Star 历史趋势
数据来源: GitHub API · 生成自 Stargazers.cn
README.md

CryoNet.Discovery

CryoNet.Discovery identifies protein structures from cryo-EM density maps across resolutions through cross-modal recognition

CryoNet.Discovery is a deep learning framework for identifying protein structural domains directly from cryo-EM density maps without requiring sequence information or atomic model fitting.

It enables robust structural identification at low resolutions by embedding density maps and protein domains into a shared representation space and performing similarity search against a Density-aware structure embedding Database (DateDB).

Try it online: Use the CryoNet.Discovery web interface to submit a density map and explore retrieval results without setting up the local environment.

DateDB is a prebuilt retrieval database containing precomputed atomic structure embeddings for protein domains annotated under CATH or SCOPe.

Overview of CryoNet.Discovery.

CryoNet.Discovery uses a dual-encoder framework:

  • DensityTower encodes cryo-EM density maps into density embeddings
  • AtomTower encodes protein domains into atomic structure embeddings
  • The search module retrieves structurally similar domains from DateDB using FAISS

The architecture of CryoNet.Discovery.


Overview

CryoNet.Discovery provides:

  • Sequence-independent structure identification
  • Domain-level structural recognition
  • Robust performance on low-resolution cryo-EM maps (6-10 Å)
  • Fast similarity search with FAISS
  • Precomputed Density-aware structure embedding Database (DateDB)

Quick start

# Step 1: Clone the repository
git clone https://github.com/zhangqf-lab/CryoNet.Discovery.git
cd CryoNet.Discovery

# Step 2: Download and install Miniconda
wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
bash Miniconda3-latest-Linux-x86_64.sh

# Step 3: Create the DensityTower_search environment
cd env
bash setup_env_DensityTower_search.sh
cd ..

# Step 4: Download checkpoints and DateDB, then place them under:
'''
ckpt/AtomTower.pth
ckpt/DensityTower.pth
ckpt/cath/ (Full CATH DateDB)
ckpt/scope/ (Full SCOPe DateDB)
ckpt/cath_benchmark/
ckpt/scope_benchmark/
'''

# Step 5: Activate the DensityTower_search environment
conda activate DensityTower_search

# Step 6: Run inference
bash code/run.sh ./data example.map cath

# Step 7: Check Output
'''
results/example_densityembedding.txt density map embedding
results/example_result_uniq.tsv ranked list (after merging) ← recommended this
results/example_result.tsv ranked list (before merging by domain type)
'''

Repository Structure

CryoNet.Discovery/
│
├── ckpt/                                   # pretrained model checkpoints
│   ├── cath/                               # CATH DateDB
│   ├── cath_benchmark/                     # CATH benchmark DateDB
│   ├── scope/                              # SCOPe DateDB
│   ├── scope_benchmark/                    # SCOPe benchmark DateDB
│
├── code/
│   ├── AtomTower/                          # AtomTower code
│   ├── DensityTower/                       # DensityTower code
│   ├── search/                             # Search module
│   └── run.sh                              # main inference script
│
├── data/                                   # input data directory
│   └──  example.map
│
├── data_list/                              # dataset metadata
│   ├── intermediatetohigh_cath_bench.txt   # intermediate- to high-resolution CATH benchmark density map list
│   ├── low_cath_bench.txt                  # low CATH benchmark density map list
│   ├── intermediatetohigh_scope_bench.txt  # intermediate- to high-resolution SCOPe benchmark density map list
│   ├── low_scope_bench.txt                 # low SCOPe benchmark density map list
│   ├── cath_DateDB_full.txt                # CATH DateDB list
│   ├── cath_DateDB_bench.txt               # CATH benchmark DateDB list
│   ├── scope_DateDB_full.txt               # SCOPe DateDB list
│   └── scope_DateDB_bench.txt              # SCOPe benchmark DateDB list
│
├── data_process/                           # data preprocessing tools
│   ├── cut_domain_map.py
│   ├── cut_domain.py
│   ├── kludo/                              # KluDo code
│   └── CryoNet.Fit/                        # rigid-body fitting of structures into cryo-EM maps
│
├── figs/                                   # figures
│
├── results/                                # inference outputs
│
├── env/                                    # environment setup scripts
│   ├── setup_env_DensityTower_search.sh    # Conda environment for DensityTower and search module
│   ├── setup_env_AtomTower.sh              # Conda environment for AtomTower
│   ├── setup.py                            # setup script for installing third-party modules required by AtomTower
│   └── attn_core/                          # attn_core implementation adapted from OpenFold
│
└── README.md

Installation

1. Clone repository

git clone https://github.com/zhangqf-lab/CryoNet.Discovery.git
cd CryoNet.Discovery

2. Setup Conda Environments

This project consists of multiple modules, each requiring a separate Conda environment.

EnvironmentPurposeRequired
DensityTower_searchDensity model inference + Retrieval module✅ Yes
AtomTowerDateDB construction❌ Optional

If you only want to run inference with the provided pretrained models and downloadable DateDB, you only need DensityTower_search.

Create Environments

Environment Setup

Download the Miniconda installer and install it in the current directory (./):

wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
bash Miniconda3-latest-Linux-x86_64.sh

DensityTower environment + Retrieval environment:

cd env
bash setup_env_DensityTower_search.sh

This environment installs third-party modules via setup.py. The attn_core implementation follows the AlphaFold2 design and is adapted from the OpenFold codebase. If you encounter installation or compilation issues, please refer to the OpenFold repository for additional guidance: https://github.com/aqlaboratory/openfold.

AtomTower environment (Optional):

Only required if you need to build a new DateDB.

cd env
bash setup_env_AtomTower.sh

Environment Requeirments

The following environments have been tested and verified to run successfully.

🖥️ Environments 1

  • OS: Ubuntu 24.04.3 LTS
  • GPU: NVIDIA H20 (sm_90, 96 GB VRAM)
  • NVIDIA Driver: 570.172.08
  • System CUDA Driver Support: 12.8

🖥️ Environments 2

  • OS: Ubuntu 24.04.3 LTS
  • GPU: NVIDIA A800-SXM4-40GB (sm_80, 40 GB VRAM)
  • NVIDIA Driver: 570.158.01
  • System CUDA Driver Support: 12.8

Additional Requirements

  • CUDA Toolkit (nvcc) is required for compiling custom CUDA extensions
  • Internet connection is required for dependency installation
  • Conda is recommended for environment management
  • A GPU with at least 20 GB VRAM is recommended for inference.

3. Download pretrained checkpoint and DateDB

Download the complete checkpoint and DateDB bundle (ckpt.zip) from the latest GitHub release, then extract it in the repository root.

FileSize
DensityTower & AtomTower checkpoints1.40 GB
CATH benchmark DateDB34.49 MB
SCOPe benchmark DateDB7.66 MB
Full CATH DateDB306.83 MB
Full SCOPe DateDB115.02 MB

🚨 Usage notes

  • Only DensityTower and the DensityTower_search environment are required to perform retrieval with the provided DateDB.
  • For exhaustive search coverage, please download and use the Full CATH DateDB or Full SCOPe DateDB.
  • If you want to build or extend DateDB yourself, please download AtomTower and install the AtomTower environment.
  • The checkpoints and DateDB require approximately 2 GB of disk space.

4. Expected directory layout

CryoNet.Discovery/
│
├── ckpt/
│   ├── AtomTower.pth
│   └── DensityTower.pth
│
├── cath/
├── scope/
├── cath_benchmark/
└── scope_benchmark/

5. Usage

Run inference using:

bash code/run.sh {input_data_dir} {input_data_name} {search_dataset}

Arguments:

ArgumentDescription
input_data_dirdirectory containing density map
input_data_namedensity map filename (.map or .mrc)
search_datasetcath or scope

cath: search against the Full CATH DateDB

scope: search against the Full SCOPe DateDB

Example

bash code/run.sh ./data example.map cath

Input

The input file should be a cryo-EM density map in .map or .mrc format.

Before running inference, the input map should be pre-segmented to approximately cover a single domain region.

Output

Results are saved in results/

Typical output includes:

results/
├── example_densityembedding.txt     # density map embedding
├── example_result_uniq.tsv          # rank list after merging by domain type
└── example_result.tsv               # rank list before merging by domain type

For most users, example_result_uniq.tsv is the recommended output file because it merges redundant hits by domain type.


Preprocessing for training data

Domain annotation and boundary prediction

Domain-level atomic structures were derived from either curated domain annotations or predicted domain boundaries, depending on data availability.

Labeled data

For labeled atomic structures, domain annotations were collected from established protein domain classification systems:

  • CATH: the primary annotations were taken from CATH v20230207, with CATH v20250115 used as a supplementary update.
  • SCOPe: annotations were obtained from SCOPe release 2.08 (2021-07-29).

Unlabeled data

For unlabeled atomic structures, domain boundaries were predicted using KluDo. The implementation is provided in ./data_process/kludo, and the original codebase is also available at https://github.com/taherimo/kludo.

Domain structure splitting

python data_process/cut_domain.py
  • Used to split a full atomic structure into domain-level atomic structures.

Density region segmentation

python data_process/cut_domain_map.py
  • Used to segment the corresponding density region according to the location of each domain-level atomic structure.

Atomic structure fitting

CryoNet.Fit performs rigid-body fitting of atomic structures into cryo-EM density maps using FFT-based global search and local map-guided refinement. It supports single-structure, multi-structure, and job-directory workflows.

The implementation, installation instructions, and usage examples are available in data_process/CryoNet.Fit.


Citation

If you use CryoNet.Discovery, please cite:

CryoNet.Discovery

Dai, M., Xu, K., Fei, Y., Zhang, Y., Yang, R., Lei, J., & Zhang, Q. C. (2026). CryoNet.Discovery identifies protein structures from cryo-EM density maps across resolutions through cross-modal recognition. LangTaoSha Preprint Server. https://doi.org/10.65215/LTSpreprints.2026.07.27.000296

@article{dai2026cryonetdiscovery,
  title={CryoNet.Discovery Identifies Protein Structures from Cryo-EM Density Maps across Resolutions through Cross-Modal Recognition},
  author={Dai, Muzhi and Xu, Kui and Fei, Yuhan and Zhang, Yunjian and Yang, Ruiyun and Lei, Jianlin and Zhang, Qiangfeng Cliff},
  journal={LangTaoSha Preprint Server},
  year={2026},
  doi={10.65215/LTSpreprints.2026.07.27.000296}
}

CryoDomain

@inproceedings{dai2025cryodomain,
  title={CryoDomain: Sequence-free Protein Domain Identification from Low-resolution Cryo-EM Density Maps},
  author={Dai, Muzhi and Dong, Zhuoer and Fu, Weining and Xu, Kui and Zhang, Qiangfeng Cliff},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={39},
  number={1},
  pages={119--127},
  year={2025},
  doi={https://doi.org/10.1609/aaai.v39i1.31987}
}

关于 About

CryoNet.Discovery code

语言 Languages

Python97.3%
Shell2.0%
Cuda0.5%
C++0.2%
C0.0%

提交活跃度 Commit Activity

代码提交热力图
过去 52 周的开发活跃度
14
Total Commits
峰值: 10次/周
Less
More

核心贡献者 Contributors