An ensemble framework for robust drug response prediction and large-scale sensitivity profiling across cancers

Predicting how cancers respond to treatment remains difficult because tumors are diverse and computational models often perform poorly on unfamiliar drugs or samples. Here we show that the accuracy of deep learning models varies across molecular inputs, drug descriptions, model designs, and evaluation settings, with the largest losses when models are tested on previously unseen drugs or distinct datasets. We therefore develop Drug Response Integration and Voting Ensemble, a meta-learning framework that combines models using predictions from held-out data. A nine-model ensemble improves predictive accuracy, drug ranking, and stability over individual models and ensemble approaches. We implement the framework as a reproducible Nextflow workflow and use it to build a resource of measured and predicted drug responses across cancer cell lines, patient-derived samples, and compound libraries. In colorectal cancer, the framework identifies treatments with confirmed antitumor activity in cell and animal models. This approach provides a scalable strategy for therapeutic discovery. We benchmark drug response models and develop DRIVE, a meta-learning ensemble for large-scale cancer drug sensitivity prediction. DRIVE improves robustness and prioritizes compounds with validated antitumor activity in colorectal cancer.