Prediction of Cancer Cell Sensitivity to Drugs∗

We consider the problem of using high-dimensional genomic covariates to predict the drug response of cancer cell lines. The cell lines are grouped according to tissue type, which we show may be an important factor in determining the dependence of drug response on genomic covariates. We develop predictors using l1-penalized linear regression models, and develop a novel variant which we call the indicator lasso which exploits inherent group structure. The superior performance of this new method in simulations over other methods which neglect group structure is illustrated. We finish by presenting the drug response prediction results.