Identifying biomarkers of anti-cancer drug synergy using multi-task learning

Combining anti-cancer drugs has the potential to increase treatment efficacy. Because patient responses to drug combinations are highly variable, predictive biomarkers of synergy are required to identify which patients are likely to benefit from a drug combination. To aid biomarker identification, the DREAM challenge consortium has recently released data from a screen containing 85 cell lines and 167 drug combinations. The main challenge of these data is the low sample size: per drug combination, a median of 14 cell lines have been screened. We found that widely used methods in single drug response prediction, such as Elastic Net regression per drug, are not predictive in this setting. Instead, we propose to use multi-task learning: training a single model simultaneously on all drug combinations, which we show results in increased predictive performance. In contrast to other multi-task learning approaches, our approach allows for the identification of biomarkers, by using a modified random forest variable importance score, which we illustrate using artificial data and the DREAM challenge data. Notably, we find that mutations in MYO15A are associated with synergy between ALK / IGFR dual inhibitors and PI3K pathway inhibitors in triple-negative breast cancer.Author summaryCombining drugs is a promising strategy for cancer treatment. However, it is often not known which patients will benefit from a particular drug combination. To identify patients that are likely to benefit, we need to identify biomarkers, such as mutations in the tumor’s DNA, that are associated with favorable response to the drug combination. In this work, we identified such biomarkers using the drug combination data released by the DREAM challenge consortium, which contain 85 tumor cell lines and 167 drug combinations. The main challenge of these data is the extremely low sample size: a median of 14 cell lines have been screened per drug combination. We found that traditional methods to identify biomarkers for monotherapy response, which analyze each drug separately, are not suitable in this low sample size setting. Instead, we used a technique called multi-task learning to jointly analyze all drug combinations in a single statistical model. In contrast to existing multi-task learning algorithms, which are black-box methods, our method allows for the identification of biomarkers. Notably, we find that, in a subset of breast cancer cell lines, MYO15A mutations associate with response to the combination of ALK / IGFR dual inhibitors and PI3K pathway inhibitors.