Datasets:

williamTLmiller
/

bindingdb-chembl-binding-affinity

pretty_name: >-
  BindingDB + ChEMBL Binding Affinity Dataset: Ki, Kd, IC50, EC50 Measurements
  (Normalized)
license: cc-by-sa-3.0
language:
  - en
multilinguality:
  - monolingual
size_categories:
  - n<1K
task_categories:
  - tabular-regression
  - tabular-classification
annotations_creators:
  - found
language_creators:
  - found
source_datasets:
  - original
tags:
  - bindingdb
  - chembl
  - bioactivity
  - binding-affinity
  - drug-discovery
  - drug-target-interaction
  - protein-ligand
  - ki
  - kd
  - ic50
  - ec50
  - cheminformatics
  - chemistry
  - biology
  - medicinal-chemistry
  - tabular
  - matrixnorm
configs:
  - config_name: measurements
    data_files:
      - split: train
        path: bindingnorm-bindingdb-202511-chembl_ccbysa/measurements.csv
    default: true

BindingDB + ChEMBL Binding Affinity Dataset: Ki, Kd, IC50, EC50 Measurements (Normalized)

Release name: BindingDB ChEMBL-Sourced Slice — Normalized Dataset

Normalized binding-affinity measurements for the ChEMBL-derived portion of BindingDB (BindingDB's re-exported ChEMBL records). This repo holds a free toy sample of 1,000 measurements in one CSV file.

Free toy sample

This public repo is a deliberately tiny, CSV-only toy sample: 1,000 measurements in one table (no linked tables). It is meant for looking at the schema and running a pipeline once (a toy training run), not for training a real model. The full ChEMBL-sourced slice (1,576,007 measurements) is not in this repo; it is kept in the private repo williamTLmiller/bindingdb-chembl-binding-affinity-full (contact energea@energeatech.com). Because it derives from CC BY-SA 3.0 data, any copy of it that is shared stays under CC BY-SA 3.0, the same license as this sample, and the underlying records are also freely available from BindingDB and ChEMBL.

Measurements were picked by a fixed hash of id: the first 5 (in hash order) of every endpoint and qualifier combination (so each of the 12 combinations appears), then the rest in hash order. The sample is stable between releases, and its endpoint and qualifier mix differs slightly from the full slice.

License and attribution

The underlying measurement data derives from BindingDB's ChEMBL-sourced slice and is redistributed here under CC BY-SA 3.0, following BindingDB's license terms for this portion of its data:

BindingDB: https://www.bindingdb.org — The BindingDB project, University of California, San Diego / Skaggs School of Pharmacy.
ChEMBL: https://www.ebi.ac.uk/chembl/ — EMBL-EBI. The original bioactivity data originate in ChEMBL (itself CC BY-SA 3.0).

If you use this dataset, please cite BindingDB and ChEMBL per their respective citation requirements, and distribute derivatives under CC BY-SA 3.0. The full CC BY-SA 3.0 legal code is in LICENSE; NOTICE states the credits and what williamTLmiller added.

file	rows	what it holds
`bindingnorm-bindingdb-202511-chembl_ccbysa/measurements.csv`	1,000	One row per binding measurement: endpoint, value as reported, value in nM and mol/L, the -log10 affinity, qualifier and censoring fields.

endpoint	measurements
`IC50`	610
`Ki`	246
`EC50`	105
`Kd`	39

relation	censor_kind	measurements	meaning
`eq`	(empty)	787	exact value
`gt`	`lower_censored`	183	reported as `>` a bound: the true value is above `censor_bound`
`lt`	`upper_censored`	30	reported as `<` a bound: the true value is below `censor_bound`

Columns

id: stable measurement id.
source_key: the BindingDB Reactant_set_id of the source row; use it to look up the ligand, target, assay and citation in BindingDB (this sample spans 50,001,636 to 51,572,151).
complex_kind: protein_small_molecule for every row; evidence_kind: experimental for every row.
endpoint: Ki, Kd, IC50 or EC50.
raw_value, raw_unit: the value as BindingDB reports it (with any > or < prefix), in nM.
relation: eq, gt or lt, from that prefix. value, unit: the number in nM.
molar_value: the value in mol/L.
p_value: the -log10 affinity on the molar scale (pKi, pKd, pIC50 or pEC50), not a statistical p-value; p_relation is the qualifier on that scale (it flips, since a larger nM value is a smaller p_value).
censor_kind, censor_bound, p_censor_bound: for qualified values, the kind of bound and the bound in nM and on the p_value scale.
lower, upper, p_lower, p_upper (range bounds) and reported_p_value, reported_p_discrepancy (a source-reported -log10 value and any disagreement with the computed one) are empty in this slice.

In this sample p_value runs from −0.33 to 12.00 (median 6.16). One row carries BindingDB's placeholder value 2147483647 nM (p_value −0.33); it is kept as reported, so filter it before any modelling.

Full slice

The full ChEMBL-sourced slice (BindingDB release 2025-11) has 1,576,007 measurements: IC50 952,432, Ki 400,454, EC50 140,119 and Kd 83,002; 1,310,680 exact, 251,508 > and 13,819 <. In the full slice 4 rows carry the placeholder value 2147483647 nM, and 141 more carry even larger values (for example 3311311214826 nM and 99999999999999 nM); filter both before modelling. The full slice is not part of this repo and stays under CC BY-SA 3.0 (see above); for access, contact energea@energeatech.com.

Scope note

This package is BindingDB's representation of ChEMBL-derived measurements — not a native ChEMBL database dump. Assay context is limited to what BindingDB re-exports; richer assay entities live in native ChEMBL releases.

Quality checks

Run on the packaged file before upload: unique id, molar_value equals the nM value times 1e-9, p_value equals -log10(molar_value), censor_kind is set exactly for > and < values, CSV re-parse.