Code of the paper Physics-informed geometric regularization of heterogeneous reconstructions in cryo-EM.

Data preparation

1. The MD-computed conformation trajectory files for the ADK and Nsp13 datasets are in ./data.
2. Convert each conformation of the trajectory to its own .pdb file.
3. Run parakeet on the pdb files using the configuration data/parakeet_config.yaml.
4. Crop the produced micrographs around the particles, and extract the 3D pose of each particle from the output of Parakeet.
5. Put the resulting tensors in a dict with keys "imgs", "poses", "conf_id" and save as picked_particles_<electron_dose>e.pickle. The classes AdkDataset and CovDataset can load this file.

Please contact me (victor.prins [at] outlook [dot] com) if you need help with any of these steps. I can also provide the complete preprocessed image datasets (~20GB in total) that were used for the paper.

Installation

Install the dependencies with pip install -r requirements-gpu.txt. The code is tested with Python 3.9. Run wandb login on the command line to enable login to your Weights & Biases account.

Training the network

1. Modify any of the values in config in train.py as appropriate for your use case.
2. Run python train.py.

Citation

If you find this useful, please cite our paper:

@inproceedings{prins2024physics,
  title={Physics-informed geometric regularization of heterogeneous reconstructions in cryo-EM},
  author={Prins, Victor and Diepeveen, Willem and Bekkers, Erik J and {\"O}ktem, Ozan},
  booktitle={ICLR 2024 Workshop on Generative and Experimental Perspectives for Biomolecular Design}
}