scPDB-Enhanced: An expansion of scPDB including all experimentally observed binding sites

Overview

The scPDB is one of the most widely used datasets for training ligand binding site (LBS) predictors. However, it has a significant limitation: it includes only one ligand per structure, usually focusing on the primary orthosteric site and omitting other valid binding cavities.

To overcome this, we cross-referenced scPDB entries with AHoJ-DB to identify all ligands, and thus all binding sites, experimentally observed across all available conformations of a given protein. By aggregating these observations, we provide a ground truth that includes sites often missed in the original dataset.

Files Description

1. scPDB-sites.csv

A list of individual binding sites. Each row represents a single site identified in a specific protein structure. 

• Format: PDB_ID;Chain;Ligand_Name;Residue_Indices

• Example: 1uh5;B;NAD;104 105 106 110 111 112 131 134 167 168 169 170 215 ...

• Note: Ligand_Name=UNKNOWN denotes binding sites from the original scPDB dataset. 

The residue indices labelling is auth (i.e., auth_seq_id)

 2. scPDB-train.csv

The primary file for model training. This file merges all identified binding sites for each unique protein and maps them directly onto the primary sequence.

• Format: PDB_ID;Chain;Binding_Residues;Full_Sequence

• Note: Binding residues are listed as ResidueType + Index (e.g., G8) The residue indices use 0-based indexing, allowing for direct indexing into the sequence string (e.g., an index of 8 refers to the 9th position in the sequence).

• Example: 1uh5;B;G8 I9 G10 G14 Y15 ...;NEDICFIAGIGDTNGYGWGIAKEL...