Ty-Perez/LBA

Overview

The LBA (Ligand Binding Affinity) dataset is designed for predicting the binding affinity between ligands and proteins based on co-crystallized protein-ligand complex structures. This dataset enables machine learning models to learn structure-activity relationships for drug discovery applications.

Task: Regression - predict experimentally measured binding affinity as pK (defined as -log(Ki) or -log(Kd))

Size: 4,463 protein-ligand complexes

LBA: Ligand Binding Affinity Dataset - Source Descirption

In this task, we predict the binding affinity of ligands to their corresponding proteins based on the co-crystallized structure of the protein-ligand complex. We predict experimentally measured binding affinity as pK, defined as -log(Ki) or -log(Kd), depending on which measurement is available.

We derive crystal structures and ligand binding data from PDBBind (Wang et al., 2004), a widely-used curated database of protein-ligand complexes with experimental affinities derived from literature. We use the 2019 update of the so-called "refined set", a subset of complexes selected based on the quality of the structures and the affinity data. After filtering ligands which could not be read by RDKit due to invalid bonding data, our final dataset consists of 4,463 complexes.

Dataset Structure

The dataset is organized with the following directory structure:

dataset/
├── data/                    # Processed .pt files
│   ├── {pdb_id}_protein.pt  # Full protein structure
│   ├── {pdb_id}_pocket.pt   # Binding pocket (within 6Å of ligand)
│   └── {pdb_id}_ligand.pt   # Ligand structure with bonds
├── summary.csv              # Dataset metadata and file mappings
├── seq-id-30.json          # 30% sequence identity splits
└── seq-id-60.json          # 60% sequence identity splits

Data Splits

Two splitting strategies are provided to prevent data leakage:

• seq-id-30: No proteins with >30% sequence identity in the same split
• seq-id-60: No proteins with >60% sequence identity in the same split

Each split contains:

• train: Training set
• val: Validation set
• test: Test set

Data Fields

Protein/Pocke/Ligand Files

Each protein/pocket .pt file contains:

• pos (torch.Tensor): 3D atomic coordinates [N, 3]
• z (torch.Tensor): Atomic numbers [N]
• res_name (list): Three-letter amino acid codes
• res_idx (torch.Tensor): Residue indices [N]
• bfactor (torch.Tensor): B-factor values (thermal motion) [N]
• is_alpha_carbon (torch.Tensor): Boolean mask for CA atoms [N]
• chain (list): Chain identifiers
• pdb_id (str): PDB identifier

Ligand Files

Each ligand .pt file contains all protein/pocket fields plus:

• edge_index_1d (torch.Tensor): Bond connectivity [2, E]
• edge_attr (torch.Tensor): Bond types (1.0=single, 2.0=double, 3.0=triple, 1.5=aromatic)
• smiles (str): SMILES representation
• neglog_aff (float): Target binding affinity (-log(Ki/Kd))

Usage

Loading with the Dataset Class

from lba_dataset import LBADataset_base

# Load dataset
dataset = LBADataset_base(
    split='train',
    task_name='seq-id-30', 
    output=['pocket', 'ligand', 'neglog_aff']
)

# Access samples
pocket, ligand, affinity = dataset[0]
print(f"Pocket atoms: {pocket['pos'].shape[0]}")
print(f"Ligand atoms: {ligand['pos'].shape[0]}")
print(f"Binding affinity: {affinity}")

Available Outputs

Configure what data to load with the output parameter:

• 'protein': Full protein structure (large files)
• 'pocket': Binding pocket only (recommended)
• 'ligand': Ligand structure with bonds
• 'neglog_aff': Binding affinity target
• 'smiles': Ligand SMILES string

Dataset Statistics

• Total complexes: 4,463
• Binding affinity range: Varies (pK values)
• Structure quality: High (PDBBind refined set)
• Ligand diversity: Filtered for RDKit compatibility
• Average pocket size: ~150-300 atoms
• Average ligand size: ~20-50 atoms

Data Processing

For the original raw data and processing scripts, see the source repository here.

https://zenodo.org/records/4914718