Hybrid protein-ligand binding residue prediction with protein language models: Does the structure matter?

This is the code repository for our paper "Hybrid protein-ligand binding residue prediction with protein language models: Does the structure matter?" by Hamza Gamouh, Marian Novotny, and David Hoksza.

Environment setup

To run the scripts, please install the Conda environement by following these steps :

1. Create a Conda environment --> conda create --name plm-gnn python=3.10
2. Activate the environment --> conda activate plm-gnn
3. If you have GPUs install pytorch for CUDA --> conda install pytorch torchvision torchaudio pytorch-cuda=11.7 cuda -c pytorch -c "nvidia/label/cuda-11.7.1"
4. If you have GPUs, install dgl for CUDA pip install dgl==1.0.1+cu117 -f https://data.dgl.ai/wheels/cu117/repo.html
5. Install other pip requirements --> pip install -r requirements.txt

Datasets

The original sequence datasets designed by Yu et al. are in datasets/yu_merged. The datasets contain protein sequences organized by their protein PDB ID and chain ID, as well as their true binding residues. Originally, there were multiple entries of binding residues due to the existence of multiple instances of a ligand, we merged the binding residues for each unique protein sequence. We also extracted the corresponding structures from BioLip, you can download them using this link.
The structure files are organized in three folders :

• receptor : Structure files extracted directly from the BioLip database and where the PDB ID and chain ID of the original sequence files match exactly.
• obsolete : Structure files missing from BioLip and downloaded from PDB RCSB where the PDB ID in the sequence files is obsolete. (they are in .cif format)
• other_pdbs : Remaining structure files missing from BioLip and downloaded from PDB RCSB.

Embeddings computation

For embeddings computation, please note that the bio-embeddings package works best in Python 3.7. Therefore, please follow these steps :

1. Create another Conda environment --> conda create --name protein_embs python=3.7
2. Activate the environment --> conda activate protein_embs
3. Install pip requirements --> pip install -r embs_requirements.txt
4. To compute the embeddings for the whole dataset please run python scripts/create_dataset_embeddings.py
5. Switch to the main environment by running conda deactivate and then conda activate plm-gnn

Protein graph construction

To construct the input data to the models, please run scripts/preprocessing_datasets.py. This script :

1. Starts with a sequence from the Yu dataset
2. Parses the corresponding PDB files to extract the sequences and atom coordinates
3. Matches the parsed sequence with the original sequence by applying manual modifications defined in scripts/manual_modifications.py
4. Loads pre-computed embeddings
5. Constructs the residue contact map (using the CUTOFF distance parameter) and converts it to a DGL graph with embedding features and binary labels for residues.

Training models

To train and test the models please run python scripts/train_model.py.

Results

To generate our result tables, please run python scripts/create_result_tables.py

Models

We trained two major architectures : Graph Convolutional Networks (GCN) and Graph Attention Networks (GAT) for all cutoff distances 4, 6, 8 and 10 Angstroms. You can download our trained models here.

Try a prediction by our GAT ensemble model

To run a prediction using our GAT ensemble model (cutoffs : 4,6,8,10), please run the following commands:

1. Change directory to be in "scripts" folder cd scripts
2. Run inference by providing your PDB file python inference.py --pdb_file your_pdb.pdb

You can try a PDB example python inference.py --pdb_file 1a2b.pdb