Transferring Preclinical Drug Response to Patient via Tumor Heterogeneity-Aware Alignment and Perturbation Modeling

Abstract

Accurate prediction of personalized drug response is critical for precision oncology, yet limited clinical data forces reliance on preclinical datasets. However, fundamental biological differences between preclinical cell lines and patient tumors hinder direct knowledge transfer. In this work, we introduce THERAPI, a novel tumor heterogeneity-aware Domain Adaptation (DA) framework that represents patient tumors as weighted combinations of multiple cell lines with tissue-specific context. Along with our comprehensive gene expression modeling by integrating drug-induced perturbation-based and rank-based representations, THERAPI outperforms both DA-free and DA-based models and generalizes robustly to an external dataset, highlighting its potential for applications in precision medicine.