Extrapolating Foundation Generative Models with Physics: A Case Study of Exploring Peptide Conformations under...

Area of Science:

• Computational chemistry
• Biophysics
• Materials science

Background:

• Foundation generative models excel at predicting molecular and material structures.
• Training these models requires extensive data, which is often scarce for specific applications like protein-environment interactions (PEIs).
• Predicting protein conformations influenced by external factors (e.g., organic linkers, material surfaces) remains a challenge.

Purpose of the Study:

• To develop a method for extending the domain of foundation models without retraining.
• To enable accurate prediction of protein conformations under specific protein-environment interactions (PEIs).
• To provide a generalizable approach for applying foundation models to proteins interacting with diverse environments.

Main Methods:

• Coupling a pretrained deep generative model with explicit, physics-based interaction potentials.
• Guiding the sampling process to conform to external constraints imposed by PEIs.
• Validating the method on cyclic peptides with organic linkers and peptides adsorbed on a gold surface.

Main Results:

• Accurate and efficient prediction of protein conformations in the presence of PEIs.
• Demonstrated success in modeling cyclic peptides with organic linkers.
• Successfully predicted conformations of peptides adsorbed onto a gold surface.

Conclusions:

• The proposed physics-guided approach effectively extrapolates foundation models beyond their training data.
• This method allows for accurate conformation prediction of proteins with system-specific environmental interactions.
• The generated structures serve as high-quality inputs for subsequent simulations, offering a systematic way to enhance foundation models.