PCP-GC-LM: single-sequence-based protein contact prediction using dual graph convolutional neural network and...

Area of Science:

• Computational Biology
• Bioinformatics
• Structural Biology

Background:

• Deep learning and evolutionary information have advanced protein contact prediction.
• Existing methods face challenges with orphan proteins and computational burden in single-sequence prediction.
• Graph neural networks offer advantages in capturing protein structural topology and hierarchical features.

Purpose of the Study:

• To develop a novel single-sequence-based protein contact map predictor.
• To address limitations of existing methods, including prediction accuracy and computational efficiency.
• To leverage the strengths of graph neural networks for enhanced protein structure analysis.

Main Methods:

• Proposed PCP-GC-LM, a predictor utilizing dual-level graph neural networks and convolution networks.
• Employed single-sequence data for training and prediction.
• Validated performance against existing single-sequence predictors and on homodimer protein datasets.

Main Results:

• PCP-GC-LM demonstrated superior performance compared to other single-sequence-based predictors in independent tests.
• The method showed effectiveness in predicting contacts for complex protein structures, including homodimers.
• Ablation experiments confirmed the necessity and contribution of the dual graph network architecture.

Conclusions:

• The proposed framework offers new modules for accurate prediction of inter-chain contact maps in proteins.
• PCP-GC-LM provides an effective solution for single-sequence protein contact prediction, overcoming previous bottlenecks.
• The method is valuable for analyzing interactions in various protein complexes and advancing structural biology.