Transferring Cancer Drug Response Prediction Through Multi-Label Heuristic Domain Adaptation.

Abstract

Accurate prediction of cancer drug response (CDR) is critical for oncology drug screening. Given the scarcity of labeled patient data, domain adaptation has emerged as a promising method to transfer supervision signals from the label-rich cell line data (source domain) into the patient prediction (target domain). However, existing efforts usually treat each drug's response prediction as an independent task, requiring model retraining when drugs change, and struggle to capture reliable domain-invariant information between cell lines and patients through regular alignment strategies. Herein, we propose AdaptCDR, a multi-label heuristic domain adaptation model for transferring CDR prediction from cell lines to patients. Specifically, we pre-train an autoencoder to extract shared features across two domains, and a multi-label association classifier to map these features to response outcomes for various drugs simultaneously in the source domain. Then, we fine-tune the autoencoder to capture domain-invariant representations by distinguishing domain-specific counterparts as heuristics, thereby bridging the domain gap and enabling robust transfer of multi-label classifier to the target domain. Extensive experiments conducted on the TCGA and PDX datasets demonstrate that AdaptCDR boosts transfer accuracy and efficiency compared to previous methods. Subsequent case studies involving unmeasured patient samples further highlight its potential for reliable screening of clinical anti-cancer drugs.