoms, because the problem becomes intractable really fast as you go higher. It's worth remembering that the GDB people actually threw out over 98% of their scaffolds because they represented potential ring structures that are too strained to be very stable. And they only considered C, N, O and F as heavy atoms (even adding sulfur was considered too much to deal with, computationally). Then they tossed out another 98 or 99% of the structures that emerged from that enumeration as reactive and/or unstable. Relax your standards a bit, allow another atom or two, bump up the molecular weight, do any of those and you're going to exceed anyone's computational capacity. Update: the Berne group has just taken a crack at it, and managed a reasonable set up to 13 heavy atoms, with various simplifying assumptions to ease the burden. If you want to mess around with it, it's here, free of charge).

No, there are a lot of compounds out there. And if you look at the really big ones - and maitotoxin is nothing if not a really big one - there are whole universes contained just in each of them. (Bonus points for guessing the source of the name of the post, by the way).

Comments (25) + TrackBacks (0) | Category: Chemical News | In Silico

July 15, 2009

Why Does Screening Work At All? (Free Business Proposal Included!)

Email This Entry

Posted by Derek

I've been meaning to get around to a very interesting paper from the Shoichet group that came out a month or so ago in Nature Chemical Biology. Today's the day! It examines the content of screening libraries and compares them to what natural products generally look like, and they turn up some surprising things along the way. The main question they're trying to answer is: given the huge numbers of possible compounds, and the relatively tiny fraction of those we can screen, why does high-throughput screening even work at all?

The first data set they consider is the Generated Database (GDB), a calculated set of all the reasonable structures with 11 or fewer nonhydrogen atoms, which grew out of this work. Neglecting stereochemistry, that gives you between 26 and 27 million compounds. Once you're past the assumptions of the enumeration (which certainly seem defensible - no multiheteroatom single-bond chains, no gem-diols, no acid chlorides, etc.), then there are no human bias involved: that's the list.

The second list is everything from the Dictionary of Natural Products and all the metabolites and natural products from the Kyoto Encyclopedia of Genes and Genomes. That gives you 140,000+ compounds. And the final list is the ZINC database of over 9 million commercially available compounds, which (as they point out) is a pretty good proxy for a lot of screening collections as well.

One rather disturbing statistic comes out early when you start looking at overlaps between these data sets. For example, how many of the possible GDB structures are commercially available? The answer: 25,810 of them - in other words, you can only buy fewer than 0.01% of the possible compounds with 11 heavy atoms or below, making the "purchasable GDB" a paltry list indeed.

Now, what happens when you compare that list of natural products to these other data sets? Well, for one thing, the purchasable part of the GDB turns out to be much more similar to the natural product list than the full set. Everything in the GDB has at least 20% Tanimoto similarity to at least one compound in the natural products set, not that 20% means much of anything in that scoring system. But only 1% of the GDB has a 40% Tanimoto similarity, and less than 0.005% has an 80% Tanimoto similarity. That's a pretty steep dropoff!

But the "purchasable GDB" holds up much better. 10% of that list has 100% Tanimoto similarity (that is, 10% of the purchasable compounds are natural products themselves). The authors also compare individual commercial screening collections. If you're interested, ChemBridge and Asinex are the least natural-product-rich (about 5% of their collections), whereas IBS and Otava are the most (about 10%).

So one answer to "why does HTS ever work for anything" is that compound collections seem to be biased toward natural-product type structures, which we can reasonably assume have generally evolved to have some sort of biological activity. It would be most interesting to see the results of such an analysis run from inside several drug companies against their own compound collections. My guess is that the natural product similarities would be even higher than the "purchasable GDB" set's, because drug company collections have been deliberately stocked with structural series that have shown activity in one project or another.

That's certainly looking at things from a different perspective, because you can also hear a lot of talk about how our compound files are too ugly - too flat, too hydrophobic, not natural-product-like enough. These viewpoints aren't contradictory, though - if Shoichet is right, then improving those similarities would indeed lead to higher hit rates. Compared to everything else, we're already at the top of the similarity list, but in absolute terms there's still a lot of room for improvement.

So how would one go about changing this, assuming that one buys into this set of assumptions? The authors have searched through the various databases for ring structures, taking those as a good proxy for structural scaffolds. As it turns out 83% of the ring scaffolds among the natural products are unrepresented among the commercially available molecules - a result that I assume that Asinex, ChemBridge, Life Chemicals, Otava, Bionet and their ilk are noting with great interest. In fact, the authors go even further in pointing out opportunities, with a table of rings from this group that closely resemble known drug-like ring systems.

But wait a minute. . .when you look at those scaffolds, a number of them turn out to be rather, well, homely. I'd be worried about elimination to form a Michael acceptor in compound 19, for example. I'm not crazy about the N,S acetal in 21 or the overall stability of the acetals in 15, 17 and 31. The propiolactone in 23 is surely reactive, as is the quinone in 25, and I'd be very surprised if that's not what they owe their biological activities to. And so on.
All that said, there are still some structures in there that I'd be willing to check out, and there must be more of them in that 83%. No doubt a number of the rings that do sneak into the commercial list are not very well elaborated, either. I think that there is a real commercial opportunity here. A company could do quite well for itself by promoting its compound collection as being more natural-product similar than the competition, with tractable molecules, and a huge number of them unrepresented in any other catalog.

Now all you'd have to do is make these things. . .which would require hiring synthetic organic chemists, and plenty of them. These things aren't easy to make, or to work with. And as it so happens, there are quite a few good ones available these days. Anyone want to take this business model to heart?

Comments (13) + TrackBacks (0) | Category: Drug Assays | Drug Industry History | In Silico

July 7, 2009

What's So Special About Ribose?

Email This Entry

Posted by Derek

While we're on the topic of hydrogen bonds and computations, there's a paper coming out in JACS that attempts to answer an old question. Why, exactly, does every living thing on earth use so much ribose? It's the absolute, unchanging carbohydrate backbone to all the RNA on Earth, and like the other things in this category (why L amino acids instead of D?), it's attracted a lot of speculation. If you subscribe to the RNA-first hypothesis of the origins of life, then the question becomes even more pressing.

A few years ago, it was found that ribose, all by itself, diffuses through membranes faster than the other pentose sugars. This results holds up for several kinds of lipid bilayers, suggesting that it's not some property of the membrane itself that's at work. So what about the ability of the sugar molecules to escape from water and into the lipid layers?

Well, they don't differ much in logP, that's for sure, as the original authors point out. This latest paper finds, though, by using molecular dynamic simulations that there is something odd about ribose. In nonpolar environments, its hydroxy groups form a chain of hydrogen-bond-like interactions, particularly notable when it's in the beta-pyranose form. These aren't a factor in aqueous solution, and the other pentoses don't seem to pick up as much stabilization under hydrophobic conditions, either.

So ribose is happier inside the lipid layer than the other sugars, and thus pays less of a price for leaving the aqueous environment, and (both in simulation and experimentally) diffuses across membranes ten times as quickly as its closely related carboyhydate kin. (Try saying that five times fast!) This, as both the original Salk paper and this latest one note, leads to an interesting speculation on why ribose was preferred in the origins of life: it got there firstest with the mostest. (That's a popular misquote of Nathan Bedford Forrest's doctrine of warfare, and if he's ever come up before in a discussion of ribose solvation, I'd like to hear about it).

Comments (9) + TrackBacks (0) | Category: Biological News | In Silico | Life As We (Don't) Know It

Another Thing We Don't Know

Email This Entry

Posted by Derek

Hydrogen bonds are important. There, that should be an sweepingly obvious enough statement to get things started. But they really are - hydrogen bonding accounts for the weird properties of water, for one thing, and it's those weird properties that are keeping us alive. And leaving out the water (a mighty big step), internal hydrogen bonding is still absolutely essential to the structure of large biological molecules - proteins, complex carbohydrates, DNA and RNA, and so on.

But we don't understand hydrogen bonds all that well, dang it all. It's not like we're totally ignorant of them, for sure, but there are a lot of important things that we don't have a good handle on. One of these may just have been illustrated by this paper in Nature Structural and Molecular Biology by a group from Scripps. They've been working on understanding the fact that all hydrogen bonds are not created equal. By carefully going through a lot of protein mutants, they have evidence for the idea that H-bonds that form in polar environments are weaker than ones that form in nonpolar ones.

That makes sense, on the face of it. One way to think of it is that a hydrogen bond in a locally hydrophobic area is the only game in town, and counts for more. But this work claims that such bonds can be worth as much as 1.2 kcal/mole more than the wimpier ones, which is rather a lot. Those kinds of energy differences could add up very quickly when you're trying to understand why a protein folds up the way it does, or why one small molecule binds more tightly than another one.

Do we take such things into account when we're trying to compute these energies? Generally speaking, no, we do not - well, not yet. If these folks are right, though, we'd better start.

Update: note that the paper itself doesn't suggest that this is a new idea - they reference work going back to 1963 (!) on the topic. What they're trying to do is put more real numbers into the mix. And that's what my last paragraph above is trying to state (and perhaps overstate): it's difficult to account for these thing computationally, since they vary so widely, and since we don't have that good a computational handle on hydrogen bonds in general. The more real world data that can be fed back into the models, the better.

Comments (7) + TrackBacks (0) | Category: In Silico

July 2, 2009

Jargon Will Save Us All

Email This Entry

Posted by Derek

Moore's Law: number of semiconductors on a chip doubling every 18 months or so, etc. Everyone's heard of it. But can we agree that anyone who uses it as a metaphor or perscription for drug research doesn't know what they're talking about?

I first came across the comparison back during the genomics frenzy. One company that had bought into the craze in a big way press-released (after a rather interval) that they'd advanced their first compound to the clinic based on this wonderful genomics information. I remember rolling my eyes and thinking "Oh, yeah", but on a hunch I went to the Yahoo! stock message boards (often a teeming heap of crazy, then as now). And there I found people just levitating with delight at this news. "This is Moore's Law as applied to drug discovery!" shouted one enthusiast. "Do you people realize what this means?" What it meant, apparently, was not only that this announcement had come rather quickly. It also meant that this genomics stuff was going to discover twice as many drugs as this real soon. And real soon after that, twice as many more, and so on until the guy posting the comment was as rich as Warren Buffet, because he was a visionary who'd been smart enough to load himself into the catapult and help cut the rope. (For those who don't know how that story ended, the answer is Not Well: the stock that occasioned all this hyperventilation ended up dropping by a factor of nearly a hundred over the next couple of years. The press-released clinical candidate was never, ever, heard of again).

I bring this up because a reader in the industry forwarded me this column from Bio-IT World, entitled, yes, "Only Moore's Law Can Save Big Pharma". I've read it three times now, and I still have only the vaguest idea of what it's talking about. Let's see if any of you can do better.

The author starts off by talking about the pressures that the drug industry is under, and I have no problem with him there. That is, until he gets to the scientific pressures, which he sketches out thusly:

Scientifically, the classic drug discovery paradigm has reached the end of its long road. Penicillin, stumbled on by accident, was a bona fide magic bullet. The industry has since been organized to conduct programs of discovery, not design. The most that can be said for modern pharmaceutical research, with its hundreds of thousands of candidate molecules being shoveled through high-throughput screening, is that it is an organized accident. This approach is perhaps best characterized by the Chief Scientific Officer of a prominent biotech company who recently said, "Drug discovery is all about passion and faith. It has nothing to do with analytics."

The problem with faith-based drug discovery is that the low hanging fruit has already been plucked, driving would be discoverers further afield. Searching for the next miracle drug in some witch doctor's jungle brew is not science. It's desperation.

The only way to escape this downward spiral is new science. Fortunately, the fuzzy outlines of a revolution are just emerging. For lack of a better word, call it Digital Chemistry.

And when the man says "fuzzy outline", well, you'd better take him at his word. What, I know you're all asking, is this Digital Chemistry stuff? Here, wade into this:

Tomorrow's drug companies will build rationally engineered multi-component molecular machines, not small molecule drugs isolated from tree bark or bread mold. These molecular machines will be assembled from discrete interchangeable modules designed using hierarchical simulation tools that resemble the tool chains used to build complex integrated circuits from simple nanoscale components. Guess-and-check wet chemistry can't scale. Hit or miss discovery lacks cross-product synergy. Digital Chemistry will change that.

Honestly, if I start talking like this, I hope that onlookers will forgo taking notes and catch on quickly enough to call the ambulance. I know that I'm quoting too much, but I have to tell you more about how all this is going to work:

But modeling protein-protein interaction is computationally intractable, you say? True. But the kinetic behavior of the component molecules that will one day constitute the expanding design library for Digital Chemistry will be synthetically constrained. This will allow engineers to deliver ever more complex functional behavior as the drugs and the tools used to design them co-evolve. How will drugs of the future function? Intracellular microtherapeutic action will be triggered if and only if precisely targeted DNA or RNA pathologies are detected within individual sick cells. Normal cells will be unaffected. Corrective action shutting down only malfunctioning cells will have the potential of delivering 99% cure rates. Some therapies will be broad based and others will be personalized, programmed using DNA from the patient's own tumor that has been extracted, sequenced, and used to configure "target codes" that can be custom loaded into the detection module of these molecular machines.

Look, I know where this is coming from. And I freely admit that I hope that, eventually, a really detailed molecular-level knowledge of disease pathology, coupled with a really robust nanotechnology, will allow us to treat disease in ways that we can't even approach now. Speed the day! But the day is not sped by acting as if this is the short-term solution for the ills of the drug industry, or by talking as if we already have any idea at all about how to go about these things. We don't.

And what does that paragraph up there mean? "The kinetic behavior. . .will be synthetically constrained"? Honestly, I should be qualified to make sense of that, but I can't. And how do we go from protein-protein interactions at the beginning of all that to DNA and RNA pathologies at the end, anyway? If all the genomics business has taught us anything, it's that these are two very, very different worlds - both important, but separated by a rather wide zone of very lightly-filled-in knowledge.

Let's take this step by step; there's no other way. In the future, according to this piece, we will detect pathologies by detecting cell-by-cell variations in DNA and/or RNA. How will we do that? At present, you have to rip open cells and kill them to sequence their nucleic acids, and the sensitivities are not good enough to do it one cell at a time. So we're going to find some way to do that in a specific non-lethal way, either from the outside of the cells (by a technology that we cannot even yet envision) or by getting inside them (by a technology that we cannot even envision) and reading off their sequences in situ (by a technology that we cannot even envision). Moreover, we're going to do that not only with the permanent DNA, but with the various transiently expressed RNA species, which are localized to all sort of different cell compartments, present in minute amounts and often for short periods of time, and handled in ways that we're only beginning to grasp and for purposes that are not at all yet clear. Right.

Then. . .then we're going to take "corrective action". By this I presume that we're either going to selectively kill those cells or alter them through gene therapy. I should note that gene therapy, though incredibly promising as ever, is something that so far we have been unable, in most cases, to get to work. Never mind. We're going to do this cell by cell, selectively picking out just the ones we want out of the trillions of possibilities in the living organism, using technologies that, I cannot emphasize enough, we do not yet have. We do not yet know how to find most individual cells types in a complex living tissue; huge arguments ensue about whether certain rare types (such as stem cells) are present at all. We cannot find and pick out, for example, every precancerous cell in a given volume of tissue, not even by slicing pieces out of it, taking it out into the lab, and using all the modern techniques of instrumental analysis and molecular biology.

What will we use to do any of this inside the living organism? What will such things be made of? How will you dose them, whatever they are? Will they be taken up though the gut? Doesn't seem likely, given the size and complexity we're talking about. So, intravenous then, fine - how will they distribute through the body? Everything spreads out a bit differently, you know. How do you keep them from sticking to all kinds of proteins and surfaces that you're not interested in? How long will they last in vivo? How will you keep them from being cleared out by the liver, or from setting off a potentially deadly immune response? All of these could vary from patient to patient, just to make things more interesting. How will we get any of these things into cells, when we only roughly understand the dozens of different transport mechanisms involved? And how will we keep the cells from pumping them right back out? They do that, you know. And when it's time to kill the cells, how do you make absolutely sure that you're only killing the ones you want? And when it's time to do the gene therapy, what's the energy source for all the chemistry involved, as we cut out some sequences and splice in the others? Are we absolutely sure that we're only doing that in just the right places in just the right cells, or will we (disastrously) be sticking in copies into the DNA of a quarter of a per cent of all the others?

And what does all this nucleic acid focus have to do with protein expression and processing? You can't fix a lot of things at the DNA level. Misfolding, misglycosylation, defects in transport and removal - a lot of this stuff is post-genomic. Are we going to be able to sequence proteins in vivo, cell by cell, as well? Detect tertiary structure problems? How? And fix them, how?

Alright, you get the idea. The thing is, and this may be surprising considering those last few paragraphs, that I don't consider all of this to be intrinsically impossible. Many people who beat up on nanotechnology would disagree, but I think that some of these things are, at least in broad hazy theory, possibly doable. But they will require technologies that we are nowhere close to owning. Babbling, as the Bio-IT World piece does, about "detection modules" and "target codes" and "corrective action" is absolutely no help at all. Every one of those phrases unpacks into a gigantic tangle of incredibly complex details and total unknowns. I'm not ready to rule some of this stuff out. But I'm not ready to rule it in just by waving my hands.

Comments (46) + TrackBacks (0) | Category: Drug Industry History | General Scientific News | In Silico | Press Coverage

April 1, 2009

Mexican Lemons To the Rescue

Email This Entry

Posted by Derek

Thanks to a comment on this post, I’ve had a chance to read this interesting article from Stephen Johnson of Bristol-Myers Squibb, entitled “The Trouble with QSAR (Or How I Learned to Stop Worrying And Embrace Fallacy)”. (As a side note, it’s interesting to see that people still make references to the titling of Dr. Strangelove. I’ve never met Johnson, but I’d gather from that that he can’t be much younger than I am).


The most arresting part of the article is the graph found in its abstract. No mention is made of it in the text, but none has to be. It’s a plot of the US highway fatality rate versus the tonnage of fresh lemons imported from Mexico, and I have to say, it’s a pretty darn straight line. I’ve seen a lot shakier plots used to justify some sweeping conclusions, and if those were justified, well, then I’m forced to conclude that Mexican lemons have improved highway safety a great deal. The vitamin C, maybe? The fragrance? Bioflavanoids?

None of the above, of course. Correlation, tiresomely, once again refuses to imply causation, even when you ask it nicely. And that’s the whole point of the article. QSAR, for those outside the business, stands for Quantitative Structure-Activity Relationship(s), an attempt to rationalize the behavior of a series of drug candidate compounds through computational means. The problem is, there are plenty of possible variables (size, surface area, molecular weight, polarity, solubility, charge, hydrogen bond donors and acceptors, and as many structural representation parameters as you can stand). As Johnson notes dryly:

” With such an infinite array of descriptions possible, each of which can be coupled with any of a myriad of statistical methods, the number of equivalent solutions is typically fairly substantial.”

That it is. And (as he rightly mentions) one of the other problems is that all these variables are discontinuous. Some region of the molecule can get larger, but only up to a point. When it’s too large to fit into the binding site any more, activity drops off steeply. Similarly, the difference between forming a crucial hydrogen bond and not forming one is a big difference, and it can be realized by a very small change in structure and properties. (Thus the “magic methyl” effect).

But that’s not the whole problem. Johnson takes many of his fellow computational chemists to task for what he sees as sloppy work. Too many models are advanced just because they’ve shown some (limited) correlations, and they’re not tested hard enough afterwards. Finding a model with a good “fitness score” becomes an end in itself:

”We can generate so many hypotheses, relating convoluted molecular factors to activity in such complicated ways, that the process of careful hypothesis testing so critical to scientific understanding has been circumvented in favor of blind validation tests with low resulting information content. QSAR disappoints so often, not only because the response surface is not smooth but because we have embraced the fallacy that correlation begets causation.”

Comments (33) + TrackBacks (0) | Category: In Silico

March 26, 2009

The Motions of a Protein

Email This Entry

Posted by Derek

So, people like me spend their time trying to make small molecules that will bind to some target protein. So what happens, anyway, when a small molecule binds to a target protein? Right, right, it interacts with some site on the thing, hydrogen bonds, hydrophobic interactions, all that – but what really happens?

That’s surprisingly hard to work out. The tools we have to look at such things are powerful, but they have limitations. X-ray crystal structures are great, but can lead you astray if you’re not careful. The biggest problem with them, though (in my opinion) is that you see this beautiful frozen picture of your drug candidate in the protein, and you start to think of the binding as. . .well, as this beautiful frozen picture. Which is the last thing it really is.

Proteins are dynamic, to a degree that many medicinal chemists have trouble keeping in mind. Looking at binding events in solution is more realistic than looking at them in the crystal, but it’s harder to do. There are various NMR methods (here's a recent review), some of which require specially labeled protein to work well, but they have to be interpreted in the context of NMR’s time scale limitations. “Normal” NMR experiments give you time-averaged spectra – if you want to see things happening quickly, or if you want to catch snapshots of the intermediate states along the way, you have a lot more work to do.

Here’s a recent paper that’s done some of that work. They’re looking at a well-known enzyme, dihydrofolate reductase (DHFR). It’s the target of methotrexate, a classic chemotherapy drug, and of the antibiotic trimethoprim. (As a side note, that points out the connections that sometimes exist between oncology and anti-infectives. DHFR produces tetrahydrofolate, which is necessary for a host of key biosynthetic pathways. Inhibiting it is espccially hard on cells that are spending a lot of their metabolic energy on dividing – such as tumor cells and invasive bacteria).

What they found was that both inhibitors do something similar, and it affects the whole conformational ensemble of the protein:

". . .residues lining the drugs retain their μs-ms switching, whereas distal loops stop switching altogether. Thus, as a whole, the inhibited protein is dynamically dysfunctional. Drug-bound DHFR appears to be on the brink of a global transition, but its restricted loops prevent the transition from occurring, leaving a “half-switching” enzyme. Changes in pico- to nanosecond (ps-ns) backbone amide and side-chain methyl dynamics indicate drug binding is “felt” throughout the protein.

There are implications, though, for apparently similar compounds having rather different effects out in the other loops:

. . .motion across a wide range of timescales can be regulated by the specific nature of ligands bound. Occupation of the active site by small ligands of different shapes and physical characteristics places differential stresses on the enzyme, resulting in differential thermal fluctuations that propagate through the structure. In this view, enzymes, through evolution, develop sensitivities to ligand properties from which mechanisms for organizing and building such fluctuations into useful work can arise. . .Because the affected loop structures are primarily not in contact with drug, it is reasonable to envision inhibitory small-molecule drugs that act by allosterically modulating dynamic motions."

There are plenty of references in the paper to other investigations of this kind, so if this is your sort of thing, you'll find plenty of material there. One thing to take home, though, is to remember that not only are proteins mobile beasts (with and without ligand bound to them), but that this mobility is quite different in each state. And keep in mind that the ligand-bound state can be quite odd compared to anything else the protein experiences otherwise. . .

Comments (3) + TrackBacks (0) | Category: Biological News | Cancer | Chemical News | In Silico

February 24, 2009

Structure-Activity: Lather, Rinse, and Repeat

Email This Entry

Posted by Derek

Medicinal chemists spend a lot of their time exploring and trying to make sense of structure-activity relationships (SARs). We vary our molecules in all kinds of ways, have the biologists run them through the assays, and then sit down to make sense of the results.

And then, like as not, we get up again after a few minutes, shaking our heads. Has anyone out there ever worked on a project where the entire SAR made sense? I’ve always considered it a triumph if even a reasonable majority of the compounds fit into an interpretable pattern. SAR development is a perfect example of things not quite working out the way that they do in textbooks.

The most common surprise when you get your results back, if that phrase “common surprise” makes any sense, is to find that you’ve pushed some trend a bit too far. Methyl was pretty good, ethyl was better, but anything larger drops dead. I don’t count that sort of thing – those are boundary conditions, for the most part, and one of the things you do in a med-chem program is establish the limits under which you can work. But there are still a number of cases where what you thought was a wall turns out to have a secret passage or two hidden in it. You can’t put any para-substituents on that ring, sure. . .unless you have a basic amine over on the other end of the molecule, and then you suddenly can.

I’d say that a lot of these get missed, because after a project’s been running a while, various SAR dogmas get propagated. There are features of the structure space that “everybody knows”, and that few people want to spend their time violating. But it’s worth devoting a small (but real) amount of effort to going back and checking some of these after the lead molecule has evolved a bit, since you can get surprised.

Some projects I’ve worked on have so many conditional clauses of this sort built into their SAR that you wonder whether there are any boundaries at all. This works, unless you have this, but if you have that over there it can be OK, although there is that other compound which didn’t. . .making sense of this stuff can just be impossible. The opposite situation, the fabled Perfectly Additive SAR, is something I’ve never encountered in person, although I’ve heard tales after the fact. That’s the closest we come to the textbooks, where you can mix and match groups and substituents any way you like, predicting as you go from the previous trends just how they’ll come out. I have to think that any time you can do this, that it has to be taking place in a fairly narrow structure space – surely we can always break any trend like this with a little imagination.

Another well-known bit of craziness is the Only Thing That Works There. You’ll have whole series of compounds that have to have a a methyl group at some position, or they’re all dead. Nothing smaller, nothing larger, nothing with a different electronic flavor: it’s methyl or death. (Or fluoro, or a thiazole, or what have you – I’ve probably seen this with methyl more than with other groups, but it can happen all over the place). A sharp SAR is certainly nothing to fear; it’s probably telling you that you really are making good close contacts with the protein target somewhere. But it can be unnerving, and sometimes there’s not a lot of room left on the ledge when you have more than one constraint like this.

Why does all this go on? Multiple binding modes, you have to think. Proteins are flexible beasts, and they've got lots of ways to react to ligands. And it's important never to forget that we can't predict their responses, at least not yet and not very well. And of course, in all this discussion, we've just been considering one target protein. When you think about the other things your molecule might be hitting in cells or in a whole animal, and that the SAR relationships for those off-target things are just as fluid and complicated as for your target, well. . .you can see why medicinal chemistry is not going away anytime soon. Or shouldn't, anyway.

Comments (40) + TrackBacks (0) | Category: Drug Assays | In Silico | Life in the Drug Labs

December 10, 2008

Floppiness Is Not Your Friend: Who Knew?

Email This Entry

Posted by Derek

There’s a trick that every medicinal chemist learns very early, and continues to apply every time its feasible: take two parts of your compound, and tie them together into a ring.

The reason that works so well may not be immediately obvious if you’re not a medicinal chemist, so let me expand on them a bit. The first thing to know is that this method tends to work either really well or not at all – it’s a “death or glory” move. And that gives you a clue as to what’s going on. The idea is that the rotatable bonds in your molecule are, under normal conditions, doing just that: rotating. Any molecule the size of a normal drug has all kinds of possible shapes and rotational isomers, and room temperature is an energetic enough environment to populate a lot of them.

But there’s only one of them that’s the best for fitting into your drug target, most likely. So what are the odds? As your molecule approaches its binding pocket, there’s a complicated energetic dance going on. Different parts of your drug candidate will start interacting with the target (usually a protein), and that starts to tie down all that floppy rotation. The question is, does the gain resulting from these interactions cancel out the energetic price that has to be paid for them? Is there a pathway that leads to a favorable tight-binding situation, or is your molecule going to approach, flop around a bit, and dance away?

Several things are at work during that shall-we-dance period. The different conformations of your compound vary in energy, depending on how much its parts are starting to bang into each other, and how much you’re asking the bonds to twist around. The closer that desired drug-binding shape is to the shape your molecule wants to be in anyway, the better off you are, from that perspective. So tying back the molecule and making a ring in the structure does one thing immediately: it cuts down on the range of conformations it can take, in the same way that tying a rope between your ankles cuts down on your ability to dance. You’ve handcuffed your molecule, which would probably be cruel if they were sentient, but then, a lot of organic chemistry would be pretty unspeakable if molecules had feelings.

That’s why this method tends to be either a big winner or a big loser. If the preferred binding mode of your compound is close to the shape it takes when you tie it down, then you’ve suddenly zeroed in on just the thing you want, and the binding affinity is going to take a big leap. But if it’s not, well, you’ve now probably made it impossible for the thing to adopt the conformation it needs, and the binding affinity is going to take a big leap over a cliff.

There’s another effect to reducing the flexibility of your compound, and that has to do with entropy. All that favorable-interaction business is one component of the energy involved, namely the enthalpy, but entropy is the other. Loosely speaking, the more disordered a system, the higher its entropy. A floppy molecule, when it binds to a drug target, has to settle down into a much tighter fit, and entropically, that’s unfavorable. Energetically, you’re paying to do that. But if your molecule is already much less flexible, there’s not much of a toll as it fits into the pocket. If loss-of-floppiness is a bad thing, then don’t start out with so much of it.

So, how much do I and my medicinal chemistry colleagues think about this stuff, day to day? A fair amount, but there are parts of it that we probably don’t pay enough attention to. Entropy gets less respect from us than it deserves, I think. It’s easy to imagine molecules bumping into each other, sticking and unsticking, but the more nebulous change-in-disorder part of the equation is just as important. And it doesn’t just apply to our drug molecules – proteins get less disordered as they bind those molecules (or more disordered, in some cases), and those entropic changes can mean a lot, too.

I also mentioned molecules finding a pathway to binding, and that’s something that we don’t think about as much, either. We probably make things all the time that would be potent binders, if they just could get past some energetic hump and wedge themselves into place. But there are no crowbars available; our drug candidates have to be able to work their way in on their own. The can’t-get-there-from-here cases come back from the assays as inactive. The tendency is to imagine these in the binding site already, and to try to think of what could be going wrong in there – but it may be that they’d be fine, but that their structures won’t allow them to come in for a landing.

Picturing this accurately is very hard indeed. We have enough trouble with good representations of static pictures of our molecules bound to their targets, so making a movie of the process is a whole different story. Each frame is on a femtosecond scale – molecules flip around rather quickly – and every frame would have to be computed accurately (drug structure, protein structure, and the energetics of the whole system) for the resulting video clip to make sense. It’s been done, but not all that often, and we’re not good at it.

Comments (13) + TrackBacks (0) | Category: In Silico | Pharma 101

September 25, 2008

Protein Folding: Complexity to Make More Complexity?

Email This Entry

Posted by Derek

Want a hard problem? Something to really keep you challenged? Try protein folding. That'll eat up all those spare computational cycles you have lounging around and come back to ask for more. And it'll do the same for your brain cells, too, for that matter.

The reason is that a protein of any reasonable size has a staggering number of shapes it can adopt. If you hold a ball-and-stick model of one, you realize pretty quickly that there are an awful lot of rotatable bonds in there (not least because they flop around while you're trying to hold the model in your hands). My daughter was playing around with a toy once that was made of snap-together parts that looked like elbow macaroni pieces, and I told her that this was just like a lot of molecules inside her body. We folded and twisted the thing around very quickly to a wide variety of shapes, even though it only had ten links or so, and I then pointed out to her that real proteins all had different things sticking off at right angles in the middle of each piece, making the whole situation even crazier.

There's a new (open access) paper in PNAS that illustrates some of the difficulties. The authors have been studying man-made proteins that have substantially similar sequences of amino acids, but still have different folding and overall shape. In this latest work, they've made it up to two proteins (56 amino acids each) that have 95% sequence identity, but still have very different folds. It's just a few key residues that make the difference and kick the overall protein into a different energetic and structural landscape. The other regions of the proteins can be mutated pretty substantially without affecting their overall folding, on the other hand. (In the picture, the red residues are the key ones and the blue areas are the identical/can-be-mutated domains).
This ties in with an overall theme of biology - it's nonlinear as can be. The systems in it are huge and hugely complicated, but the importance of the various parts varies enormously. There are small key chokepoints in many physiological systems that can't be messed with, just as there are some amino acids that can't be touched in a given protein. (Dramatic examples include the many single-amino-acid based genetic disorders).

But perhaps the way to look at it is that the complexity is actually an attempt to overcome this nonlinearity. Otherwise the system would be too brittle to work. All those overlapping, compensating, inter-regulating feedback loops that you find in biochemistry are, I think, a largely successful attempt to run a robust organism out of what are fundamentally not very robust components. Evolution is a tinkerer, most definitely, and there sure is an awful lot of tinkering that's been needed.

Comments (8) + TrackBacks (0) | Category: General Scientific News | In Silico

September 4, 2008

X-Ray Structures: Handle With Care

Email This Entry

Posted by Derek

X-ray crystallography is wonderful stuff – I think you’ll get chemists to generally agree on that. There’s no other technique that can provide such certainty about the structure of a compound – and for medicinal chemists, it has the invaluable ability to show you a snapshot of your drug candidate bound to its protein target. Of course, not all proteins can be crystallized, and not all of them can be crystallized with drug ligands in them. But an X-ray structure is usually considered the last word, when you can get one – and thanks to automation, computing power, and to brighter X-ray sources, we get more of them than ever.

But there are a surprising number of ways that X-ray data can mislead you. For an excellent treatment of these, complete with plenty of references to the recent literature, see an excellent paper coming out in Drug Discovery Today from researchers at Astra-Zeneca (Andy Davis and Stephen St.-Gallay) and Uppsala University (Gerard Kleywegt). These folks all know their computational and structural biology, and they’re willing to tell you how much they don’t know, either.

For starters, a small (but significant) number of protein structures derived from X-ray data are just plain wrong. Medicinal chemists should always look first at the resolution of an X-ray structure, since the tighter the data, the better the chance there is of things being as they seem. The authors make the important point that there’s some subjective judgment involved on the part of a crystallographer interpreting raw electron-density maps, and the poorer the resolution, the more judgment calls there are to be made:

Nevertheless, most chemists who undertake structure-based design treat a protein crystal structure reverently as if it was determined at very high resolution, regardless of the resolution at which the structure was actually determined (admittedly, crystallographers themselves are not immune to this practice either). Also, the fact that the crystallographer is bound to have made certain assumptions, to have had certain biases and perhaps even to have made mistakes is usually ignored. Assumptions, biases, ambiguities and mistakes may manifest themselves (even in high-resolution structures) at the level of individual atoms, of residues (e.g. sidechain conformations) and beyond.

Then there’s the problem of interpreting how your drug candidate interacts with the protein. The ability to get an X-ray structure doesn’t always correlate well with the binding potency of a given compound, so it’s not like you can necessarily count on a lot of clear signals about why the compound is binding. Hydrogen bonds may be perfectly obvious, or they can be rather hard to interpret. Binding through (or through displacement of) water molecules is extremely important, too, and that can be hard to get a handle on as well.

And not least, there’s the assumption that your structure is going to do you good once you’ve got it nailed down:

It is usually tacitly assumed that the conditions under which the complex was crystallised are relevant, that the observed protein conformation is relevant for interaction with the ligand (i.e. no flexibility in the active-site residues) and that the structure actually contributes insights that will lead to the design of better compounds. While these assumptions seem perfectly reasonable at first sight, they are not all necessarily true. . .

That’s a key point, because that’s the sort of error that can really lead you into trouble. After all, everything looks good, and you can start to think that you really understand the system, that is until none of your wonderful X-ray-based analogs work out they way you thought they would. The authors make the point that when your X-ray data and your structure-activity data seem to diverge, it’s often a sign that you don’t understand some key points about the thermodynamics of binding. (An X-ray is a static picture, and says nothing about what energetic tradeoffs were made along the way). Instead of an irritating disconnect or distraction, it should be looked at as a chance to find out what’s really going on. . .

Comments (15) + TrackBacks (0) | Category: Analytical Chemistry | Drug Assays | In Silico

May 23, 2008

Up Close and Personal

Email This Entry

Posted by Derek

Something that’s come up in the last few posts around here is the way that we chemists think about the insides of enzymes. It’s a tricky subject, because when you picture things on that scale, the intuition you have for objects starts to betray you.

Consider water. We humans have a pretty good practical understanding of how water behaves in the bulk phase; we have the experience. But what about five water molecules sitting in the pocket of an enzyme? That’s not exactly a glass from the tap. These guys are interacting with the protein as much (or more) than they’re interacting with each other, and our intuition about water molecules is based on how they act when it’s surrounded by plenty of their own.

And if five water molecules are hard to handle, how about one? There’s no hope of seeing any bulk properties now, because there’s no bulk. We’re more used to having trouble in the other direction, predicting group behavior from individuals: you can’t tell much about a thousand-piece jigsaw puzzle from one piece that you found under the couch, and you wouldn’t be able to say much about the behavior of an ant colony from observing one ant in a jar. And neither of those are worth very much, compared to their group. But with molecules, the single-ant-in-a-jar situation is very important (that’s a single water molecule sitting in the active site of an enzyme), and knowledge of ant social behavior or water’s actions in a glass doesn’t help much.

Larger molecules than water are our business, of course, and those are tricky, too. We can study the shape and flexibility of our drug candidates in solution (by NMR, to pick the easiest method), and in the solid phase, surrounded by packed arrays of themselves (X-ray crystal structures). But the way that they look inside an enzyme's active site doesn't have to be related to either of those, although you might as well start there.

As single-molecule (and single-atom) techniques have become more possible, we're starting to get an idea of how small clusters of them have to be before they stop acting like tiny pieces of what we're used to, and starts acting like something else. But these experiments are usually done in isolation, in the gas phase or on some inert surface. The inside of a protein is another thing entirely; molecules there are the opposite of isolated. And studying them in those small spaces is no small task.

Comments (4) + TrackBacks (0) | Category: In Silico

May 1, 2008

O Pioneers!

Email This Entry

Posted by Derek

Drug Discovery Today has the first part of an article on the history of the molecular modeling field, this one covering about 1960 to 1990. It’s a for-the-record document, since as time goes on it’ll be increasingly hard to unscramble all the early approaches and players. I think this is true for almost any technology; the early years are tangled indeed.

As you would imagine, the work from the 1960s and 1970s has an otherwordly feel to it, considering the hardware that was available. And that brings up another thing common to the early years of new technologies: when you look back on them from their later years, you wonder how these people could possibly have even tried to do these things.

I mean, you read about, say, Richard Cramer establishing the computer-aided drug design program at Smith, Kline and French in nineteen-flipping-seventy-one, and on one level you feel like congratulating his group for their farsightedness. But mainly you just feeling like saying “Oh, you poor people. I am so sorry.” Because from today's perspective, there is just no way that anyone could have done any meaningful molecular modeling for drug design in 1971. I mean, we have enough trouble doing it for a lot of projects in 2008.

Think about it: big ol’ IBM mainframe, with those tape drives that for many years were visual shorthand for Computer System but now look closer to steam engines and water wheels. Punch cards: riffling stacks of them, and whole mechanical devices with arrays of rods to make and troubleshoot stiff pieces of paper with holes in them. And the software – written in what, FORTRAN? If they were lucky. And written in a time when people were just starting to say, well, yes, I suppose that you could, in fact, represent attractive and repulsive molecular forces in terms that could be used by a computer program. . .hmm, let’s see about hydrogen bonds, then. . .

It gives a person the shudders. But that must be inevitable – you get the same feeling when you see an early TV set and wonder how anyone could have derived entertainment from a fuzzy four-inch-wide grey screen. Or see the earliest automobiles, which look to have been quite a bit more trouble than a horse. How do people persevere?

Well, for one thing, by knowing that they’re the first. Even if technology isn’t what you might dream of it being some day, you’re still the one out on the cutting edge, with what could be the best in the world as it is. They also do it by not being able to know just what the limits to their capabilities are, not having the benefit of decades of hindsight. The molecular modelers of the early 1970s did not, I’m sure, see themselves as tentatively exploring something that would probably be of no use for years to come. They must have thought that there was something good just waiting right there to be done with the technology they had (which was, as just mentioned, the best ever seen). They may well have been wrong about that, but who was to know until it was tried?

And all of this – the realizations that there’s something new in the world, that there are new things that can be done with it, and (later) that there’s more to it (both its possibilities and difficulties) than was first apparent – all of this comes on gradually. If it were to hit you all at once, you’d be paralyzed with indecision. But the gap in the trees turns into a trail, and then into a dirt path before you feel the gravel under your feet, speeding up before you realize that you’re driving down a huge highway that branches off to destinations you didn’t even know existed.

People are seeing their way through to some of those narrow footpaths right now, no doubt. With any luck, in another thirty years people will look back and pity them for what they didn’t and couldn’t know. But the people doing it today don’t feel worthy of pity at all – some of them probably feel as if they’re the luckiest people alive. . .

Comments (8) + TrackBacks (0) | Category: Drug Industry History | In Silico | Who Discovers and Why

March 27, 2008

Start Small, Start Right

Email This Entry

Posted by Derek

There’s an excellent paper in the most recent issue of Chemistry and Biology that illustrates some of what fragment-based drug discovery is all about. The authors (the van Aalten group at Dundee) are looking at a known inhibitor of the enzyme chitinase, a natural product called argifin. It’s an odd-looking thing – five amino acids bonded together into a ring, with one of them (an arginine) further functionalized with a urea into a sort of side-chain tail. It’s about a 27 nM inhibitor of the enzyme.

(For the non-chemists, that number is a binding affinity, a measure of what concentration of the compound is needed to shut down the enzyme. The lower, the better, other things being equal. Most drugs are down in the nanomolar range – below that are the ulta-potent picomolar and femtomolar ranges, where few compounds venture. And above that, once you get up to 1000 nanomolar, is micromolar, and then 1000 micromolar is one millimolar. By traditional med-chem standards, single-digit nanomolar = good, double-digit nanomolar = not bad, triple-digit nanomolar or low micromolar = starting point to make something better, high micromolar = ignore, and millimolar = can do better with stuff off the bottom of your shoe.

What the authors did was break this argifin beast up, piece by piece, measuring what that did to the chitinase affinity. And each time they were able to get an X-ray structure of the truncated versions, which turned out to be a key part of the story. Taking one amino acid out of the ring (and thus breaking it open) lowered the binding by about 200-fold – but you wouldn’t have guessed that from the X-ray structure. It looks to be fitting into the enzyme in almost exactly the same way as the parent.

And that brings up a good point about X-ray crystal structures. You can’t really tell how well something binds by looking at one. For one thing, it can be hard to see how favorable the various visible interactions might actually be. And for another, you don’t get any information at all about what the compound had to pay, energetically, to get there.

In the broken argifin case, a lot of the affinity loss can probably be put down to entropy: the molecule now has a lot more freedom of movement, which has to be overcome in order to bind in the right spot. The cyclic natural product, on the other hand, was already pretty much there. This fits in with the classic med-chem trick of tying back side chains and cyclizing structures. Often you’ll kill activity completely by doing that (because you narrowed down on the wrong shape for the final molecule), but when you hit, you hit big.

The structure was chopped down further. Losing another amino acid only hurt the activity a bit more, and losing still another one gave a dipeptide that was still only about three times less potent than the first cut-down compound. Slicing that down to a monopeptide, basically just a well-decorated arginine, sent the activity down another sixfold or so – but by now we’re up to about 80 micromolar, which most medicinal chemists would regard as the amount of activity you could get by testing the lint in your pocket.

But they went further, making just the little dimethylguanylurea that’s hanging off the far end. That thing is around 500 micromolar, a level of potency that would normally get you laughed at. But wait. . .they have the X-ray structures all along the way, and what becomes clear is that this guanylurea piece is binding to the same site on the protein, in the same manner, all the way down. So if you’re wondering if you can get an X-ray structure of some 500 micromolar dust bunny, the answer is that you sure can, if it has a defined binding site.

And the value of these various derivatives almost completely inverts if you look at them from a binding efficiency standpoint. (One common way to measure that is to take the minus log of the binding constant and divide by the molecular weight in kilodaltons). That’s a “bang for the buck” index, a test of how much affinity you’re getting for the weight of your molecule. As it turns out, argifin – 27 nanomolar though it be – isn’t that efficient a binder, because it weighs a hefty 676. The binding efficiency index comes out to just under 12, which is nothing to get revved up about. The truncated analogs, for the most part, aren’t much better, ranging from 9 to 15.

But that guanylurea piece is another story. It doesn’t bind very tightly, but it bats way above its scrawny size, with a BEI of nearly 28. That’s much more impressive. If the whole argifin molecule bound that efficiently, it would be down in the ten-to-the-minus nineteenth range, and I don’t even know the name of that order of magnitude. If you wanted to make a more reasonably sized molecule, and you should, a compound of MW 400 would be about ten femtomolar with a binding efficiency like that. There’s plenty of room to do better than argifin.

So the thing to do, clearly, is to start from the guanylurea and build out, checking the binding efficiency along the way to make sure that you’re getting the most out of your additions. And that is exactly the point of fragment-based drug discovery. You can do it this way, cutting down a larger molecule to find what parts of it are worth the most, or you can screen to find small fragments which, though not very potent in the absolute sense, bind very efficiently. Either way, you take that small, efficient piece as your anchor and work from there. And either way, some sort of structural read on your compounds (X-ray or NMR) is very useful. That’ll give you confidence that your important binding piece really is acting the same way as you go forward, and give you some clues about where to build out in the next round of analogs.

This particular story may be about as good an illustration as one could possibly find - here's hoping that there are more that can work out this way. Congratulations to van Aalten and his co-workers at Dundee and Bath for one of the best papers I've read in quite a while.

Comments (12) + TrackBacks (0) | Category: Analytical Chemistry | Drug Assays | In Silico

March 5, 2008

Smaller, Wetter, Harder to Work With

Email This Entry

Posted by Derek

There’s an interesting article coming out in J. Med. Chem. on antibiotic compounds, which highlights something that’s pretty clear if you spend some time looking at the drugs in that area. We make a big deal (or have made one over the last ten years) about drug-like properties – all that Rule-of-Five stuff and its progeny. Well, take a look at the historically best-selling antibiotic drugs: you’ve never seen such a collection of Rule of Five violators in your life.

That’s partly because a lot of structures in that area have come from natural products, but hey, natural products are drugs, too. Erythromycin, the aminoglycosides, azithromycin, tetracycline: what a crew! But they’ve helped an untold number of people over the years. It’s true that the fluoroquinolones are much more normal-looking, but those are balanced out by weirdo one-shots like fosfomycin. I mean, look at that thing – would you ever believe that that’s a marketed drug? (And with decent bioavailability, too?)

No, you have to be broad-minded if you’re going to beat up on bacteria, and I think some broad-mindedness would do us all good in other therapeutic areas, too. I don’t mean we should ignore what we’ve learned about drug-like properties: our problem is that we tend to make allowances and exceptions on the greasy high-molecular weight end of the scale, since that’s where too many of our compounds end up. It wouldn’t hurt to push things on the other end, because I think that you have a better chance of getting away with too much polarity than you have of getting away with too little.

One reason for that might be that there are a lot of transporter proteins in vivo that are used to dealing with such groups. It’s easy to forget, but a great number of proteins are decorated with carbohydrate residues, and they’re on there for a lot of reasons. And a lot of extremely important small molecules in biochemistry are polar as well – right off the top of my head, I don’t know what the logD or polar surface area of things like ATP or NAD are, but I’ll bet that they’re far off the usual run of drugs. Admittedly, those aren’t going to reach good blood levels if you dose them orally; we’re trying to do something that’s rather unnatural as far as the body’s concerned. But we could still usefully take advantage of some of the transport and handling systems for such molecules.

But that’s not always easy to do. We all talk about making our compounds more polar and more soluble, but we balk at some of the things that will do that for us. Sure, you can slap a couple of methoxyethoxys on your ugly flat molecule, or hang a morpholine off the end of a chain to drag things into the water layer. But slap five or six hydroxyls on your molecule, and you’ll be lucky not to have the security guards show up at your desk.

There are, to be sure, some good reasons why they might. Hydroxyls and such tend to introduce chiral centers, which can make your synthesis difficult and dramatically increase the amount of work needed to fill out the structural possibilities of your lead series. That’s why these things tend to be (or derive from) natural products. Some bacterium or fungus has done most of the heavy lifting already, both in terms of working out the most active isomers and in synthesizing them for you. Erythromycin’s a fine starting material when you can get it by fermentation, but no one would ever, ever consider it if it had to be made by pure total synthesis.

There’s another consideration, which gets you right at the bench level. For an organic chemist, working with charged, water-soluble compounds is no fun. A lot of our lab infrastructure is built for things that would rather dissolve in ethyl acetate than water. A constant run of things with low logD values would mean that we’d all have to learn some new skills (and that we’d all probably have to spend a lot of time on the lyophilizer). Ion-exchange resins, gel chromatography, desalting columns – you might as well be a biochemist if you’re going to work with that stuff. But in the end, perhaps we might be better off, at least part of the time, if we were.

Comments (13) + TrackBacks (0) | Category: Drug Industry History | In Silico | Infectious Diseases

April 22, 2007

Melting Keys and Squishy Locks

Email This Entry

Posted by Derek

Pretty much the only thing that an interested lay person has heard about ligand binding is the "lock and key" metaphor. I'm not saying that you could walk down the sidewalk getting nods of recognition with it, but if someone's heard anything about how enzymes or receptors work (well, anything correct), that's probably what they've heard.

And there's a lot to it. Many proteins are really, really good at picking out their ligands from crowds of similar compounds. (If they were perfect at it, on the other hand, we drug company types would be out of business). But the lock-and-key metaphor makes the listener believe that both the ligand and the protein are rigid objects, which they most definitely are not. There's no everyday analog to the way that two conformationally mobile objects fit to each other - well, OK, maybe there is, but it's not one that you can safely use for illustrative purposes. Ahem.

The other big breakdown of the lock and key is that it doesn't deal well with the numerous proteins that can recognize more than one ligand for their binding sites. Particularly impressive are the nuclear receptors and the CYP metabolizing enzymes. Both those classes bind a bewildering number of not-very-similar compounds, and they can do it impressively well. They manage the trick by having binding pockets that can drastically change their shapes and charge distributions, as parts of the proteins themselves slide, twist, and flip around. I can't come up with even a vulgar metaphor for that process.

I'm thinking of doing several posts on the limits of metaphor and simplification in science, and if I do, this will be the first. It's a constant struggle not to mistake the picture for the real thing, particularly if the simplification is a pretty useful one. But eventually, no matter how good, the metaphor will thin out on you, and you'll be in the position of a Greek bird pecking at some painted fruit and wondering why it's still hungry.

Comments (29) + TrackBacks (0) | Category: In Silico | Metaphors, Good and Bad

March 12, 2007

No Shortcuts

Email This Entry

Posted by Derek

I wanted to link tonight to the "Milkshake Manifesto" over at OrgPrep Daily. It's a set of rules for med-chem, and looking them over, I agree with them pretty much across the board. There's a general theme in them of getting as close to the real system as you can, which is a theme I've sounded many times.

That applies to things like "Rule of Five" approximations and docking scores - useful, perhaps if you're sorting through a huge pile of compounds that you have to prioritize, not so useful if you've already got animal data.

He also takes a shot at Caco-2 cells and other such approximations to figure out membrane and tissue penetration. I've never yet seen an in vitro assay for permeability that I would trust - it's just too complicated, and it may never yield to a reductionist approach.

I'm a big fan of reductionism, don't get me wrong, but it's not the tool for every job. Living systems are especially tricky to pare down, and you can simplify yourself right out of any useful data if you're not very careful. The closer to the real world, the better off you are. It isn't easy, and it isn't cheap, but nothing good ever came easy or cheap, did it?

Comments (6) + TrackBacks (0) | Category: Drug Assays | Drug Development | In Silico

February 27, 2007

Wrong, But Still Convincing

Email This Entry

Posted by Derek

SciTheory has a post, complete with links to the relevant articles in Science, etc., on a recent batch of trouble in structural biology. Geoffrey Chang and his group at Scripps have been working on the structures of transporter proteins, which sit in the cell membrane and actively move nonpermeable molecules in and out. There are a heap of these things, since (as any medicinal chemist will tell you) a lot of reasonable-looking molecules just won't get into cells without help. It's even tougher at a physiological level, because (from a chemist's perspective) many of the things that need to be shuttled around aren't very reasonable-looking at all - they're too small and polar or too large and greasy.

Many of these transportersm especially in bacteria, fall into a large group known as the ABC transporters, which have an ATP binding site in them for fuel. (For the non-scientists in the audience, ATP is the molecule used for energy storage in everything living on Earth. Thinking of an ATP-binding site as a NiCad battery pack gets you remarkably close to the real situation). Chang solved the structure of one of these, the bacterial protein MsbA, by X-ray crystallography back in 2001, and it was quite an accomplishment. Getting good X-ray diffraction data on proteins which spend their lives stuck in the cell membrane is rather a black art.

How dark an art is now apparent - here's the original paper's abstract in PubMed, but if you look just above the abstract, you'll see a retraction notice, and it's not alone. Five papers on various structures have been withdrawn. As SciTheory says, anyone who doubted the original MsbA structure had some real food for thought last year when another bacterial transporter was solved at the ETH in Zurich. These two should have looked more similar than they did, to most ways of thinking, but they were quite divergent.

And now we know why. Chang's group was done in by some homebrew software which swapped two columns of data. In a structure this large and complicated, you can have such disruptive things happen and still be able to settle down on a final protein picture - it's just that it'll be completely wrong. And so it was. The same software seems to have undermined the other determinations, too.

This is important (as well as sad and painful) on several levels. For one thing, transporters are essential to understanding resistance to antibiotics and cancer therapies, and they're vital parts of a lot of poorly understood processes in normal cells. We're not going to be able to get a handle on the often-inscrutable distribution of drug candidates in living systems until we know more about these proteins, but now some of what we thought we knew has evaporated on us.

Another point that people shouldn't miss is the trouble with relying too much on computational methods. There's really no alternative to them in protein crystallography, of course, but there always has to be a final "Does that make sense?" test. The difficulty is that many perfectly valid protein structures show up with odd and surprising features. Alternately, it's unnerving that the data for these things can be so thoroughly hosed and still give you a valid-looking structure, but that just serves to underline how careful you have to be.

And we're talking about X-ray data, which (done properly) is considered to be pretty solid stuff. So what does this say about basing research programs on the higher levels of abstraction found in molecular modeling and docking progams?

Comments (21) + TrackBacks (0) | Category: In Silico

December 14, 2006

Love and Anger

Email This Entry

Posted by Derek

Glenn Reynolds gave the pharma industry a much-appreciated thank-you card over at Instapundit:

Only a moron would want to live in a society where people are ashamed to work for drug companies. And yet, I'm not surprised to see that resulting from the demagogy that abounds among politicians and "public interest" types who are not serving the public interest whatsoever.

I'm thinking of having that first sentence engraved on something expensive. Glenn's post prompted Dean Esmay to write a short post on the ethics of drug companies, though, and he's rather less positive. I suppose I shouldn't be surprised, given some of the things he's gone in for in the past. As usual, some of the problem is the difficulty that people have coming to terms with the fact that drug discovery is a for-profit industry.

One comment on his post came from Jerry Kindall, which is mostly favorable to the industry, but nonetheless contains this paragraph:

Drug discovery used to be a total crap-shoot but it's getting more and more targeted as the years go by thanks to ever more sophisticated computer modeling. They are now able to say "okay, this is the chemical receptor that we think we need to address, let's design a molecule that fits into it." This is essentially a nanotechnology, although not the type most people think of when they hear the term.

Ay, would that it were true. As my industry readers know, and as I've been ranting abouit here fairly often, drug discovery is just as much of a crap-shoot as it's ever been. And wouldn't it be great if "sophisticated computer modeling" helped that much? Instead, we get things like this. No, I think what's happening here is that we're being underestimated by our enemies and overestimated by our friends. . .

Comments (32) + TrackBacks (0) | Category: In Silico | Why Everyone Loves Us

September 11, 2006

Enzymes Do Whatever They Want To

Email This Entry

Posted by Derek

It's been a while since I wrote about the neuraminidase inhibitors (Tamiflu and Relenza, oseltamavir and zanamivir). As we start to head into fall, though, I'm sure that avian flu will invade the headlines again, if nothing else (and I hope it's nothing else).

There's an interesting report in Nature (subscriber link) on how these drugs work. Bird flu is a Type A influenza, but there are two broad groups inside that class, which are defined by what variety of neuraminidase enzyme they express. (There are actually nine enzyme variants known, but four of them fall into one group and five into the other).

The drugs were developed against group-2 enzymes, but they're also effective against group-1 influenzas. Since the X-ray crystal structures showed the the drugs bound in the same way to all the group-2 neuraminidases, and since the active sites of all the subtypes across the two groups are extremely similar, no one ever thought that their binding modes would be different. Well, until last month, anyway, when the X-ray crystallographic data came in.

And what it showed was that the active sites of the group-1 enzymes, sequence homology be damned, have a much different structure than the group-2s. As it turns out, though, they can adopt a similar shape when an inhibitor binds to them, which is why the marketed inhibitors still work on them, but they're fundamentally quite different.

I can't resist the urge to use this example to illustrate some of the real problems in our current state of the art for computation and modeling. The differences between these two enzymes are due to their different amino acid residues far away from the active site, which makes modeling them much, much more difficult (and makes the error bars much, much wider when you do). That's why no one realized how far off the group-1 and group-2 neuraminidases were until the X-ray structure was available: modeling couldn't tell you. Any modeling efforts that tried would probably have decided, incorrectly, that the two groups were nearly identical. Why shouldn't they be?

But if we'd had that X-ray data from the start, modeling would very likely have told you, incorrectly, that there was little chance that either Relenza or Tamiflu would work on the group-1 enzyme variants. Why should they? The "induced fit" binding modes, where the enzyme changes shape significantly as the ligand binds, are understandably very difficult to model. There are just too many possibilities, too many of which are within each other's computational error bars.

Now, it's true that this latest work isn't based on molecular modeling at all. (You have to wonder how close these guys got, though). But plenty of projects that are using it are just as much in the dark as a neuraminidase team would have been, and they may not even realize it. Most molecular modelers are well aware of these limitations, but not all of them - or all of the managers over them - are willing to accept them. And when you get out to investors or the general public, it's all too easy for modelers or managers to act as if things are perfectly under control, when in reality they're lurching around in the dark. Like the rest of us. . .

Comments (11) + TrackBacks (0) | Category: In Silico | Infectious Diseases

March 23, 2006

Crystals of Doubt

Email This Entry

Posted by Derek

Here's a limits-to-knowledge post for you. On Wednesday, when I was cranking out a batch of an intermediate we're using these days, I needed to separate two fairly closely related compounds (which I'll call A and B) from each other. One surefire way to have done that was chromatography, but I just didn't have time for that. While I was rota-vapping down the mixture, I noticed that some white crystals were starting to come out of the methylene chloride solution, so I took the flask off and checked a small sample of the solid. Sure enough, it was pretty pure A, so I filtered that off and continued.

Taking out all the solvent left me with more white stuff, which was mostly B, with some A still hanging in there. In the past, we'd purified B by crystallizing it from another solvent mixture (ethyl acetate/hexane, the first combination the lazy - or just plain experienced - organic chemist reaches for). So I tried that out, dissoving the solid in a small-to medium amount of hot ethyl acetate, then adding hexane while it was still warm. I cooled the solution down by dipping the flask in ice water until it had come down to about room temperature, and was swirling it around when suddenly it starting snowing white powder. Ta-daa! A check of this stuff showed that it was almost completely pure B. The solution, for its part, was now a majority of A with some B left around. I took what I had and ran with it - this was one of the bird-in-the-hand situations, because people were waiting on this stuff.

My point is that such things are almost completely empirical. I've never heard of anyone who could predict from first principles what solvent system to use to get something to crystallize. I'd be tremendously impressed if anyone could take the structures of my two compounds, feed them into a dissolvo-matic program and announce "Yep, methylene chloride for A, and ethyl acetate-hexane for B. That'll do the trick."

As far as I know, there's no such thing, and no one is even close. I'd be glad to hear if I'm wrong. But if we can't predict, even just in rank order, what solvents will dissolve (or crash out) a given molecule, just how good is our molecular modeling, anyway?

Comments (8) + TrackBacks (0) | Category: In Silico

November 1, 2005

Molecular Modeling Cage Match

Email This Entry

Posted by Derek

I mentioned an interesting paper that's coming out in the Journal of Medicinal Chemistry on molecular modeling. It's a long one from a large group of people scattered across GlaxoSmithKline's worldwide research facilities, entitled "A Critical Assessment of Docking Programs and Scoring Functions." And that's what it is, all right.

For the non-med-chem readers, those are two of the key techniques in computational molecular modeling. Docking refers to taking a modeled version of your small molecule and trying to fit it into a similarly modeled version of the binding site of your protein target. The program ties to take into account the size and shape of the molecule and the binding site, of course, as well as more subtle interactions between the various functional groups. Scoring functions are what the programs use to try to rate how well the docking procedure went for a given compound, and to compare it to others in a given data set.

The GSK team did a very thorough job, evaluating ten different docking programs. They started with seven varying types of protein targets, mostly different classes of enzymes, all of which are known drug targets. An expert computational chemist took each one and polished up the model of the binding site. At the same time, lists of between one and two hundred potential binding compounds were put together for each target, including several series of related compounds. Another modeling chemist took these structures and got them ready for docking. They made sure that a crystal structure of each structural class was known for each case (to check the accuracy of the modeling later on), and also made sure that the binding affinity of the compounds ranged over at least four orders of magnitude (from pretty darn good, in other words, to pretty darn awful). The goal was to make the whole exercise as real-world as possible. Then each of those binding site models and their associated lists of potential ligands were turned over to separate chemists with experience in the various docking programs, and they told them to have at it. As the paper puts it:

"To optimize the performance of each docking program, computational chemists with expertise in a particular program were identified from the worldwide GSK computational chemistry community. Each program expert was given complete freedom and sufficient time to maximize the performance of the docking program. . .No time deadlines were imposed so that even low-throughput docking programs could be evaluated. Indeed, no constraints whatsoever were placed on the level of agonizing over details of how each docking program was applied."

It's important to remember that the results of this paper come from experienced users who had a great deal of knowledge about the targets, and all the time they needed to mess with them. The aformentioned agonizing was devoted to three typical kinds of question that such software is designed to answer: The first was: what is the conformation (the 3-D physical "pose") of a small molecule once it's in a binding site? This is why they picked all these things with known crystal structures, since those provide a check with real data. Results of this test were OK, in some cases fairly good. Some of the target proteins seemed to have binding sites that were more suited for the capabilities of the programs, which could take the majority of the compounds in their list and fit them pretty close (within two angstroms) to the known crystal structures.

And every target had at least one program that could take at least a third or so of the test compounds and dock them fairly well. But the problem was, no one program could do that for more than 35% of the binding modes. The best performances were scattered among the different software packages, and there seems to be absolutely no way to know in advance whether a given program is going to perform well on a new target. The other problem, and it's a big one, was that the scoring functions couldn't reliably identify when the program had hit on one of the good answers. There wasn't much correlation between what the program thought was a well-docked conformation and its resemblance to the known crystal structure.

The second question they looked at was: given a list of molecules (some active, some inactive), how well can the software pick out some active ones? This process is often known as "virtual screening". Again, the results were fairly good, but with some significant problems. For all but one of the targets, at least one of the programs could find at least half of the top 10% of the active compounds. (I know, that sounds like a lot of defensive hedging compared to what some people think these programs can do, but that's the real world for you). The programs also did pretty well at pulling a variety of structures out, and not just making their total by grabbing only the members of one particular class.

But that fairly-decent performance is for the programs as a group. As before, though, the best performances were scattered through all the software packages, with no real standout. Most of the programs, at one point or another, had to grind through a significant amount of a compound lists to do the job, too, which is something you really don't want in real-world use. Another disturbing result was that some of the scoring functions seemed to be picking the right compounds for the wrong reasons – that is, based on incorrect binding modes.

Now we're ready for the third question, a hard one which (in my experience) is one of the ones that medicinal chemists most would like molecular modeling software to answer: given a list of compounds, can the program rank-order them according to their expected affinity for the target? Unfortunately, the answer is "absolutely not." No scoring function in any of the software packages could even come close. The compounds that the programs ranked as winners were just as likely to stink, and the ones that they put into the discard heap were just as likely to be fine.

My way of looking at the first two tests is to say that if you have just one molecular modeling package, it is guaranteed to mislead you a fair amount of the time. And you have no way of knowing when it's doing that. If you have more than one program to work with, though, then they are guaranteed to disagree with each other a fair amount of the time, and you have no way of knowing which one of them is right – if either. I'll let the authors have last word on the third test, and on the software in general:

". . .in the area of rank-ordering or affinity prediction, reliance on a scoring function alone will not provide broadly reliable or useful information. . .This study demonstrates unequivocally that significant improvements are needed before compound scoring by docking algorithms will routinely have a consistent and major impact on lead optimization. . .it is not completely obvious by what means these improvements will arise. . ."

Comments (5) + TrackBacks (0) | Category: In Silico

September 29, 2005

The Hazards of Molecular Modeling

Email This Entry

Posted by Derek

A comment to the last post really gave me the shivers:

"I like to think of modelling as the "silent killer". It is easy to rely on it for quick answers, and easy to forget that there is no substitute for an actual experiment. . .

I remember asking a fellow scientist if a particular molecule performed as hypothesized, the response was: " I don't know. It did not dock well into the enzyme, so I didn't make it."

I've made this point before, but it needs to be made again: molecular modeling is not reality. Most models are not that good, or only good around a limited group of rather similar compounds. If you as a medicinal chemist are crossing out easy-to-make compounds in unexplored chemical space just because the software doesn't like it, you are handcuffing yourself and tying your thumbs together. Stop it, stop it for your own good, or you may never discover anything unexpected or useful.

"The silent killer": I like that phrase a lot. I get the occasional testy e-mail from the computational types when I talk like this, but I'm sticking to my beliefs here. Molecular models based on numerous high-resolution X-ray structures are, I think, sort of useful, sometimes. Models based on only one X-ray structure are to be approached with great caution. And binding models that are just calculated up de novo should be treated as hazardous to your scientific health, unless you have a great deal of evidence to make you think otherwise.

OK, you silicon jockeys, go ahead and flood my in-box. I've earned it.

Comments (7) + TrackBacks (0) | Category: In Silico

September 28, 2005

Clamping Down, or Loosening Up?

Email This Entry

Posted by Derek

We medicinal chemists spend our days trying to make small molecules that bind to targets in living systems. Almost all of those targets are proteins of one sort or another, and most of them have binding pockets already built into them, which we're trying to hijack for our own purposes. Molecular modelers try to figure out how these things fit together, but there are still a lot of unknowns in what would seem so basic a process.

I'm willing to bet that most chemists and biologists have a mental picture of a small molecule ligand fitting into a binding site which involves the protein sort of folding down around things - gently biting down on the ligand, as it were. It seems intuitively obvious that a protein's motions would settle down once it complexes with its target molecule.

And like a lot of intuitively obvious things in drug research, that idea appears to be mistaken. There's a recent study in the Journal of Medicinal Chemistry from a group at Michigan that tackles this question in a rigorous manner. They looked through the X-ray crystal structure data banks for proteins that have had high-quality structures determined both with and without small molecules bound in them. After controlling for experimental conditions (the temperature that the X-ray structure was taken at, among other things) and for the way the data were processed, they still had a few dozen closely matched pairs.

What they found was that in most of these structures, at least some of the atoms in and near the binding site are more mobile when there's a ligand bound. At times, the effect was pretty dramatic, with the entire binding site becoming more flexible, weirdly enough. Examples where everything got less mobile were found, but that only happened in a minority of the cases. The proteins the authors studied were scattered across a wide range of structural and functional classes, and there's no reason to think that they hit on an anomalous data set.

So, we're going to have to adjust our mental pictures, and the molecular modelers will have to adjust their simulations. I'd like to know just how many of those in silico models of binding would have predicted this greater flexibility. I fear that the answer is "darn near none of them". We have a long way to go.

Comments (9) + TrackBacks (0) | Category: In Silico

September 6, 2005

Crossing Your Fingers, Authoritatively

Email This Entry

Posted by Derek

I recall a project earlier in my career where we'd all been beating on the same molecular series for quite a while. Many regions of the molecule had been explored, and my urge was often to leave the reservation. I put some time into extending the areas we knew about, but I wanted to go off and make something that didn't look like anything that we'd done before.

Which I did sometimes, and then I'd often get asked: "Why did you make that compound?" My answer was simply "Because no one had ever messed with that area before, and I wanted to see what would happen." Reactions to that approach varied. Some folks found that a perfectly reasonable answer, sufficient by itself. Others didn't care for it much. "You have to have a hypothesis in mind," they'd say. "Are you trying to improve the pharmacokinetics? Fix a metabolic problem? Pick up a binding interaction that you think is out there in the XYZ loop of the protein? You can't just. . .make stuff."

I respected the people in that first group a lot more than I did the ones in the second. I thought then, and think now, that you can just go make stuff. In fact, you not only can, but you should. You probably don't want to spend all your time doing that, but if you never do it at all, you're going to miss the best surprises.

I take issue with the idea that there has to be a specific hypothesis behind every compound. That supposes amounts of knowledge that we just don't have. Most of the time, we don't know why our PK is acting weird, and we're not sure about the metabolic fate of the compounds. And we sure don't know their binding mode well enough to sit at our desks and talk about what amino acids in the protein backbone we're reaching out for. (OK, if you've got half a dozen X-ray structures of your ligands bound in the active site of your target, you have a much better idea. But if your next compound breaks new structural ground, off you may well go into a different binding mode, and half your presuppositions will go, too.)

I like to think that I've come to realize just how ignorant I am in issues of drug discovery. (In case you have any doubt, I'm very ignorant indeed.) But I still hear people confidently sizing up new analog ideas on the blackboard, though: No, that one won't bind well in the Whoozat region. Doesn't have the right spacing. And that one should be able to reach out to that hydrophobic pocket we all know about. Let's make that one first. (These folks are talking without X-ray structures in hand, mind you.)

Well, if it makes you feel better, then go ahead, I suppose. But this kind of thing is one tiny step up from lucky rabbit feet, for which there is still a market.

Comments (4) + TrackBacks (0) | Category: In Silico | Who Discovers and Why

August 17, 2004

Kinases and Their Komplications

Email This Entry

Posted by Derek

I'm going to take off from another comment, this one from Ron, who asks (in reference to the post two days ago): "would it not be fair to say that cellular biochemistry gets even more complicated the more we learn about it?

It would indeed be fair. I think that as a scientific field matures it goes through several stages. Brute-force collection of facts and observations comes early on, as you'd figure. Then the theorizing starts, with better and better theories being honed by more targeted experiments. This phase can be mighty lengthy, depending on the depth of the field and the number of outstanding problems it contains. A zillion inconsistent semi-trivialities can take a long time to sort out (think of the mathematical proof of the Four-Color Theorem), as can a smaller number of profound headscratchers (like, say, a reconciliation of quantum mechanics with relativity as they deal with gravity.)

If the general principles discovered are powerful enough, things can get simpler to understand. Think of the host of problems that early 20th-century physics had, many of which resolved themselves as applications of quantum mechanics. Earlier, chemistry went through something similar earlier, on a smaller scale, with the adoption of the stereochemical principles of van't Hoff. Suddenly, what seemed to be several separate problems turned out to be facets of one explanation: that atoms had regular three-dimensional patterns of bonding to other atoms. (If that sounds too obvious for such emphasis, keep in mind that this notion was fiercely ridiculed at resisted at the time.)

Cell biology is up to its pith helmet in hypotheses, and is nowhere near out of the swamps of fact collection. As in all molecular biology, the sheer number of different systems is making for a real fiesta. Your average cell is a morass of interlocking positive and negative feedback loops, many of which only show up fleetingly, under certain conditions, and in very defined locations. Some general principles have been established, but the number of things that have to be dealt with is still increasing, and I'm not sure when it's going to level out.

For example, the other day a group at Sugen (now Pfizer) published a paper establishing just how many genes there are in mice that code for protein kinase enzymes. Through adding phosphoryl groups, these enzymes are extremely important actors in the activation, transport, and modulation of the activities of thousands upon thousands of other proteins, and it turns out that there are exactly 540 of them. (Doubtless there are some variations as they get turned into proteins, but that's how many genes there are.) And that's that.

Now, that earlier discovery of protein phosphorylation as a signaling mechanism was a huge advance, and it has been appropriately rewarded. And knowing just how many different kinase enzymes there are is a step forward, too. But figuring out all the proteins they interact with, and when, and where, and what happens when they do - well, that's first cousin to hard work.

Comments (0) + TrackBacks (0) | Category: Biological News | In Silico

August 15, 2004

FullCell 1.0?

Email This Entry

Posted by Derek

Reader Maynard Handley, in a comment to the most recent post below, asks:

". . .how far are we from doing at least a substantial fraction of this stuff in silico? I've read that some amazing computational models of full cells now exist, but even so, this author didn't expect that drugs could be usefully tested computationally until 2030 which seems awfully far out."

I don't know the article that he's referring to, but "awfully far out" pretty much sums up my reaction, too. I just don't think we have enough data to do any real whole-cell modeling yet. It's coming, and perhaps for a few very well-worked-out subsystems we could do it now, but I'm sceptical even of that.

A few days reading the current cell biology literature will illustrate the problem. All sorts of proteins are found, all the time, to be players in systems that no one suspected them of being involved it. Kinases are found to phosphorylate things that no one had seen them do before, lipases are found to accept substrates that no one had realized they could. A given signaling peptide is gradually found to have more uses than a Swiss army knife. We don't even really understand the basic mechanisms (like G-protein-coupled receptor signaling) enough to model them to any useful level.

The process of finding these things out doesn't seem like it's going to end soon, and there have to be many fundamental surprises waiting for us. Modeling the system in their absence is going to be risky - interesting, no doubt, and potentially lucrative (if you find a useful approximation), but risky. It's going to take some pretty convincing stuff for the drug industry to ever depend on it.

And all of this applies to single cells, which come in, naturally, an uncounted variety, each with its own peculiarities, the great majority of which we don't have any clue about. And then you come to the interactions between cells, which are highly significant and (in many ways) a closed book to us at present. If we knew more about these things, we'd be able, for example, to culture human cell lines that acted just like their primary tissue progenitors - but we can't do it, not yet.

No, although I have every belief that these things are susceptible to modeling, I just don't think we'll see it (on a useful scale) any time soon. Over the next twenty years, I'd expect to have some of the easier-to-handle cellular subsystems worked out to give robust in silico treatments, but a whole cell? And all the types of whole cells? Much longer than that. More than that I can't guess.

Comments (3) + TrackBacks (0) | Category: In Silico

April 15, 2004

The March of Folly

Email This Entry

Posted by Derek

Thinking about molecular modeling, as I did in the last post, brings up another topic: when you go back to the late 1980s, in the real manic phase of the technological hype, what brings you up short is realizing that these folks were planning on doing all this with 1980s hardware.

That puts things in perspective. Here we are in 2004, and we still can't just sit down and design a drug from first principles. Don't believe anyone who tells you that we can, either - if that were possible, there would be a lot more drugs out there. I'm not saying that molecular modeling never makes a contribution (I know better, and from personal experience.) It's just that it hasn't (yet) caught up to the hallucinations of fifteen or twenty years ago, which is entirely the fault of the people who were doing the hallucinating.

You can make the same comments about other waves of hype that have broken over the pharmaceutical world (combinatorial chemistry comes immediately to mind.) What I'm wondering is: what's the hype of today? There's bound to be a hot new idea that's going to solve our problems, but will end up changed beyond recognition after twenty years of the real world. Any votes on what's going to look faintly ridiculous to us in 2024? As you'd guess, I have some candidates of my own. . .

Comments (2) | Category: Drug Industry History | In Silico

April 14, 2004

Reality's Revenge

Email This Entry

Posted by Derek

Molecular modeling is a technology with a past. Specifically, it's a past of overoptimistic predictions (often made, to be fair, by people who didn't understand what they were talking about.) Back in the late 1980s, when I started in the drug industry, modeling was going to take over the world and pretty darn soon, too. Several companies were founded to take advantage of this brave new world that had such software in it, and they raised serious money with tales of how they were just going to zzzzzip right to the drug structures. No dead ends, no detours, no cast of thousands - just a few chemists standing by to make the structure as it printed out for them. This has not quite worked out.

For those not in the business, modeling is the attempt to figure out molecular shapes, properties, and interactions by computation. There are many levels, some more successful than others. The ones I'm speaking of involve predicting three-dimensional shapes of molecules (and their target binding sites), and deciding which ones are more likely to fit well. It sounds like just what we need. It also sounds reasonably doable, in the same way that Hercules was probably told at first that he was going to just have to round up a few stray animals.

Predicting the shapes involves modeling the individual chemical bonds, and the interactions as the atoms and functional groups rotate around them, banging into each other or sticking through various forces. Originally, these things were calculated as if they were in interstellar space, with nothing around them. Later (and ever since) a number of methods to add some real-world solvent effects have been tried.

Another set of programs evaluates intermolecular fits, trying to work out the energies in play when a drug molecule slides into its binding site. Many tricky refinements have been added to those packages over the years, too, taking advantage of the latest insights into how various groups stack, pack, and interact.

And often enough, it just isn't enough. Many times the structures we have for our binding sites aren't accurate - the best ones are from X-ray crystallography, and plenty of good stuff just doesn't crystallize. (There are other cases where the crystal structure doesn't bear much relation to what's going on inside the real system, too, just to keep everyone on their toes.) Modeling goes haywire for all kinds of reasons.

One of the companies that emerged back in the change-the-world era of modeling was Vertex, up in Cambridge. It was founded by Joshua Boger, a Merck chemist who wanted a piece of the new thing and wasn't sure that Merck was taking it seriously enough. Well, coming soon in the Journal of Medicinal Chemistry (it's in the web preprint section now) is a paper from Vertex which gives us all some idea of why things didn't work out quite as planned.

The Vertex guys went back over about 150 cases, and found that in the majority of them, the structure of the small molecule in its binding pocket wasn't the structure you would have predicted as the best (read: lowest-energy.) In many of them, it isn't even close. You'd literally never have picked some of these conformations to start a modeling effort - they look very disfavored, and if you're going to pick things that far from the ground state then there's no end to it. The number of structures gets worse very rapidly as you move away from the local energy minima.

We in the business had suspected as much, and everyone knew of an example or two, but this is a quantitative look at just how bad the situation is. When you add in the cases where the binding site changes its conformation unexpectedly in response to the ligand, it's a wonder that any modeling efforts work at all. (Frankly, in my experience, they mostly don't, but I'm willing to stipulate that my experience has been more negative than the average.)

I like to say that molecular modeling is a magic wand, one that we keep waving in the hope that sparks will eventually start to shoot out of it. Someday they will. But there's a lot more hard work ahead, and no shortcuts in sight.

Comments (0) | Category: Drug Industry History | In Silico