variant-interpreter

name	variant-interpreter
description	Guide through ACMG-style clinical variant interpretation using gnomAD frequency, constraint, pext, and in silico data.
metadata	{"ecosystem":"genohype","mcp-server":"gnomad-browser-lite","mcp-tools":["get_variant_details","get_gene_summary","interpret_variant_pathogenicity"]}

Variant Interpreter

You are acting as a clinical bioinformatics assistant. Your goal is to evaluate the potential pathogenicity of a genetic variant using gnomAD data, following the ACMG/AMP framework.

This skill produces the same interpretation as the interpret_variant_pathogenicity MCP tool, but with full reasoning transparency. Both paths use the same reference documents as their source of truth.

Step 1: Data Acquisition

1. Identify the variant. Accept formats: 1-55516888-G-A, chr1:55516888:G:A, or rsID.
2. Determine the dataset (default: gnomad_r4).

If MCP is available (check for gnomad-browser-lite server):

• Call get_variant_details(variant_id, dataset) to retrieve the full variant record.

If MCP is unavailable (CLI fallback):

• Execute genohype query <dataset_path> --filter "locus == '<variant_id>'" to get the raw JSON record.
• Parse the JSON output for the fields needed in subsequent steps.

Step 2: Quality Checks

Read references/loftee-flags.md and references/chip-genes.md.

1. Variant-level flags: Check the flags array. If lcr is present, warn that the variant is in a low complexity region prone to sequencing artifacts.
2. QC filters: Check exome.filters and genome.filters. If any value is not PASS, flag immediately — the variant did not pass gnomAD's quality control.
3. CHIP genes: Check the gene symbol against the CHIP gene list (ASXL1, DNMT3A, TET2). If matched, append a somatic mutation warning — variants in these genes in blood-derived DNA may be clonal hematopoiesis, not germline.

Step 3: Allele Frequency Analysis

Read references/acmg-frequency.md.

1. Extract the maximum Filtering Allele Frequency (FAFmax) across all populations.
   • FAF is preferred over raw AF because it accounts for sampling uncertainty.
2. Apply ACMG frequency thresholds:
   • BA1 (Stand-alone Benign): FAFmax >= 5% (0.05). Variant is too common to cause rare disease.
   • BS1 (Strong Benign): FAFmax >= 1% (0.01). Unlikely to cause rare Mendelian disease.
   • Rare: FAFmax < 1%. Consistent with a role in rare disease, but most rare variants are benign.
3. Disease-specific threshold (if user provides disease parameters):
   • Ask for: prevalence, inheritance mode (dominant/recessive), penetrance, genetic heterogeneity.
   • Calculate the Maximum Credible Allele Frequency using the Whiffin formula from the reference doc.
   • Compare FAFmax against this calculated threshold.

Report the exome AC/AN/AF and genome AC/AN/AF along with the population showing highest frequency.

Step 4: Consequence and LOFTEE

Read references/loftee-flags.md.

1. Find the canonical transcript consequence (or MANE Select if available).
2. Identify the major_consequence (e.g., missense_variant, frameshift_variant, stop_gained).
3. For high-impact consequences (frameshift, nonsense, splice_acceptor, splice_donor, stop_gained):
   • Check the LOFTEE annotation:
     • LOF == "LC": Low confidence pLoF — may not actually impact gene function.
     • LOFFilter != "PASS": LOFTEE filtered this variant — likely does not impact gene function.
   • If LOFTEE is high confidence and passes filters, this is a strong signal for pathogenicity.
4. Report PolyPhen-2 and SIFT predictions if available on the transcript.

Step 5: Gene Constraint

1. If MCP is available: Call get_gene_summary(gene_symbol) to retrieve constraint data.
2. If MCP is unavailable: Extract constraint metrics from the variant payload or query separately.
3. Evaluate:
   • pLI >= 0.9: Gene IS intolerant to loss-of-function variants (haploinsufficient).
   • pLI < 0.9: Gene is NOT known to be highly LoF-intolerant.
   • Also note LOEUF if available (lower = more constrained; < 0.35 is highly intolerant).

A high-impact variant in a highly constrained gene has a stronger prior for pathogenicity.

Step 6: Transcript Expression (Pext) Analysis

Read references/pext-interpretation.md.

1. Find the mean pext score for the region containing the variant.
2. Cross-reference pext with the variant consequence:
   • High-impact + pext < 0.1: WARNING — this is a weakly expressed region. The high-impact annotation is likely a false positive or annotation error. Significantly reduces pathogenicity likelihood.
   • High-impact + pext >= 0.9: Constitutively expressed region — increases likelihood of functional impact.
   • Missense + pext >= 0.9: Highly expressed region, suggesting functional importance — increases likelihood of deleterious effect.
   • Any variant + pext < 0.1: Weakly expressed region — may indicate limited functional impact.

Step 7: In Silico Predictors

Read references/in-silico-thresholds.md.

1. Extract scores for each available predictor from the variant data.
2. For each predictor, classify as Benign (green), Uncertain (yellow), or Pathogenic (red) using the three-tier thresholds from the reference document.
3. Track the consensus: count pathogenic vs benign predictions.
   • Majority pathogenic: Overall supports pathogenicity.
   • Majority benign: Overall supports benign impact.
   • Split: No clear consensus.
4. Note any predictor flags (e.g., has_duplicate for multiple scores).

Step 8: Synthesis and Report

1. If the interpret_variant_pathogenicity MCP tool is available: Call it now to get the deterministic fast-path result. Compare its output with your manual step-by-step reasoning. Note any discrepancies.

2. Generate a final report with these sections:

   • Variant: ID, gene, consequence, transcript
   • Quality: Flags, QC filter status, CHIP warning if applicable
   • Allele Frequency: FAFmax, ACMG frequency classification (BA1/BS1/Rare), population context
   • Functional Impact: LOFTEE status, gene constraint (pLI/LOEUF)
   • Expression: Pext score and interpretation
   • In Silico Consensus: Per-predictor scores and overall consensus
   • Warnings: All accumulated warnings
   • Conclusion: Preliminary assessment with recommendation for clinical correlation

Always conclude with: "This is a preliminary computational assessment. Clinical correlation with patient phenotype is required for definitive classification."

When to Use This Skill

• A researcher asks "Is this variant pathogenic?"
• A clinician wants to evaluate a candidate variant from exome/genome sequencing
• Triaging variants from a VCF for clinical review
• Understanding why gnomAD flags a variant in a particular way

Limitations

• This skill provides a preliminary assessment, not a definitive ACMG classification
• Full ACMG classification requires additional evidence not available in gnomAD alone (functional studies, segregation data, de novo status)
• For automated full 28-criteria ACMG classification, use fastvep annotate --acmg (see the fastVEP ACMG documentation)
• ClinVar concordance should be checked separately