Benchmarking AI-generated structural ensembles of membrane proteins against physics-based modelling

Proteins can transition between multiple conformations, with understanding these dynamics being crucial for comprehending protein function and designing drugs. While molecular dynamics (MD) simulations can provide insights into protein conformational ensembles, they can be computationally demanding, especially for sampling rare states.

Recently, AI-based methods like the Biomolecular Emulator (BioEmu) have been developed to generate protein conformational ensembles, but it is unclear if these methods can accurately capture conformational landscapes, particularly for membrane proteins. In this study, BioEmu was evaluated for its ability to model the conformational dynamics of the bacterial rhomboid intramembrane protease GlpG.

The results showed that BioEmu generates a variety of conformations representing both open and closed states of the rhomboid lateral gate, including states involved in the catalytic cycle. These conformations align with states observed during microsecond-timescale MD simulations. However, BioEmu does not reproduce the complete conformational landscape observed with MD.

Additionally, BioEmu captures significant conformational heterogeneity within the soluble domains of rhomboids, which are flexible and not well-represented in experimental structures. The findings indicate that BioEmu can produce plausible conformational ensembles for large, multi-pass membrane proteins, sampling rare states at a fraction of the cost compared to traditional MD simulations.

These results suggest that AI-based ensemble generation could be a practical approach to exploring membrane protein dynamics and could complement conventional molecular modeling techniques.