MHGCL: Multimodal heterogeneous graph contrastive learning for cancer patient classification

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in medicine

Background:

• Cancer's heterogeneity necessitates accurate patient subgroup classification for precision medicine.
• Integrating multimodal data (genomic, clinical) with graph representation learning shows promise but faces limitations.

Purpose of the Study:

• To address limitations in dynamic feature selection and multimodal representation learning for cancer patient classification.
• To propose a novel Multimodal Heterogeneous Graph Contrastive Learning (MHGCL) framework.

Main Methods:

• MHGCL utilizes dynamic graph learning to construct patient similarity networks (PSNs) via adaptive feature selection.
• A multimodal heterogeneous graph contrastive learning approach aligns representations.
• A dual-view attention mechanism integrates multimodal information capturing global and local interactions.

Main Results:

• MHGCL significantly outperformed twelve state-of-the-art methods across four public datasets.
• Demonstrated strong generalizability for both binary and multi-class classification tasks.
• Achieved notable improvements in AUC, F1 score, and accuracy on LUSC and BRCA datasets.

Conclusions:

• The proposed MHGCL framework offers a robust approach for cancer patient subgroup classification.
• MHGCL effectively leverages multimodal data and graph learning for improved precision medicine.
• The framework shows significant potential for advancing cancer research and clinical applications.