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

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

Citation

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

@article {Chen2022.05.19.492741,
    author = {Chen, Xiao and Morehead, Alex and Liu, Jian and Cheng, Jianlin},
    title = {DProQ: A Gated-Graph Transformer for Protein Complex Structure Assessment},
    elocation-id = {2022.05.19.492741},
    year = {2022},
    doi = {10.1101/2022.05.19.492741},
    publisher = {Cold Spring Harbor Laboratory},
    URL = {https://www.biorxiv.org/content/early/2022/05/20/2022.05.19.492741},
    eprint = {https://www.biorxiv.org/content/early/2022/05/20/2022.05.19.492741.full.pdf},
    journal = {bioRxiv}
}

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: DockQ scores and CAPRI class label

Installation

1. Download this repository

   git clone https://github.com/BioinfoMachineLearning/DProQ.git

2. Set up conda environment locally

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

3. Activate conda environment

   conda acitvate DPRoQ

Usage

Here is the inference.py script parameters’ introduction.

python inference.py
-c --complex_folder     Raw protien complex complex_folder
-w --work_dir           Working directory to save all intermedia files and folders, it will created if it is not exits
-r --result_folder      Result folder to save two ranking results, it will created if it is not exits
-r --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

DProQ requires GPU. We provide few protein complexes in example folder for test. 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 you onw dataset for evaluation, the data folder should look like:

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

Main results

Following four tables show DProQ’s consistent best result on HAF2 and DBM55-AF2 test sets in terms of hit rate and ranking loss. The best result is highlighted on bold.

HAF2 test set

Table 1: Hit rate performance on the HAF2 dataset. The BEST column represents each target’s best-possible Top-10 result. The SUMMARY row lists the results when all targets are taken into consideration.

Table 2: Ranking loss performance on the HAF2 dataset. The BEST row represents the mean and standard deviation of the ranking losses for all targets.

DBM55-AF2 test set

Table 3: Hit rate performance on DBM55-AF2 dataset. The BEST column represents each target’s best-possible Top-10 result. The SUMMARY row lists the results when all targets are taken into consideration.

Table 4: Ranking loss performance on the DBM55-AF2 dataset. The BEST row represents the mean and standard deviation of the ranking losses for all targets.

GitHub

View Github