Sequence-based Drug-Target Complex Pre-training Enhances Protein-Ligand Binding Process Predictions Tackling Crypticity

Environment Setup

The environment can be built based on the NVIDIA container image for PyTorch, release 23.01. Running NVIDIA container image requires NVIDIA Container Toolkit installed to run with GPU support.

Inside the docker container, following commands are needed to install additional required packages:

pip install fair-esm
pip install 'dllogger @ git+https://github.com/NVIDIA/dllogger.git'
pip install 'openfold @ git+https://github.com/aqlaboratory/openfold.git@4b41059694619831a7db195b7e0988fc4ff3a307'
pip install --verbose git+https://github.com/pyg-team/pyg-lib.git
pip install --verbose torch_scatter
pip install --verbose torch_sparse
pip install biopython rdkit==2023.9.5 pdbeccdutils pytorch-lightning==1.5.10 wandb einops==0.6.1
git clone https://github.com/dptech-corp/Uni-Core/tree/8f0e59dcaea1619a5c79d6086072fb9d24872f44
cd ./Uni-Core
python setup.py install
pip install pymatgen==2023.8.10 unimol_tools

Datasets

Chemical Component Dictionary:

Download mmCIF file. Unzip Components-rel-alt.cif under ./data.

scPDB:

Download scPDB dataset. Unzip folder scPDB under ./data.

Compute ESM-2 embeddings for scPDB based on the fasta information provided:

python esm2_infer_fairscale_fsdp_cpu_offloading.py esm2_t33_650M_UR50D ./data/scPDB_unique_fasta.fasta ./data/esm/scPDB --repr_layers 33 --include per_tok contacts --truncation_seq_length 1600 --toks_per_batch 1600

COACH420 and HOLO4K:

Download protein fasta files and ligand mol2 files. Unzip folders p2rank-datasets and DeepPocket_data under ./data.

Compute ESM-2 embeddings for COACH420 and HOLO4K based on the fasta information provided:

python esm2_infer_fairscale_fsdp_cpu_offloading.py esm2_t33_650M_UR50D ./data/coach420_unique_fasta.fasta ./data/esm/coach420 --repr_layers 33 --include per_tok contacts --truncation_seq_length 1500 --toks_per_batch 1500

python esm2_infer_fairscale_fsdp_cpu_offloading.py esm2_t33_650M_UR50D ./data/holo4k_unique_fasta.fasta ./data/esm/holo4k --repr_layers 33 --include per_tok contacts --truncation_seq_length 1500 --toks_per_batch 1500

PocketMiner:

Compute ESM-2 embeddings for PocketMiner based on the fasta information provided:

python esm2_infer_fairscale_fsdp_cpu_offloading.py esm2_t33_650M_UR50D ./data/pocketminer_all_fasta.fasta ./data/esm/pocketminer --repr_layers 33 --include per_tok contacts --truncation_seq_length 1500 --toks_per_batch 1500

PDBbind v2020:

Compute ESM-2 embeddings for PDBbind v2020 based on the fasta information provided:

python esm2_infer_fairscale_fsdp_cpu_offloading.py esm2_t33_650M_UR50D ./data/pdbbind2020/pdbbind2020_fasta.fasta ./data/esm/pdbbind2020 --repr_layers 33 --include per_tok contacts --truncation_seq_length 3000 --toks_per_batch 3000

Curated binding kinetics dataset:

Compute ESM-2 embeddings for binding kinetics dataset based on the fasta information provided:

python esm2_infer_fairscale_fsdp_cpu_offloading.py esm2_t33_650M_UR50D ./data/binding_kinetics/binding_kinetics_fasta.fasta ./data/esm/binding_kinetics --repr_layers 33 --include per_tok contacts --truncation_seq_length 700 --toks_per_batch 700

How to Run

General binding site prediction

Training on scPDB for inference on COACH420:

python train_general_bs.py --task 'coach420' --lr 5e-5 --batch_size 8 --epochs 10 --seed 705 --gpu 0 --wandb

Training on scPDB for inference on HOLO4K:

python train_general_bs.py --task 'holo4k' --lr 5e-5 --batch_size 8 --epochs 10 --seed 705 --gpu 0 --wandb

Inference on COACH420:

python inference_coach420.py --lr 5e-5 --saved saved_model --seed 705 --gpu 0 --wandb

Inference on HOLO4K:

python inference_holo4k.py --lr 5e-5 --saved saved_model --seed 705 --gpu 0 --wandb

Cryptic binding site prediction

Training on scPDB, with validation and testing on PocketMiner:

python train_cryptic_bs.py --lr 1e-6 --batch_size 8 --epochs 40 --seed 705 --gpu 0 --data_txt 'scPDB_subset_pocketminer.txt' --max_length 1280 --wandb

Pre-training on scPDB for downstream binding affinity and kinetics prediction

python train_cryptic_bs.py --lr 1e-6 --batch_size 8 --epochs 40 --seed 705 --gpu 0 --data_txt 'scPDB_full.txt' --max_length 1600 --wandb

Binding Affinity prediction

Inference on PDBbind v2020 using pre-trained ProMoSite to obtain scores:

python inference_affinity_score.py --lr 1e-6 --saved saved_model --seed 705 --gpu 0 --wandb

Training (fine-tuning), validation and testing on PDBbind v2020 using ProMoBind:

python train_affinity.py --lr 5e-5 --batch_size 256 --epochs 300 --seed 705 --gpu 0 --wandb

Binding Kinetics prediction

Inference on binding kinetics dataset using pre-trained ProMoSite to obtain scores:

python inference_kinetics_score.py --lr 1e-6 --saved saved_model --seed 705 --gpu 0 --wandb

Training (fine-tuning), validation and testing on binding kinetics dataset using ProMoBind:

python train_kinetics.py --lr 5e-6 --batch_size 4 --epochs 500 --seed 705 --gpu 0 --wandb