Covalent Ligand Dataset for Protein-Ligand Docking

This repository contains tools and datasets for generating and curating covalent protein-ligand complexes from the Protein Data Bank (PDB). The dataset is designed for benchmarking covalent docking methods and structure prediction tools by removing redundancy and train-test data leakage.

Table of Contents

• Overview
• Dataset Description
• JSON File Formats
• Installation
• Usage

Overview

This project provides a curated dataset of covalent protein-ligand complexes extracted from the PDB, along with tools to:

1. Extract covalent ligand structures from PDB with quality filters
2. Remove train-test leakage based on sequence and ligand similarity
3. Cluster and deduplicate structures to create non-redundant datasets
4. Generate cross-docking pairs for evaluating cross-docking scinario
5. Prepare input files for various docking and structure prediction tools (GalaxyCDock, AutoDock4, DOCK6)

Dataset Description

Training Datasets

1. Scarpino Set (train_data/scarpino_set.json)

   • Covalent complexes from the Scarpino et al. benchmark
   • Only Cystein-targetting ligands
   • Pre-2018 structures
   • Used as training/reference set

2. PDBbind Data (train_data/pdbbind_data.json)

   • Non-covalent protein-ligand complexes from PDBbind
   • Used as training set of GalaxyDock-DL, which is utilized for rescoring in GalaxyCDock web server.

Test Datasets

1. Covalent Test Set (test_data/cov_set.json)

   • Newly curated covalent complexes (pre-2025)
   • Filtered for quality and non-redundancy within the test set
   • Remove redundancy from the Scarpino set
   • Extended Target residues including Cys, Ser, Lys, Tyr, and His
   • Main benchmark set

2. Cross-Docking Set (cross_set/cross_covset.json)

   • Pairs of structures with same protein but different ligands
   • Used for cross-docking validation
   • Includes ligand similarity cutoff of maximum 0.5

Directory Structure

`` cdock_dataset/ ├── README.md # This file ├── requirements.txt # Python dependencies ├── check_binding_site.py # Filter structures by binding site quality ├── clustering.py # Clustering and leakage removal ├── gen_testset.py # Generate test set from PDB ├── gen_input.py # Generate input files for docking tools ├── test_data/ │ ├── cov_set.json # Main covalent test set │ ├── covset_seq.fasta # Protein sequences for test set │ └── cov_rcsb_19950101_20250919.txt # PDB IDs list (downloaded from RCSB PDB) ├── train_data/ │ ├── scarpino_set.json
│ ├── pdbbind_data.json
│ ├── train_seq.fasta # Training set sequences of the Scarpino and PDBbind sets. │ └── scarpino.txt # Scarpino PDB IDs └── cross_set/ ├── cross_covset.json # Cross-docking pairs of the main test set. └── cross_scarpino.json # Scarpino cross-docking pairs

## JSON File Formats

### Covalent Ligand Dataset Format

Files: `cov_set.json`, `scarpino_set.json`

Each entry represents a covalent protein-ligand complex:

```json
{
  "pdb": "1b12",                    // PDB ID (4-character code)
  "date": "1999-11-24",             // Deposition date (YYYY-MM-DD)
  "exp_type": "X-RAY DIFFRACTION",  // Experimental method
  "resolution": 1.95,                // Resolution in Angstroms
  "uniprot_id": "P00803",           // UniProt accession of protein
  
  // Protein residue information (covalently bonded residue)
  "res_name": "SER",                // Residue name (3-letter code)
  "res_atom": "OG",                 // Atom name forming covalent bond
  "res_atom_alt_id": "?",           // Alternate location indicator
  "res_num": "15",                  // Residue number (label)
  "res_num_auth": "90",             // Residue number (author)
  "res_chain": "A",                 // Chain ID (label)
  "res_chain_auth": "A",            // Chain ID (author)
  
  // Ligand information
  "lig_name": "1PN",                // Ligand name (CCD code)
  "lig_atom": "C7",                 // Atom name forming covalent bond
  "lig_atom_alt_id": "?",           // Alternate location indicator
  "lig_num": ".",                   // Ligand number (label)
  "lig_num_auth": "1001",           // Ligand number (author)
  "lig_chain": "E",                 // Chain ID (label)
  "lig_chain_auth": "A",            // Chain ID (author)
  "lig_atom_num": 20,               // Number of heavy atoms in ligand
  "lig_n_tor": 7,                   // Number of rotatable bonds
  "smiles": "C[C@@H](OC(C)=O)..."   // SMILES string of ligand
}

PDBbind Dataset Format

File: pdbbind_data.json

Non-covalent complexes with binding affinity data:

{
  "pdb": "2tpi",                    // PDB ID
  "lig_name": "2-mer",              // Ligand name
  "resolution": "2.10",             // Resolution (Angstroms)
  "year": 1982,                     // Deposition year
  "smiles": "CC[C@H](C)[C@H]...",   // SMILES string
  "affinity": 4.309803919971486,    // -log(affinity) value
  "affinity_type": "Kd",            // Affinity type (Kd, Ki, IC50)
  "affinity_sign": "=",             // Equality sign (=, >, <, ~)
  "uniprot": "P00760"               // UniProt ID (legacy field name)
}

Cross-Docking Set Format

Files: cross_covset.json, cross_scarpino.json

Pairs for cross-docking experiments:

{
  "lig_id": "5ydm_DUW_B",           // Ligand structure ID (pdb_ligname_chain)
  "rec_id": "5ydl_DUV_A",           // Receptor structure ID
  "uniprot_id": "A0A0E3JLZ0",       // UniProt ID (same protein)
  "tanimoto": 0.3611111111,         // Ligand Tanimoto similarity
  "source": "additional"            // Source ("existing" or "additional")
}

Key Fields:

• lig_id: Structure to use as ligand (format: {pdb}_{lig_name}_{chain})
• rec_id: Structure to use as receptor
• tanimoto: Ligand similarity (0-1, higher = more similar)
• source:
  • "existing": Both structures in original covalent dataset. The receptor structure is in the bound-form with a covalent ligand.
  • "additional": The receptor structure is in the bound-form with non-covalent ligand

Usage: Dock ligand from lig_id structure into protein from rec_id structure

Installation

Requirements

• Python 3.8+
• MMseqs2 (for sequence clustering)
• OpenBabel (for molecular format conversion)

Setup

1. Clone the repository:

git clone https://github.com/yourusername/cdock_dataset.git
cd cdock_dataset

2. Install Python dependencies:

pip install -r requirements.txt

3. Install external tools:

MMseqs2 (sequence clustering):

# Ubuntu/Debian
sudo apt-get install mmseqs2

# macOS
brew install mmseqs2

# From source
git clone https://github.com/soedinglab/MMseqs2.git
cd MMseqs2
mkdir build && cd build
cmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=. ..
make && make install

OpenBabel (molecular format conversion):

# Ubuntu/Debian
sudo apt-get install openbabel

# macOS
brew install open-babel

# Conda
conda install -c conda-forge openbabel

Usage

1. Generate Covalent Test Set from PDB

Extract covalent ligands from PDB files:

python gen_testset.py \
  --input-list test_data/cov_rcsb_19950101_20250919.txt \
  --output test_data/cov_set_raw.json \
  --cif-dir /path/to/cif/files \
  --min-date 2018-01-01 \
  --max-resolution 2.5

Options:

• --input-list: Text file with PDB IDs (one per line) downloaded from RCSB PDB
• --output: Output JSON file path
• --cif-dir: Directory containing mmCIF files
• --min-date: Minimum deposition date (YYYY-MM-DD)
• --max-resolution: Maximum resolution in Angstroms

2. Filter by Binding Site Quality

Remove structures with poor binding site geometry:

python check_binding_site.py \
  --input test_data/cov_set_raw.json \
  --output test_data/cov_set_filtered.json \
  --cif-dir /path/to/cif/files \
  --cutoff 0.65

Options:

• --cutoff: Minimum contact ratio (default: 0.65)
  • Higher values = stricter quality filter
  • Recommended range: 0.6-0.75

3. Remove Train-Test Leakage

Filter test set to avoid similar proteins/ligands in training set:

python clustering.py remove-leakage \
  --test-fasta test_data/covset_seq.fasta \
  --train-fasta train_data/train_seq.fasta \
  --output test_data/cov_set_no_leak.json \
  --seq-threshold 0.3 \
  --lig-threshold 0.5

Options:

• --seq-threshold: Maximum sequence identity (default: 0.3)
• --lig-threshold: Maximum Tanimoto similarity (default: 0.5)

4. Remove Redundancy

Create non-redundant dataset by clustering:

python clustering.py remove-redundancy \
  --input-fasta test_data/covset_seq.fasta \
  --input-json test_data/cov_set_no_leak.json \
  --output test_data/cov_set_final.json \
  --seq-threshold 0.3 \
  --lig-threshold 0.5

Options:

• --seq-threshold: Sequence identity for clustering (default: 0.3)
• --lig-threshold: Ligand similarity for clustering (default: 0.5)

5. Generate Input Files for Docking

Prepare structure files for docking programs:

python gen_input.py \
  --input test_data/cov_set_final.json \
  --cif-dir /path/to/cif/files \
  --output-dir docking_inputs

This generates:

• Protein PDB/CIF files
• Ligand MOL2 files
• Configuration files for AutoDock4, DOCK6, etc.

Workflow

Complete Pipeline for Dataset Curation

# 1. Extract covalent complexes from PDB
python gen_testset.py \
  --input-list test_data/pdb_list.txt \
  --output test_data/cov_set_raw.json \
  --cif-dir cif_files/ \
  --min-date 2018-01-01

# 2. Filter by binding site quality
python check_binding_site.py \
  --input test_data/cov_set_raw.json \
  --output test_data/cov_set_filtered.json \
  --cif-dir cif_files/ \
  --cutoff 0.65

# 3. Remove train-test leakage
python clustering_refactored.py remove-leakage \
  --test-fasta test_data/covset_seq.fasta \
  --train-fasta train_data/train_seq.fasta \
  --output test_data/cov_set_no_leak.json

# 4. Remove redundancy
python clustering_refactored.py remove-redundancy \
  --input-fasta test_data/covset_seq.fasta \
  --input-json test_data/cov_set_no_leak.json \
  --output test_data/cov_set_final.json

# 5. Generate docking inputs
python gen_input.py \
  --input test_data/cov_set_final.json \
  --cif-dir cif_files/ \
  --output-dir docking_inputs/

Dataset Statistics

Covalent Test Set (cov_set.json)

• Total structures: 647 complexes
• Date range: 1995-2025
• Resolution: ≤ 2.5 Å
• Experiment types: X-ray crystallography, Cryo-EM
• Covalent residues: CYS (55%), SER (24%), LYS (18%), others (3%)
• Unique proteins: 346 (based on UniProt ID)

Quality Filters

Structural Quality

• X-ray or Cryo-EM, resolution ≤ 2.5 Å
• Single covalent bond per ligand
• Binding site contact ratio ≥ 0.65

Chemical Properties (Ligands)

• Heavy atom count: 8-50 atoms
• Molecular weight: ≤ 700 Da
• Rotatable bonds: ≤ 20
• cLogP: 0-7.5
• H-bond donors: ≤ 5
• H-bond acceptors: ≤ 10
• Polar surface area: ≤ 200 Ų
• Allowed elements: H, C, N, O, F, P, S, Cl, Br, I

Redundancy Removal

• Sequence identity: < 30%
• Ligand Tanimoto similarity: < 0.5
• Same protein + same ligand → keep best resolution

File Naming Conventions

Structure IDs

Format: {pdb}_{lig_name}_{lig_chain}

Examples:

• 5ydm_DUW_B: PDB 5ydm, ligand DUW, chain B
• 1b12_1PN_A: PDB 1b12, ligand 1PN, chain A

Generated Files

From gen_input.py:

• {pdb}_{lig}_{chain}_prot.pdb: Protein structure
• {pdb}_{lig}_{chain}_lig.mol2: Ligand structure
• {pdb}_{lig}_{chain}_lig_ca.mol2: Ligand + CA-CB of covalent residue (for GalaxyCDock input)
• {pdb}_{lig}_{chain}_comp.cif: Complete complex