Scan approved drugs for repurposing candidates against a disease

Given a disease name or EFO/MONDO ID, produce a ranked shortlist of approved or late-clinical drugs whose targets are genetically or mechanistically tied to the disease, with potency, indication, and interaction context attached to each candidate.

   
Problem class Knowledge synthesis
Subject areas Drug Repurposing and Discovery, Translational Medicine
Evidence level Proposed
Complexity Multi-tool harness
Availability Subscription required
Compute Laptop

Problem

The opening move of a repurposing project is to compress thousands of approved compounds into a handful worth following up. The data needed is well-known — disease-target evidence (genetics, expression, mouse KO, literature), known bioactivity of approved drugs against those targets, off-target and interaction liabilities, and the legal indication landscape — and lives across at least three databases (Open Targets, ChEMBL or PubChem, DrugBank). The cost is not the lookups; it is reconciling identifiers (ENSG vs UniProt vs ChEMBL target ID; DrugBank vs ChEMBL compound ID), filtering for clinical-stage molecules, and writing it up. Solved looks like: one prompt, a 10–20 candidate table with cited evidence per row, in under ten minutes of wall-clock.

  1. Install the three components. Open Targets and ChEMBL come from the same marketplace; DrugBank is a community MCP that needs the user-supplied DrugBank XML (a separate license).

    /plugin marketplace add anthropics/life-sciences
/plugin install open-targets@life-sciences
/plugin install chembl@life-sciences

    Then install DrugBank MCP per its catalog page (clone, npm run download:db with your DrugBank XML, register the stdio server). If your institution does not have a DrugBank licence, substitute the PubChem MCP — you lose mechanism and interaction queries but keep bioactivity and structure.

  2. Drive the scan with a single multi-step prompt. A minimal version:

    Find drug-repurposing candidates for idiopathic pulmonary fibrosis (EFO_0000768).

Use:
- open-targets: query target-disease associations for EFO_0000768
  ordered by overall association score; return the top 25 targets
  with their ENSG IDs, UniProt accessions, and tractability flags
  (small_molecule, antibody, clinical_compounds).
- chembl: for each of those 25 targets, call compound_search /
  get_bioactivity to find approved or late-clinical compounds
  (max_phase >= 3) with reported activity (IC50/Ki/Kd < 10 uM).
  Skip endogenous ligands and chemical probes.
- drugbank: for each surviving compound, pull indication,
  mechanism_of_action, and known drug-drug interactions.

Render as a Markdown table with one row per (target, compound) pair:
gene | drug | mechanism | approved indication | pChEMBL | OT
association score | OT tractability | DDI flags | source IDs.
Sort by (OT association score desc, pChEMBL desc).
Cite each row with the Open Targets target ID, ChEMBL ID, and
DrugBank ID.

  3. Read the table critically. A high Open-Targets score + an approved drug whose label is in a totally unrelated indication is the repurposing signal. Same-indication hits are not repurposing (they are confirmation). Filter out endogenous ligands (insulin, EGF, etc.) before showing the table to a clinician — they pass the bioactivity filter but are not deployable drugs.

  4. Spot-check the top three. For each candidate, paste the (target, drug) pair into a fresh Claude session and ask for the supporting literature (PubMed via the PubMed MCP if installed). If no human-evidence paper exists in the last 10 years, demote.

  5. Save the prompt as a slash command. Once the scan is right, parameterize on the EFO/MONDO ID — /repurpose-scan <efo_id> — and reuse for every new indication.

Why this assembly

Rung 3 of the simplicity ladder. The scan needs three heterogeneous evidence axes: disease-target ranking (Open Targets), quantitative bioactivity tied to approved compounds (ChEMBL), and indication / mechanism / interaction context for those compounds (DrugBank). No single MCP covers all three end-to-end at the granularity repurposing needs. Open Targets does include a knownDrugs block per target, but it surfaces clinical compounds without bioactivity values and without interaction data — that is why ChEMBL and DrugBank earn their seats. Rung 2 (Open Targets alone) under-resolves the candidate shortlist; rung 4 (a full autonomous system like Biomni) is overkill for what is fundamentally a ranked-join across three databases.

Availability

Subscription required, driven by DrugBank. The DrugBank XML is license-gated — academic licences are typically free but require institutional sign-off; commercial licences are paid. Open Targets and ChEMBL are CC0 / CC-BY-SA-3.0 with no auth. If you cannot get a DrugBank licence, swap in PubChem (Fully open substitution but you lose the curated mechanism-of-action and DDI fields). The Anthropic life-sciences marketplace itself is free.

Compute requirements

Laptop-sufficient. All three components are read-only API or local-SQLite lookups; the DrugBank stdio server uses ~50–100 MB RAM. The orchestration time is dominated by Claude’s tool-calling latency — expect 3–10 minutes for a 25-target × 5-compound scan. No GPU.

Evidence

Proposed. No published end-to-end benchmark of this exact three-MCP composition is known. The closest documented analogue is DeepDrug (Li et al., Scientific Reports 2025), which integrates DrugBank (v5.1.10), DrugCentral, ChEMBL (v31), and BindingDB into a signed directed heterogeneous biomedical graph for Alzheimer’s drug repurposing — confirming that the database combination and the join pattern (target evidence + bioactivity + approved-drug metadata) are the right primitives. DeepDrug returned a five-drug combination (tofacitinib, niraparib, baricitinib, empagliflozin, doxercalciferol) operating across 7,379 drug-target edges. Robin (Ghareeb et al., Nature 2026) shows that an LLM-agent system can identify a viable repurposing candidate end-to-end — ripasudil for dry age-related macular degeneration, validated in vitro with a 1.89-fold increase in RPE phagocytosis — though Robin’s component stack is FutureHouse-internal (PaperQA2 + Finch), not the open MCPs used here. On the LLM-validation side, Zunzunegui Sanz et al. (bioRxiv 2025-06-13) benchmarked four LLMs (GPT-4o, Claude-3, Gemini-2, DeepSeek) on a DREBIOP dataset of pathway-based drug-repurposing cases and reported significantly higher accuracy with structured prompts (p < 0.001) — supporting the structured-prompt approach in step 2. The exact Open-Targets + ChEMBL + DrugBank MCP composition has not been independently benchmarked.

Alternatives considered

  • Open Targets alone (rung 2). The Open Targets knownDrugs field on each target returns clinical compounds and trial phase already. Reach for this when you only need a candidate list — no potency comparison, no interaction screen. It is the first thing to try if you want to skip DrugBank licensing.
  • DrugBank-only by indication search. Useful when you already have a candidate drug and want to ask “what else does this hit”. Inverts the target-first flow; appropriate when polypharmacology is the question, not target-driven repurposing.
  • Biomni (rung 4). The Biomni paper (Huang et al. 2025) wires Open Targets, ChEMBL-class bioactivity, and a DrugBank-class drug-knowledge graph into one autonomous agent. Reach for Biomni when the scan is one step inside a larger autonomous loop (e.g., scan → hypothesis → bench experiment → re-scan). For a one-shot repurposing scan against a defined disease, the rung-3 toolbelt is more transparent and easier to audit per row.
  • Robin (rung 4). Reach for Robin specifically when wet-lab validation and an iterative dry-AMD-style closed loop are part of the scope. Robin is open source but heavier to set up than the three MCPs above.

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.