1128-DrEval: Evaluation of Drug Response Prediction Models

This article introduces DrEval, an open-source benchmarking framework designed to address the reproducibility crisis and performance inflation in cancer drug response modeling. The authors identify critical flaws in current research, such as data leakage, biased evaluation metrics, and the failure of complex deep learning models to outperform simple tree-based baselines. By utilizing application-aware data splits and robust statistical testing, the pipeline reveals that many state-of-the-art models primarily memorize drug means rather than learning genuine biological insights. The study demonstrates that these models struggle to generalize to unseen drugs, new tissue types, or clinical patient data. Ultimately, DrEval provides a standardized, extensible environment to foster more transparent and rigorous progress in personalized medicine.

References:

• Bernett J, Iversen P, Picciani M, et al. Critical evaluation of drug response prediction models with DrEval[J]. Nature Communications, 2026, 17(1): 4238.