Structured penalized regression for drug sensitivity prediction

Abstract

Large-scale {\it in vitro} drug sensitivity screens are an important tool in personalized oncology to predict the effectiveness of potential cancer drugs. The prediction of the sensitivity of cancer cell lines to a panel of drugs is a multivariate regression problem with high-dimensional heterogeneous multi-omics data as input data and with potentially strong correlations between the outcome variables which represent the sensitivity to the different drugs. We propose a joint penalized regression approach with structured penalty terms which allow us to utilize the correlation structure between drugs with group-lasso-type penalties and at the same time address the heterogeneity between omics data sources by introducing data-source-specific penalty factors to penalize different data sources differently. By combining integrative penalty factors (IPF) with tree-guided group lasso, we create the IPF-tree-lasso method. We present a unified framework to transform more general IPF-type methods to the original penalized method. Because the structured penalty terms have multiple parameters, we demonstrate how the interval-search Efficient Parameter Selection via Global Optimization (EPSGO) algorithm can be used to optimize multiple penalty parameters efficiently. Simulation studies show that IPF-tree-lasso can improve the prediction performance compared to other lasso-type methods, in particular for heterogenous data sources. Finally, we employ the new methods to analyse data from the Genomics of Drug Sensitivity in Cancer project.