Efficient AI Generation of Protein Folding Ensembles from Limited Coarse-Grained MD Simulations

Characterizing protein folding transition states is crucial for understanding how proteins fold to carry out their biological functions, elucidating misfolding mechanisms, developing therapeutic interventions for degenerative diseases, and advancing protein engineering and design. Despite significant experimental progress, direct observations of transition states remain challenging due to their transient and sparsely populated nature while characterization via molecular dynamics (MD) simulations are computationally expensive. To address these challenges, we evaluated diffusion models, variational autoencoders, and generative adversarial networks to generate accurate transition state structures from coarse-grained SMOG MD simulations using GROMACS.

Free energy profiles of representative single domain all-alpha, alpha/beta, and all-beta two-state folding proteins were calculated using the weighted histogram analysis method (WHAM) to identify the unfolded, transition, and folded states. We present our results on leveraging short simulations for compiling training datasets for generating novel transition state ensembles.