DRACO: A Denoising-reconstrution Autoencoder for Cryo-EM

Accepted by NeurIPS 2024 | project page | hugging face

Yingjun Shen, Haizhao Dai, Qihe Chen, Yan Zeng, Jiakai Zhang, Yuan Pei, Jingyi Yu

TODOS

• More compatibility with modern softwares, i.e., supporting .star file output
• More downstream tasks (particle level)

Setup

First, download the codebase:

git clone https://github.com/Cellvers/draco.git -o draco && cd draco

Second, install the dependencies by manually installing them:

• Install dependencies manually:

  conda create -n draco python=3.11
  conda activate draco
  conda install pytorch torchvision pytorch-cuda=12.1 -c pytorch -c nvidia
  conda install h5py ipykernel matplotlib "numpy<2.0.0" pandas rich scipy
  pip install fvcore mrcfile omegaconf timm opencv-python pycocotools

Pretrained models

We have pretrained DRACO with two different parameter size. You can adapt DRACO's encoder to your own downstream task. If you want to use the large model parameters, please refer to Licence.

Model name	Model size
Draco/B	Base

Please refer to the jupyter script for further instructions.

Usage

This repository contains the inferencing code of the downstream tasks mentioned in the paper. We also provide a .h5 format data sample, which is from Empiar-10096, for inferencing. If you want to use the large model parameters, please refer to Licence.

1. Particle picking

For particle picking, we have finetuned two models with different parameter size using pretrained draco model.

Model name	Model size
Detectron/B	Base

Please refer to the jupyter script for further instructions.

2. Denoising

For micrograph denoising, we further finetune our model on base parameter size.

Model name	Model size
Denoise/B	Base

Please refer to the jupyter script for further instructions. We also provide hugging face demo with more controllability.

Acknowledgements

DRACO uses detectron2 as the detection framework for particle picking.

Licence

DRACO source code is released under the Creative Commons Attribution-Non-Commercial ShareAlike International License, Version 4.0 (CC-BY-NC-SA 4.0) (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at https://github.com/Cellverse/draco/blob/main/LICENSE.

We only provide base model parameters for users to freely download and use them for non-commercial purposes. If users want to use the large model parameters, please send an email to contact@cellverse.tech with a brief description of your intended research use and your organization.

Use restrictions for all model parameters:

• They are provided solely for non-commercial use by non-commercial organizations, and redistribution or use for other purposes is prohibited.
• You cannot publish or share DRACO model parameters outside your organization. However, sharing internally for approved non-commercial use is allowed.
• You may share and adapt DRACO output under these terms, with requirements for clear notice of modifications.

Citation

@article{shen2024draco,
  title={DRACO: A Denoising-Reconstruction Autoencoder for Cryo-EM},
  author={Shen, Yingjun and Dai, Haizhao and Chen, Qihe and Zeng, Yan and Zhang, Jiakai and Pei, Yuan and Yu, Jingyi},
  journal={Advances in Neural Information Processing Systems},
  volume={38},
  year={2024}
}