MTGCL: Multi-Task Graph Contrastive Learning for Identifying Cancer Driver Genes From Multi-Omics Data.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Identifying cancer driver genes is essential for understanding cancer mechanisms.
• Existing graph convolutional network methods face limitations like biased predictions and sparse data.
• There's a need for improved methods integrating network topology and biological features for accurate driver gene identification.

Purpose of the Study:

• To propose a novel method, Multi-Task Graph Contrastive Learning (MTGCL), for enhanced cancer driver gene identification.
• To address limitations of existing graph convolutional network approaches by integrating structural and feature information.
• To leverage semi-supervised learning to improve driver gene identification using both labeled and unlabeled data.

Main Methods:

• Developed a new graph convolutional layer structure to integrate graph topology and node features.
• Implemented a semi-supervised graph contrastive learning task as a regularizer within a multi-task learning framework.
• Utilized somatic mutation data, including features from different mutation types, to enhance identification.

Main Results:

• MTGCL demonstrated effectiveness in identifying cancer driver genes across pan-cancer and specific cancer types.
• The integrated approach significantly improved prediction performance compared to existing methods.
• Features from different mutation types were found to be particularly beneficial for specific cancer types.

Conclusions:

• MTGCL offers a robust and effective approach for cancer driver gene identification.
• Integrating network structure, biological features, and semi-supervised learning enhances predictive accuracy.
• Somatic mutation type features provide valuable insights for targeted cancer driver gene discovery.