Towards Stable Representations for Protein Interface Prediction

This is the implementation of the paper ATProt submitted to NeurIPS 2024

Dependencies

ATProt needs the following environment:

python==3.7
numpy==1.22.4
torch-geometric==2.2.0
cuda==10.2
torch==1.11.0
dgl==0.8.1
biopandas==0.4.1
dgllife==0.2.9
joblib==1.1.0
prody==2.4.0

Dataset Curation

First, generate the required graph structured data for complex with our code. The curator includes two datasets:

• Docking Benchmark 5.5 (DB5.5).
• Database of Interacting Protein Structures (DIPS).

For data preparations, you can choose the configuration as follows:

• data. ["dips","db5"]: Datasets will be processed separately, so please choose one.
• graph_cutoff. If the physical distance between two residues in a protein is less than this value, they will be assigned an edge in the KNN graph.
• graph_max_neighbor. It means the maximum number of neighbors for each central residue.
• pocket_cutoff. If the physical distance between inter-protein residues is less than this value, they will be considered in the pocket.

You can preprocess the raw data as follows for DB5.5:

python src.preprocess_raw_data.py -data db5 -graph_cutoff 20 -graph_max_neighbor 10 -pocket_cutoff 8

After this, use the following script for generating ESMFold structures

python data.esmfold_pro.py 

How to run

You can find a detailed explanation of the parameters in ./src/utils/args.py.

To reproduce the results in the paper, you can run the following for native-bound and ESMFold inference settings.

python /src/train.py -inf_data nativebound

python /src/train.py -inf_data esmfold

How to cite

@article{gao2024towards,
  title={Towards stable representations for protein interface prediction},
  author={Gao, Ziqi and Liu, Zijing and Li, Yu and Li, Jia},
  journal={Advances in Neural Information Processing Systems},
  volume={37},
  pages={73079--73097},
  year={2024}
}