MHGCL: Multimodal heterogeneous graph contrastive learning for cancer patient classification

Given the high heterogeneity of cancer, accurate patient subgroup classification based on biological characteristics and disease progression is critical for advancing precision medicine. Recent advances indicate that integrating multimodal data (e.g., genomic and clinical data) and employing graph representation learning can significantly improve classification accuracy. However, existing methods suffer from two limitations. First, they fail to dynamically select features highly relevant to downstream tasks, leading to suboptimal performance. Second, they overlook intra- and inter-modal structural information when learning cross-modal representations. To address these issues, we propose a Multimodal Heterogeneous Graph Contrastive Learning (MHGCL) framework. Specifically, MHGCL employs a dynamic graph learning approach that adaptively constructs modality-specific patient similarity networks (PSNs) through dynamic feature selection. This enables end-to-end learning of information-rich representations via iterative PSN refinement. To effectively align multimodal representations, a multimodal heterogeneous graph contrastive learning method is proposed. Furthermore, we design a dual-view attention mechanism to integrate multimodal representations by capturing complex interactions from both global and local perspectives. Experimental evaluations on four public datasets demonstrate that MHGCL significantly outperforms twelve state-of-the-art methods, exhibiting strong generalizability for both binary and multi-class classification tasks. Notably, MHGCL surpasses the best-performing baseline by 4.4% in AUC and 4.9% in F1 score on the binary classification dataset LUSC, and achieves improvements of 1.8% in accuracy and 3.8% in macro-averaged F1 score on the multi-class classification dataset BRCA. The source code of MHGCL will be released at https://github.com/wx99538/MHGCL upon publication.