Biophysics-based protein language models for protein engineering.

Area of Science:

• Computational Biology
• Biophysics
• Machine Learning

Background:

• Protein language models (PLMs) excel at predicting protein sequence, structure, and function using evolutionary data.
• Existing PLMs often neglect crucial biophysical factors that govern protein behavior and function.

Purpose of the Study:

• To introduce Mutational Effect Transfer Learning (METL), a novel framework combining machine learning with biophysical modeling.
• To enhance the predictive power of protein language models by incorporating biophysical principles.

Main Methods:

• Pretraining transformer-based neural networks on biophysical simulation data to learn sequence-structure-energetics relationships.
• Fineting METL on experimental sequence-function data to predict protein properties like thermostability, activity, and fluorescence.

Main Results:

• METL demonstrates superior performance in protein engineering tasks, particularly in generalizing from small datasets and position extrapolation.
• The framework successfully designed functional green fluorescent protein variants using only 64 training examples.

Conclusions:

• METL offers a powerful biophysics-informed approach to protein language modeling.
• This framework shows significant potential for advancing protein engineering and design through the integration of machine learning and biophysical insights.