CryoFIRE (Fast heterogeneous ab Initio Reconstruction for cryo-EM), performs ab initio heterogeneous reconstruction with unknown poses in an amortized framework. Poses and conformation are jointly estimated by an encoder while a physics-based decoder aggregates the images into an implicit neural representation of the conformational space.
See our project page for more information: https://cryofire.cs.princeton.edu/
Amortized Inference for Heterogeneous Reconstruction in Cryo-EM
Axel Levy, Gordon Wetzstein, Julien Martel, Frédéric Poitevin, Ellen D. Zhong.
We recommend using Anaconda to set up the environment. Run the following commands:
conda env create -f environment.yml
conda activate cryofire
This repository is mainly addressed to method developers. We still recommend users interested in ab initio heterogeneous reconstruction to use the cryoDRGN software. Unlike cryoFIRE, cryoDRGN uses pose search to estimate poses. We are actively working on releasing a user-oriented version of cryoFIRE.
Example of script for downloading and preprocessing for "EMPIAR-10028: Pf80S Ribosome" Wong, W. et al. Cryo-EM structure of the Plasmodium falciparum 80S ribosome bound to the anti-protozoan drug emetine. Elife 3, e01963 (2014).:
# Download EMPIAR-10028 particles (~51GB)
ascp -QT -l 200M -P33001 -i ~/.aspera/connect/etc/asperaweb_id_dsa.openssh emp_ext3@hx-fasp-1.ebi.ac.uk:/10028 .
# Downsample dataset
cryodrgn downsample shiny_2sets.star --datadir 10028/data -D 128 -o particles.128.mrcs -D 50000
# Extract pose and ctf information from cryoSPARC refinement
cryodrgn parse_ctf_csparc cryosparc_P11_J4_003_particles.cs -o ctf.pkl
cryodrgn parse_pose_csparc cryosparc_P11_J4_003_particles.cs -D 360 -o poses.pkl
More datasets available at https://github.com/zhonge/cryodrgn_empiar.
The parameters of your experiment must be saved in a .json file in the configs directory.
Run your experiment with
python src/commands/train.py name_of_your_config
We provide an example of config file and script for reconstructing a pre-catalytic spliceosome from the EMPIAR-10028 dataset. In the scripts directory, run
bash run_empiar10028.sh
Experiments can also be launched from a jupyter notebook. See notebooks/template.ipynb for an example.
Your experiment can be monitored using tensorboard.
The path to the summaries must be specified in the config file.
To monitor your experiment, run
tensorboard --logdir path_to_log_directory --port XXXX --bind_all
You can generate volumes from a CryoFIRE model and a set of conformation variables using src/commands/generate_volumes.py.
See notebooks/template.ipynb for a template.
particles: Particle stack file (.mrcs, .star, .txt)
outdir: Output directory to save model
norm: Data normalization as shift, 1/scale (default: mean, std of dataset)
load: Initialize training from a checkpoint
log-interval: Logging interval in number of images
log-heavy-interval: Heavy logging interval (poses, images, testing) and checkpoint saving in epochs
seed: Random seed
ctf: CTF parameters (.pkl)
ind: Filter indices (.pkl)
pose: GT poses (.pkl)
colors: GT colors (.pkl)
relion31: Flag if relion3.1 star format
lazy: Activates lazy data loading
verbose-time: Print runtimes
num-workers: Number of CPUs for data loading
test-particles: Particle stack file for testing (.mrcs)
test-ctf: Test CTF parameters (.pkl)
test-pose: Test GT poses (.pkl)
test-colors: Test GT colors (.pkl)
num-epochs: Number of training epochs
batch-size: Minibatch size
wd: Weight decay in Adam optimizer
lr: Learning rate in Adam optimizer
beta-conf: Choice of beta schedule or a constant for KLD weight for conf VAE
no-trans: Inference over image rotation only
sym-loss: Activate symmetric loss
sym-loss-factor: Symmetric loss factor ([2])
loss-scale: Total loss multiplicative factor
pose-only-phase: Number of images for the pose-only phase
output-mask: Type of output mask (['circ', 'frequency_marching'])
add-one-frequency-every: Number of images between the addition of one frequency to the output mask
use-gt-poses: Use ground truth poses
depth-cnn: Number layers in the shared cnn
channels-cnn: Number of channels in the first layer of the shared cnn
kernel-size-cnn: Kernel size in the shared cnn
variational-het: Sets the conf VAE in variational mode
z-dim: Dimension of latent variable
input-mask: Type of input mask (['none'])
std-z-init: Standard deviation of z during pose-only phase
hypervolume-layers: Number of hidden layers fot hypervolume
hypervolume-dim: Number of nodes in hidden layers for hypervolume
pe-type: Type of positional encoding (['gaussian'])
pe-dim: Num features in positional encoding
feat-sigma: Scale for random Gaussian features
hypervolume-domain: Implicit representation domain (['hartley'])
@article{levy2022amortized,
title={Amortized Inference for Heterogeneous Reconstruction in Cryo-EM},
author={Levy, Axel and Wetzstein, Gordon and Martel, Julien and Poitevin, Frederic and Zhong, Ellen D},
journal={arXiv preprint arXiv:2210.07387},
year={2022}
}