[Paper] [Website]

CryoBoltz is a method for fitting atomic structures into cryo-EM density maps of dynamic proteins. It is built on top of Boltz-1, a state-of-the-art structure prediction model for biomolecular complexes. Through a multi-stage guidance mechanism that modifies the Boltz diffusion trajectory at inference time, CryoBoltz recovers diverse conformations from input cryo-EM data.

We recommend installing CryoBoltz in a clean conda environment -- first clone the git repository, and then use pip to install the package from the source code:

conda create --name cryoboltz python=3.10
conda activate cryoboltz
git clone https://github.com/ml-struct-bio/cryoboltz.git
cd cryoboltz
pip install -e .

Run the following to check that the installation was successful:

boltz predict --help

CryoBoltz requires the following inputs:

1. The sequence to be modeled, in FASTA or YAML format. See the Boltz prediction instructions for more details.

2. A cryo-EM density map, in MRC format. The prediction may be sped up by cropping background regions of the map first -- see the Example section below. It is also recommended to have an estimated map resolution, and choose a map threshold that excludes background noise.

3. An existing structure that is roughly aligned with the map, in CIF format. This structure aids in pre-aligning the CryoBoltz prediction to the map during the diffusion sampling process, but is not otherwise used as a template. This may be, e.g., an unguided Boltz prediction aligned to the map, or a partially built model from ModelAngelo. Please ensure that the chain labels in the input structure match those of the sequence file.

Once the inputs have been prepared, prediction can be run with a single command:

boltz predict <SEQ_PATH> --density_map <MRC_PATH> --aligned_model <CIF_PATH> --res <RESOLUTION> --thresh <THRESHOLD> --use_msa_server --out_dir results

Below are the available arguments specific to CryoBoltz. For a full list of options of the base Boltz model, see its prediction instructions.

See the Boltz prediction instructions for a description of all output files.

Here we walk through the modeling of a Pma1 monomer in its autoinhibited state (PDB:9UGC).

Download the FASTA file from PDB:9UGC. Following the Boltz formatting convention, and restricting ourselves to a single monomer, we rewrite the header. The file (rcsb_pdb_9UGC.fasta) should looks as follows:

>A|protein
MTDTSSSSSS...

Download and decompress the map file from EMDB:64136. We note that the reported resolution is 3.52 Å. Furthermore, by visual inspection (e.g. in ChimeraX), we choose 0.35 as the map threshold. The prediction can be sped up by cropping out background regions of the map. Here, we crop to the density at the 0.35 threshold with 10 voxels of padding on each side:

python scripts/preproc_map.py emd_64136.map --thresh 0.35 --pad 10 -o map_cropped.mrc

Alternatively, the --dim option can be used to crop the map to a fixed size.

For the input structure, we will use an unguided Boltz prediction that has been aligned to the map. We have provided such a structure in examples/input_struct.cif. To create your own, run Boltz without cryo-EM inputs:

boltz predict rcsb_pdb_9UGC.fasta --diffusion_samples 1 --use_msa_server --out_dir unguided_results

Then align the prediction (unguided_results/boltz_results_rcsb_pdb_9UGC/predictions/rcsb_pdb_9UGC/rcsb_pdb_9UGC_model_0.cif) to the map, e.g. using the fitmap function in ChimeraX.

With the inputs ready, we can now run Boltz with map guidance:

boltz predict rcsb_pdb_9UGC.fasta --diffusion_samples 5 --use_msa_server --density_map map_cropped.mrc --aligned_model examples/input_struct.cif --res 3.52 --thresh 0.35 --out_dir cryoboltz_results

The predicted structures are ordered by decreasing confidence score. We suggest visually assessing the fit of each candidate structure to the map. To select candidates based on measures of map-model fit, we refer to the tools available in Phenix and other validation packages.

For any feedback, questions, or bugs, please open a Github issue.

@article{raghu2025cryoboltz,
  title={Multiscale Guidance of Protein Structure Prediction with Heterogeneous Cryo-EM Data},
  author={Raghu, Rishwanth and Levy, Axel and Wetzstein, Gordon and Zhong, Ellen D.},
  journal={Advances in neural information processing systems},
  year={2025},
}

@article{wohlwend2024boltz1,
  author = {Wohlwend, Jeremy and Corso, Gabriele and Passaro, Saro and Reveiz, Mateo and Leidal, Ken and Swiderski, Wojtek and Portnoi, Tally and Chinn, Itamar and Silterra, Jacob and Jaakkola, Tommi and Barzilay, Regina},
  title = {Boltz-1: Democratizing Biomolecular Interaction Modeling},
  year = {2024},
  doi = {10.1101/2024.11.19.624167},
  journal = {bioRxiv}
}