---
license: mit
tags:
- binding-affinity
- biology
- chemistry
pretty_name: Binding Affinity
configs:
- config_name: default
  data_files:
  - split: train
    path: "affinity-data-combined.parquet"
  - split: davis
    path: "davis.parquet"
  - split: davis_filtered
    path: "davis-filtered.parquet"
  - split: kiba
    path: "kiba.parquet"
  - split: pdbbind_2020_combined
    path: "pdbbind-2020-combined.parquet"
  - split: pdbbind_2020_refined
    path: "pdbbind-2020-refined.parquet"
  - split: bindingdb_ic50
    path: "bindingdb-ic50.parquet"
  - split: bindingdb_ki
    path: "bindingdb-ki.parquet"
  - split: bindingdb_kd_filtered
    path: "bindingdb-kd-filtered.parquet"
  - split: bindingdb_kd
    path: "bindingdb-kd.parquet"
  - split: glaser
    path: "glaser.parquet"
---

# Binding Affinity Dataset

## Overview

This dataset is a comprehensive collection of protein-ligand binding affinity data, compiled from multiple sources. The dataset is structured with multiple splits, each corresponding to a specific source:

- BindingDB IC50 Split
- BindingDB Kd Split
- BindingDB Kd Filtered Split
- BindingDB Ki Split
- Davis Split
- Davis Filtered Split
- KIBA Split
- PDBBind 2020 Combined Split
- PDBBind 2020 Refined Split
- Glaser Split

In addition to these source-specific splits, a main training split is provided that combines and aggregates data from all these sources.

## Training Dataset Composition

The training split is a comprehensive aggregation of multiple molecular binding datasets:

- Davis-filtered dataset
- PDBBind 2020 Combined dataset
- BindingDB IC50 dataset
- BindingDB Ki dataset
- BindingDB Kd Filtered dataset
- Glaser dataset

## Preprocessing Steps

1. **Dataset Merging**: All specified datasets were combined into a single dataset.
2. **Duplicate Removal**: Duplicate entries were dropped to ensure data uniqueness.
3. **Binding Affinity Normalization**: 
   - Entries with a binding affinity of 5 were reduced
   - For duplicate protein-ligand pairs, the mean binding affinity was calculated

## Data Sources

| Dataset | Source | Notes |
|---------|--------|-------|
| bindingdb_ic50.parquet | [TDC Python Package](https://tdcommons.ai/) | Therapeutic Data Commons |
| bindingdb_kd.parquet | [TDC Python Package](https://tdcommons.ai/) | Therapeutic Data Commons |
| bindingdb_kd_filtered.parquet | Manually Filtered | See `standardize_data.ipynb` |
| bindingdb_ki.parquet | [TDC Python Package](https://tdcommons.ai/) | Therapeutic Data Commons |
| davis.parquet | [TDC Python Package](https://tdcommons.ai/) | Therapeutic Data Commons |
| davis_filtered.parquet | [Kaggle Dataset](https://www.kaggle.com/datasets/christang0002/davis-and-kiba) | Filtered Davis dataset |
| kiba.parquet | [TDC Python Package](https://tdcommons.ai/) | Therapeutic Data Commons |
| pdbbind_2020_combined.parquet | [PDBBind](https://www.pdbbind.org.cn/) | Combined PDBBind 2020 dataset |
| pdbbind_2020_refined.parquet | [PDBBind](https://www.pdbbind.org.cn/) | Refined PDBBind 2020 dataset |
| glaser.parquet | [HuggingFace Dataset](https://huggingface.co/datasets/jglaser/binding_affinity) | Glaser binding affinity dataset |

## Dataset Columns

| Column | Description |
|--------|-------------|
| `seq` | Protein sequence |
| `smiles_can` | Canonical SMILES representation of the ligand |
| `affinity_uM` | Binding affinity in micromolar (µM) concentration |
| `neg_log10_affinityM` | Negative logarithm (base 10) of the affinity in molar concentration |
| `affinity_norm` | Normalized binding affinity |
| `affinity_mean` | Mean binding affinity for duplicate protein-ligand pairs |
| `affinity_std` | Standard deviation of binding affinity for duplicate protein-ligand pairs |