KEGCL: Knowledge-Enhanced Graph Contrastive Learning for Protein Complex Identification.

Area of Science:

• Computational Biology
• Bioinformatics
• Systems Biology

Background:

• Protein complexes are crucial for cellular functions and understanding disease.
• Current protein-protein interaction (PPI) network methods struggle with data sparsity, false positives/negatives, and capturing complex topology.
• Existing approaches fail to fully leverage biological resources and diverse neighborhood information for accurate complex identification.

Purpose of the Study:

• To develop a novel knowledge-enhanced graph contrastive learning (KEGCL) framework for accurate protein complex identification.
• To overcome limitations of existing methods in handling sparse PPI data and preserving network topology.
• To improve the representation of diverse protein interactions within complexes.

Main Methods:

• Constructed a knowledge-enhanced PPI network by integrating external biological priors.
• Applied a spatiotemporal constraint-guided perturbation strategy to enhance semantic diversity in graph views.
• Utilized graph convolutional encoders with randomized propagation depths to capture multi-level interaction patterns.

Main Results:

• KEGCL achieved competitive performance against state-of-the-art methods on multiple real-world PPI datasets.
• Enrichment analyses validated the biological relevance of the protein complexes identified by KEGCL.
• The framework effectively captures both core and peripheral protein interactions within complexes.

Conclusions:

• KEGCL offers a robust and effective approach for protein complex identification by integrating knowledge and advanced graph learning techniques.
• The proposed method enhances the understanding of cellular functions and disease mechanisms through accurate complex identification.
• KEGCL provides a valuable tool for computational biologists and bioinformaticians, with open-source code available.