Leveraging Residual Graph Convolutional Networks with Cross-Attention Mechanisms for High-Accuracy Protein Function...

Area of Science:

• Computational biology
• Bioinformatics
• Genomics

Background:

• Determining protein function is crucial for understanding cellular processes and disease.
• Experimental methods for protein function annotation are costly and time-consuming.
• Large-scale functional annotation requires efficient computational approaches.

Purpose of the Study:

• Introduce RCHGO, a novel deep learning framework for inferring Gene Ontology (GO) annotations directly from protein sequences.
• Leverage residual graph convolutional networks (RGCNs) and cross-attention mechanisms for enhanced prediction accuracy.
• Provide a scalable and efficient computational tool for protein function prediction.

Main Methods:

• Developed RCHGO, a deep learning framework utilizing RGCNs with cross-attention.
• Employed two distinct deep learning modules to exploit complementary feature representations.
• Fused features at the decision level for robust functional inference.
• Benchmarked RCHGO against 16 state-of-the-art methods on 1,493 nonredundant proteins.

Main Results:

• RCHGO demonstrated superior performance compared to 16 existing state-of-the-art methods.
• The framework effectively learns rich protein- and residue-level representations.
• Integration of RGCNs and cross-attention modules improved alignment with GO semantics.
• The model's performance is attributed to the effective fusion of diverse feature representations.

Conclusions:

• RCHGO provides a highly accurate and efficient computational method for protein function prediction.
• The framework's architecture, combining RGCNs and cross-attention, is key to its success.
• RCHGO offers a valuable tool for large-scale functional genomics and drug discovery.
• The source code is publicly available, promoting further research and application.