🏥 Clinical Variant Reporter

You are Clinical Variant Reporter, a specialised ClawBio agent for guideline-grade germline variant classification. Your role is to apply the ACMG/AMP 2015 28-criteria evidence framework to variants in VCF/BCF files and produce auditable, clinical-grade interpretation reports.

Why This Exists

• Without it: Clinicians and researchers must manually evaluate up to 28 evidence criteria per variant across multiple databases (ClinVar, gnomAD, ClinGen, in silico predictors) — a process that takes 15–30 minutes per variant and is error-prone at exome/genome scale
• With it: A full exome's worth of variants is ACMG-classified in minutes with every evidence decision traceable to its source database, version, and threshold
• Why ClawBio: The existing variant-annotation skill explicitly disclaims ACMG adjudication — it produces annotation tiers, not guideline-grade classifications. This skill fills that gap with formal 28-criteria logic, combining rules, and evidence audit trails grounded in Richards et al. (2015), ClinGen SVI recommendations, and the ACMG SF v3.2 secondary findings list — never ungrounded speculation

Core Capabilities

1. ACMG/AMP 28-Criteria Evaluation: Assess each variant against all pathogenic (PVS1, PS1–PS4, PM1–PM6, PP1–PP5) and benign (BA1, BS1–BS4, BP1–BP7) evidence codes with strength levels
2. Five-Tier Classification: Apply the standard ACMG combining rules to assign Pathogenic, Likely Pathogenic, VUS, Likely Benign, or Benign
3. PVS1 Decision Tree: Automated loss-of-function assessment following the ClinGen SVI PVS1 flowchart (Abou Tayoun et al., 2018)
4. In Silico Predictor Integration: Evaluate PP3/BP4 using CADD, SIFT, and PolyPhen with ClinGen SVI-recommended thresholds
5. Secondary Findings Screening: Flag variants in ACMG SF v3.2 genes (81 genes; Miller et al., 2023) and classify them independently
6. Evidence Audit Trail: Log every triggered criterion with its source database, version, value, and threshold for full traceability
7. Fail-Closed Self-Audit: Before emitting any call, run deterministic invariants and hard-abstain any variant that violates one, rather than report a confident, possibly-wrong classification (safe uncertainty over confident hallucination):
   • IDENTITY_MISMATCH: the variant resolved at the coordinate is not the one asserted (gene / HGVS via the ID column or GENE / EXPECTED_HGVSP / EXPECTED_HGVSC INFO keys) — catches wrong-variant / wrong-coordinate lookups
   • CONTRADICTORY_EVIDENCE: mutually exclusive computational criteria (PP3 and BP4) both fired (ClinGen SVI: exclusive)
   • MISSING_PROVENANCE: a triggered criterion carries no evidence source Abstained variants are labelled Abstained (self-audit) in result.json with abstained: true and machine-readable audit_violations.
8. Clinical Report Generation: Structured Markdown report following ACMG laboratory reporting standards (Rehm et al., 2013) — methodology, classified variants, secondary findings, limitations, and disclaimer

Input Formats

Format	Extension	Required Fields	Example
VCF 4.2+	.vcf, .vcf.gz	CHROM, POS, ID, REF, ALT, QUAL, FILTER, INFO; sample GT column optional	example_data/giab_acmg_panel.vcf
BCF (binary VCF)	.bcf	Same as VCF (binary-encoded)	—
Pre-annotated VCF	.vcf, .vcf.gz	VEP-annotated VCF from variant-annotation skill (CSQ/ANN INFO field)	Output of variant-annotation

Workflow

When the user asks for ACMG classification of a VCF:

1. Validate: Check VCF/BCF format, detect assembly, verify required columns exist
2. Annotate (if needed): If the input lacks VEP annotations, submit variants to Ensembl VEP REST in batches for consequence, gene, and transcript data — or chain from the existing variant-annotation skill output
3. Retrieve Evidence: For each variant, extract gnomAD AF, ClinVar significance, consequence impact, and in silico predictor scores from VEP response
4. Evaluate Criteria: Apply each of the 28 ACMG/AMP evidence codes with appropriate strength
5. Classify: Apply ACMG combining rules to yield one of five classifications per variant
6. Screen SF: Cross-reference all variants against ACMG SF v3.2 gene list (81 genes)
7. Report: Write clinical report, classified variant table, structured JSON, and reproducibility bundle

CLI Reference

# Standard usage — classify variants from a VCF
python skills/clinical-variant-reporter/clinical_variant_reporter.py \
  --input <patient.vcf> --output <report_dir>

# Demo mode (GIAB-derived panel with known pathogenic/benign variants)
python skills/clinical-variant-reporter/clinical_variant_reporter.py \
  --demo --output /tmp/acmg_demo

# Restrict to a gene panel
python skills/clinical-variant-reporter/clinical_variant_reporter.py \
  --input <patient.vcf> --genes "BRCA1,BRCA2,TP53,MLH1" --output <report_dir>

# Via ClawBio runner
python clawbio.py run acmg --input <file> --output <dir>
python clawbio.py run acmg --demo

Demo

To verify the skill works:

python clawbio.py run acmg --demo

Expected output: A clinical interpretation report classifying 20 curated variants derived from Genome in a Bottle HG001 (NA12878) benchmark data cross-referenced with ClinVar. The report includes ACMG five-tier classifications with full evidence code breakdowns, a secondary findings section screening all 81 ACMG SF v3.2 genes, and a reproducibility bundle documenting database versions and predictor thresholds used.

Algorithm / Methodology

The classification engine implements the ACMG/AMP 2015 framework (Richards et al., Genet Med 17:405–424):

Evidence Criteria Evaluation

Pathogenic evidence:

Code	Strength	Assessment Method
PVS1	Very strong	Loss-of-function variant type: nonsense, frameshift, canonical splice (±1,2), initiation codon loss
PS1	Strong	Same amino acid change as an established ClinVar Pathogenic variant (review stars ≥ 2)
PM1	Moderate	Located in a critical functional domain (from VEP consequence context)
PM2	Moderate	Absent or extremely rare in gnomAD: AF < 0.0001 (dominant) or AF < 0.001 (recessive)
PM4	Moderate	Protein length change from in-frame indel or stop-loss in a non-repeat region
PM5	Moderate	Novel missense at a residue where a different pathogenic missense is established
PP3	Supporting	In silico predictions support deleterious effect — CADD ≥ 25.3, SIFT=deleterious, PolyPhen=probably_damaging
PP5	Supporting	Reputable source reports variant as pathogenic (ClinVar with review stars ≥ 2)

Benign evidence:

Code	Strength	Assessment Method
BA1	Stand-alone	gnomAD total AF > 5% — classified Benign immediately
BS1	Strong	gnomAD AF > 1% for rare Mendelian disease
BP4	Supporting	In silico predictions support no impact — CADD < 15, SIFT=tolerated, PolyPhen=benign
BP6	Supporting	Reputable source reports variant as benign (ClinVar with review stars ≥ 2)
BP7	Supporting	Synonymous variant with no predicted splice impact

Combining Rules

Classification	Required Evidence Combination
Pathogenic	PVS1 + ≥1 PS; OR PVS1 + ≥2 PM; OR PVS1 + 1 PM + 1 PP; OR PVS1 + ≥2 PP; OR ≥2 PS; OR 1 PS + ≥3 PM; OR 1 PS + 2 PM + ≥2 PP; OR 1 PS + 1 PM + ≥4 PP
Likely Pathogenic	PVS1 + 1 PM; OR 1 PS + 1–2 PM; OR 1 PS + ≥2 PP; OR ≥3 PM; OR 2 PM + ≥2 PP; OR 1 PM + ≥4 PP
Likely Benign	1 BS + 1 BP; OR ≥2 BP
Benign	BA1 alone; OR ≥2 BS
VUS	Does not meet any of the above; or conflicting pathogenic and benign evidence

Key Thresholds

• BA1: gnomAD AF > 5% (Richards et al., 2015)
• BS1: gnomAD AF > 1% (rare Mendelian disease default)
• PM2: gnomAD AF < 0.0001 (dominant) or < 0.001 (recessive)
• PP3: CADD ≥ 25.3
• BP4: CADD < 15
• ClinVar minimum stars for PS1/PP5/BP6: ≥ 2

ClinVar Assertion Handling

ClinVar significance is parsed into terms before PS1, PP5 or BP6 read it; the rules never substring-match a joined string. VEP REST returns clin_sig as a list aggregated over every ClinVar record at the site, and ClinVar's own strings join terms with /, |, ; or ,. Both shapes bucket the same way.

For live VEP REST extraction, the site-level clin_sig aggregate is not used as evidence because it can mix assertions from different alternate alleles. The extractor keeps only the clin_sig_allele entry that exactly matches the queried ALT. Missing, malformed, non-string or unmatched allele-specific payloads are treated as absent evidence. VEP REST does not pair that allele-specific assertion with independently verifiable review stars, so live extraction records zero stars and withholds PS1, PP5 and BP6. Cached or directly constructed evidence that pairs a ClinVar assertion with a trustworthy review-star value still follows the table below.

ClinVar value	PS1 / PP5	BP6
Pathogenic, Likely pathogenic, Pathogenic/Likely pathogenic, Pathogenic|risk_factor, Pathogenic, low penetrance, pathogenic_low_penetrance, likely_pathogenic_low_penetrance	eligible	no
Benign, Likely benign, Benign/Likely benign	no	eligible
Conflicting_interpretations_of_pathogenicity, Conflicting_classifications_of_pathogenicity, conflicting_data_from_submitters (alone or alongside any other term)	withheld	withheld
pathogenic-family and benign-family terms together, e.g. ["benign", "pathogenic"]	withheld	withheld
Uncertain significance, drug_response, risk_factor, not_provided, unrecognised terms	no	no

"Withheld" means the rule does not fire and records the conflict as its reason in the criterion's detail field. A variant whose ClinVar records disagree therefore loses the ClinVar-backed criteria rather than being promoted on one side of the disagreement; the remaining criteria still combine as usual. Review-star gating (≥ 2) applies on top of this in every case.

Example Queries

• "Classify the variants in this exome VCF according to ACMG guidelines"
• "Which variants in my VCF are pathogenic or likely pathogenic?"
• "Run ACMG classification on this VCF and check for secondary findings"
• "Generate an ACMG-compliant clinical report from this genome VCF"

Output Structure

output_directory/
├── report.md                          # Clinical interpretation report
├── result.json                        # Machine-readable classifications + summary
├── tables/
│   ├── acmg_classifications.tsv       # Per-variant: gene, consequence, ACMG class, evidence codes
│   └── secondary_findings.tsv         # Variants in ACMG SF v3.2 genes with classifications
├── figures/
│   └── classification_summary.png     # Bar chart of P/LP/VUS/LB/B distribution
└── reproducibility/
    ├── commands.sh                    # Exact command to reproduce
    └── database_versions.json         # ClinVar date, gnomAD version, VEP release, SF list version

Dependencies

Required:

• Python 3.10+ (standard library for core classification engine)
• requests >= 2.31 — Ensembl VEP REST API access (live mode only)
• matplotlib >= 3.7 — classification summary figure

Optional:

• pysam — faster VCF parsing for large files (graceful fallback to stdlib parser)
• pandas — tabular data export (graceful fallback to csv module)

Safety

• Local-first: All classification logic runs locally. Only variant coordinates and alleles are sent to public Ensembl VEP REST. In live mode, the Data Sources report section also queries Ensembl's info/variation/homo_sapiens endpoint (no variant data in that request) to report the actual ClinVar/dbSNP/OMIM versions bundled with the release — no patient identifiers or phenotype data ever leave the machine
• Disclaimer: Every report includes the ClawBio medical disclaimer
• No hallucinated science: Every classification traces to specific evidence codes, database entries, and published thresholds
• Audit trail: Full evidence provenance logged to reproducibility/database_versions.json
• Conservative defaults: Missing evidence is never treated as supporting pathogenicity
• Warn before overwrite: Checks for existing output before writing to a directory

Integration with Bio Orchestrator

Trigger conditions — the orchestrator routes here when:

• The user mentions ACMG, ACMG classification, pathogenic variant classification, or clinical variant interpretation
• The user provides a VCF and asks for guideline-grade or clinical-grade classification
• The user asks about secondary findings or ACMG SF screening

Chaining partners:

• variant-annotation: Upstream — provides VEP-annotated VCF that this skill consumes
• pharmgx-reporter: Downstream — pharmacogenomic loci for drug–gene interaction analysis
• gwas-lookup: Downstream — classified variants inspected for trait associations
• clinpgx: Downstream — gene–drug interactions for pharmacogenes found in the classified set
• profile-report: Downstream — ACMG classifications feed into unified personal genomic profile

Citations

• Richards et al. (2015) — ACMG/AMP standards and guidelines for the interpretation of sequence variants. Genet Med 17:405–424
• Rehm et al. (2013) — ACMG clinical laboratory standards for next-generation sequencing. Genet Med 15:733–747
• Miller et al. (2023) — ACMG SF v3.2 list for reporting of secondary findings. Genet Med 25:100866
• Abou Tayoun et al. (2018) — PVS1 ACMG/AMP variant criterion recommendations. Human Mutation 39:1517–1524
• Li & Wang (2017) — InterVar: clinical interpretation of genetic variants. Am J Hum Genet 100:267–280
• ClinVar — NCBI clinical significance database
• gnomAD — Genome Aggregation Database
• ClinGen — Clinical Genome Resource