chembl

Query the ChEMBL database web services for measured bioactivity data, compound records and calculated properties, targets, assays, mechanisms of action, drug indications and warnings. Use this skill t

SKILL.md

GitHub

ChEMBL

ChEMBL is the curated database of measured bioactivity from the medicinal-chemistry literature — about 24.5 million activity values over 2.9 million compounds and 18,500 targets. It is where you find out what has already been made against a target, how potent it was, and how much of that is trustworthy.

Base URL: https://www.ebi.ac.uk/chembl/api/data — REST, no key. Docs: interactive schema · ChEMBL home Checked against: ChEMBL_37, released 2026-05-01.

Read references/api-reference.md before writing a request by hand, references/data-curation.md before modelling anything you pull, and references/entity-fields.md when you need to know what a field means or what type it really is.

The one thing to get right

Raw ChEMBL rows are not a dataset. activity.json?target_chembl_id=CHEMBL203 returns everything ever published against that id: censored > values, rows ChEMBL itself flags as wrong, assays whose target assignment is a guess, mutant-protein assays, and the same compound measured eight times across four papers. Modelling that directly is the most common way to produce a QSAR model that scores well and predicts nothing.

target_activities.py applies the curation and prints the attrition:

python skills/chembl/scripts/target_activities.py --uniprot P00533 \
    --standard-type Ki --min-confidence 8 --out egfr_ki.tsv
# resolved P00533 -> CHEMBL203 (Epidermal growth factor receptor, SINGLE PROTEIN); 15 other ChEMBL target(s) share this component
# CHEMBL203 Ki assay_type=B: 538 activity rows
# fetching confidence scores for 78 assays
# input rows: 538
# dropped 23: potential_duplicate flag
# kept rows: 515
# warning: 78 molecule(s) have replicate pChEMBL values spanning >= 1.0 log units -- review before modelling
# wrote 321 rows to egfr_ki.tsv
molecule_chembl_id  pchembl_median  pchembl_spread  n_measurements  n_documents  inconsistent
CHEMBL4100860       10.26           0.52            2               1            false
CHEMBL3883534       9.85            0               1               1            false
CHEMBL3758502       9.22            1.00            5               1            true

515 measurements collapse to 321 molecules, 78 of which carry replicates that disagree by more than a log unit. That last number is the one people skip.

The curation, and why each step exists:

Step Why
One SINGLE PROTEIN target A UniProt accession maps to family and complex targets too; pooling them mixes "inhibits EGFR" with "inhibits something in that family"
One standard_type IC50, Ki, Kd, EC50 are different physical quantities
standard_relation = '=' A > row means "inactive at the top dose", not a measurement
Drop data_validity_comment ChEMBL's own flag for implausible or mis-converted values
Require pchembl_value The one column where unit conversion is already done and checked
confidence_score >= 8 Below 5, the target assignment is inferred rather than stated
Median per molecule + spread Replicates disagree; a compound with pIC50 5.1 and 8.9 has no median worth using

Add --list-targets to see the target choices before committing, and --raw for one row per measurement instead of per molecule.

Compound lookup

python skills/chembl/scripts/compound_lookup.py id CHEMBL941 CHEMBL939
python skills/chembl/scripts/compound_lookup.py name imatinib
python skills/chembl/scripts/compound_lookup.py smiles "CC(=O)Oc1ccccc1C(=O)O"
python skills/chembl/scripts/compound_lookup.py inchikey BSYNRYMUTXBXSQ      # skeleton only
python skills/chembl/scripts/compound_lookup.py similar "CC(=O)Oc1ccccc1C(=O)O" --threshold 80
python skills/chembl/scripts/compound_lookup.py substructure "c1ccc2c(c1)ncnc2"
python skills/chembl/scripts/compound_lookup.py mechanism --target CHEMBL203

Structure searches put the query in the URL path, so a SMILES with #, +, /, or \ must be percent-encoded — unencoded, # truncates the URL at the fragment marker and you get a 404 or the wrong molecule. The script encodes for you; a hand-written curl will not.

Exact lookup defaults to flexmatch, which is salt-, charge-, and tautomer-tolerant — usually what "is this compound in ChEMBL?" means. --exact requires a byte-identical canonical SMILES and will miss the parent of a hydrochloride salt.

Anything else

python skills/chembl/scripts/chembl_query.py endpoints
python skills/chembl/scripts/chembl_query.py status
python skills/chembl/scripts/chembl_query.py count activity --filter target_chembl_id=CHEMBL203
python skills/chembl/scripts/chembl_query.py fetch drug_warning \
    --filter molecule_chembl_id=CHEMBL941
python skills/chembl/scripts/chembl_query.py fetch molecule \
    --filter molecule_properties__full_mwt__lte=300 --filter max_phase=4 \
    --limit 500 --out small_approved.tsv

Filters use Django-style lookups: field__gte, field__in, field__isnull, field__icontains, related__field. Multiple filters AND together; there is no OR across fields.

Four ways this API fails quietly

  1. limit caps at 1000 without saying so. Ask for 2000, get 1000, with page_meta.limit == 1000 and no error. Follow page_meta.next; the bundled paged() does.
  2. A lookup written with one underscore is read as a field name. pchembl_value_gte=8 filters nothing and returns the whole table. parse_filters rejects that spelling.
  3. Numbers are strings. "standard_value": "41.0", "max_phase": "4.0". Sorting as text puts "9" above "41". Convert on ingest.
  4. molecule_structures and molecule_properties are null for biologics (structure_type SEQ). Guard the access or every antibody in your set crashes the loop.

Also current as of ChEMBL_37: the ChemAxon-derived properties (cx_logp, cx_logd, cx_most_apka, molecular_species, hba_lipinski) no longer exist. Filtering on one now returns 400. Compute logD and pKa yourself — rowan for macro-pKa, rdkit for approximations.

When to stop using the API

Above roughly 10⁵ activities, page the FTP dump instead: https://ftp.ebi.ac.uk/pub/databases/chembl/ChEMBLdb/latest/ ships Postgres, MySQL, SQLite, and SDF. A whole-database question ("every kinase inhibitor under 400 Da with a measured Ki") is one SQL query locally and tens of thousands of HTTP requests remotely.

The official chembl_webresource_client Python package is the other route. It is convenient, but it hides paging behaviour and caches across releases; the scripts here stay standard-library only so their behaviour is inspectable.

Composing with the rest of the bundle

  • open-targets — which targets are worth pulling chemistry for; its clinical_precedence evidence names the ChEMBL drugs.
  • medchem / rdkit / datamol — standardise, deduplicate, and triage what you pulled. Do this before modelling: the same compound appears under several ChEMBL ids.
  • pytdc — scaffold splits. A random split on ChEMBL data reports a fantasy R², because congeneric series from one paper land on both sides.
  • molfeat / deepchem — featurisation and model fitting once the set is clean.
  • uniprot-rcsb — the structure to dock the actives into.
  • chemical-space — ChEMBL is what has been measured; ZINC-22 and Enamine REAL are what can be bought. Different questions, and a hit list needs both.
  • patent-landscape — SureChEMBL ids cross-reference through UniChem, joining measured bioactivity to patent chemistry.
  • clinicaltrials / openfda — what happened to these compounds after the assay.

For bioassay coverage far wider than ChEMBL's curated set, PubChem's PUG-REST is the complement — many more assays, much less curation, 5 requests/second and a 30-second timeout.

Reporting results honestly

Say the release (ChEMBL_37), the target id and type, the measurement type, and the filters you applied. Report how many rows you started from and how many survived — a pipeline that loses nothing has not filtered. Absence of data in ChEMBL means nobody published it in an indexed journal, not that a compound is inactive.