"Drugs and diseases are characterized in completely different scientific languages," said Todd Golub, director of the Broad Institute of MIT and Harvard's cancer program. And that means that because little is known about the biological states that underlie diseases, the best way to treat them often is elusive, reducing doctors to symptom management.
In a paper in the Sept. 29, 2006, issue of Science, senior author Golub, first author Justin Lamb, and their colleagues at the Broad described a project they hope will be able to describe drugs and diseases in a common language, and so help "establish the relation among diseases, physiological processes, and the action of small-molecule therapeutics," as the authors put it. And two papers published online in Cancer Cell on Sept. 28 show the first practical results of the effort, by connecting drugs to cancers in novel ways.
Gene expression profiling is not new, of course, but Golub and his colleagues wanted to broaden the techniques' horizons, studying the cellular response to a wide variety of what they termed 'perturbagens.'
Given the immense number of cell types, perturbagens, and possible cellular responses that could have been studied, the work is reminiscent of Otto von Bismarck's description of politics as "the art of the possible."
In this case, what was possible was to take four cell lines and treat them with 164 small molecules; The scientists picked the compounds from cell biological categories, such as histone deacetylase inhibitors and estrogen receptor modulators, as well as clinical ones like antidiabetics and antipsychotics.
In each case, the researchers compared the gene expression signature of placebo-treated cells to those treated with a given molecule, and rank-ordered 22,000 genes according to how much their expression levels changed after the treatments.
In some ways, analyzing the data proved trickier than generating it. The researchers developed their own pattern-matching strategy to analyze their signatures; Lamb said that the usual approach, known as hierarchical clustering, would not work because it uses absolute values of gene expression levels, and the team was studying relative values to prevent the drug effects from being completely overpowered by background noise.
The strategy they devised, which they termed Gene Set Enrichment Analysis or GSEA, had the added benefit of not being tied to any technology platform; Lamb said that researchers could compare their own data to the map, which the Broad Institute has made publicly available, regardless of the type of microarrays used.
Even with a custom-made analysis, statistical significance proved surprisingly difficult to pin down. Lamb said that "we consulted with easily a dozen statisticians, and each of them had a completely different idea," including ideas that were mutually exclusive, about how to generate "p" values from the data. The problem, Lamb said, is that "it's not clear that the multiple samples that we are looking at are truly independent in the statistical sense."
The team ended up settling on a so-called permutation "p" value for their analyses, but also cautioned that "there is no standard approach for estimating the statistical significance of the connections observed."
The team reported a variety of findings they generated by using the map, but when asked about his favorite applications of the map to date, Lamb pointed not to anything in the Science paper, but to the two papers published online in Cancer Cell.
In one paper, the researchers used the map to understand the molecular actions of gedunin, which were basically unknown despite a long history of medicinal use of the compound. The connectivity map suggested that gedunin as well as another compound, celastrol, inhibit androgen receptor mediated signaling. Androgen receptor signaling plays a critical role in the progression of prostate cancer, but there are very few clinically effective inhibitors.
Lamb said the paper illustrated a strength of the connectivity map: the ability to "develop a hypothesis about the mechanism of action of an unknown small molecule" - in the case of gedunin, the connectivity map was able to do that in "literally 10 minutes."
The second Cancer Cell paper identified the immunosuppressant sirolimus (rapamycin) as a therapeutic candidate for overcoming drug resistance to glucocorticoids in acute lymphoblastic leukemia (ALL). The researchers compared samples of drug-sensitive and drug-resistant ALL cells to the connectivity map, and found that the expression profile induced by rapamycin matched that of glucocorticoid-sensitive ALL cells, suggesting that rapamycin might be able to restore such sensitivity to ALL patients. Lamb termed the finding a "double win: Not only have you come up with a hypothesis, but the drug is already FDA-approved," making clinical trials in a new indication that much easier.
Though the connectivity map's uses are not restricted to drug discovery, Lamb said that the team's next priority is to add all FDA-approved drugs to the database.
"In the short term, that looks like a sweet spot to us," he said.