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

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 inferencetime approach allows us to build atomic models into heterogeneous cryo-EM maps across a variety of dynamic biomolecular systems including transporters and antibodies.

Recent exploratory lines of work have attempted to address this outstanding challenge. MSA subsampling methods, for example, rely on randomly masking input sequence data to broaden the diversity of output structures [87,23,43,36]. Despite results showing improved diversity on specific systems, MSA subsampling methods remain an active area of research, with no clear consensus yet regarding their performance. Moreover, these methods are not well suited for complexes that can adopt many different conformational states or a continuum of conformational states. Other works, including AlphaFlow [40] and BioEmu [50], investigated incorporating physics-based molecular dynamics simulation as additional training data. These works also showed greater variability among output structures but were mainly demonstrated on small peptides, additionally requiring costly training and relying on simulations that may not capture realistic atomic motions.

Here we introduce a method, CryoBoltz, that leverages heterogeneous cryo-EM data to guide the sampling process of a diffusion-based structure prediction algorithm (Figure 1). Our implementation is based on Boltz-1 [88], an open-source sequence-conditioned diffusion model heavily inspired by the state-of-the-art model AlphaFold3 [1]. Through a multiscale guidance mechanism, CryoBoltz combines the structural information learned by the pretrained diffusion model with experimentallycaptured data, producing structures consistent with the cryo-EM data. Importantly, our method does not require an additional training step, while effectively mitigating the single-structure bias of current structure prediction models. We demonstrate results on both synthetic and real cryo-EM maps of dynamic biomolecular complexes.

Recent advances in protein structure prediction from sequences are exemplified by major breakthroughs such as AlphaFold2 [41] and AlphaFold3 [1]. While AlphaFold2 predicts static structures with remarkable accuracy, AlphaFold3 introduces a diffusion modeling head within its structure module, enabling generative sampling of different conformations, conditioned on the same sequence.