Modeling cryo-EM structures in alternative states with generative AI and density-guided simulations

Modeling atomic coordinates into a target cryo-electron microscopy map is a crucial step in structure determination. Despite recent advances, proteins with multiple functional states remain a challenge - particularly when suitable molecular templates are unavailable for certain states, and the map resolution is not high enough to build de novo models. This is a common scenario, for example, among pharmacologically relevant membrane-bound receptors and transporters. Here, we introduce a refinement approach in which i) several initial models are generated by stochastic subsampling of the multiple sequence alignment (MSA) space in AlphaFold2, ii) the resulting models are subjected to structure-based clustering, iii) density-guided molecular dynamics simulations are performed from the centroid structures, and iv) a final model is selected on the basis of both map fit and model quality. This approach improves fitting accuracy compared to single starting point scenarios for three membrane proteins (the calcitonin receptor-like receptor, L-type amino acid transporter and alanine-serine-cysteine transporter which undergo substantial conformational transitions between functional states. Our results indicate that ensemble construction using generative AI combined with simulation-based refinement facilitates building of alternative states in several families of membrane proteins.