Learning Protein Conformational Transitions for Reference-Guided Functional Conformational Modelling

Abstract

Abstract Accurate protein structure prediction does not necessarily identify the conformational state relevant to a specific biological or pharmacological function. In practice, structural biologists often resolve this ambiguity by drawing on homologous, family-related or ligand-bound reference structures, but this process remains largely heuristic and expert-dependent. Here we formalize such reference-guided structural reasoning as a quantitative framework for functional conformational modelling. Transition Diffusion Dynamics (TDD) learns reusable molecular-dynamics-derived structural transformation priors, whereas TDD-DE combines these priors with reference-supported structural information to model specified functional states. Across approximately 600 held-out proteins, TDD-DE improved ligand-pocket reconstruction from AlphaFold-derived structures while maintaining global fold agreement and outperformed molecular dynamics and representative generative conformational models at the pocket level. Prospective target-hidden applications in C3aR, rhodopsin, D2R and SIRT3 demonstrated that related structural references could be transferred to previously unresolved target states and support experimental discovery. These results move protein modelling beyond static structure prediction toward quantitative modelling of functionally relevant conformational states, providing a framework for linking structural prediction to mechanistic interpretation and structure-guided drug discovery.