Is Chembl Data Free For Machine Learning Data: Dataset Structure and Field Coverage

Every activity row ties a molecule to a target through a specific assay drawn from a source document, which is the shape most property- and activity-prediction work needs. Because the database is literature-derived, coverage runs from compounds documented in the 1940s through current publications, sourced globally across international journals and patents.

That structure supports the hit-triage workflow Datadory tags for the industry: pull bioactivity records for one target from the 24,527,044 activities, then cross-reference candidate structures against PubChem's ~179.5 million compounds to widen the search space. It equally supports target-side models, because the 18,552 targets arrive with assay provenance rather than bare identifiers - see the bioactivity data terminology for how these fields fit together.

One caveat belongs in any model card. Curation is manual, so the label surface reflects the measurements scientists chose to run and publish, not a uniform screen. That sampling bias is worth stating explicitly when you report performance.

Expect the counters to move, too. Every figure on this page - 2,921,148 molecules, 18,552 targets, 24,527,044 activities, 1,970,438 assays - came from the live API in August 2026 and will rise at the next quarterly drop, so quote them as of August 2026 rather than as permanent constants.