Boltz-Jump: Accelerated Sampling of the Conformational Landscape of Biomolecular Structure Prediction Models

Conformational ensembles provide valuable insight into the properties and function of biomolecules beyond what can be obtained from single structures. Diffusion-based biomolecular structure models such as AlphaFold 3 and Boltz-2 support generation of an ensemble of structures via repeated diffusion sampling. However, these models have two main limitations for this purpose. First, these models are mostly trained on static 3D structures; even with perfect sampling, the model distribution is not expected to exactly match the distribution from molecular dynamics simulations. Second, sampling ensembles with these models is computationally expensive due to the many function evaluations required to generate the reverse diffusion process. Here, we address the second issue by introducing Boltz-Jump, a method that accelerates the generation of conformational ensembles by up to $10\times$ using the Boltz-2 model without additional training using walk-jump sampling. Ensembles generated by Boltz-Jump also show significantly improved ability over Boltz-2 ensembles to replicate ensemble properties such as predicting exposed residues, weak and transient contacts in the ATLAS and mdCATH datasets. Since Boltz-Jump directly leverages the open-source Boltz-2 model, it supports sampling ensembles for all biomolecular systems (protein, small molecules, and nucleic acids) supported by Boltz-2, as well as steering the sampling process using user-defined potentials (e.g. for guidance towards physically realistic structures). Finally, Boltz-Jump enables the sampling of folding and unfolding trajectories for small proteins, unlocking new capabilities beyond the base model.