Knowledge Distillation of a Protein Language Model Yields a Foundational Implicit Solvent Model.

Area of Science:

• Computational chemistry
• Biophysics
• Structural biology

Background:

• Implicit solvent models (ISMs) aim for explicit solvent accuracy at lower computational cost.
• Traditional ISMs struggle with accuracy for protein folding and intrinsically disordered proteins.
• Developing transferable, data-driven ISMs is a key challenge.

Purpose of the Study:

• To develop a novel, data-driven implicit solvent model.
• To overcome limitations of traditional ISMs in protein simulations.
• To create a unified model for both folded and disordered proteins.

Main Methods:

• Distilled evolutionary information from protein language model (ESM3) into a graph neural network (GNN).
• Trained GNN potential on effective energies from ESM3.
• Combined GNN potential with a standard electrostatics term for molecular dynamics simulations.

Main Results:

• The GNN potential drives stable, long time-scale molecular dynamics simulations.
• The hybrid model accurately reproduces protein folding free energy landscapes.
• The model predicts structural ensembles of intrinsically disordered proteins.
• Achieved a single, transferable model for folded and disordered protein states.

Conclusions:

• Successfully distilled evolutionary knowledge into a physical potential for ISMs.
• Developed a foundational ISM that overcomes limitations of conventional approaches.
• The novel ISM accelerates the development of predictive large-scale simulation tools for proteins.