Feature selection translates drug response predictors from cell lines to patients

Targeted therapies and chemotherapies are prevalent in cancer treatment. Identification of predictive markers to stratify patients who will respond these remains challenging because patient drug response data limited. As large amounts have been generated by cell lines, methods efficiently translate cell-line-trained predictors human tumors be useful clinical practice. Here, we propose versatile feature selection procedures that can combined with any classifier. For demonstration, the a (linear) logit model (non-linear) K-nearest neighbor trained on lines result LogitDA KNNDA, respectively. We show LogitDA/KNNDA significantly outperforms existing methods, e.g., logistic deep learning method thousands genes, prediction AUC (0.70–1.00 for seven ten drugs tested) is interpretable. This may due fact sample sizes often limited area prediction. further derive novel adjustment cutoff yield accuracy 0.70–0.93 drugs, including erlotinib cetuximab, whose pathways relevant anti-cancer also uncovered. These results indicate our into tumors.