Your Pipeline Called the Variant. Now What? Why Clinical Genomics Keeps Failing at Interpretation

Your Pipeline Called the Variant. Now What?

A clinical genomics lab runs a validated variant calling pipeline on a whole-genome sample. The run completes. The VCF lands. QC metrics pass. And then the report says something like "a pathogenic variant was identified" — which tells a clinician almost nothing they can act on.

After variant calling and QC, interpretation remains the step that converts a VCF into a clinically actionable report. Variant calling and QC are comparatively more amenable to automation than clinical interpretation, and the part that determines whether a patient actually gets treated correctly is still often done by a human being reading a spreadsheet at 6pm.

Variant calling is a solved problem in the sense that widely used software pipelines exist. Variant interpretation is not solved at all, and it can create substantial labor, quality-control, and clinical-risk burdens.

The ACMG/AMP variant-classification framework defines 28 evidence criteria and uses a five-tier scheme for classifying variants as pathogenic, likely pathogenic, of uncertain significance, likely benign, or benign.[1] Applying these criteria correctly to a single rare-disease variant requires reading multiple databases, checking population frequencies, reconciling functional assay results, and applying phenotype-specific relevance judgments. Nobody does that reliably by memory, but the standard of care in a regulated lab still expects analysts to apply the framework consistently.

Why interpretation breaks down in practice

The criteria are weighted wrong by default. ACMG/AMP evidence codes are designed as combining rules — Bayesian-style odds for pathogenicity — with strengths spanning from supporting through moderate and strong to very strong.[2] In practice, some implementations treat the criteria mechanically, counting codes without explicitly modeling their probabilistic rationale. That discards the actual structure of the framework: different combinations of strong, moderate, and supporting evidence yield different posterior odds, and a variant with one very strong criterion plus several supporting criteria can deserve more confidence than a variant with several moderate criteria piled up.

Population frequency requires judgment about the population. PM2 concerns rarity or absence in control populations, and its application requires attention to ancestry, dataset composition, and disease-specific frequency thresholds.[3] What counts as "absent"? Not present in gnomAD as of a given release? Not present above an allele-frequency threshold that was itself chosen for a different ancestry? A variant seen at 0.1% in an under-represented ancestry group is common in that population and rare in the reference population, and a frequency filter that ignores ancestry structure will misclassify it in both directions.

Functional evidence is the most fragile layer. PS3 and BS3 rely on well-established in vitro or in vivo functional studies, and the strength of this evidence depends on how closely the assay reflects the relevant disease mechanism, assay class, analytical validity, and variant-level application.[4] A missense variant with a modest but reproducible effect in an assay that does not reflect the relevant disease mechanism is not strong evidence of pathogenicity, and there is no universal formula for the acceptable signal-to-noise of a reporter assay or how much of an assay's dynamic range constitutes meaningful evidence.

Phenotype matching is doing more work than it gets credit for. PP3 and BP4 capture computational predictive evidence, while PM1 and PS3 capture features such as location in a mutational hotspot and functional impact; their application depends on gene- and disease-specific specifications and on how well the submitted phenotype matches the disease model.[1] A variant that is clearly pathogenic for a metabolic disorder may be benign for an unrelated indication. Workflow variability can arise when the clinical indication changes: the same variant, in the same patient, can receive different classifications depending on which clinical indication code is on the requisition.

And the reanalysis problem never closes. ACMG guidance recommends that laboratories establish and periodically review policies for variant-level reevaluation and case-level reanalysis, and cites a two-year interval as an example for policy review rather than a universal requirement for every classification.[5] In practice, the frequency and completeness of reanalysis vary, and laboratories that lack explicit policies, triggers, and workflows risk letting classifications age silently while new population data and functional evidence accumulate.

What a defensible interpretation layer actually looks like

A defensible interpretation layer combines automation with structured human review. It means putting a structured, auditable, versioned decision process in front of the analysts, so that human judgment is applied consistently and the reasoning is reconstructable.

Encode the criteria as code, with the Bayesian structure intact. Model the evidence as likelihood ratios per ACMG code and combine them properly, rather than counting. Keep the evidence table itself as the artifact of record — every code applied, with the source, the retrieval date, and the rationale. The code is an enforcement mechanism, not the deliverable.

Automate the evidence gathering, not the judgment. Pull population frequencies, and compute ancestry-aware frequency estimates rather than accepting the raw database query. Retrieve ClinVar assertions with their review status and expert-panel evidence separated from single-submitter assertions.[6] Query the functional assay literature automatically. The analyst should spend their time on cases with conflicting evidence, uncertain penetrance, or phenotype mismatch, not on tabulation.

Route by difficulty, not by queue. Score cases by evidence ambiguity — conflicting assertions, uncertain penetrance, phenotype mismatch — and give cases with conflicting evidence, uncertain penetrance, or phenotype mismatch the time they need. Case complexity varies, and treating every case as if it were maximally difficult is how turnaround times and costs get out of control.

Make reanalysis a scheduled pipeline, not a project. New gnomAD releases, new ClinVar submissions, and new functional publications are dated events. A reanalysis queue keyed to those releases, with automatic detection of affected variant classifications, turns ad hoc reanalysis into a scheduled workflow with defined triggers and ownership — which is also what a regulated lab needs to evidence.

Keep the audit trail versioned. Every interpretation decision should be reconstructable: which evidence, which version of which source, which analyst, which date, and which clinical context. This is both a quality control mechanism and the artifact you will need if the interpretation is ever challenged.

The honest ROI framing

Sequencing is only one component of total clinical-genomics cost; interpretation, reporting, validation, data management, and quality processes can also be substantial cost drivers. Programs should track both operational measures such as cost per sample and samples per week and interpretation measures such as review time, classification concordance, reanalysis burden, and report quality.

The programs that scale are the ones that reduce unnecessary analyst time without reducing interpretive rigor, and that make every interpretation auditable. That is not a genomics problem in the narrow sense; it is a data infrastructure problem with a bioinformatics surface.

Pii Data Science Solutions builds the interpretation infrastructure that clinical genomics programs need to scale. Our team works with clinical genomics programs on variant calling and QC pipelines for NGS and RNA workflows, evidence-structured interpretation systems with full provenance and audit trails, and reanalysis pipelines keyed to reference database releases that support a defined operational cadence appropriate to the laboratory’s workload, evidence-update frequency, and quality system. If your lab's turnaround time is limited by analyst interpretation, or your classification database is aging without a refresh path, Pii Data Science Solutions can assess whether pipeline automation, provenance controls, or scheduled reanalysis would address these bottlenecks.

Sources

[1] Genetics in Medicine — "Standards and guidelines for the interpretation of sequence variants" — https://pmc.ncbi.nlm.nih.gov/articles/PMC4908185/

[2] Genetics in Medicine — "Modeling the ACMG/AMP variant classification guidelines as a Bayesian classification framework" — https://www.nature.com/articles/gim2017210

[3] ClinGen Platelet Disorders Expert Panel — "Specifications to the ACMG/AMP Variant Interpretation Guidelines Version 2.1" — https://www.clinicalgenome.org/site/assets/files/7223/clingen_platelet_acmg_specifications_v2_1.pdf

[4] ClinGen Sequence Variant Interpretation Working Group — "Recommendations for application of the functional evidence PS3/BS3 criterion using the ACMG/AMP sequence variant interpretation framework" — https://clinicalgenome.org/docs/recommendations-for-application-of-the-functional-evidence-ps3-bs3-criterion-using-the-acmg-amp-sequence-variant-interpretation/

[5] Genetics in Medicine / American College of Medical Genetics and Genomics — "Points to consider in the reevaluation and reanalysis of genomic test results: a statement of the American College of Medical Genetics and Genomics" — https://www.nature.com/articles/s41436-019-0478-1

[6] Clinical Genome Resource (ClinGen) — "ClinGen Variant Classification Guidance" — https://www.clinicalgenome.org/tools/clingen-variant-classification-guidance/