MHGCN: A Multi-Channel Hybrid Graph Convolutional Neural Network for Cancer Drug Response Prediction

Due to the heterogeneity of cancer cells, personalized treatment plans for cancer patients remain a continuous concern. High-throughput drug screening technology has led to the development of deep learning models that generate personalized therapies. However, most existing models fail to account for the topological relationships between cell line-drug pair (CDP) nodes, thereby ignoring their intrinsic connections. This paper proposes a multi-channel hybrid graph convolutional neural network (MHGCN) for predicting cancer drug response (CDR). First, we define CDPs by integrating gene expression and drug molecular fingerprints. These CDPs are refined through denoising autoencoders to eliminate noise. Second, we compute pairwise cosine similarities among CDPs to build a similarity network, while simultaneously establishing a heterogeneous response graph connecting cell lines and drugs. Third, MHGCN processes the CDP network via graph convolutional layers and generates the response matrix through linear projection. Concurrently, a heterogeneous graph convolutional neural network learns the response heterogeneous network. Following data augmentation, we derive feature embeddings for cell lines and drugs, then compute their similarity matrix. Finally, CDR predictions are generated through weighted matrix fusion of these components. To the best of our knowledge, MHGCN represents the first framework explicitly incorporating CDP topology into CDR prediction. Experiments demonstrate MHGCN's statistically significant improvements over state-of-the-art methods.