Interpret a clinical variant from a natural-language query

Hand Claude a variant in whatever notation the source happens to use; get back a single-page report with ClinVar significance, gnomAD frequency, CADD/PolyPhen/SIFT predictions, related ClinicalTrials.gov entries, and a short literature citation block.

   
Problem class Knowledge synthesis
Subject areas Translational Medicine, Molecular and Cellular Biology
Evidence level Proposed
Complexity One skill or MCP
Availability Fully open
Compute Laptop

Problem

Variant interpretation is mostly database choreography: convert the input notation to a canonical identifier, hit MyVariant.info for the consolidated annotation, read off ClinVar significance and gnomAD frequency, sanity-check with in-silico predictors, then pull relevant PubMed and ClinicalTrials.gov records. Every step is fast; the cost is the swivel-chairing across portals and the format-coercion when the input is “BRAF V600E” instead of an HGVS string. Solved looks like: paste any of {rsID, HGVS-c, HGVS-p, gene + protein change}, get one page of cited evidence, no manual format conversion.

  1. Install BioMCP. One MCP, one install:

    uv tool install biomcp-cli
claude mcp add --transport stdio biomcp -- biomcp run

  2. Prompt with the variant and the report template. A minimal version:

    Build a clinical-grade variant report for: BRAF V600E.

Use BioMCP. Specifically:
1. variant_searcher to locate the variant (try gene + hgvsp first;
   if that fails, fall back to genomic coordinates or rsID).
2. variant_getter for the full MyVariant.info payload.
3. Extract and tabulate:
     - ClinVar clinical_significance, review_status, last_evaluated
     - gnomAD_exome and gnomAD_genome allele frequency
     - CADD phred, PolyPhen-2 prediction, SIFT prediction
     - Associated conditions from ClinVar
     - Cancer context from cBioPortal (if returned)
4. pubmed_searcher for the 5 most relevant papers on this variant
   in the last 5 years; return PMIDs and one-line summaries.
5. trial_searcher with recruiting_status=OPEN and the variant as
   an other_term; return the top 3 trial NCT IDs.

Render as a single Markdown page. Cite every fact with its source
database (ClinVar accession, PMID, NCT ID).

  3. Handle the no-result case. If variant_searcher returns nothing, the input notation likely doesn’t match an indexed variant. Re-prompt to convert via gene + protein change (hgvsp="p.V600E") or, if that also fails, drop to genomic coordinates with the assembly explicit (chromosome="7", start=140453136, end=140453136, assembly="hg38").

  4. For oncology workflows, add the trial step explicitly. The trial search alone is the trial-matching recipe; chain it for variant-driven enrollment hunts.

Why this assembly

Rung 2. One MCP server covers every database the report needs (MyVariant.info, ClinVar via MyVariant.info, gnomAD, CADD, PolyPhen, ClinicalTrials.gov, PubMed). Adding a second component buys nothing. Claude Code alone (rung 1) cannot do this — it has no live database access and will confabulate gnomAD frequencies. A rung-3 toolbelt (separate MyVariant + ClinVar + gnomAD MCPs) is redundant since BioMCP already federates these via MyVariant.info. Rung 4 (Biomni) is overkill for a single-variant report.

Availability

Fully open. BioMCP is MIT-licensed. MyVariant.info, ClinVar, gnomAD, and ClinicalTrials.gov are public APIs. PubMed access is free; an NCBI API key raises the rate limit but is not required.

Compute requirements

Laptop-sufficient. All steps are read-only API calls; a single variant report typically completes in under a minute including five PubMed lookups. No GPU.

Evidence

Proposed. No published benchmark of the exact BioMCP-driven variant-report assembly is known. The closest documented analogue is MARRVEL-MCP (bioRxiv 2025-11-26), which equips LLMs with 39 tools spanning gene/variant utilities, pathogenicity databases (dbNSFP, ClinVar, gnomAD), and literature APIs, and benchmarks variant-interpretation accuracy on 45 expert-curated rare-disease tasks; a 20B-parameter model with the MCP reaches 95% accuracy versus 33% without tools. MARRVEL-MCP and BioMCP cover overlapping data sources (ClinVar via different transports, gnomAD, literature), so the MARRVEL-MCP result is the closest evidence that the assembly class works. Independent component evidence: MyVariant.info itself is the canonical aggregator behind dozens of published variant pipelines; CADD/PolyPhen/SIFT are standard in-silico predictors. The exact BioMCP-only composition has not been independently benchmarked.

Alternatives considered

  • MARRVEL-MCP directly. If the workflow is rare-Mendelian-disease focused and you need MARRVEL’s curated cross-species evidence, the MARRVEL-MCP server is purpose-built. It is not currently catalogued in this repo; until it is, BioMCP is the catalogued substitute.
  • Claude Code alone (rung 1). Insufficient — the model cannot fetch live ClinVar or gnomAD records.
  • Biomni (rung 4). Worthwhile only if variant interpretation is one step of a larger autonomous loop (e.g., variant → mechanism hypothesis → wet-lab experiment proposal). For a one-shot report, the rung-2 MCP is enough.

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.