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

In recent years, many foundation generative models have been developed to predict structures of molecules and materials. Although these foundation models have achieved great success, it is challenging to collect enough data to train foundation generative models. One such example is to predict protein conformations with protein-environment interactions (PEIs), such as interactions introduced by organic linkers or material surfaces. We propose a physics-guided route to extrapolate foundation models beyond their training domain. Our method couples a pretrained deep generative model with explicit, physics-based interaction potentials for PEIs, steering sampling toward conformations consistent with external constraints without any retraining or fine-tuning. We demonstrate accurate and efficient conformation prediction of (i) cyclic peptide with organic linkers and (ii) peptide adsorbed on the gold surface. The generated structures serve as high-quality initial conditions for downstream simulations, providing a general, systematic approach to extend foundation models to proteins under system-specific environmental interactions.