ACMG/AMP Variant Classification Criteria

A variant caller finishes running on an exome or genome and hands back a VCF full of candidate variants. Most of that list is noise — sequencing artifacts, common population variation, calls that will never matter to the biological question being asked. The ACMG/AMP variant classification framework is the standardized system researchers and clinical laboratories both use to sort the calls that remain into a defensible tier of clinical or functional significance, instead of relying on a single test or a gut call from whoever is reviewing the pipeline output.

What the ACMG/AMP framework actually is

ACMG/AMP refers to the joint standards published by the American College of Medical Genetics and Genomics (ACMG) and the Association for Molecular Pathology (AMP) for interpreting sequence variants identified by clinical and research sequencing. The framework sorts a variant into one of five tiers:

Pathogenic — evidence strongly supports a causal role in disease.
Likely Pathogenic — evidence favors a causal role, but not to the same evidentiary bar as Pathogenic.
Variant of Uncertain Significance (VUS) — the evidence available doesn’t clear the bar in either direction.
Likely Benign — evidence favors no causal role.
Benign — evidence strongly supports no causal role.

These same five tiers are the classification vocabulary used in ClinVar, NCBI’s public archive of variant-phenotype relationship claims, which is the most common place a researcher will actually encounter ACMG/AMP tier labels attached to a specific variant.

How classification actually works: weighted evidence, not one test

The framework does not classify a variant from a single piece of evidence. It combines multiple independently-scored evidence codes, each carrying a defined strength (very strong, strong, moderate, or supporting evidence toward pathogenicity; strong or supporting evidence toward benignity), across several evidence categories:

Population data — how common the variant is in large reference cohorts. A variant that’s genuinely rare in the general population (checked against a resource like gnomAD) supports pathogenicity; one that’s common argues against it.
Computational and predictive data — in-silico tools that predict a variant’s effect on protein structure or function, plus conservation across species.
Functional data — results from wet-lab assays (e.g., a functional study showing the variant disrupts the protein’s normal activity).
Segregation data — whether the variant co-segregates with disease within affected families, across multiple meioses.
De novo data — whether the variant arose new in an affected individual, with both parents confirmed unaffected and non-carriers.
Allelic data — the variant’s relationship to other known variants on the same or opposite chromosome (in cis or in trans), including whether it’s ever been observed with a known pathogenic variant in a way that’s informative.
Other data — case-level or database evidence that doesn’t fit cleanly into the categories above, including prior classifications reported in a resource like ClinVar.

Each applicable evidence code is combined using the framework’s defined combining rules to arrive at one of the five tiers. This is deliberately a rules-based aggregation rather than a single decisive test, because no individual line of evidence — not even a strong computational prediction — is reliably sufficient on its own to call a variant pathogenic or benign.

Why this matters specifically for research variant calls

It’s tempting to treat ACMG/AMP classification as strictly a clinical-diagnostic concern — something that matters for a genetic counselor reporting results to a patient, not for a research pipeline. That’s a mistake. Research variant calls from exome and genome sequencing studies routinely surface candidate variants that get reported in a paper, a poster, or a returned-results protocol as a “candidate disease-causing variant” — and that claim carries the same evidentiary weight, and the same risk of being wrong, whether it comes from a clinical lab or a research pipeline.

A variant caller (see DeepVariant vs. GATK HaplotypeCaller for how two of the most widely used callers differ) tells you a variant exists and estimates its genotype quality. It says nothing about whether that variant matters biologically. Applying ACMG/AMP-style evidence review before a variant gets promoted from “called” to “candidate finding” is what separates a defensible research claim from an overstated one — and it’s increasingly expected by journals, IRBs, and any downstream clinical confirmation pipeline a research finding might feed into.

The VUS category is a real result, not a failure to classify

The single most common way ACMG/AMP classification gets misused in a research context is treating a Variant of Uncertain Significance as a weak positive — “probably pathogenic, we just don’t have enough evidence yet.” That’s not what VUS means. VUS is a genuine, information-honest outcome: the evidence gathered so far doesn’t clear the bar in either direction, and the correct action is to report it as uncertain, not to round it toward pathogenic because it happens to sit in a gene of interest.

Over-interpreting VUS calls — reporting them with more certainty than the evidence supports, or letting a researcher’s prior belief about a gene push an uncertain call toward “likely pathogenic” in a manuscript or a returned result — is exactly the failure mode ACMG/AMP’s tiered, evidence-coded structure exists to prevent. The category will keep shrinking over time as population databases grow and more variants accumulate functional and segregation evidence; that reclassification is a feature of the system, not evidence the original VUS call was a mistake.

Practical research-workflow notes

A few things are worth knowing before applying ACMG/AMP classification to a real research variant-calling pipeline:

Automated tools handle initial evidence assignment, not final classification. Tools such as InterVar apply the ACMG/AMP evidence codes computationally against available annotation — population frequency, prior ClinVar entries, computational predictors — but only cover a subset of the full criteria set automatically; the remainder (functional data, segregation data, case-level and family evidence) require a human curator with access to information the tool can’t see.
ClinGen’s Sequence Variant Interpretation (SVI) working group and gene-specific Variant Curation Expert Panels publish refined, gene- and disease-specific specifications of the general ACMG/AMP criteria, because a criterion’s exact evidentiary weight is not identical across every gene. A research pipeline classifying variants in a gene with a published VCEP specification should use that specification rather than the unmodified general-purpose criteria.
Final classification requires expert review. Automated evidence-code assignment is a starting point for a curator, not a substitute for one — especially before a variant is reported externally as a candidate finding, returned to a study participant, or submitted to ClinVar.
Document which evidence codes were applied, and by whom. A classification without a recorded evidentiary basis can’t be re-evaluated later as new population or functional data becomes available, which defeats the purpose of using a structured, reproducible framework in the first place.

Frequently asked questions

Is ACMG/AMP classification only for clinical labs, or does it apply to research sequencing too?

The framework itself makes no clinical/research distinction — it’s a general standard for interpreting sequence variants. Research variant-calling pipelines that report candidate findings, especially anything that might be returned to a participant or submitted to a public archive like ClinVar, benefit from the same evidence-based rigor a clinical lab is expected to apply.

Does a VUS classification mean the variant is probably harmless?

No. VUS means the evidence gathered so far doesn’t clear the bar for either Likely Pathogenic/Pathogenic or Likely Benign/Benign — it is a statement about the current evidence, not a lean in either direction. A VUS can be reclassified later as new evidence accumulates.

Can InterVar alone produce a final ACMG/AMP classification?

No. InterVar and similar tools automate a subset of the evidence codes from available annotation data, but criteria that depend on functional assays, segregation in families, or case-level judgment require expert curator review before a classification is finalized.

ACMG/AMP Variant Classification Criteria for Research Variant Calls