DProQA: A Gated-Graph Transformer for Protein Complex Structure Assessment

ISMB2023

Our work has been accepted by ISMB 2023, here is the latest manuscript.

CASP15 result

In terms of TMscore ranking loss, DProQA (team: MULTICOM_egnn) ranked 9th among all 26 methods and 3rd place among all single-model methods.

---

DProQA is a Gated-Graph Transformer model for end-to-end protein complex structure's quality evaluation. DProQA achieves significant speed-ups and better quality compared to the current baseline method. If you have any questions or suggestions, please contact us at [email protected]. We are happy to help!

Citation

If you think our work is helpful, please cite our work by:

@article{10.1093/bioinformatics/btad203,
    author = {Chen, Xiao and Morehead, Alex and Liu, Jian and Cheng, Jianlin},
    title = "{A gated graph transformer for protein complex structure quality assessment and its performance in CASP15}",
    journal = {Bioinformatics},
    volume = {39},
    number = {Supplement_1},
    pages = {i308-i317},
    year = {2023},
    month = {06},
    issn = {1367-4811},
    doi = {10.1093/bioinformatics/btad203},
    url = {https://doi.org/10.1093/bioinformatics/btad203},
    eprint = {https://academic.oup.com/bioinformatics/article-pdf/39/Supplement\_1/i308/50741583/btad203.pdf},
}

Installation

1. Download this repository

   git clone https://github.com/jianlin-cheng/DProQA.git

2. Set up conda environment locally

   cd DProQA
   conda env create --name DProQA -f environment.yml

3. Activate conda environment

   conda activate DProQA

Dataset

Benchmark sets

We provide our benchmark tests HAF2 and DBM55-AF2 for download by:

wget https://zenodo.org/record/6569837/files/DproQ_benchmark.tgz

Each dataset contains:

1. decoy folder: decoys files
2. native folder: native structure files
3. label_info.csv: DockQA scores and CAPRI class label

Build own training data

Due to the larger data size, we are not able to provide direct download. But we provided our training, validation targets information in Data/data_list folder. You could to use any protein complex structure prediction tool to generate predicted structures (pdb format) first, then use our data generation script to generate protein graph (.dgl) for training.

customer_data_folder
├── pdbcode1.pdb
├── pdbcode2.pdb
├── pdbcode3.pdb
├── pdbcode4.pdb
└── pdbcode5.pdb

python ./src/data_generator.py
--input_pdb_folder -p input pdbs folder
--dgl_save_folder -d dgl files save folder
--cores  -c number of cores for parallel processing

# example code
python ./src/data_generator.py -p /example_pdbs_folder/ -d /dgl_save_folder -c 10

Usage

Here is the inference.py script parameters' introduction.

python inference.py
-c --complex_folder     Raw protein complex complex_folder
-w --work_dir           Working directory to save all intermedia files and folders, it will be created if it is not exit
-r --result_folder      Result folder to save two ranking results, it will be created if it is not exit
-t --threads            Number of threads for parallel feature generation and dataloader, default=10
-s --delete_tmp         Set False to save work_dir and intermedia files, otherwise set True, default=False

Use provided model weights to predict protein complex structures' quality

DProQA requires GPU. We provide a few protein complexes in example folder for testing. The evaluation result Ranking.csv is stored in result_folder.

python ./inference.py -c ./examples/6AL0/ -w ./examples/work/ -r ./examples/result

You can build your own dataset for evaluation, the data folder should look like this:

customer_data_folder
├── decoy_1.pdb
├── decoy_2.pdb
├── decoy_3.pdb
├── decoy_4.pdb
└── decoy_5.pdb