Chembl MCP Server

What chembl-mcp-server does

chembl-mcp-server exposes ChEMBL’s curated relationships between compounds, protein targets, and measured bioactivity through MCP. It also covers drug pharmacology, including mechanisms of action, molecular targets, action types, first-approval years, and clinical indications across alternate drug forms.

The Biology & Bioinformatics server supports both direct lookups and multi-step research workflows. An agent can find a molecule, resolve a protein or gene to a ChEMBL target, retrieve measurements for either entity or their intersection, and inspect the assay behind a result.

How it works

Molecule searches accept names, synonyms, ChEMBL IDs, InChIKeys, or SMILES-based structure searches. Exact, similarity, and substructure modes are available; similarity results include Tanimoto percentages. Search results expose properties such as maximum clinical phase, molecular weight, AlogP, Lipinski rule-of-five violations, QED, and parent compound relationships.

Target searches resolve protein names, gene symbols, or UniProt accessions to ChEMBL target IDs. Those IDs can then be passed to the bioactivity tool. Bioactivity queries may filter by molecule, target, compound-target pair, standard type, minimum pChEMBL value, assay type, organism, or variant. Measurements are ranked by pChEMBL value when applicable, and alternate salts or forms can be included for a molecule family.

Large bioactivity responses can be staged on a canvas. Read-only SQL queries can then rank, group, deduplicate, or aggregate the staged rows. Resources at chembl://molecule/{chemblId} and chembl://target/{chemblId} provide molecule and target records directly.

Setup and configuration

The server supports STDIO and Streamable HTTP. It can be installed through the npm package @cyanheads/chembl-mcp-server, including an npx-based MCP configuration, or run from the published Docker image. A hosted endpoint is available at https://chembl.caseyjhand.com/mcp.

The chembl_dataframe_drop tool is opt-in. Set CHEMBL_DATAFRAME_DROP_ENABLED=true to expose it; otherwise staged tables are reclaimed through their time-to-live behavior and the tool is omitted from the tool list.

Tools and capabilities

• chembl_search_molecules searches by text or identifier and performs exact, similarity, or substructure searches from SMILES.
• chembl_search_targets maps protein names, gene symbols, and UniProt accessions to ChEMBL target IDs.
• chembl_get_bioactivities retrieves and filters molecule, target, or pairwise measurements, including alternate molecular forms.
• chembl_get_drug_info returns mechanisms, targets, action types, approval year, and indications for drugs and their forms.
• chembl_get_assay provides assay type, target, organism, and ChEMBL’s confidence score for a bioactivity record.
• chembl_dataframe_describe lists staged tables and columns, while chembl_dataframe_query runs read-only SQL over staged bioactivity data.
• chembl_dataframe_drop removes a named staged table when explicitly enabled.

Limitations and notes

pChEMBL rankings are comparable only within the same standard type. Standard-type matching is case-sensitive, so values such as Kd and KD may represent separate ChEMBL records. Variant filters distinguish wild-type measurements from recorded mutations and match complete recorded variant strings.

Name searches stop counting and paging at 10,000 matches, while substructure searches cap at 30,000. A capped result must be narrowed to reach a smaller result set. Structure searches can fail with an invalid-structure response when ChEMBL cannot parse the supplied SMILES. Bioactivity results use cursors for pagination, and cursor requests must retain the filters that created them.