Genome Analysis Readout: Turn a Raw Variant Report into a Clinician-Ready Summary with Claude

The task

Cardiology and internal medicine clinicians increasingly get patients walking in with raw whole-genome or exome output — a VCF export, a lab PDF, or a direct-to-consumer dump — and ask, "what does this mean for me?" This workflow turns a messy variant list into a structured, plain-language readout you can review, correct, and paste into your visit note. It's prompted by a Stanford cardiologist's STAT First Opinion describing how fast — and how standards-free — AI genome analysis has become.

Before AI

Today you either (a) refer out to genetic counseling and wait weeks, or (b) manually look up each flagged variant in ClinVar, cross-reference ACMG classifications, check gene-disease validity in ClinGen, and hand-write a summary. For a report with 10–20 flagged variants, that's an hour-plus before you've even thought about the patient's phenotype. Most clinicians skip it and defer entirely.

The workflow

Important before you start: strip all PHI from the variant file before pasting. No name, MRN, DOB, address, or accession number. The input should be variants + demographics (age range, sex, self-reported ancestry) only. If your organization hasn't signed a BAA with your model vendor, treat the model as a non-covered tool and de-identify per HIPAA Safe Harbor. The STAT piece is explicit that we need standards to ensure we're ready for consumers using AI to analyze their genomes — until those exist, you own the guardrails.

Step 1 — Normalize and triage the variant list

Paste the de-identified variant text. This prompt forces a consistent internal format and flags anything the model can't confidently parse, which matters because older VCFs reference out-of-date genome builds.

You are assisting a board-certified clinician. The user will paste a de-identified variant report (VCF-style rows, a lab table, or a consumer genomics export) plus minimal demographics.

Do the following and nothing else:
1. Detect and state the reference genome build (GRCh37/hg19, GRCh38, or unknown). If unknown, say so and stop — do not guess coordinates.
2. Parse every variant into a table with columns: Gene | HGVS (c. and p. if available) | Zygosity | Reported classification (if any) | Source line.
3. In a separate list, flag any row you could not parse cleanly, with the reason (truncated, ambiguous notation, non-standard gene symbol, etc.).
4. Do not look up or infer pathogenicity yet. Do not add variants that are not in the input.

End with: "Parsed N variants, flagged M for review. Ready for clinical triage."

De-identified variant export — patient demographics: 46yo male, self-reported Northern European ancestry, referred for family history of early-onset cardiomyopathy (father, dx age 42).

Reference: GRCh38
Source: in-house clinical exome panel (cardio-150), returned variants only.

CHROM POS       GENE    HGVS_C             HGVS_P        ZYG   LAB_CALL
14    23412901  MYH7    c.1988G>A          p.Arg663His   het   Likely Pathogenic
11    47364249  MYBPC3  c.1504C>T          p.Arg502Trp   het   Pathogenic
1     201328373 TNNT2   c.421C>T           p.Arg141Trp   het   VUS
10    112572223 TCF7L2  c.-?               -             het   Risk allele (T2D GWAS)
19    11200235  LDLR    c.1646G>A          p.Gly549Asp   het   VUS
13    32914438  BRCA2   c.9976A>T          p.Lys3326Ter  het   Benign (per ClinVar)
X     31224699  DMD     c.?                -             hemi  Not reported — low coverage flag
2     47635523  MSH2    c.942+3A>T         splice        het   Likely Pathogenic

Step 2 — Clinical interpretation against the phenotype

Now the model does the real work: cross-referencing each parsed variant against the stated reason for referral, and separating actionable findings from noise.

Using the parsed table from the previous step and the patient's referral reason (family history of early-onset cardiomyopathy), produce a clinical interpretation with these sections:

A. PRIMARY FINDINGS — variants in genes with established gene-disease validity for the referral indication. For each: gene, variant, zygosity, lab classification, inheritance pattern of the associated condition, and a one-sentence clinical implication written for a physician (not a patient).

B. SECONDARY/INCIDENTAL FINDINGS — variants in ACMG SF v3.x reportable genes unrelated to the referral. Note whether patient consent for secondary findings should be confirmed before disclosure.

C. UNCERTAIN OR LOW-CONFIDENCE — VUS, low-coverage calls, variants in genes with limited evidence. Do not overstate significance.

D. NOT CLINICALLY ACTIONABLE — benign calls, common risk alleles with small effect sizes (e.g., GWAS T2D hits). One line each, then move on.

Rules:
- Do NOT invent allele frequencies, ClinVar star ratings, or paper citations. If you don't know, write "requires ClinVar/literature confirmation."
- If a lab call conflicts with your clinical reasoning, surface the conflict explicitly rather than resolving it silently.
- Flag any variant that warrants a genetic counseling referral before patient disclosure.

Step 3 — Build the clinician-ready readout and a patient-facing draft

Produce two deliverables based on the interpretation above.

DELIVERABLE 1 — Clinician note (for the chart, under "Genetics review"):
- 150-250 words
- Structured: Indication / Key findings / Recommended next steps / Items requiring confirmation
- Use clinical register. Include specific gene and variant notation.
- End with an explicit "AI-assisted interpretation; not independently verified against ClinVar/ClinGen on [DATE]. Review before signing." disclaimer.

DELIVERABLE 2 — Patient-facing explainer (for the after-visit summary):
- 8th-grade reading level
- Explain what was found, what it may mean for their heart, what it does NOT mean, and what the next step is (likely: referral to cardiovascular genetics)
- No probabilities or percentages unless they appeared in the source report
- Do not include incidental findings in the patient draft unless consent for secondary findings is confirmed — instead, add a note to the clinician that these need a separate conversation.

Output both in full, labeled clearly.

Gotchas

• Reference build drift. Older VCFs are hg19; most current annotation databases default to GRCh38. If Step 1 can't confirm the build, stop — coordinates reported against the wrong build are worse than no report. The STAT author flags this himself: his genome file is more than a decade old, was generated using an earlier version of the human reference genome, and lacks information available in newer sequencing formats.
• The model will hallucinate ClinVar entries if you let it. The prompts above explicitly forbid invented citations and star ratings. Still spot-check the primary findings against ClinVar and gene-disease validity against ClinGen before signing.
• Secondary findings consent. Dumping ACMG SF incidental findings into an after-visit summary without prior consent is both an ethical and a workflow failure. Step 3 segregates them — don't collapse that back together.
• This is not a diagnostic. The STAT piece exists because consumers are about to start using AI to analyze their whole genomes and there is no professional standard yet. Treat the output as a structured first pass that saves lookup time, not as a report you'd defend without a geneticist's eyes on it.
• PHI leakage. If your prompt accidentally includes a patient name or MRN pulled from a header row, you've just sent PHI to a third party. Grep your paste before you hit send.

Time saved

Rough estimate: ~45 minutes per patient readout. Manual triage of a 10–20 variant panel plus writing both a clinician note and a patient explainer typically runs 60–75 minutes; this workflow brings it to ~15–20 minutes of model time plus clinician review. That estimate assumes you already have a de-identification habit and are reviewing — not rubber-stamping — the output. If you're building a repeatable intake process for a genetics-adjacent clinic, the savings compound; if you see one of these a quarter, the setup cost may not pay off versus just referring out.

What we got when we ran it

We executed every prompt above, in order, against the sample input (model: claude-sonnet-4-6). Outputs below are real and unedited apart from length truncation.

Step 1 output

Step 1 — Reference Genome Build **GRCh38** — explicitly stated in the report header. --- ## Step 2 — Parsed Variant Table | # | Gene | HGVS (c.) | HGVS (p.) | Zygosity | Reported Classification | Source Line | |---|------|-----------|-----------|----------|------------------------|-------------| | 1 | MYH7 | c.1988G>A | p.Arg663His | Heterozygous | Likely Pathogenic | CHROM 14, POS 23412901 | | 2 | MYBPC3 | c.1504C>T | p.Arg502Trp | Heterozygous | Pathogenic | CHROM 11, POS 47364249 | | 3 | TNNT2 | c.421C>T | p.Arg141Trp | Heterozygous | VUS | CHROM 1, POS 201328373 | | 4 | TCF7L2 | c.-? | — | Heterozygous | Risk allele (T2D GWAS) | CHROM 10, POS 112572223 | | 5 | LDLR | c.1646G>A | p.Gly549Asp | Heterozygous | VUS | CHROM 19, POS 11200235 | | 6 | BRCA2 | c.9976A>T | p.Lys3326Ter | Heterozygous | Benign (per ClinVar) | CHROM 13, POS 32914438 | | 7 | DMD | c.? | — | Hemizygous | Not reported | CHROM X, POS 31224699 | | 8 | MSH2 | c.942+3A>T | splice (p. not provided) | Heterozygous | Likely Pathogenic | CHROM 2, POS 47635523 | --- ## Step 3 — Flagged Rows **Flag 1 — Row 4 (TCF7L2, c.-?)** - The c. notation is explicitly incomplete (`c.-?`); no nucleotide change is specified. No p. consequence is given. This variant cannot be characterized at the sequence level from the data provided. Additionally, TCF7L2 is a type 2 diabetes

…output truncated at 250 words.

Step 2 output

# Clinical Interpretation Report Patient: 46yo male | Northern European ancestry | Referral: family history of early-onset cardiomyopathy (father, dx age 42) Reference: GRCh38 | Panel: cardio-150 | Interpreted by: AI-assisted triage (requires clinician sign-off) > ⚠️ This output is a clinical decision-support aid. All findings must be reviewed and authorized by the responsible board-certified clinician and/or laboratory director before any patient communication or clinical action. --- ## A. PRIMARY FINDINGS Genes with established gene–disease validity for hypertrophic or dilated cardiomyopathy --- ### 1. MYH7 — c.1988G>A, p.Arg663His | Heterozygous | Lab: Likely Pathogenic - Gene–disease relationship: MYH7 (β-myosin heavy chain) has definitive gene–disease validity for hypertrophic cardiomyopathy (HCM) and is among the most frequently implicated genes; autosomal dominant inheritance. - Inheritance pattern: Autosomal dominant. - Clinical implication: A heterozygous likely pathogenic variant in MYH7 in a patient with a first-degree relative diagnosed with cardiomyopathy at age 42 is highly concordant with an autosomal dominant HCM pedigree and warrants expedited cardiac evaluation including echocardiography and cardiac MRI if not already performed. - Conflicts/caveats: None identified in the supplied data; confirmatory classification against current ClinVar evidence and functional literature is recommended before finalizing report. - 🔴 Genetic counseling referral indicated prior to cascade family testing disclosure. --- ### 2. MYBPC3 — c.1504C>T, p.Arg502Trp | Heterozygous | Lab: Pathogenic - Gene–disease relationship: MYBPC3 (cardiac myosin-binding protein C) has definitive gene–disease validity for HCM; autosomal dominant; the most common HCM-associated gene. - Inheritance pattern: Autosomal dominant. - Clinical implication: A heterozygous pathogenic variant

…output truncated at 250 words.

Step 3 output

# DELIVERABLE 1 — Clinician Chart Note Section: Genetics Review --- Indication: 46-year-old male presenting with family history of early-onset cardiomyopathy (father, diagnosed age 42). Cardio-150 clinical exome panel performed; variants below represent returned findings requiring clinical integration. Key Findings: Primary — Cardiomyopathy-related: Two sarcomere gene variants with pathogenic/likely pathogenic classifications were identified: (1) MYH7 c.1988G>A, p.Arg663His (het, Likely Pathogenic) and (2) MYBPC3 c.1504C>T, p.Arg502Trp (het, Pathogenic). Co-occurrence of pathogenic variants in two independent sarcomere genes constitutes a dual sarcomere genotype, associated with earlier onset and more severe HCM phenotype in published cohorts. A third variant, TNNT2 c.421C>T, p.Arg141Trp (het, VUS), was identified in a third sarcomere gene; clinical significance remains uncertain. The triple sarcomere variant constellation warrants Sanger confirmation of all three loci prior to finalization. Secondary findings (consent status unconfirmed — do not disclose to patient pending verification): MSH2 c.942+3A>T (het, Likely Pathogenic, splice) — Lynch syndrome gene, ACMG SF v3.x reportable. LDLR c.1646G>A, p.Gly549Asp (het, VUS) — familial hypercholesterolemia gene, ACMG SF v3.x reportable. DMD locus flagged for low coverage; no interpretable sequence-level result available. Recommended Next Steps: 1. Urgent referral to cardiovascular genetics for formal counseling; cascade testing planning for first-degree relatives. 2. Echocardiography and/or cardiac MRI if not recently performed. 3. Sanger sequencing confirmation of MYH7, MYBPC3, and TNNT2 variants. 4. Confirm informed consent for secondary findings before any disclosure of MSH2 or LDLR results; expedite Lynch syndrome referral (gastroenterology/oncology) if consent confirmed. 5. Discuss DMD low-coverage region with laboratory director; consider MLPA if phenotypically

…output truncated at 250 words.

---

This content is for informational purposes only and is not medical advice. AI tools used with patient data must meet your organization's HIPAA and privacy requirements.