Skip to content

ACMG Variant Classifier

Track: D - Clinical Difficulty: Advanced Time estimate: 2-4 hours

What You'll Build

A ClawBio skill that takes variant information (gene, consequence, population frequency, functional data) and applies the ACMG/AMP 2015 guidelines to produce a five-tier classification: Pathogenic, Likely Pathogenic, Variant of Uncertain Significance (VUS), Likely Benign, or Benign. Each classification includes the specific evidence codes that support it.

Why This Matters

ACMG classification is the international standard for clinical variant interpretation, but manually applying all 28 criteria is time-consuming and error-prone. Automating the framework ensures consistent, auditable classifications that meet clinical-grade standards.

Inputs and Outputs

Input: JSON or dict with fields: gene, hgvs_c, hgvs_p, consequence, gnomad_af, clinvar_entries, functional_assay_result, inheritance_pattern, computational_predictions Output: Classification (Pathogenic/LP/VUS/LB/Benign), list of triggered evidence codes (e.g. PVS1, PM2, PP3), and a plain-language explanation of the reasoning

Key APIs / Data Sources

Getting Started

  1. Create your skill folder: skills/acmg-classifier/
  2. Implement the 28 evidence criteria as individual functions that return True/False
  3. Build the combining logic: PVS1 + PM1 = Likely Pathogenic; 2x Strong = Pathogenic; etc.
  4. Wire up automatic checks: population frequency for BA1/BS1/PM2, consequence type for PVS1, computational tools for PP3/BP4
  5. Return the final classification with all triggered codes and explanations

Domain Decisions for SKILL.md

  • BA1 stand-alone benign threshold: allele frequency > 0.05 (5%) in any gnomAD population
  • BS1 benign strong threshold: allele frequency > 0.01 (1%) for dominant conditions
  • PM2 supporting threshold: absent or below 0.0001 in gnomAD
  • PVS1 applies to null variants (nonsense, frameshift, canonical splice) in genes where loss of function is a known disease mechanism
  • Use ClinGen-curated gene-specific rule modifications where available
  • Computational predictions (PP3/BP4): require consensus from at least 3 of 5 tools (REVEL, CADD, MetaSVM, BayesDel, GERP)

Demo Data

Test with well-characterised variants: BRCA1 c.68_69delAG (Pathogenic, founder mutation), TP53 R175H (Pathogenic, hotspot), and a common benign polymorphism like rs1801133 (MTHFR C677T, gnomAD AF 0.33). Include at least one VUS to demonstrate intermediate classification.

Stretch Goals

  • Add ClinGen gene-specific rule modifications for BRCA1, TP53, RASopathy genes
  • Support automated gnomAD and ClinVar lookups (not just pre-supplied data)
  • Generate a visual evidence summary showing each criterion's status
  • Implement the 2023 ClinGen quantitative framework for PP3/BP4 (calibrated REVEL thresholds)