Impact of between-tissue differences on pan-cancer predictions of drug sensitivity

Abstract

Pharmacogenomic databases of drug response and genomics data in tumor cell lines allow the development of pan-cancer (i.e. tissue-agnostic) predictions of drug response. However, it is unclear whether between-tissue differences in both drug response and molecular characteristics bias pan-cancer drug response predictions. Using two datasets containing 346 and 504 cell lines with MEK inhibitor (MEKi) response data and RNA, SNP, and CNV data, we show between-tissue differences produced confounding effects that increase apparent performance in pan-cancer MEKi response predictions. We estimate that 45-58% of the variance in pan-cancer prediction scores was driven by accurate betweentissue predictions, rather than inter-individual predictions within a tissue. Nevertheless, MEKi response is predicted as well or better by pan-cancer approaches compared to tissue-specific approaches. Our results demonstrate strong between-tissue effects in tissue-agnostic drug response prediction models, but also highlight how drug response may invoke shared regulatory mechanisms although different cancer types arise in tissue-specific routes.