Enhancing Generalizability in Protein-Ligand Binding Affinity Prediction with Multimodal Contrastive Learning

Requirements

To set up your environment to run the code, install the following packages: Install python 3.8.16 using conda with conda install:
pytorch==2.2 (see in pytorch official website instructions).
pymol-open-source==2.5.0

Install GVP-GNN:
pip install pyg_lib torch_scatter torch_sparse torch_cluster torch_spline_conv torch_geometric -f https://data.pyg.org/whl/torch-2.2.0+cu121.html
follow instructions here.
git clone https://github.com/drorlab/gvp-pytorch.git
cd gvp-pytorch
pip install -e .

Finally the rest of dependencies (copy it in a requirements.txt and run pip install -r requirements.txt):

matplotlib==3.7.2
networkx==3.1
numpy==1.24.3
pandas==2.0.3
rdkit==2022.09.5
scikit_learn==1.3.0
scipy==1.5.2
seaborn==0.12.2
tqdm==4.63.0
biopython==1.78

Usage

1. Prepare a Dataset

We provide a toy dataset to demonstrate the training and testing of our model.

Dataset structure:

data/
toy_set/
    ligand/
    ligand_1.sdf
    ligand_2.sdf
    ...
    protein/
    protein_1.pdb
    protein_2.pdb
    ...

CSV file format:

pdb,affinity
3uri,9
4m0z,5.19
4kz6,3.1
4jxs,4.74
2r9w,5.1
...

Preprocessing steps:

1. Run the preprocessing script: python preprocessing.py

2. Prepare the dataset: python dataset_ConBAP.py

2. Model Training

Contrastive Learning with Redocked 2020 Dataset:

• Process the redocked 2020 dataset. (The processed data sets are available at here.)
• Run the pretraining script: python pretrain.py.

Fine-Tuning with PDBbind Dataset:

• A checkpoint for contrastive learning is available in ./unsupervised/model.
• Run the training script: python train_ConBAP.py (The processed data sets are available at here.) (Note: Modify file paths based on your directory structure.)

3. Model Testing

• Testing checkpoints are located in ./supervised/model.
• Run the prediction script: python predict_single.py.
• If you want to test the docking power or screening power in CASF-2016:
• Run the test script: python casf_docking_single.py casf_screening_single.py.
• If you want to use this model on your own dataset, Run the test script: predict.py.

(Note: Modify file paths based on your directory structure.)

Reference

• Ding Luo, Dandan Liu, Xiaoyang Qu, Lina Dong, and Binju Wang*, J. Chem. Inf. Model. 2024, 64, 6, 1892–1906. Enhancing Generalizability in Protein–Ligand Binding Affinity Prediction with Multimodal Contrastive Learning