Multiscale guidance of protein structure prediction with heterogeneous cryo-EM data

TL;DR

CryoBoltz presents an inference-time framework that guides a pretrained diffusion-based protein structure predictor using heterogeneous cryo-EM data to sample conformational ensembles reflecting experimental evidence. It combines global shape constraints with a physics-informed local forward-model term in a multiscale guidance pipeline, enabling atomic models to be built into heterogeneous density maps without retraining. Across synthetic and real datasets, CryoBoltz increases sampling diversity, improves fit in flexible regions such as antibody CDR loops, and outperforms existing diffusion baselines and model-building methods. The approach offers a practical path to exploring biomolecular conformational landscapes while leveraging large-scale structural priors.

Abstract

Protein structure prediction models are now capable of generating accurate 3D structural hypotheses from sequence alone. However, they routinely fail to capture the conformational diversity of dynamic biomolecular complexes, often requiring heuristic MSA subsampling approaches for generating alternative states. In parallel, cryo-electron microscopy (cryo-EM) has emerged as a powerful tool for imaging near-native structural heterogeneity, but is challenged by arduous pipelines to transform raw experimental data into atomic models. Here, we bridge the gap between these modalities, combining cryo-EM density maps with the rich sequence and biophysical priors learned by protein structure prediction models. Our method, CryoBoltz, guides the sampling trajectory of a pretrained biomolecular structure prediction model using both global and local structural constraints derived from density maps, driving predictions towards conformational states consistent with the experimental data. We demonstrate that this flexible yet powerful inference-time approach allows us to build atomic models into heterogeneous cryo-EM maps across a variety of dynamic biomolecular systems including transporters and antibodies. Code is available at https://github.com/ml-struct-bio/cryoboltz .

Figures (12)

• Figure 1: Problem statement. Diffusion-based structure prediction models af3boltz1 can sample different conformations, but the generated ensemble is generally peaked around a single conformation. A cryo-EM experiment probes the full conformational landscape but current reconstruction algorithms only provide density maps, not atomic models. With CryoBoltz, we guide the diffusion process in atomic space with experimental cryo-EM measurements to increase sample diversity and more faithfully reflect the true conformational ensemble.
• Figure 2: Overview of the guidance mechanism. The diffusion process starts with a "warm up" stage where the score model is only conditioned on the sequence. During "global guidance", the structure is guided to minimize its distance to a point cloud representation of the input cryo-EM map. During "local guidance", it is guided to maximize its consistency with the original input map. Finally, the last "relaxation" steps are unguided and allow the model to correct fine-grained details.
• Figure 3: Results on STP10 stp10.a) CryoBoltz can predict both the outward and inward states, when guided with the respective cryo-EM density map. Best sample (lowest all-atom RMSD across all model replicates) is shown. b, c) Current structure prediction models are biased towards one of these states. Boltz-1 boltz1 only predicts the outward state (b) while AlphaFold3 af3 only predicts the inward state (c). Five best samples relative to the outward and inward PDB structures, respectively, are shown.
• Figure 4: Results on CH67 antibody antibody. CryoBoltz fits the CDR H3 loop more accurately than Boltz-1 boltz1 and AlphaFold3 af3. Three best samples (lowest all-atom RMSD across all model replicates) are shown on the right.
• Figure 5: Results on P-glycoprotein glycoprotein. CryoBoltz recovers distinct states of the protein using four experimental cryo-EM maps. ModelAngelo jamali2024automated produces highly incomplete models.