Overcoming extrapolation challenges of deep learning by incorporating physics in protein sequence-function modeling.

Area of Science:

• Computational Biology
• Biophysics
• Machine Learning

Background:

• Understanding protein sequence-to-function relationships is vital for genetics, evolution, and engineering.
• Deep learning models excel at sequence-to-function mapping but struggle with extrapolation to unseen data.
• Complex correlations and dynamics in protein sequences pose challenges for current predictive models.

Purpose of the Study:

• To enhance the extrapolation capabilities of deep learning models for protein function prediction.
• To investigate the integration of biophysics-based modeling with neural networks.
• To improve predictions for protein variants with unseen mutations or positions.

Main Methods:

• Developed biophysics-based models to quantify mutation energetic effects.
• Integrated physical energetics into convolutional neural networks (CNNs) and graph CNNs.
• Evaluated model performance on predicting functional effects of unseen variants.

Main Results:

• Models incorporating biophysical features significantly improved positional and mutational extrapolation.
• Physics-informed neural networks outperformed models lacking biophysical insights.
• Demonstrated enhanced prediction accuracy for variants outside the training data distribution.

Conclusions:

• Integrating physical knowledge, specifically protein energetics, effectively overcomes data scarcity limitations in deep learning models.
• Biophysics-guided machine learning offers a robust approach for accurate protein function prediction and variant effect analysis.
• This strategy enhances the reliability of models in biological and engineering applications.