Abstract

The growing integration of machine learning into industries like digital health systems has made interest in computational models for the prediction of drug response using high-dimensional molecular data more intense. The mathematical form, interpretability and strength of such models are, however, important issues. We provide a biomathematical analysis of machine-learning models of drug response prediction using a standardized cross-study pharmacogenomic benchmark of drug-cell line pairs with molecular-feature representations and log-normalized IC50 values in this study. A comparative framework was used to test regularized linear models (Ridge, Lasso, ElasticNet), robust regression and a few of the selected nonlinear baselines with identical data splits and preprocessing conditions. The error-based and rank-based measures of model performance were used, with the support of sparsity analysis, residual diagnostics, and bootstrap stability analysis. The findings indicate that sparsity-favoring linear models can be used to predict drug response as well as or better than more complicated nonlinear models, and find a smaller effective feature space. The residual and stability tests also show that the model has strong and unbiased predictive power and has minimal heteroscedasticity. These results indicate that mathematically constrained, interpretable frameworks are useful to predict drug responses, which justifies the usefulness of biomathematical machine-learning models as accurate elements of digital health decision-support systems.