Coherent Blending of Biophysics-Based Knowledge with Bayesian Neural Networks for Robust Protein Property Prediction.

Area of Science:

• Computational Biology
• Biophysics
• Machine Learning

Background:

• Machine learning (ML) models for protein property prediction often struggle with data distribution shifts.
• Biophysics-based models offer uniform accuracy but may be less precise near training data.

Purpose of the Study:

• To develop a scalable method for integrating biophysical knowledge into neural networks.
• To improve the accuracy and generalizability of protein property prediction models.

Main Methods:

• A Bayesian formulation was used to incorporate biophysical knowledge into neural networks (BNNs).
• A novel probabilistic method was devised to bridge the gap between BNN weight priors and biophysical function-value priors.
• The approach leverages BNN epistemic uncertainty to dynamically balance reliance on biophysical priors versus neural network predictions.

Main Results:

• Predictions adaptively favor biophysical information when BNN uncertainty is high and neural network information when uncertainty is low.
• The method demonstrates intuitive and effective integration of diverse data sources.
• Successful application on synthetic data, protein fluorescence and binding prediction, and small molecule property prediction.

Conclusions:

• The developed method offers a practical and scalable solution for enhancing ML models with biophysical insights.
• This hybrid approach improves prediction accuracy across different regions of the protein space.
• The findings have implications for protein engineering and understanding biological systems.