Example: Variant Triage — BRCA1 c.5266dupC

Status: Illustrative. This transcript shows the shape of a real Claude Desktop session against alphafold-sovereign-mcp for a well-characterised pathogenic BRCA1 variant. Output fields are consistent with what the server's generate_variant_clinical_report tool emits when its upstream clients (Ensembl VEP, AlphaMissense, ClinVar, gnomAD, Open Targets, UniProt) return data we have unit tests for. End-to-end validation against the live upstreams is on the v1.2.0 roadmap (see STATUS.md) — until then, treat the numbers below as schematic, not as a clinical reference.

Setup

// claude_desktop_config.json
{
  "mcpServers": {
    "alphafold-sovereign": {
      "command": "alphafold-sovereign-mcp",
      "args": []
    }
  }
}

After installing from source (see top-level README) and restarting Claude Desktop, the server's tools become available in the conversation.

User prompt

Triage the BRCA1 c.5266dupC variant for me. Tell me whether it's pathogenic, what the gnomAD frequency is, the AlphaMissense score, the ClinVar status, and the predicted protein consequence.

What the model calls

// MCP tool call
{
  "tool": "generate_variant_clinical_report",
  "params": {
    "gene_symbol": "BRCA1",
    "hgvs_c": "c.5266dupC",
    "include_alphafold_structure": true,
    "include_clinvar": true,
    "include_gnomad": true,
    "include_alphamissense": true
  }
}

What the server returns (schema-accurate; values illustrative)

{
  "input": {
    "gene_symbol": "BRCA1",
    "hgvs_c": "c.5266dupC",
    "transcript": "NM_007294.4"
  },
  "ensembl_vep": {
    "predicted_consequence": "frameshift_variant",
    "impact": "HIGH",
    "canonical_transcript": "ENST00000357654",
    "protein_consequence": "p.Gln1756ProfsTer74",
    "exon": "20/23"
  },
  "alphamissense": {
    "score": null,
    "note": "AlphaMissense scores apply to missense variants only; this is a frameshift."
  },
  "clinvar": {
    "variation_id": "17661",
    "clinical_significance": "Pathogenic",
    "review_status": "criteria provided, multiple submitters, no conflicts",
    "last_evaluated": "2024-08-21",
    "n_submissions": 31,
    "conditions": [
      {"name": "Hereditary breast and ovarian cancer syndrome", "id": "MONDO:0019303"},
      {"name": "Familial cancer of breast", "id": "MONDO:0017180"}
    ]
  },
  "gnomad": {
    "allele_frequency": 0.0000142,
    "allele_count": 4,
    "allele_number": 281618,
    "homozygotes": 0,
    "filter": "PASS",
    "popmax_population": "nfe",
    "popmax_af": 0.0000301
  },
  "alphafold": {
    "uniprot_id": "P38398",
    "model_url": "https://alphafold.ebi.ac.uk/files/AF-P38398-F1-model_v4.pdb",
    "residue": 1756,
    "plddt_at_residue": 38.4,
    "plddt_note": "Low pLDDT (<50) — residue lies in a disordered region; structural inference unreliable.",
    "pae_neighbourhood_p95_residues_within_8A": [1750, 1751, 1752, 1755, 1757, 1758, 1761]
  },
  "acmg_evidence_draft": {
    "PVS1": "applied",
    "PS1": "not applied",
    "PM2_supporting": "applied (gnomAD AF < 0.0001)",
    "PP5": "not applied (deprecated)",
    "draft_classification": "Pathogenic",
    "draft_classification_note": "Auto-derived ACMG/AMP criteria draft. NOT independently validated by a clinical geneticist. See LIMITATIONS.md L1."
  },
  "sources_cited": [
    "Ensembl VEP REST",
    "ClinVar VCV00017661",
    "gnomAD v4.1.0",
    "AlphaFold DB v4 (P38398)"
  ],
  "report_generated_at": "2026-05-11T14:23:47Z"
}

What Claude says back (paraphrased)

BRCA1 c.5266dupC is a well-known pathogenic frameshift variant. Here is the triage summary:

• Predicted effect: frameshift at codon 1756 (p.Gln1756ProfsTer74), high-impact.
• ClinVar: classified Pathogenic (Variation ID 17661), with 31 submissions and no conflicts (last reviewed August 2024).
• gnomAD v4.1.0: extremely rare (allele frequency 1.42×10⁻⁵ globally; popmax 3.0×10⁻⁵ in non-Finnish European). No homozygotes.
• AlphaMissense: not applicable — the variant is a frameshift, not a missense.
• AlphaFold structural context: residue 1756 of BRCA1 (P38398) sits in a low-confidence region (pLDDT 38.4 of 100). Structural inference there is unreliable, which is expected — the C-terminal region is partly disordered. The variant truncates the BRCT domain.
• Auto-derived ACMG: PVS1 + PM2_supporting → Pathogenic — consistent with the ClinVar consensus.
• ⚠ Disclaimer: The ACMG draft surfaced by this server has not been validated by an independent clinical geneticist. For clinical reporting, re-derive each criterion from the cited primary sources.

What the same call looks like in --self-test mode

alphafold-sovereign --self-test runs an offline, deterministic subset of this flow against a built-in BRCA1:c.5266dupC fixture that does not require live network access. It exercises only the pure functions of the server (HGVS parsing, ACMG criterion mapping, report assembly), and prints PASS if the report shape and the deterministic ACMG draft match the golden file.

Reproducing this example for real

1. Install: see top-level README.md.
2. Configure Claude Desktop as shown above.
3. Restart Claude Desktop.
4. Paste the user prompt at the top of this document.

Files

• transcript.jsonl — machine-readable MCP exchange (one line per request / response / tool call / tool result).
• expected_output_schema.json — JSON Schema for the report shape so regressions are catchable in CI.