AACDR: Integrating Graph Isomorphism Networks and Asymmetric Adversarial Domain Adaptation for Cancer Drug Response Prediction
Yi Zhang, Tingfang Wu, Liangpeng Nie, Yelu Jiang, Geng Li, Jia Xu, Zhenglong Zhou, Yiwei Chen, Lijun Quan, Qiang Lyu · Journal of Chemical Information and Modeling · 2025
Predicting cancer drug response is critical for the development of personalized treatments, as it helps identify optimal therapeutic options for cancer patients. Due to the limited scale of clinical trial data, previous research has focused on preclinical data, often overlooking the distribution discrepancies between these data. To address this issue, a cancer drug response prediction method with asymmetric adversarial domain adaptation (AACDR) is introduced. This approach addresses the limitations of conventional adversarial domain adaptation methods by pushing distribution of target domain toward that of source, achieving greater accuracy and stability in predictions of target task. Drug feature extraction is enhanced by using graph isomorphism networks, enabling a more comprehensive data representation. Experiments on cancer patient data sets demonstrate effective knowledge transfer from cell lines to patients, outperforming existing methods. Further validation on patient-derived xenograft (PDX) data set highlights the generalizability of AACDR across various distributional discrepancies. Additionally, testing on copy number variation (CNV) data set demonstrates its adaptability to different cancer representation methods. The model not only accurately predicts therapeutic results for real clinical records, but also recommends potential therapeutic options for specific cancer patients, supported by relevant studies, underscoring its practical importance in delivering personalized treatment strategies.