Protein conformational dynamics and ligand binding processes are fundamental to biological function, yet their systematic sampling and thermodynamic characterization remain challenging. Here, we present a distance-restraint-guided inference method that extends AlphaFold3-like diffusion model frameworks to predict protein structures at specified conformational states. By restraining intergroup distances─defined as geometric centroid distances between atom groups─during the reverse diffusion process, our method enables systematic sampling along reaction coordinates without model retraining. We implemented this approach in Boltz-2 and demonstrated its effectiveness on three model proteins that undergo open-closed conformational transitions, as well as on a protein-peptide dissociation pathway. Compared with conventional approaches that induce conformational diversity by manipulating the input multiple sequence alignments, our method achieved more uniform coverage of conformational space while maintaining high structural quality as assessed by both learning-based confidence metrics and stereochemistry-based validation. By combining distance-restrained sampling with molecular dynamics simulations, we constructed free energy landscapes and quantitatively estimated binding free energies. Altogether, our approach bridges deep learning-based structure prediction and physics-based simulations, providing an efficient strategy for characterizing the dynamics of biomolecules.
Hori et al. (Thu,) studied this question.