DAMCDR: Dual Attention Mechanism for Cross-Domain Cancer Drug Response Prediction

Patients vary in their responses to anticancer drugs due to individual differences and variations in drug properties. To optimize personalized treatment strategies, developing high-precision models to predict patient drug responses is crucial. Although clinical drug response data from cancer patients is limited, abundant cancer cell line data provide a valuable resource for such research. However, distributional differences between cell line and patient data make it difficult to directly apply models trained on cell lines to real-world scenarios. Appropriate transfer learning strategies can effectively bridge this gap by utilizing labeled cell line data to predict patient responses. Existing methods fail to consider the influence of the varying degrees of association between cell lines and patients on feature alignment process, and ignore the complex interactions between drugs and samples. In this study, we propose a deep transfer learning method based on a dual attention mechanism to address the cross-domain adaptation problem in cancer drug response prediction, termed DAMCDR. DAMCDR employs a cross-domain co-attention mechanism to extract domain-invariant features from both cell lines and patients. It leverages their mutual associations to achieve deep feature fusion and dynamic alignment. Prior to concatenating drug and sample embeddings, DAMCDR uses cross-modal coattention to integrate their interaction patterns into respective representations, thereby enriching knowledge for subsequent response prediction. To improve prediction accuracy for unlabeled patient instances, we suggest incorporating an unsupervised component into the supervised classification loss. Experiments on multiple drugs show that DAMCDR surpasses state-of-the-art methods in cross-domain cancer drug response prediction, especially in accurately identifying patients who respond to specific drugs. This provides valuable guidance for tailoring more precise clinical cancer treatment regimens.