DrugCLIP: Contrastive Protein-Molecule Representation Learning

The --homoaug_dir contains both the original data and the generated augmented pairs.

Note: If a pocket-ligand pair is unsuitable (e.g., the pocket is composed of multiple chains), the program will skip it. Only valid results are saved in the extend/ directory.

Directory Structure:

homoaug_dir
├── PDBID
│   ├── PDBID.fasta
│   ├── PDBID.pdb
│   ├── PDBID_pocket.pdb
│   ├── PDBID_pocket_chain.pdb
│   ├── PDBID_ligand.pdb
│   ├── PDBID_pocket_position.txt
│   ├── rotation_matrix
│   │   └── (TMalign output files)
│   └── *extend*
│       ├── AugmentedPDBID1
│       │   ├── AugmentedPDBID1_protein.pdb
│       │   ├── AugmentedPDBID1_pocket.pdb
│       └── ...

To use HomoAug, you must install the ray Python package and ensure the following HMMER and alignment tools are available in your system's PATH:

1. Python Package: Install ray version 1.12.0 via pip.
2. HMMER Tools: jackhmmer and esl-reformat.
3. Alignment Tool: TM-align.

The executables for jackhmmer, esl-reformat, and TM-align are provided in the bin folder of the repository. You must add this bin directory to your environment's PATH variable.

pip install ray==1.12.0
# Ensure the 'bin' directory from the repo is in your PATH
export PATH=$PATH:/path/to/drugclip/HomoAug/bin

DrugCLIP requires the same dependencies as Uni-Mol.

Critical Requirement: You must use rdkit version 2022.9.5.

HomoAug generates pseudo-pocket-ligand pairs by leveraging the similarity of ligand binding sites in homologous proteins. Execute the augmentation process using the run_HomoAug.py script with the required input files and configuration parameters.

python run_HomoAug.py \
    --id_file your_id_file \
    --homoaug_dir your_homoaug_dir \
    --fasta_file your_fasta_file \
    --protein_pdb_dir your_protein_pdb_dir \
    --pocket_pdbs_dir your_pocket_pdbs_dir \
    --jackhmmer_output_dir your_jackhmmer_output_dir \
    --n_thread your_n_thread \
    --database_fasta_path your_database_fasta_path \
    --database_pdb_dir your_database_pdb_dir \
    --max_extend_num your_max_extend_num \
    --TMscore_threshold your_TMscore_threshold \
    --Match_rate_threshold your_Match_rate_threshold

The training datasets (train.lmdb and valid.lmdb) can be read using the utility script py_scripts/lmdb_utils.py. The LMDB files contain the following keys:

• atoms: atom types for each atom in the ligand
• coordinates: 3D coordinates for each atom in the ligand generated by RDKit (Max 10 conformations)
• pocket_atoms: atom types for each atom in the pocket
• pocket_coordinates: 3D coordinates for each atom in the pocket
• mol: RDKit molecule object for the ligand
• smi: SMILES string for the ligand
• pocket: pdbid of the pocket

To start the training process, execute the provided bash script from the root directory.

bash drugclip.sh

To evaluate the model on test datasets, execute the provided bash script.

bash test.sh

To perform retrieval tasks, execute the provided bash script. Example files for pocket.lmdb and mols.lmdb are available in the retrieval directory within the provided Google Drive data folder.

bash retrieval.sh

The esl-reformat tool supports two main categories of output: Unaligned and Aligned formats.

Unaligned formats

• fasta
• hmmpgmd

Aligned formats

• a2m
• afa
• clustal
• clustallike
• pfam
• phylip
• phylips
• psiblast
• selex
• stockholm

Output format choices:
Unaligned:      fasta, hmmpgmd
Aligned:        a2m, afa, clustal, clustallike, pfam, phylip, phylips, psiblast, selex, stockholm