The Broad Institute‘s Connectivity Map is a systematic effort to characterise cellular responses to chemical challenge, particularly by small molecules (small that is, in relation to things like proteins and chromosomes). The initial report, in this week’s Science, details the pilot phase of the project: essentially, a database of gene expression profiles taken from cells challenged with 164 substances, mostly drawn from libraries of therapeutics and candidate drugs.
The way it works is deceptively simple: each treatment of cells is coupled to a control experiment, where no treatment is administered. Genes are then ranked by the magnitude of differential expression, relative to control, to create signatures of the treatment’s effect. You can then use these signatures to ask a host of questions:
- Do drugs of the same class have the same signature? Comparing signatures will tell you a lot about similarities and differences in their mechanisms of action.
- I’ve done an experiment – does my treatment signature look like anything in the database? That will tell me a lot about this new compound that I’m working on.
- I’ve done another experiment – does the signature I get for disease X look like anything in the database? That tells me a lot about the mechanism of the disease, and maybe even suggests a few drug targets.
- I’m developing a new drug for disease X, which shows some promise. I’m not quite sure how it works: does it target the same pathway(s) as previous drugs, or is it new? Does that tell me something new about the disease’s biology?
The authors use previously reported signatures of compounds similar to some in the database to stress-test the process. They find they can correctly match up cancer drugs, hormones, and anti-psychotics to cognates in the database, despite the original signatures having been derived in different cell types under different experimental conditions, and that they can assign a mechanism of action to a previously uncharacterised compound which interferes with prostate cancer cell proliferation. They also show that a signature derived from fat cells of rats made obese through diet correlates well with several antidiabetic drugs known to promote weight gain in humans. Most intriguingly, they show that a drug whose signature is inversely correlated with drug resistance in leukaemia patients can counter that resistance, making cancerous cells once again susceptible to treatment. This is a relevant problem in cancer therapy where cancers frequently become resistant to treatments over time.
From a methodological pov, one of the compelling things about this work is that very simple statistics are used to do the matching: signatures appear to be robust enough to pick out easily, rather than requiring the involved mathematical sifting that has become the staple of gene expression studies. I suspect this is an attribute of expression data, where data aggregates such as signatures are meta-stable: even if individual gene measurements are inaccurate, the pattern stands out clearly. If you’re going to spend the next couple of years of your life following up on some of this, that’s very comforting…