Sampling in structure-token space enables accurate prediction of multiple protein conformations

Abstract Proteins often function through multiple conformations, yet most structure-prediction methods are biased toward a single dominant state. Here we present MSFold, a training-free framework that explores the discrete, sequence-conditioned structure-token space of ESM3 using parallel tempering, coupling high-temperature exploration with low-temperature refinement to cross barriers between conformational states. Across 312 multi-conformation protein pairs, MSFold recovered both conformations in 51.6% of cases, compared with 38.1% for MSA Cluster and 37.8% for AlphaFold3. This performance advantage persisted on a post-ESM3-cutoff evaluation set, arguing against simple structural memorization. We further introduce sequence log-likelihood, a reference-free confidence score based on sequence–structure consistency, which improves ranking of sampled conformations. Together, these results provide a practical means of accessing the conformational information encoded by protein language models and establish MSFold as an effective framework for multi-conformation prediction.