Overview

- Please note that the  repository is for reference purposes only. We do not guarantee its active functionality. A user-friendly version is currently under development.

Folding-Docking-Affinity (FDA) is a framework which folds proteins, determines protein-ligand binding conformations, and predicts binding affinities.

Dependencies

The Folding part was tested with Python 3.10.13 and CUDA 12.3 on Ubuntu 20.04, with access to Nvidia Tesla V100 (32GB RAM), Intel(R) Xeon(R) Platinum 8168 CPU @ 2.70GHz, and 1.5TB RAM. Please follow localcolabfold to install the working environment.

The Docking and Affinity parts were tested with Python 3.9.18 and CUDA 11.5 on CentOS Linux 7 (Core), with access to Nvidia A100 (80GB RAM), AMD EPYC 7352 24-Core Processor, and 1TB RAM. Run the following to create a conda environment, FDA.

conda create --name FDA python=3.9
conda activate FDA
conda install conda-forge::pymol-open-source
conda install pytorch==1.11.0 torchvision==0.12.0 torchaudio==0.11.0 cudatoolkit=11.3 -c pytorch
pip install scipy
pip install --no-index pyg_lib torch_scatter torch_sparse torch_cluster torch_spline_conv -f https://data.pyg.org/whl/torch-1.11.0+cu113.html
pip install torch_geometric
python -m pip install PyYAML scipy "networkx[default]" biopython rdkit-pypi e3nn spyrmsd pandas biopandas
!!!! need to install openbabel, but seems to have conflicts.

Datasets

Create a directory /data, and download the processed data for replicating benchmark, ablation study, and binding pose augmentation results from zenodo and decompress the files

git clone git@github.com:ZhiGroup/FDA.git
cd FDA
mkdir data
cd data
wget https://zenodo.org/records/15058571/files/benchmark.tar.gz?download=1
wget https://zenodo.org/records/15058571/files/ablation_study.tar.gz?download=1
tar -xvzf benchmark.tar.gz?download=1
tar -xvzf ablation_study.tar.gz?download=1
cd ../

File structure

|--data
   |--ablation_study
      |--crystal_crystal
      |--crystal_diffdock
      |--colabfold_diffdock
      |--train.csv
      |--valid.csv
      |--test.csv
   |--benchmark
      |--davis_colabfold_protein
      |--davis_colabfold_protein_rank_2
      |--davis_colabfold_protein_rank_3
      |--davis_complex_colabfold_diffdock
      |--davis_complex_colabfold_rank2_diffdock
      |--davis_complex_colabfold_rank3_diffdock
      |--davis_ligand
      |--davis_data.tsv
      |--davis_cluster_id50_cluster.tsv
      |--kiba_colabfold_protein
      |--kiba_colabfold_protein_rank_2
      |--kiba_colabfold_protein_rank_3
      |--kiba_complex_colabfold_diffdock
      |--kiba_complex_colabfold_rank2_diffdock
      |--kiba_complex_colabfold_rank3_diffdock
      |--kiba_ligand
      |--kiba_data.tsv
      |--kiba_cluster_id50_cluster.tsv

Replicate results

Affinity prediction benchmark

Folding

Use ColabFold to generate three-dimensional protein structures. Please follow localcolabfold to install the working environment. Or directly download the processed data from zenodo and place them in /data directory and jump to the last step.

python folding/create_davis_protein_input.py
colabfold_batch --templates --amber folding/input/davis_protein.csv folding/output/davis_colabfold_protein --use-gpu-relax --num-relax 3 --gpu 0

Docking

Create protein-ligand complex directories

python docking/create_dir.py --data_dir data/benchmark/ --complex_dir_name davis_complex_colabfold_diffdock --protein_dir_name davis_colabfold_protein --ligand_dir_name davis_ligand

Download ESM2 embedding from zenodo and place the file in docking/DiffDock/data/ for the input of DiffDock. The process of generating ESM2 embedding could refer DiffDock or follow the following commands.

cd docking/DiffDock
python datasets/esm_embedding_preparation.py --protein_path ../../data/benchmark/davis_colabfold_protein --out_file data/davis_colabfold_protein.fasta
git clone https://github.com/facebookresearch/esm.git
cd esm
python scripts/extract.py esm2_t33_650M_UR50D ../data/davis_colabfold_protein.fasta ../data/davis_colabfold_protein_embedding_output --repr_layers 33 --include per_tok --truncation_seq_length 4096
cd ../
python datasets/esm_embeddings_to_pt.py --esm_embeddings_path data/davis_colabfold_protein_embedding_output --output_path data/esm2_3billion_embeddings_davis_colabfold_protein.pt

Implement DiffDock to generate ligand binding poses

cd docking/DiffDock
python -m affinity.dataset_davis_colabfold --run_name davis_colabfold --split_train ../../data/benchmark/davis_data.tsv --data_dir ../../data/benchmark --protein_dir_name davis_colabfold_protein --ligand_dir_name davis_ligand --esm_embeddings_path data/esm2_3billion_embeddings_davis_colabfold_protein.pt --inference_steps 20 --samples_per_complex 10 --batch_size 10 --ns 12 --nv 6 --num_conv_layers 3 --dynamic_max_cross --scale_by_sigma --dropout 0.2 --remove_hs --c_alpha_max_neighbors 24 --receptor_radius 15 --gpu_num 0

Update DiffDock-generated ligand poses into original complex directories.

cd ../../
python docking/update_dir.py --diffdock_ligand_path docking/DiffDock/data/cacheNew/davis_colabfold_protein_davis_data_split_train_limit_0/ligand_positions_rank1.pkl --complex_path data/benchmark/davis_complex_colabfold_diffdock

Affinity

Pre-process protein-ligand complexes and generate inputs for GIGN.

python affinity/GIGN/preprocessing.py --data_df data/benchmark/davis_data.tsv --complex_path data/benchmark/davis_complex_colabfold_diffdock
cd affinity/GIGN
python affinity/GIGN/dataset_GIGN_benchmark.py --data_df data/benchmark/davis_data.tsv --complex_path data/benchmark/davis_complex_colabfold_diffdock --mmseqs_seq_clus_df data/benchmark/davis_cluster_id50_cluster.tsv

Train GIGN to predict binding affinity under different split_methods (drug, protein, both, and seqid).

cd affinity/GIGN
python train_GIGN_benchmark.py --job_name benchmark_davis_drug --split_method drug --gpu 0 --data_df ../../data/benchmark/davis_data.tsv --complex_path ../../data/benchmark/davis_complex_colabfold_diffdock

For the KIBA dataset, the process remains the same, with only the input being changed. The input can also be found in /data/benchmark.

Ablation study

Download the processed data from zenodo and place them in /data. Train GIGN to predict binding affinity under three different scenarios (crystal_crystal, crystal_diffdock, and colabfold_diffdock).

cd affinity/GIGN
python train_GIGN_ablation.py --scenario crystal_crystal --gpu 0

Binding poses augmentation

Use the same dataset as the affinity prediction benchmark. The dataset can be found in /data/benchmark. The following commands are used to train GIGN with different binding poses augmentation strategies.

F-5D-A: The top five binding poses generated by DiffDock are used to augment the training set.

python train_GIGN_benchmark.py --split_method drug --job_name benchmark_davis_pose_augmentation5_protein_rank1_drug --gpu 0 --data_df ../../data/benchmark/davis_data.tsv --complex_path ../../data/benchmark/davis_complex_colabfold_diffdock --mmseqs_seq_clus_df ../../data/benchmark/davis_cluster_id50_cluster.tsv --seeds 0 1 2 3 4 --top_n 5

F-10D-A: The top ten binding poses generated by DiffDock are used to augment the training set

python train_GIGN_benchmark.py --split_method drug --job_name benchmark_davis_pose_augmentation10_protein_rank1_drug --gpu 0 --data_df ../../data/benchmark/davis_data.tsv --complex_path ../../data/benchmark/davis_complex_colabfold_diffdock --mmseqs_seq_clus_df ../../data/benchmark/davis_cluster_id50_cluster.tsv --seeds 0 1 2 3 4 --top_n 10

2F-5D-A: For each protein-ligand pair, five binding poses were selected from both the rank-1 and rank-2 protein conformations generated by ColabFold, resulting in a total of 10 distinct binding poses

python train_GIGN_benchmark.py --split_method drug --job_name benchmark_davis_pose_augmentation10_protein_rank1_2_drug --gpu 0 --data_df ../../data/benchmark/davis_data.tsv --complex_path ../../data/benchmark/davis_complex_colabfold_diffdock --mmseqs_seq_clus_df ../../data/benchmark/davis_cluster_id50_cluster.tsv --seeds 0 1 2 3 4 --top_n 5 --protein_rank 1 2 

3F-5D-A: For each protein-ligand pair, five binding poses were selected from the rank-1, rank-2, and rank-3 protein conformations generated by ColabFold, resulting in a total of 15 distinct binding poses.

python train_GIGN_benchmark.py --split_method drug --job_name benchmark_davis_pose_augmentation15_protein_rank1_2_3_drug --gpu 0 --data_df ../../data/benchmark/davis_data.tsv --complex_path ../../data/benchmark/davis_complex_colabfold_diffdock --mmseqs_seq_clus_df ../../data/benchmark/davis_cluster_id50_cluster.tsv --seeds 0 1 2 3 4 --top_n 5 --protein_rank 1 2 3

For the KIBA dataset, the process remains the same, with only the input being changed. The input can also be found in /data/benchmark.

Citation

Wu, MH., Xie, Z., & Zhi, D. Protein-ligand binding affinity prediction: Is 3D binding pose needed?. bioRxiv (2024). https://doi.org/10.1101/2024.04.16.589805