Advancing cancer driver gene detection via Schur complement graph augmentation and independent subspace feature...

Area of Science:

• Computational biology
• Bioinformatics
• Genomics

Background:

• Identifying cancer driver genes (CDGs) is critical for targeted cancer therapies.
• Existing deep learning models struggle with the complexity and heterogeneity of cancer gene regulatory networks.
• Accurate CDG identification requires advanced computational approaches.

Purpose of the Study:

• To develop an advanced deep learning model, SCIS-CDG, for precise prediction of cancer driver genes.
• To overcome limitations of current models in handling complex and heterogeneous cancer gene networks.
• To improve the accuracy and efficiency of identifying both known and novel CDGs.

Main Methods:

• Utilized Schur complement graph augmentation within a graph contrastive learning framework for enhanced network representation.
• Employed independent subspace feature extraction to reduce dependence and increase model expressiveness.
• Integrated a feature expansion component and a learnable attention mechanism in a graph neural network (GNN) encoder.
• Implemented rapid randomization of the Schur complement strategy for improved generalization.

Main Results:

• SCIS-CDG demonstrated high efficiency in identifying known cancer driver genes.
• The model successfully uncovered potential novel cancer driver genes in external datasets.
• Significant performance improvements were observed compared to conventional GNN models.
• The approach effectively handles complex gene networks and node feature limitations.

Conclusions:

• SCIS-CDG offers a robust and effective method for cancer driver gene identification.
• The model's architecture addresses key challenges in cancer gene network analysis.
• This work advances the computational tools available for precision oncology research.
• Publicly available code and data facilitate further research and application.