tooluniverse-acmg-variant-classification

ACMG/AMP Variant Classification

ACMG Reasoning

Each criterion (PS, PM, PP for pathogenic; BS, BP for benign) contributes a weighted piece of evidence for or against pathogenicity. The classification is the COMBINATION of all activated criteria, not any single criterion. Do not overweight a single finding.

The hierarchy is: PVS1 (very strong) > PS (strong) > PM (moderate) > PP (supporting). On the benign side: BA1 (stand-alone) > BS (strong) > BP (supporting). A frameshift in a LOF-intolerant gene (PVS1) plus a ClinVar expert-panel pathogenic entry (PS1) is pathogenic. A single PP criterion alone is not. The combination rule is what matters.

Two common errors to avoid: (1) seeing a "Pathogenic" ClinVar entry and stopping — that is PP5 (supporting) unless it has expert-panel review, not automatic confirmation; (2) dismissing a variant because one predictor says "tolerated" — discordant predictors mean neither PP3 nor BP4 applies, which is neutral evidence, not benign evidence.

Always apply criteria conservatively. When evidence is ambiguous, leave the criterion unmet. Cite the source for every criterion you activate so clinicians can audit the reasoning.

KEY PRINCIPLES:

1. Criteria-driven — cite which criteria were activated and why
2. Conservative — do not upgrade a criterion when evidence is ambiguous
3. Gene-aware — adjust thresholds based on gene mechanism (LOF vs. gain-of-function)
4. Population-calibrated — use ancestry-specific gnomAD frequencies, not just global AF
5. Transparent — show evidence for each criterion
6. Source-referenced — every criterion activation must cite the database/tool source
7. English-first queries — always use English terms in tool calls; respond in user's language

LOOK UP, DON'T GUESS

When uncertain about any scientific fact, SEARCH databases first (PubMed, UniProt, ChEMBL, ClinVar, etc.) rather than reasoning from memory. A database-verified answer is always more reliable than a guess.

When to Use

• "Classify BRCA2 c.5946delT using ACMG criteria"
• "Is this VUS pathogenic? NM_000059.4:c.7397T>C"
• "Apply ACMG guidelines to rs28897743"
• "What is the pathogenicity of CFTR p.Arg117His?"
• "ACMG classification for TP53 R248W"

Tool Parameter Reference

Phase 0: Variant Validation and Normalization

Wrong HGVS or wrong transcript cascades errors through every downstream criterion. Validate first.

1. Get MANE Select transcript: VariantValidator_gene2transcripts(gene_symbol="BRCA2"). Cross-check the canonical transcript with Tark_get_mane_transcripts(gene="BRCA2") (returns the ENST↔RefSeq pairing, e.g. ENST00000380152.8 / NM_000059.4); use Tark_get_transcript(stable_id="ENST00000380152") when you need to pin a specific transcript version.
2. Validate variant: VariantValidator_validate_variant(variant_description="NM_000059.4:c.5946delT", genome_build="GRCh38", select_transcripts="mane_select")
3. Resolve gene IDs: MyGene_query_genes(query="BRCA2") — extract Ensembl ID and UniProt accession. Filter results by symbol == 'BRCA2' (first hit may not match).
4. Record: HGVS coding, HGVS protein, genomic coordinates, variant type (frameshift/missense/nonsense/splice/synonymous/in-frame indel).

Accepted inputs: HGVS coding (NM_000059.4:c.5946delT), HGVS protein (BRCA2 p.Val600Glu), rsID (rs28897743), gene+change (BRCA1 c.68_69del), genomic coordinates.

Phase 1: Population Frequency (BA1, BS1, BS2, PM2)

Population AF is among the strongest evidence in either direction. A variant at >5% in any population is almost certainly benign (BA1 — stand-alone, no further analysis needed). Absent from gnomAD supports pathogenicity (PM2, now usually applied as PM2_Supporting per ClinGen guidance).

Use ancestry-specific AF, not just global. A variant at 8% in East Asian populations but rare globally is benign in that ancestry context. For BS1, the threshold depends on disease prevalence and inheritance — the default is 1% for common diseases, 0.1% for rare.

gnomad_search_variants(query="rs28897743")          # get gnomAD variant ID
gnomad_get_variant(variant_id="...")                 # per-ancestry frequencies
gnomad_get_gene_constraints(

If gnomAD data is unavailable, note the gap and continue — absence of data is not the same as evidence of absence.

Phase 2: Computational Predictions (PP3, BP4)

No single predictor is definitive. The reasoning is: concordance across multiple independent predictors provides supporting evidence. Discordance means neither PP3 nor BP4 applies — it is neutral, not benign.

PP3 (supporting pathogenic) applies when the majority of predictors agree damaging, or when REVEL >= 0.7 alone (sufficient per ClinGen guidance). BP4 (supporting benign) requires ALL predictors to agree benign, or REVEL < 0.15 or CADD < 15. These criteria apply only to missense variants.

For non-missense variants, skip PP3/BP4 and focus on SpliceAI scores in Phase 5.

Mechanism complement for VUS-resolution narratives: PP3/BP4 is a vote count of predictor scores — it tells you whether the variant is damaging, not how. For a VUS resolution report where the verdict needs a mechanistic explanation (clinician asks "why is this damaging?"), add ESM_explain_variant_mechanism(sequence=..., position=..., ref_aa=..., alt_aa=...) — returns lost/gained ESMC-6B SAE feature categories (catalytic / ligand-binding / ptm / structural-stability / domain / etc.) with a one-line summary. This does not change PP3 strength but turns "PP3 satisfied (REVEL 0.78, AlphaMissense 0.85)" into "PP3 satisfied — and SAE shows catalytic-feature loss at position X, consistent with active-site disruption." Requires ESM_API_KEY; missense only.

Phase 3: Clinical Database Evidence (PS1, PM5, PP5, BP6)

ClinVar aggregates clinical lab classifications. The reasoning: if the same amino acid change (different nucleotide) is established pathogenic, that is strong evidence (PS1) because the mechanism is the amino acid change. If a different pathogenic missense occurs at the same residue, that is moderate evidence (PM5) — the residue is functionally important.

PP5 applies when ClinVar shows Pathogenic with >= 2-star review (criteria provided, multiple submitters). Weight by the number of concordant submitters. Conflicting ClinVar interpretations mean neither PP5 nor BP6 should be applied. ClinGen has proposed downweighting PP5/BP6 — treat them as supporting, not strong.

Phase 4: Functional Domain and Protein Analysis (PM1, PP2, BP1)

Variants in well-established functional domains with known pathogenic variant enrichment are more likely pathogenic. PM1 (moderate pathogenic) requires the variant to be in a hotspot domain with low benign variation — use InterPro domain architecture and UniProt active/binding sites to assess.

PP2 and BP1 are mutually exclusive. PP2 (supporting pathogenic) applies to missense in genes where missense is the known mechanism and benign missense rate is low (mis_z > 3.09). BP1 (supporting benign) applies to missense in genes where only truncating variants cause disease (LOF-only mechanism) — a missense in such a gene is unlikely to be pathogenic.

Phase 5: Splice Impact Assessment (PVS1)

PVS1 is the strongest single pathogenic criterion. A null variant (nonsense/frameshift/canonical splice/initiation codon) in a gene where LOF is the established mechanism can activate PVS1, but the full strength depends on context.

Apply PVS1 at full strength when: null variant + LOF is known mechanism + variant is not in the last exon or last 50bp of the penultimate exon + no rescue transcript exists. Downgrade to PVS1_Moderate if the variant is in the last exon or NMD escape is likely. Downgrade to PVS1_Supporting if a rescue transcript is possible or SpliceAI >= 0.5 but not a canonical splice site. Do NOT apply PVS1 if LOF mechanism is uncertain.

Phase 6: Literature and Functional Evidence (PS3, BS3, PP1, PP4)

Well-designed functional assays showing LOF (PS3) or normal function (BS3) can shift a classification decisively. PS3/BS3 can be downgraded (e.g., PS3_Supporting) for less rigorous assays. Not all functional assays qualify — ClinGen gene-specific guidance defines valid assays.

PP1 (co-segregation) upgrades to PP1_Strong at >= 7 informative meioses. PP4 applies when the patient's phenotype is highly specific for the gene's disease.

Criteria requiring clinical data (PS2, PS4, PM3, PM6, BS4, BP2, BP5) cannot be assessed automatically. Document as "Not Assessed" unless the user provides clinical context.

PM4 (protein length change in non-repeat region) and BP3 (in-frame indel in repeat) can be partially assessed from variant type. BP7 (synonymous, no splice impact) is assessable via SpliceAI < 0.1.

Automated Cross-Check (InterVar)

Everything in Phases 0-6 is a manual, cited, criterion-by-criterion assessment — that is this skill's core value and should remain the primary output. InterVar_classify_variant runs the equivalent ACMG/AMP 2015 rubric server-side from bare genomic coordinates and returns a classification plus all 28 criteria's activation status in one call. Use it as a fast independent cross-check on your own manual verdict, never as a substitute for it — the manual process cites its sources per criterion; InterVar does not.

Cross-validation example, verified live: TP53 chr17:7674220 C>T (GRCh38, the same variant used elsewhere in tooluniverse-variant-analysis for the GeneBe/ClinGen cross-check) returns classification="Pathogenic" with PS3+PM1+PM2+PP3+PP5 active — concordant with GeneBe's independent "Pathogenic" call on the same variant. Two independently-implemented ACMG classifiers agreeing is exactly the kind of corroboration worth citing in a report; if they disagree, treat it as a reason to re-examine your manual criteria, not a reason to silently prefer one tool's answer.

Known data-quality issue, verified live, not a guess: the tool's own JSON description field claims its first worked example (chrom="17", pos=41245466, ref="G", alt="A", labeled "BRCA1 c.5266dupC") classifies as "Likely Pathogenic with PM2+PP3+PP5." The live API actually returns classification="Benign" with BA1+BP4+BP6+BP7+BS1+BS2 active. This is expected, not a bug in the tool: c.5266dupC is a single-base duplication (an insertion), which cannot be represented as a ref="G"/alt="A" substitution — the example's own coordinates don't encode the frameshift its label claims, so the "Benign" result is InterVar correctly classifying the SNV it was actually given, not misclassifying BRCA1 c.5266dupC. Never trust a tool's documented worked example without re-verifying live; always pass your own correctly-encoded variant.

Classification Algorithm

Combine criteria at their applied strength (after upgrades/downgrades):

Pathogenic: (1) PVS1 + ≥1 Strong; (2) PVS1 + ≥2 Moderate; (3) PVS1 + 1 Moderate + 1 Supporting; (4) PVS1 + ≥2 Supporting; (5) ≥2 Strong; (6) 1 Strong + ≥3 Moderate; (7) 1 Strong + 2 Moderate + ≥2 Supporting; (8) 1 Strong + 1 Moderate + ≥4 Supporting

Likely Pathogenic: (1) PVS1 + 1 Moderate; (2) 1 Strong + 1-2 Moderate; (3) 1 Strong + ≥2 Supporting; (4) ≥3 Moderate; (5) 2 Moderate + ≥2 Supporting; (6) 1 Moderate + ≥4 Supporting

Benign: (1) BA1 stand-alone; (2) ≥2 Strong benign

Likely Benign: (1) 1 Strong benign + 1 Supporting benign; (2) ≥2 Supporting benign

VUS: Criteria do not meet any threshold above, OR pathogenic and benign evidence conflict.

Output Format

Common Patterns

Pattern 1: Known pathogenic frameshift — "Classify BRCA2 c.5946delT" Phase 0 (validate) → Phase 1 (gnomAD absent, PM2_Supporting) → Phase 3 (ClinVar Pathogenic, PP5) → Phase 4 (DNA repair domain, PM1) → Phase 5 (frameshift + LOF gene, PVS1) → Phase 6 (literature PS3) Result: Pathogenic (PVS1 + PS3 + PM1 + PM2_Supporting + PP5)

Pattern 2: Missense VUS — "Is BRCA1 p.Arg1699Gln pathogenic?" Phase 0 → Phase 1 (rare, PM2_Supporting) → Phase 2 (REVEL 0.82, CADD 26, PP3) → Phase 3 (ClinVar VUS) → Phase 4 (BRCT domain, PM1) → Phase 6 (reduced activity, PS3_Moderate) Result: Likely Pathogenic (PS3_Moderate + PM1 + PM2_Supporting + PP3)

Pattern 3: Common benign variant — "ACMG for rs1800497" Phase 1 (gnomAD AF=0.21, BA1) → short-circuit. Result: Benign (BA1 stand-alone)

Pattern 4: Deep-intronic variant — "Classify NM_000059.4:c.7977+100A>G" Phase 1 (check AF) → Phase 5 (SpliceAI < 0.1) → Result: Likely Benign or VUS depending on frequency