GNMCADS: Sampling For Protein Conformation Diversity With Gaussian Network Model Based Condition Annealed Diffusion Sampler

Proteins are dynamic molecules existing in multiple conformational states that dictate their biological functions. Current methods for predicting protein conformational ensembles mostly rely on multiple sequence alignment (MSA) modifications to increase the diversity of single-state predicting models, or incorporate molecular dynamics (MD) simulations during training to mimic conformational variability. However, MSA modifications can be insufficient for capturing conformational changes, particularly in proteins with low sequence homology such as de novo designed proteins and in protein complexes, where MSAs are constructed on a per-chain basis. Similarly, MD trained methods can fail to capture the conformational changes that happen in large timescales or large protein complexes as it is infeasible to perform MD simulations for these cases. Here, we propose a new diffusion sampling strategy that can be applied to any diffusion protein generative model that uses pairwise conditioning signals without any additional training. Our method GNMCADS, enhances structural diversity by increasing the uncertainty between the dynamical domains derived by the Gaussian Network Model (GNM), using Condition Annealed Diffusion Sampler (CADS). As GNM naturally extends to multi chain systems and does not depend on MSA, GNMCADS enables the sampling of long timescale conformational changes of proteins, including those with low evolutionary informations and of multimeric complexes. We further implement GNMCADS on AlphaFold3 (AF3) and compare it with current state-of-the-art MSA modification or MD based conformational sampling methods in 38 cases.