InFo-DRP: Integrating Invariant Learning and Foundation Representations for Unseen Drug Response Prediction.

Abstract

Accurate prediction of cancer cell responses to chemical compounds remains a fundamental challenge in precision oncology and drug discovery. Although several drug response prediction (DRP) models have been proposed, most were evaluated using random splits of drug-cell line pairs, allowing the same drugs to appear in both training and test sets. In practice, models must predict responses to previously unseen compounds, necessitating robustness under drug- blind and scaffold- blind splits. However, many existing models show substantial performance degradation under these conditions, indicating limited robustness to chemical distribution shift. To address this challenge, we propose InFo-DRP that integrates invariant learning with foundation-model drug representations. InFo-DRP incorporates multi-branch drug encoders (MolBERT, Chemformer, and ECFP4) with physicochemical descriptors, separates drug and cell representations into invariant and spurious components, and applies dual-path fusion for prediction. Across two Genomics of Drug Sensitivity in Cancer (GDSC) benchmarks, InFo-DRP outperforms strong baselines under drug- blind and scaffold- blind evaluations. On five non-cancerous cell lines, InFo-DRP maintains predictive performance under scaffold shift, suggesting potential utility beyond cancer-specific settings. In external validation on the Cancer Cell Line Encyclopedia dataset, InFo-DRP achieves the highest correlation among the compared methods without any adaptation. Overall, InFo-DRP provides a practical approach for generalizing DRP to previously unseen chemical space.