Background: Recent landmark studies have profiled cancer cell lines for molecular features, along with measuring the corresponding growth inhibitory effects for specific drug compounds. These data present a tool for determining which subsets of human cancer might be more responsive to particular drugs. To this end, the NCI-DREAM-sponsored DREAM7: Drug Sensitivity Prediction Challenge (sub-challenge 1) set out to predict the sensitivities of 18 breast cancer cell lines to 31 previously untested compounds, on the basis of molecular profiling data and a training subset of cell lines.
Methods and results: With 47 teams submitting blinded predictions, team Creighton scored third in terms of overall accuracy. Team Creighton's method was simple and straightforward, incorporated multiple expression data types (RNA-seq, gene array, RPPA), and incorporated all profiled features (not only the "best" predictive ones). As an extension of the approach, cell line data, from public datasets of expression profiling coupled with drug sensitivities (Barretina, Garnett, Heiser) were used to "predict" the drug sensitivities in human breast tumors (using data from The Cancer Genome Atlas). Drug sensitivity correlations within human breast tumors showed differences by expression-based subtype, with many associations in line with the expected (e.g. Lapatinib sensitivity in HER2-enriched cancers) and others inviting further study (e.g. relative resistance to PI3K inhibitors in basal-like cancers).
Conclusions: Molecular patterns associated with drug sensitivity are widespread, with potentially hundreds of genes that could be incorporated into making predictions, as well as offering biological clues as to the mechanisms involved. Applying the cell line patterns to human tumor data may help generate hypotheses on what tumor subsets might be more responsive to therapies, where multiple cell line datasets representing various drugs may be used, in order to assess consistency of patterns.