Michael Shultz of Novartis is back with more thoughts on how we assign numbers to drug candidates. Previously, he's written about the mathematical wrongness of many of the favorite metrics (such as ligand efficiency), in a paper that stirred up plenty of comment.
His new piece in ACS Medicinal Chemistry Letters is well worth a look, although I confess that (for me) it seemed to end just when it was getting started. But that's the limitation of a Viewpoint article for a subject with this much detail in it.
Shultz makes some very good points by referring to Daniel Kahneman's Thinking, Fast and Slow, a book that's come up several times around here as well (in both posts and comments). The key concept here is called "attribute substitution", which is the mental process by which we take a complex situation, which we find mentally unworkable, and try to substitute some other scheme which we can deal with. We then convince ourselves, often quickly, silently, and without realizing that we're doing it, that we now have a handle on the situation, just because we now have something in our heads that is more understandable. That "Ah, now I get it" feeling is often a sign that you're making headway on some tough subject, but you can also get it when you're understanding something that doesn't help you with it at all.
And I'd say that this is the take-home for this whole Viewpoint article, that we medicinal chemists are fooling ourselves when we use ligand efficiency and similar metrics to try to understand what's going on with our drug candidates. Shultz go on to discuss what he calls "Lipinski's Anchor". Anchoring is another concept out of Thinking Fast and Slow, and here's the application:
The authors of the ‘rules of 5’ were keenly aware of their target audience (medicinal chemists) and “deliberately excluded equations and regression coefficients...at the expense of a loss of detail.” One of the greatest misinterpretations of this paper was that these alerts were for drug-likeness. The authors examined the World Drug Index (WDI) and applied several filters to identify 2245 drugs that had at least entered phase II clinical development. Applying a roughly 90% cutoff for property distribution, the authors identified four parameters (MW, logP, hydrogen bond donors, and hydrogen bond acceptors) that were hypothesized to influence solubility and permeability based on their difference from the remainder of the WDI. When judging probability, people rely on representativeness heuristics (a description that sounds highly plausible), while base-rate frequency is often ignored. When proposing oral drug-like properties, the Gaussian distribution of properties was believed, de facto, to represent the ability to achieve oral bioavailability. An anchoring effect is when a number is considered before estimating an unknown value and the original number signiﬁcantly inﬂuences future estimates. When a simple, specific, and plausible MW of 500 was given as cutoff for oral drugs, this became the mother of all medicinal chemistry anchors.
But how valid are molecular weight cutoffs, anyway? That's a topic that's come up around here a few times, too, as well it should. Comparisons of the properties of orally available drugs across their various stages of development seem to suggest that such measurements converge on what we feel are the "right" values, but as Shultz points out, there could be other reasons for the data to look that way. And he makes this recommendation: "Since the average MW of approved oral drugs has been increasing while the failure rate due to PK/biovailability has been decreasing, the hypothesis linking size and bioavailability should be reconsidered."
I particularly like another line, which could probably serve as the take-home message for the whole piece: "A clear understanding of probabilities in drug discovery is impossible due to the large number of known and unknown variables." I agree. And I think that's the root of the problem, because a lot of people are very, very uncomfortable with that kind of talk. The more business-school training they have, the less they like the sound of it. The feeling is that if we'd just use modern management techniques, it wouldn't have to be this way. Closer to the science end of things, the feeling is that if we'd just apply the right metrics to our work, it wouldn't have to be that way, either. Are both of these mindsets just examples of attribute substitution at work?
In the past, I've said many times that if I had to work from a million compounds that were within rule-of-five cutoffs versus a million that weren't, I'd go for the former every time. And I'm still not ready to ditch that bias, but I'm certainly ready to start running up the Jolly Roger about things like molecular weight. I still think that the clinical failure rate is higher for significantly greasier compounds (both because of PK issues and because of unexpected tox). But molecular weight might not be much of a proxy for the things we care about.
This post is long enough already, so I'll address Shultz's latest thoughts on ligand efficiency in another entry. For those who want more 50,000-foot viewpoints on these issues, though, these older posts will have plenty.