Free Energy Calculation Method Based on Enhanced Sampling of Diverse Protein Conformations Predicted by Artificial Intelligence

Recent advances in artificial intelligence (AI) have enabled the rapid and accurate prediction of diverse protein structures. The predicted conformations are specified as the initial structures for molecular dynamics simulations in their conformational sampling because they are frequently distributed near transition paths. Therefore, their conformational sampling promotes structural transitions toward neighboring metastable states, leading to the efficient calculation of the free-energy landscapes (FELs) of proteins. To establish this framework, our original enhanced conformational sampling method, Outlier FLOODing (OFLOOD), was integrated with AI models as AI-Assisted OFLOOD. As a demonstration, the AI-Assisted OFLOOD was implemented with AlphaFold2 by reducing multiple sequence alignments, and it was applied to soluble and membrane proteins. Consequently, the AI-Assisted OFLOOD successfully identified the transition-like and metastable states of both proteins in their FELs, demonstrating that the present framework is a reliable and efficient FEL calculation method for quantitatively evaluating the stability of AI-predicted structures.