Transferring Cell-level Drug Response to Patient via Tumor Heterogeneity-Aware Alignment and Gene-level Foundational Models

ICML 2025 Workshop FM4LS Submission33 Authors

Keywords: Precision medicine, Drug response prediction, Transfer learning, Domain Adaptation, Foundational model

TL;DR: We propose THERAPI, a deep learning framework that models tumor heterogeneity by aligning transcriptomes and leverages gene representations derived from foundation models to predict patient drug response.

Abstract: Prediction of patient-level drug response is critical for precision oncology but remains limited by the scarcity of labeled clinical data. While machine learning models trained on cancer cell lines offer a scalable alternative, biological differences introduce domain shifts that hinder direct translation to patient tumors. We present THERAPI (Tumor Heterogeneity-aware Embedding for Response Adaptation and Patient Inference), a deep learning framework designed to bridge this gap. THERAPI aligns patient and cell line transcriptomes via an attention-based aggregation of cell lines guided by tissue context, enabling tumor heterogeneity-aware modeling. For drug response prediction, it further transfers gene-level knowledge from foundation models based on drug-induced perturbations and rank-based representations. Our approach outperforms nine baselines in predicting patient drug responses and generalizes to external cohorts, while providing interpretable insights into tumor heterogeneity and clinical outcomes. These results highlight the promise of biological context-aware domain adaptation and gene-level knowledge integration for robust, interpretable drug response prediction in precision medicine.

Submission Number: 33