Variant Classification Drift: Why a Correct Report Becomes Wrong Without the Data Ever Changing

Zetobit.

Bioinformatics Insight Series

Variant Classification Drift: Why a Correct Report Becomes Wrong Without the Data Ever Changing

A pathogenicity classification is not a property of a variant. It is a statement about the evidence available on the date it was made. The sequence is permanent; the interpretation has a shelf life — and nothing in the pipeline announces when it expires.

By Kanna Nandakumar, PhD · Zetobit

Two previous pieces in this series looked at results that were wrong at the moment they were produced — a basecall that changes with the model, a report issued about the wrong patient. This one is different. Here the report was correct when it was signed out, correct by the standards in force, correct given everything known at the time.

It became wrong later, sitting in a chart, while nobody touched it.

This is the failure mode with the longest tail in clinical genomics. A BAM does not degrade. A VCF does not drift. What drifts is the interpretive layer sitting on top of them — the mapping from variant to clinical meaning — and that layer is rebuilt continuously by the entire field, asynchronously, without notifying anyone holding an older copy.

The magnitude is larger than most people assume

The instinct is that reclassification is an edge case affecting a small number of ambiguous variants. In some clinical areas that is roughly true. In others it is dramatically false.

The most striking published example comes from inherited arrhythmogenic syndromes. A group reanalyzed 128 rare variants they had originally classified in 2010, applying then-current ACMG guidelines a decade later. The classification changed for 71.87% of them — 92 of 128. The direction of that drift is what makes the study memorable: variants originally called likely pathogenic made up 65.62% of the set, and after reanalysis only about 18% remained in that category. Meanwhile the proportion classified as uncertain rose from 18.75% to roughly 60%. Of the variants originally called likely pathogenic, 59 were downgraded to VUS and 4 went all the way to likely benign.

Read that carefully, because it inverts the usual expectation. Reclassification is often discussed as a mechanism for solving cases — VUS becoming pathogenic, diagnoses appearing where there were none. In this cohort the dominant movement was the opposite: confident calls dissolving back into uncertainty as the evidentiary bar rose and population databases revealed that variants once thought rare were not rare enough to be causal.

Other indications drift differently, and the differences are not noise — they reflect real variation in how much evidence existed at baseline and how fast it accumulated.

Figure 1. Sequencing produces two artifacts with very different half-lives. The data layer is immutable and reproducible indefinitely. The interpretive layer above it is rebuilt as population databases grow, functional assays are published, cases accumulate, and specification standards change. Reports are snapshots of the upper layer, but they are stored and read as though they described the lower one.

Drift is not uniform, and it is not always toward the answer

Direction and speed both vary by clinical context. In germline cancer testing, the well-documented pattern is rapid resolution toward benign: reviews report that on the order of 73% to 91% of reclassified variants move from VUS to benign, with a median time to reclassification of roughly 1.17 years. That is the system working as designed — most uncertain variants are ultimately not causal, and growing population reference data settles them quickly.

Neurodevelopmental conditions behave differently. Reinterpreting 2,834 laboratory reports submitted to a research registry, one group reclassified 20.4% of variants, with 230 upgrades and 173 downgrades. Among ACMG-classified monogenic variants, 25.4% of VUS were upgraded with a median revision time of 3 years, while 4.3% were downgraded — an upgrade-dominant pattern, and a slower one.

At the program level, systematic reanalysis converts into diagnoses at a fairly consistent rate. A retrospective review of 10,416 clinical exome cases analyzed between 2011 and 2021, with reanalysis events tracked through 2023, found that reanalysis raised total diagnostic yield from 21.4% to 25.5%. Roughly four percentage points of additional diagnoses came from data that had been sitting on a server the whole time.

The sequencing was already paid for. The diagnoses were already latent in the data. What was missing was a second look at a later date.

The reanalysis literature is also clear about one operational finding: waiting for clinicians to request reanalysis does not work. In one cohort, only 5.9% of updated reports were provider-initiated on the basis of new phenotypic information, and the authors point to practical barriers — patients moving, providers changing practice — that make recontact difficult. Laboratory-initiated proactive reanalysis is the mechanism that actually reaches the cases.

Two variants, same evidence, different labs

Drift over time has a spatial analogue: at any single moment, different laboratories may hold different classifications for the same variant.

A ClinGen initiative had four clinical laboratories compare their ClinVar submissions. Of 6,169 variants interpreted by at least two participants, 88.3% were initially concordant — meaning roughly one in nine carried a difference. The encouraging part is what happened next: of 242 discordant variants reassessed, 87.2% were resolved by applying current criteria and sharing internal data, and overall concordance rose from 88.3% to 91.7%. Only 12.8% of reassessed variants remained discordant due to genuine differences in how the ACMG-AMP guidelines were applied.

The clinical consequence of this is easy to overlook. A study of patients who had undergone hereditary cancer panel testing found that 2.5% had a variant with a clinically significant conflict in ClinVar at the time their report was issued. Among patients with a VUS on their report who had such a conflict, only 36% appeared to have been counseled by a provider who was aware of the conflict.

Much of the historical discordance is also an artifact of comparing across time rather than across judgment. Analyses of ClinVar sources have found that older classifications carry greater discordance than newer ones, and that clinical testing classifications agree far more than entries derived from literature curation or research submissions. A "conflict" in an aggregate database is frequently a current classification sitting next to a stale one, not two laboratories disagreeing today.

Table 1. Reclassification behavior varies substantially by clinical context. A single reanalysis cadence applied across all indications will be too slow for some and largely wasted on others.

Context	Dominant direction	Timescale
Germline cancer panels	VUS → benign; most uncertainty resolves downward	Fast — median ~1.2 years
Neurodevelopmental conditions	Upgrade-dominant, but with substantial downgrades	Slower — median ~2–3 years
Inherited arrhythmia syndromes	Historically downgrade-dominant as evidence bar rose	Slow; large shifts visible over a decade
Exome / genome programs	Net yield gain from previously undiagnosed cases	Continuous; a few points of yield per reanalysis era

What this asks of a pipeline

Most of the burden here is clinical and regulatory rather than computational. But several parts are squarely engineering problems, and they are the parts most often left undone.

Store the classification with its date and evidence basis

A classification recorded as a bare label is not auditable. Recorded with the date, the guideline version, the database snapshot, and the specific criteria codes applied, it becomes possible to ask later whether the reasoning still holds. The question at reanalysis is rarely "what did we call this" — it is "what did we call it on the basis of, and has that basis moved."

Make reanalysis a query, not a project

If prior cases are stored such that identifying every patient carrying a given variant requires reprocessing, reanalysis will not happen at a useful cadence. A variant-indexed case store turns "this variant was just upgraded — who is affected" into a lookup. This is the single highest-leverage architectural decision for supporting reanalysis, and it has to be made before the cases accumulate.

Watch the sources, not just the variants

Classification changes are downstream of source changes — a ClinVar submission update, a new gene-disease validity assessment, a population frequency shift, a revised gene-specific specification. Monitoring those feeds and mapping changes back onto your reported variants is what turns reanalysis from periodic sweep into targeted alert.

Separate the two triggers

Reanalysis can be prompted by new knowledge about the variant or by new information about the patient — an evolving phenotype, a new family member tested, a functional study. These have different owners and different cadences. A workflow designed around only one of them will systematically miss the other.

Distinguish stale from contested

When a variant shows a conflict in an aggregate database, the operationally useful question is whether the disagreeing entries were made at comparable times under comparable standards. An old submission and a current one are not two opinions. Treating them as equivalent inflates the apparent conflict rate and buries the cases where laboratories genuinely disagree today.

The short version

A variant classification is a dated claim, not a fact about the genome. Pipelines that store classifications without their evidentiary basis, or that make finding affected prior cases expensive, convert a normal scientific process — knowledge improving — into an accumulating stock of quietly outdated reports.

The obligation the data creates

There is a genuine tension underneath all of this, and it is worth stating plainly rather than resolving glibly. Sequencing data outlives the interpretation attached to it. That is what makes reanalysis valuable — the same file yields new answers years later at essentially no marginal sequencing cost. It is also what creates an open-ended duty that the field has not fully worked out: how long, and how actively, someone remains responsible for revisiting a result that was correct when issued.

That question is not primarily a bioinformatics question. But the bioinformatics determines whether it can be answered at all. A laboratory that can identify every prior case affected by a reclassification in a single query has a policy choice to make. A laboratory that cannot has already made it.

Zetobit builds and validates CAP/CLIA-compliant NGS pipelines, including variant-indexed case architectures and reanalysis workflows designed so that reclassification becomes a query rather than a reprocessing project. If you are designing for reanalysis or auditing an accumulated case archive, we're happy to talk.

References

1. Campuzano O, Sarquella-Brugada G, Cesar S, et al. Reanalysis and reclassification of rare genetic variants associated with inherited arrhythmogenic syndromes. eBioMedicine 54:102732 (2020). doi:10.1016/j.ebiom.2020.102732.
2. Harrison SM, Dolinsky JS, Knight Johnson AE, et al. Clinical laboratories collaborate to resolve differences in variant interpretations submitted to ClinVar. Genetics in Medicine 19(10):1096–1104 (2017). doi:10.1038/gim.2017.14.
3. Improvement of variant reclassification in genetic neurodevelopmental conditions. Genetics in Medicine Open (2024). doi:10.1016/j.gimo.2024.101829.
4. Ten years of exome sequencing and reanalysis among racial, ethnic, and ancestral groups: the importance of equitable reanalysis access. Genetics in Medicine (2025). doi:10.1016/j.gim.2025.101487.
5. Impact of laboratory-driven proactive reanalysis: reclassification to positive in 5% of initially negative or uncertain exome sequencing cases. Genetics in Medicine (2025). doi:10.1016/j.gim.2025.101375.
6. Diagnostic yield of exome reanalysis over time: contribution of reevaluation type, timing, and patient phenotype. Genetics in Medicine (2026).
7. Yang S, Lincoln SE, Kobayashi Y, et al. Sources of discordance among germ-line variant classifications in ClinVar. Genetics in Medicine 19(10):1118–1126 (2017). doi:10.1038/gim.2017.60.
8. Clinical implications of conflicting variant interpretations in the cancer genetics clinic. eScholarship, University of California (2020).