From PhysOrg:
The scientists drew their data from the NIH National Center for Biotechnology Information Gene Expression Omnibus, a publicly available database that contains the results of thousands of genomic studies on a wide range of topics, submitted by researchers across the globe. The resource catalogs changes in gene activity under various conditions, such as in diseased tissues or in response to medications. Butte's group focused on 100 diseases and 164 drugs. They created a computer program to search through the thousands of possible drug-disease combinations to find drugs and diseases whose gene expression patterns essentially cancelled each other out. For example, if a disease increased the activity of certain genes, the program tried to match it with one or more drugs that decreased the activity of those genes. Many of the drug-disease matches were known, and are already in clinical use, supporting the validity of the approach. For example, the analysis correctly predicted that prednisolone could treat Crohn's disease, a condition for which it is a standard therapy.
More here.