Analysis of distance-based protein structure prediction by deep learning in CASP13

Area of Science:

• Computational biology
• Structural bioinformatics
• Deep learning in protein science

Background:

• Protein structure prediction is crucial for understanding biological function.
• Previous methods relied on template-based modeling and coevolutionary analysis.
• Deep learning, specifically residual neural networks (ResNet), showed promise in CASP12 for contact prediction.

Purpose of the Study:

• To report the performance of RaptorX servers in the CASP13 competition.
• To evaluate the effectiveness of deep convolutional ResNet for protein contact and structure prediction.
• To confirm the utility of predicted distances and integrated deep learning approaches.

Main Methods:

• Implementation of distance-based contact prediction, threading, and folding methods in RaptorX servers.
• Utilizing deep convolutional ResNet architecture for contact prediction.
• Applying methods to 32 CASP13 free-modeling (FM) targets with varying multiple sequence alignment (MSA) depths.

Main Results:

• RaptorX achieved the best contact prediction accuracy among 46 participating groups.
• RaptorX demonstrated near-best 3D structure modeling performance without extensive sampling.
• High precision in predicting long-range contacts (70% at L/5) and correct fold prediction for 18/32 FM targets.
• Successful prediction of correct folds for all FM targets with >300 residues and generation of top models for specific targets.

Conclusions:

• Deep learning, particularly ResNet, significantly enhances protein contact and structure prediction.
• Predicted distances are more informative than contacts for both template-based and free modeling.
• Integrating template and coevolutionary information via deep learning offers a powerful strategy for improving structure modeling.