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: Don't miss Derek Lowe's excellent commentary on drug discovery and the pharma industry in general at In the Pipeline

In the Pipeline

« Repurposing for Cervical Cancer | Main | Startups vs. Big Pharma »

March 10, 2014

A Blood Test for Alzheimer's?

Email This Entry

Posted by Derek

Update: more doubts on the statistical power behind this, and the coverage it's getting in the press.

There's word of a possible early diagnostic blood test for Alzheimer's. A large team (mostly from Georgetown and Rochester) has published a paper in Nature Medicine on their search for lipid-based markers of incipient disease. They say that they have a ten-lipid panel that has a 90% success rate in predicting cognitive decline within three years.

I can certainly see how this would be possible - lipids could be markers of membrane trouble and myelin trouble, and we already know that the lipoprotein ApoE4 is linked with Alzheimer's. At the same time, I'd like to see how this looks when more data are available. The absolute number of patients showing the effect isn't large. And there's always a danger, on these biomarker fishing expeditions, of finding a spurious correlation. The fact that it takes ten lipids to get the accuracy up could be OK, or it could be a sign of statistical trouble. (It's a bit like seeing a QSAR model that needs ten parameters to be predictive).

But this could indeed be real, and if it is, a larger sample will nail it down. That should also give a much better idea of the false-positive and false-negative rates, which will be very important in a diagnosis like this. It'll also be interesting to see if the time horizon can be improved past three years. The usual worries about an Alzheimer's diagnostic apply - some people will want to know, and others won't, since there's no treatment. If this works out, though, it would also seem to be very useful for future clinical trials, which are (more and more) focusing on people in the earliest stages of the disease.

Comments (21) + TrackBacks (0) | Category: Alzheimer's Disease


1. MoBio on March 10, 2014 8:31 AM writes...

A very small collection of patients (106 total entered into the 'omics' study) and not particularly diverse:

"This biomarker panel requires external validation using similar rigorous clinical classification before further development for clinical use. Such additional validation should be considered in a more diverse demographic group than our initial cohort."

Also, a large number of analytes were initially quantified:

"144 lipids simultaneously by multiple reaction monitoring. The other metabolites are resolved on the UPLC and quantified using scheduled MRMs. The kit facilitates absolute quantitation of 21 amino acids, hexose, carnitine, 39 acylcarnitines, 15 sphingomyelins, 90 phosphatidylcholines and 19 biogenic amines..."
to arrive at the final set which were used for what appears to be retrospective =based predictions.

Finally, the blood test only predicts "memory impairment in older adults" not frank Alzheimer's Dementia.

Other than that I liked the paper.

Permalink to Comment

2. Anonymous on March 10, 2014 8:53 AM writes...

This is similar to a workflow we use looking at plasma biomarkers of pancreatic cancer, where global lipid dysregulation would also be suspected. Got great hits, figured out some new lipid structures along the way. The trouble is validation in comparing cancer vs benign pancreatitis.... Lastly, good luck trying to get funding to validate that or trying to patent that biomarker.

Permalink to Comment

3. cynical1 on March 10, 2014 8:54 AM writes...

And if it works, it'll be used by the insurance companies against you. Imagine applying for long term care insurance where you have to take this test every year to requalify. This isn't a good thing. And what do you really do with this information? Forget that you're going to get Alzheimer's?

Permalink to Comment

4. Anon on March 10, 2014 11:21 AM writes...

I hate to be a negative-Nancy...but just throwing some scientific rigor out is this THAT much different than all of the array datasets out there? I'm going to take a shot in the dark and guess that the PIs that "designed" this experiment were MDs? The NIH loves to pay those guys to do correlative research!

Permalink to Comment

5. DCRogers on March 10, 2014 12:03 PM writes...

From the paper, the authors discovered these 10 variables from a collection of around ~4000 they measured.

I took 106 samples evenly split between active and inactive, and created 4000 random variables. Using a generic algorithm (GFA), I discovered one model with 10 variables with an r^2 of 0.73. A ROC plot with that model on the training data gave a ROC score of 0.98 (which means if a person was given a true active and an true inactive, and used the model to choose, you would be correct 98% of the time).

I don't have access to the original paper to see what validation they performed, but I'm hoping they are aware of the huge chance for random models when you select 10 variables out of 4000.

Permalink to Comment

6. Kevin on March 10, 2014 12:21 PM writes...

@DCRogers: That is precisely the problem that comes up over and over again. How do we get MD's to understand this?

Is it more clear to state the analysis is not explaining the probability that these 10 predictors are correlated with the observable, but that some 10 predictors out of 4000 could be correlated? Intuitively, that should make sense why the probability of observing some correlation by random chance alone is high.

Permalink to Comment

7. Sam P on March 10, 2014 12:36 PM writes...

I really wish that abstracts on scientific papers on tests would supply false positive and false negative rates, rather than "accuracy".

According to the Alzheimer's Association's 2012 Alzheimer's Disease Facts and Figures brochure, the 1 year incidence rate for people ages 65-74 for Alzheimer's and other dementias is about 53 per 1000. (It also states that it is estimated that 60-80% of these dementias are Alzheimer's) So if their test is all false positives, then we'd have 100 false positives and 159 true positives. Flip it to all false negatives and we'd get 143 true positives and 16 false negatives. So 90% accuracy isn't too bad for this base rate.

Permalink to Comment

8. Anonymous on March 10, 2014 1:56 PM writes...

@DCRogers: I'm not a statistician by any shake of the imagination. Maybe this means more to you...

"The m/z features of metabolites were normalized with log transformation that stabilized the variance, followed by a quantile normalization to make the empirical distribution of intensities the same across samples37. The metabolites were selected among all those known to be identifiable using a ROC regularized learning technique38,39 based on the LASSO penalty8,9 as implemented with the R package ‘glmnet’40, which uses cyclical coordinate descent in a path-wise fashion. We first obtained the regularization path over a grid of values for the tuning parameter λ through tenfold cross-validation. The optimal value of the tuning parameter lambda, which was obtained by the cross-validation procedure, was then used to fit the model. All the features with nonzero coefficients were retained for subsequent analysis. This technique is known to reduce overfitting and achieve similar prediction accuracy as the sparse supporting vector machine. The classification performance of the selected metabolites was assessed using area under the ROC curve (AUC). The ROC can be understood as a plot of the probability of classifying correctly the positive samples against the rate of incorrectly classifying true negative samples. So the AUC measure of an ROC plot is a measure of predictive accuracy. To maintain rigor of independent validation, the simple logistic model with the ten-metabolite panel was used, although a more refined model can yield greater AUC. The validation phase was performed in a blinded fashion such that the sample group was not known by the statistical team."

Permalink to Comment

9. DCRogers on March 10, 2014 2:41 PM writes...


Thanks for the excerpt from the paper -- it's a lot to chew on!

But it sounds to me like their 'validation phase' happens after variable selection -- so the bias introduced by variable selection is not being measured.

My gut suspicion is that they might have some markers with useful information, but the quality will be somewhat like many genetic markers: "20% increased risk of XXX", not the advertised "90% accurate".

Permalink to Comment

10. Anonymous on March 10, 2014 2:59 PM writes...

I spotted the statistical problem with this paper as soon as I heard about it:

Reading a "high" or "low" signal of 10 different lipids gives 2^10 = 1024 possible combinations, which is double the number of people actually tested. That virtually guarantees you will find a correlation. Completely useless!

Permalink to Comment

11. Jose on March 10, 2014 7:26 PM writes...

The authors DO in fact correct for so many combinations (multiple testing) using a Bonferroni correction, BUT they use 0.025 (0.05 ^2), but they should be down in the 5 E -10 range!

Permalink to Comment

12. RKN on March 10, 2014 7:47 PM writes...

Reading the link provided in Derek's update, it appears the authors never considered the impact of disease prevalence in the wider population when computing the positive predictive value of the test, which means 90% is greatly over optimistic.

Permalink to Comment

13. Allchemistry on March 11, 2014 1:14 AM writes...

The authors performed cross-validation to test the robustness of the model and to prevent for overfitting, so I think that the predictors they identified are reliable. As pointed out above, a sensitivity and a specificity of 90% is simply not sufficient for early diagnosis of Alzheimer.
On the other hand, we should not throw out the baby with the bathwater. Based on this approach it may be possible to develop simple blood tests for monitoring disease progression or -treatment efficacy in patients diagnosed with neurodegenerative diseases (MS, ALS etc).

Permalink to Comment

14. TheScarletPimple on March 11, 2014 4:41 AM writes...

Here is an interesting take on the topic that looks at the statistical issues involved (good for non-stats people like me):

Permalink to Comment

15. Anonymous on March 11, 2014 6:31 AM writes...

@14 As I linked to Derek's post at Language Log will the blogosphere disappear into a black hole?
(P.S. Apolgies to Derek for spelling his name wrongly (hangs head in shame))

Permalink to Comment

16. Anonymous on March 11, 2014 6:33 AM writes...

@14 As I linked to Derek's post at Language Log will the blogosphere disappear into a black hole?
(P.S. Apolgies to Derek for spelling his name wrongly (hangs head in shame))

Permalink to Comment

17. GladToMoveToProcess on March 11, 2014 6:58 AM writes...

From Derek's post: "The fact that it takes ten lipids to get the accuracy up could be OK, or it could be a sign of statistical trouble. (It's a bit like seeing a QSAR model that needs ten parameters to be predictive)." Back in the 60s, Prelog gave a lecture in which he quoted some professor of his: "Give me three parameters and I can fit the elephant..."

Permalink to Comment

18. petros on March 11, 2014 9:17 AM writes...

While interesting my gut feel is that you would need to replicate this study in a much larger group to get a good feel for its potential reliability.

Permalink to Comment

19. Lane Simonian on March 12, 2014 10:52 AM writes...

I cannot comment on the accuracy of the test, but lipid rafts in a sense provide the foundation for Alzheimer's disease. Peroxynitrites more readily form in these rafts and it is peroxynitrites (and hydrogen peroxide) likely via cysteine oxidation that stimulates the beta secretase that cleaves the amyloid precursor protein into a c-terminal fragment. The subsequent formation of amyloid oligomers and plaques depends on the release of intracellular calcium.

It is peroxynitrites that lead to the c-terminal fragment and not the other way around. Otherwise, the authors of the following article get it right.

J Neurochem. 2001 Jul;78(1):109-20.
C-terminal fragment of amyloid precursor protein induces astrocytosis.
Bach JH1, Chae HS, Rah JC, Lee MW, Park CH, Choi SH, Choi JK, Lee SH, Kim YS, Kim KY, Lee WB, Suh YH, Kim SS.
Author information
One of the pathophysiological features of Alzheimer's disease is astrocytosis around senile plaques. Reactive astrocytes may produce proinflammatory mediators, nitric oxide, and subsequent reactive oxygen intermediates such as peroxynitrites. In the present study, we investigated the possible role of the C-terminal fragment of amyloid precursor protein (CT-APP), which is another constituent of amyloid senile plaque and an abnormal product of APP metabolism, as an inducer of astrocytosis. We report that 100 nM recombinant C-terminal 105 amino acid fragment (CT105) of APP induced astrocytosis morphologically and immunologically. CT105 exposure resulted in activation of mitogen-activated protein kinase (MAPK) pathways as well as transcription factor NF-kappaB. Pretreatment with PD098059 and/or SB203580 decreased nitric oxide (NO) production and nuclear factor-kappa B (NF-kappaB) activation. But inhibitors of NF-kappaB activation did not affect MAPKs activation whereas they abolished NO production and attenuated astrocytosis. Furthermore, conditioned media derived from CT105-treated astrocytes enhanced neurotoxicity and pretreatment with NO and peroxynitrite scavengers attenuated its toxicity. These suggest that CT-APP may participate in Alzheimer's pathogenesis through MAPKs- and NF-kappaB-dependent astrocytosis and iNOS induction.

Also if you knockout inducible nitric oxide synthase (inducible nitric oxide combines with superoxide anions to form peroxynitrites), you essential prevent Alzheimer's disease.

J Exp Med. 2005 Nov 7;202(9):1163-9. Epub 2005 Oct 31.
Protection from Alzheimer's-like disease in the mouse by genetic ablation of inducible nitric oxide synthase.
Nathan C1, Calingasan N, Nezezon J, Ding A, Lucia MS, La Perle K, Fuortes M, Lin M, Ehrt S, Kwon NS, Chen J, Vodovotz Y, Kipiani K, Beal MF.
Author information
Brains from subjects who have Alzheimer's disease (AD) express inducible nitric oxide synthase (iNOS). We tested the hypothesis that iNOS contributes to AD pathogenesis. Immunoreactive iNOS was detected in brains of mice with AD-like disease resulting from transgenic expression of mutant human beta-amyloid precursor protein (hAPP) and presenilin-1 (hPS1). We bred hAPP-, hPS1-double transgenic mice to be iNOS(+/+) or iNOS(-/-), and compared them with a congenic WT strain. Deficiency of iNOS substantially protected the AD-like mice from premature mortality, cerebral plaque formation, increased beta-amyloid levels, protein tyrosine nitration, astrocytosis, and microgliosis. Thus, iNOS seems to be a major instigator of beta-amyloid deposition and disease progression. Inhibition of iNOS may be a therapeutic option in AD.

And if you lessen lipid rafts you lessen the activity of the inducible nitric oxide synthase.

Diabetes. 2005 Sep;54(9):2576-85.
Essential role for membrane lipid rafts in interleukin-1beta-induced nitric oxide release from insulin-secreting cells: potential regulation by caveolin-1+.

Veluthakal R1, Chvyrkova I, Tannous M, McDonald P, Amin R, Hadden T, Thurmond DC, Quon MJ, Kowluru A.

Disruption of lipid rafts (e.g., with cyclodextrin) markedly reduced IL-1beta-induced gene expression of inducible NO synthase (iNOS) and NO release from beta-cells.

So the test might not be as good as advertised, but it is certainly pointing in the right direction.

Permalink to Comment

20. Anonymous on March 12, 2014 1:49 PM writes...


I can't resist a quick comment. I do work in this field.

The point is not about the PPV/NPV in the general elderly population - which indeed is likely to be fairly uninformative. This kind of test could however be deployed in the MCI [mild cognitive impairment] population if it holds out. That is a different proposition.

Little is known about the relationship between emerging peripheral tests and the more established cetnral measures of disease prediction/progression (hippocampal MRI, PET amyloid load, tau CSF load, FDG PET etc), not least longitudinally. But still, anything non-central that has potential prognostic value - whilst acknowledging that there are as yet no disease modifying treatments - is intriguing. And so as a peripheral and hence relatively non-invasive measure, it is worth keeping a (jaundiced) eye on.

Permalink to Comment

21. Anonymous on March 26, 2014 5:34 PM writes...

At least one other group/company has thrown their hat into this ring:

They haven't published yet, but it suggests more data are there to support the idea.

Permalink to Comment


Remember Me?


Email this entry to:

Your email address:

Message (optional):

Scripps Update
What If Drug Patents Were Written Like Software Patents?
Stem Cells: The Center of "Right to Try"
Speaking of Polyphenols. . .
Dark Biology And Small Molecules
How Polyphenols Work, Perhaps?
More On Automated Medicinal Chemistry
Scripps Merging With USC?