Profile a compound’s polypharmacology from ChEMBL bioactivity data

Hand Claude Code a SMILES or ChEMBL ID; get back a target-by-target table of measured potencies, the assay descriptions behind each number, and a flagged short-list of off-targets that warrant counter-screening before the chemotype goes any further.

   
Problem class Knowledge synthesis
Subject areas Chemistry, Drug Repurposing and Discovery
Evidence level Reported
Complexity One skill or MCP
Availability Fully open
Compute Laptop

Problem

A new lead’s intended target is the easy part — that comes free with the project. The hard part is the rest of the activity profile: every other protein the molecule was ever tested against, the assays behind those numbers, and whether the off-targets are dose-relevant. Medicinal chemists need this before they invest in analogs; pharmacologists need it before they pick an animal model; safety scientists need it before tox planning. The reference data is in ChEMBL — manually curated from medicinal-chemistry literature, with IC50 / Ki / Kd values, pChEMBL scores, and the assay descriptions — but pulling it cleanly for a single compound by name or SMILES and grouping it by target takes more REST calls than a chemist wants to write.

Solved looks like: one prompt, one table, one screen. Compound name in; target / standard relation / standard value / standard units / assay description / pChEMBL / ChEMBL document ID out, sorted by potency. Off-targets within 10-fold of the intended target flagged.

  1. Install the ChEMBL connector. Two options depending on the surface:

    • Claude Code — plugin marketplace: 
      /plugin marketplace add anthropics/life-sciences
/plugin install chembl@life-sciences

      Restart, then /mcp to confirm.

    • Claude.ai (Team / Enterprise or individual) — Settings → Connectors → ChEMBL → Connect.

  2. Resolve the compound, then pull every activity. A minimal prompt:

    Compound: imatinib (or paste a SMILES / ChEMBL ID).

Use the ChEMBL connector:
  1. compound_search to resolve to a ChEMBL molecule ID.
     Print the canonical SMILES, max phase, ATC code, and synonyms.
  2. get_bioactivity for that ChEMBL ID. Keep only rows with
     standard_type in {IC50, Ki, Kd, EC50}, standard_relation = "=",
     standard_units = "nM", and pchembl_value not null.
  3. Group by target ChEMBL ID and target pref_name. For each
     target, report: median pchembl, n measurements, best (lowest)
     nM value and its document_chembl_id, and the assay_description
     of the best measurement.

Sort by median pchembl descending. Emit as a markdown table.

  3. Flag the polypharmacology. Add a follow-up:

    For the intended target (ABL1 in the imatinib example),
list every other target whose best measured value is within
10-fold of the intended target's best value. For each off-target,
give the ChEMBL document_chembl_id and a one-sentence note on
whether the off-target is in a related family (e.g., another
tyrosine kinase) or a clearly unrelated protein.

    The 10-fold cut-off is the standard rule of thumb for “this is plausibly hit in vivo at the dose that hits the intended target”; tune to project context.

  4. Cross-reference mechanism if the compound is approved. ChEMBL get_mechanism returns curated MoA. Add:

    If max phase >= 3, call get_mechanism for this molecule and
print the action_type and target. Compare to the polypharmacology
table — flag any approved mechanism target that does NOT appear
in the bioactivity hits (sign that the curated MoA target may not
be the highest-potency target measured).

  5. Persist the table. Save the markdown plus the underlying JSON to compounds/<name>_profile_<date>.md and .json. ChEMBL refreshes quarterly; re-run any time a new release lands to catch added measurements.

Why this assembly

Rung 2 of the simplicity ladder. ChEMBL is the canonical source for measured bioactivity, and the ChEMBL connector wraps six tool calls — compound_search, target_search, get_bioactivity, get_mechanism, plus cross-reference and metadata — that map directly onto this question. Plain Claude Code with raw HTTP could hit the REST API but loses the curated, scoped tool surface and the typed responses. A multi-tool harness is unnecessary — every column on the desired output table comes from ChEMBL.

Availability

Fully open. The ChEMBL connector is GA in the anthropics/life-sciences marketplace; ChEMBL data is CC-BY-SA-3.0. No subscription, no institutional licence. Any current Claude plan supports the connector.

Compute requirements

Laptop. The full bioactivity profile for a well-studied drug (imatinib has ~6 000 ChEMBL measurements) returns in <30 seconds over the HTTP MCP transport; obscure compounds with <100 measurements return in <5 seconds. No local compute beyond the Claude client.

Evidence

Reported. Anthropic’s ChEMBL Connector tutorial walks through this exact assembly — compound resolution → bioactivity pull → target grouping — for medicinal-chemistry SAR work. The underlying ChEMBL bioactivity workflow is the de-facto reference for polypharmacology profiling: Mendez et al. (Nucleic Acids Res. 2019, 47:D930) describe the curation pipeline, and ChEMBL underpins the off-target benchmarks in DeepDrug (Li et al., Sci. Rep. 2025) and the ToolUniverse polypharmacology evaluations (Lin et al., bioRxiv 2024).

No peer-reviewed benchmark of “Claude + ChEMBL connector” vs a hand-written ChEMBL REST script is known — the productivity gain is the named tool surface, not new data. ChEMBL coverage is itself a known limitation: the activities are only what literature reports, with selection bias toward published positive results.

Alternatives considered

  • PubChem MCP alone. PubChem covers a wider compound universe but its assay data is heterogeneous and lacks the pChEMBL normalisation. Use it when a compound is not in ChEMBL at all; use ChEMBL when both have records.
  • Open Targets plugin. Open Targets is target-centric (disease → target → drug), not compound-centric — wrong axis for this question.
  • DrugBank MCP. DrugBank gives mechanism, indications, and interactions for approved drugs, not raw bioactivity at off-targets. Reach for it after this recipe, not instead of it.
  • ToolUniverse harness. ToolUniverse bundles ChEMBL alongside 600+ other tools. Worth it if the workflow needs more than polypharmacology — for the single-question case here, the focused ChEMBL connector is the lower rung.

See also

Sources

Tried this recipe?

Share feedback — what worked, what didn’t, what you’d change. The form opens with this recipe pre-selected and a link back to this page.