Interpreting forces as deep learning gradients improves quality of predicted protein structures

Area of Science:

• Computational biology
• Structural biology
• Machine learning

Background:

• Deep learning models like AlphaFold2 achieve high accuracy in protein structure prediction.
• Current predictions may lack the physical realism needed for tasks like molecular docking.
• Integrating physical intuition can enhance the utility of predicted protein structures.

Purpose of the Study:

• To develop a novel method for training deep learning protein structure prediction models.
• To improve the accuracy and physical realism of predicted protein structures.
• To enable direct use of predicted structures in downstream applications.

Main Methods:

• Proposed a custom PyTorch loss function, OpenMM-Loss, representing potential energy.
• Integrated OpenMM-Loss with the SidechainNet software package for all-atom protein structures.
• Applied the method to finetune the OpenFold model.

Main Results:

• Finetuned OpenFold produced protein structures with comparable accuracy to original predictions.
• Predicted structures exhibited lower potential energy, indicating improved physical realism.
• MolProbity metrics showed enhanced structural quality in the finetuned predictions.

Conclusions:

• Training deep learning models with molecular dynamics force fields is effective.
• The OpenMM-Loss function improves the physical quality of protein structure predictions.
• This approach enhances the utility of deep learning models for structural biology applications.