Improved protein structure prediction by deep learning irrespective of co-evolution information

Area of Science:

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

Background:

• Predicting protein tertiary structure from primary sequence is a fundamental challenge.
• Deep learning and co-evolutionary analysis have significantly advanced structure prediction accuracy.
• Assessing the impact of network architecture and data integration is crucial for method development.

Purpose of the Study:

• To evaluate the performance of a large ResNet (convolutional residual neural networks) for protein structure prediction.
• To analyze the contribution of co-evolutionary data versus primary sequence information.
• To investigate the efficacy of the model on both naturally occurring and designed proteins.

Main Methods:

• Utilized a large ResNet architecture for protein structure prediction.
• Integrated and analyzed co-evolutionary data alongside primary sequence information.
• Tested model performance on CASP13 free-modeling targets and designed protein sequences.

Main Results:

• ResNet predicted correct folds for 26/32 CASP13 free-modeling targets with >80% precision for long-range contacts.
• Even without co-evolution, ResNet predicted correct folds for 18 CASP13 targets, outperforming previous methods.
• ResNet accurately predicted structures for all tested human-designed proteins using only primary sequence.

Conclusions:

• ResNet learns fundamental protein sequence-structure relationships, not just denoising co-evolution signals.
• The model demonstrates strong performance on designed proteins, potentially outperforming co-evolution-trained models.
• These findings have significant implications for protein design and engineering, particularly when co-evolutionary data is limited.