Sunday, August 28, 2011

Wishing I Had Been A Referee: A Renaissance for Tagamet?

Back from vacation & watching another wave of Hurricane Irene soak the area (curiously, the windiest times so far seem to be breaks in the rain).  August has not seen much attention paid to this space (indeed, I have one piece that has gestated nearly the whole month), so time to put the shoulder to the wheel.
Just before my vacation, a pair of papers showed up in Science Translational Medicine which describe two attempts at drug re-positioning by transcriptional profiling.  The key concept is to take expression profiles for diseases and try to find drugs which appear to generate the opposite transcriptional pattern, with the theory that the drug could nudge the disease pattern back to a normal state.  This is an idea which has been kicking around for a while, and at one time was the focus of a number of companies.  One was even trying to re-position a drug I have a small connection to (MLNM developed it from a target I spotted in an EST library), but I believe that is a dead effort.  One challenge in tracking this field is that it is rarely obvious what happened in the end; did a drug fail to pan out or did the backers just run out of cash?
Both papers are from the same group.  I'll focus on the first paper (Sirota et al), since it has the main computational focus and is the source of some complaint on my part; Dudley et al primarily focuses on results.  It must be acknowledged up front one key aspect of these papers: both have provided interesting new hypotheses, with initial animal data, of re-uses of well known existing drugs for important medical conditions.  Sirota et al proposes using the gastric acid prevention cimetidine (Tagamet, which made a fortune for Smith Kline lone ago) for lung adenocarcinoma and Dudley et al the anticonvulsant topiramate for inflammatory bowel disease.

My first complaint with Sirota et al is a production issue: a key supplementary file promised in the text is missing from the website.  This is the table listing all of the public expression datasets used to build their search patterns for diseases.  Whether fault here lies with the editors or the authors isn't clear, but I must say that such files are often missing from the packages sent out for review.  This is a critical file for anyone attempting to re-create their results.  Of course, this is something that should be correctable in short order.

But there are meatier issues that the reviewers should have flagged, and I certainly would have.  Sirota et al tout that a hierarchical clustering of drugs puts some drugs of similar mechanism-of-action (MoA) in the same neighborhood of the tree.  I perform such clustering regularly (and perhaps excesssively), but it is a problematic way to look at such associations.  In any case, Sirota et al suffers from looking at only a few cherry-picked successes and not examining a broader range of curiosities.  For example, they note that geldanamycin clusters with several other drugs known to be heat shock protein 90 (HSP90) inhibitors.  However, they don't note that two geldanamycin derivatives are also in their dataset and do not tree with the parent compound or each other (disclosure: my employer is developing a geldanamycin compound for cancer).  Inspection of the plot reveals other cases of structurally similar drugs not co-clustering (though this search is hampered both by my limited knowledge of the field and a number of drugs being shown only as code numbers), such as two different glitazones, two Cox-2 inhibitors, and metformin/phenformin are candidates I spotted.  

A similar tree is provided of diseases.    I can make claims about drugs being structurally similar and hitting the same target, but critiquing the clustering of diseases would require a bit of hubris.  Still, there are curiosities -- breast cancer being closer to "complex dental caries" than any other cancer (despite the dataset being largely cancers) and "fracture of femur" being quite distant from "fracture".  

The big payoff is a color-coded correlation matrix of diseases and genes.  As a reviewer I would have groused about staring at that diagram,.  I already look like I shop for eye wear at an astronomy supply store, and squinting at the labels on that diagram didn't help any (nor the fact that Acrobat refused to zoom on them, squawking an error instead).  Ideally the online materials would have a powerful data viewer included, but I certainly would have asked for the table of the results to be provided. The diagram also leads to another round of indiscriminate cherry-picking.  The authors note cases in which diseases seem to partner up with their treatments.  What is really missing though, and a huge missed opportunity, would be systematic analysis of how often these work out.  Given the limited number of drugs, it really would not have been impractical to identify all the cases in which one of the drugs in the study was approved for one of the diseases in the study.  Some failures of recall would be expected; the expression profiles for the drugs are all from a single cancer cell line  in culture, and so drugs acting by modulating pathways not active in that line or by modulating more complex environments would not be expected to appear.  Still, such an analysis would be useful for calibrating this approach and should have been included.

In their closing comments, the authors discuss the scientific issues with pushing these drugs to potential new indications.  Animal models too often fail to predict clinical experience, and even though both drugs have seen widespread use and therefore have extensive safety data, the safety in particular patient populations could be different.  One tack not discussed is to see whether large patient registries, both well-controlled prospective ones such as the Framingham study or the Nurses Health Study but also more eclectic ones such as PatientsLikeMe might have patients exhibiting these disease-drug pairs.  But perhaps an equally great challenge is a financial one.  It is highly unlikely that any company would be interested in these, as both drugs are long off patent.  Attempts to build companies around clever reuses of off-patent drugs have generally failed, such as NitroMed and CombinatoRx.  The very striking exception is Celgene, which re-purposed thalidomide from a legendary pharmaceutical disaster to a useful treatment for leprosy and cancer -- but only because they could patent their system for avoiding another round of disasters and thereby control the market for an unpatentable drug.  That clearly isn't available here; cimetidine is even available over-the-counter.  So if either of these drugs are to be pushed through the clinical testing obstacle course, it will need to be entirely driven by public money.  Money well spent, but someone will have to decide to do so.


Anonymous said...

Kim Stegmaier developed a similar strategy at the Broad, called gene expression based high-throughput screening (GE-HTS). I'm sure you're probably well aware of this technology, but just in case, they have also come up with some interesting findings over the years.

Keith Robison said...

Yes, thank your for pointing that out -- I meant to explicitly cite Dr. Stegmaier's work with Todd Golub which pioneered this approach -- I believe they have initiated clinical trials based on it, but I don't know if any of those panned out.

Anonymous said...

Actually both papers are essentially a minor tweak of Connectivity Map. Very surprising it got into STM.

Anonymous said...

I can't help but wonder who are these reviewers and why they didn't ask some of these same questions. Seems to be a rampant problem in this field.

atulbutte said...

Hi, thanks for noticing the papers. Yes, we cited the Connectivity Map in the papers, of course. Of course the Connectivity Map is a great resource, generously contributed to the public by a well-funded lab. I think one key that may not have been clear is that we could find so many disease profiles using free, publicly-available resources.

To me, it is incredible that, due to increasing requirements from journals and funding agencies, that molecular measurements are available in the public (i.e. NCBI GEO, EBI ArrayExpress, etc).

Several investigators have used Connectivity Map to make similar predictions, and we cite several of these, but to my knowledge, very few have actually tried their predictions in pre-clinical animal models.

You are right that we didn't provide a list of the exact experiments we downloaded from NCBI GEO, and that would have been useful. We did provide a smaller list in our previous publication, including the specific disorders used in this paper. See here for details and the list:

Best wishes -- Atul Butte