Allele Frequency Filters

Zetobit · Bioinformatics Insight Series Filtering & Reference Data

Bioinformatics Insight Series

Allele Frequency Filters

Why “Rare” Means Rare in the People Who Happened to Be Sequenced

Kanna Nandakumar, PhD · Zetobit, LLC

Every rare-disease pipeline runs the same filter early and hard. Drop anything above some allele frequency in gnomAD — 1%, 0.1%, or a gene-specific threshold — and a few million variants become a few hundred. It is the single highest-leverage step in the whole workflow, and it is usually one line of configuration.

The line encodes an assumption that is easy to miss: that the frequency in the database is a good estimate of the frequency in the patient. That holds when the patient resembles the people who were sequenced. When they do not, the same threshold behaves like a different, weaker filter — and it does so silently, because a variant that survives filtering looks exactly like a variant that deserved to.

This is a distinct problem from two the series has covered before. The population-stratification piece was about association testing, where ancestry differences between cases and controls manufacture a signal. The polygenic-score portability piece was about a score losing accuracy with genetic distance from its training data. Both concern research cohorts. This one concerns a single patient in a diagnostic pipeline, and the mechanism is not a statistical model at all — it is a lookup.

What the database is

gnomAD is the largest resource of its kind and has grown enormously. Version 4 aggregates 807,162 individuals — 730,947 exomes and 76,215 genomes — contributed by 308 data contributors from over 25 countries, and adds roughly three-fold more global diversity than earlier releases, with about 169,000 individuals of non-European genetic ancestry¹.

Those are real and substantial gains. They also sit alongside a compositional fact: v4 incorporated 416,555 individuals from UK Biobank, which increased the European proportion relative to previous releases². The database got much bigger and, on that axis, somewhat less balanced at the same time.

The consequence is that the resolution of the filter varies by group. For a European-ancestry patient, the denominator is enormous, so a variant at 0.5% frequency in that population is seen hundreds of times and removed. For a patient from a population represented by a few thousand individuals, a variant at the same true frequency may appear a handful of times or not at all — and the filter keeps it.

The threshold is a property of the pipeline. What it actually removes is a property of who was sequenced.

The measurement that makes this concrete

The gnomAD team quantified this directly when v4 was released, and the result is the clearest statement of the problem I know of.

Despite a nearly eight-fold growth in European samples, there was limited growth in the number of variants whose European allele frequencies became high enough to eliminate them as candidate Mendelian disease variants. In contrast, the three-fold increase in non-European samples produced major growth: 228,718 variants reached GroupMax allele frequencies above 0.1%, and 1,210,431 above 0.01%, in non-European populations — allowing most of them to be treated as likely benign in genes causal for Mendelian disease¹.

Read that as a statement about the previous release. Before those samples were added, more than a million variants that could have been filtered were not being filtered — and they were not being filtered specifically for patients from those populations. Adding European samples barely moved the number, because the European part of the frequency spectrum was already well measured. The filter had already converged for some patients and had not for others.

The clinical case, in full

The canonical demonstration is a 2016 study of hypertrophic cardiomyopathy that traced the failure from a database gap to reports issued to patients.

The authors identified variants previously considered causal for HCM that were overrepresented in the general population, then reviewed nearly a decade of records at a major genetic testing laboratory. Multiple patients — all of African or unspecified ancestry — had received positive reports based on variants classified as pathogenic under the understanding at the time. All of those variants were subsequently recategorised as benign, and the variants most common in the general population were significantly more common among Black Americans than White Americans³.

The counterfactual is the part worth dwelling on, because it establishes that this was a sampling problem rather than an inherently hard scientific question. Using the lower-frequency of the two variants that were actually misclassified in patients — MYBPC3 G278E, with a minor allele frequency of 0.0157 in Black Americans against 0.000122 in White Americans — the authors calculated that even if Black Americans made up just 10% of a control cohort, a cohort of only 200 people would have given a 50% chance of correctly ruling out pathogenicity³.

Two hundred people. The original studies were not underpowered in the ordinary sense; they were composed in a way that made a common variant look rare.

And the damage propagates past the index patient. A misclassification invalidates the risk assessments performed in relatives, requiring amended reports and revised management plans; relatives without the variant receive false reassurance; and for a patient with ambiguous clinical features, a spurious pathogenic call can shift the perceived benefit of an implantable defibrillator³.

Figure 1. Both directions are consequences of the same gap. The first produces a false candidate that survives to interpretation; the second produces an unresolvable one. Only the first is usually discussed, and only the second shows up in aggregate statistics.

The other direction, which is more common

A false pathogenic report is the dramatic failure. The quieter and far more frequent one is a variant that cannot be cleared.

The ACMG/AMP framework provides stand-alone benign evidence (BA1) and strong benign evidence (BS1) on the basis of allele frequency. Both require the database to show the variant is common somewhere. If the population in which it is common is thinly represented, neither criterion can fire, and the variant remains a VUS — not because the evidence points to pathogenicity, but because the evidence that would exonerate it was never collected.

This shows up consistently in outcomes. Elevated VUS rates among patients of non-European genetic ancestry are a repeated finding across hereditary cancer panels and other settings, and the cause is attributed to underrepresentation in the population databases⁴. One analysis of admixed populations found the mechanism directly: comparing a Brazilian cohort against reference populations, 43% of shared variants showed significantly different allele frequencies, and 20 VUS had a high enough frequency in the Brazilian cohort to exceed the BS1 benignity threshold while remaining rare in other populations⁵.

Twenty variants sitting in the uncertain category for want of a denominator. The evidence needed to resolve them exists in the world; it had not been sequenced into a reference.

The clinical cost is not neutral either. A VUS is not simply an absence of information: it carries documented anxiety, can alter self-perception, and occasionally drives unnecessary interventions such as risk-reducing surgery or intensified screening, particularly when non-genetics providers misread it⁴.

The statistical fix that already exists

There is a well-developed answer to part of this, and it is underused relative to how good it is.

The problem with a raw allele frequency is that it is a point estimate presented without its uncertainty. Two alleles observed in 500 people and two observed in 500,000 produce very different frequencies, but a threshold comparison treats both as measurements of equal standing. What you actually want to know is whether the variant is confidently common enough to exclude.

The filtering allele frequency does exactly that: it reports the frequency for which the upper bound of the 95% confidence interval on the observed allele count is compatible with the observation⁶. Because the tightness of the underlying distribution depends on sample size, the approach is automatically more conservative when the allele number is low — which, as its authors note, avoids inappropriately filtering variants on the basis of a chance observation of a few alleles in a smaller subpopulation⁷.

That is precisely the guard rail this problem needs, and it is already computed and displayed in gnomAD. It also pairs with a disease-specific threshold rather than a generic one: the maximum credible population allele frequency, calculated as prevalence times genetic heterogeneity divided by penetrance⁸. Applied to cardiomyopathy, this reduced the number of candidate variants per exome by roughly two-thirds⁷ — a better filter, not merely a safer one.

Two honest limits. First, gnomAD does not compute a filtering allele frequency for non-continental groupings, nor where a continental population has only a singleton observation⁸ — the method declines to answer where the data are thinnest, which is correct behaviour but means the hardest cases return no guidance rather than reassurance. Second, being statistically careful about a small sample does not create information: a group with 3,000 individuals sequenced supports a weaker claim than one with 300,000, and no confidence interval changes that.

What the same filter does for different patients

Situation	What the database shows	What the filter does	Result
Well-represented population, common variant	Frequency estimated from a large denominator	Removes it	Correct exclusion
Underrepresented population, locally common variant	Few or no observations	Keeps it	False candidate reaches interpretation
Underrepresented population, benign variant under review	Insufficient frequency evidence	Cannot apply BA1 or BS1	Variant stays a VUS
Admixed individual	Group-level frequency averages over ancestries	Applies a frequency that may fit neither component	Unpredictable in either direction
Any population, small subgroup, chance observation	Two or three alleles in a small denominator	Raw AF may exceed the threshold	Real variant filtered on noise

The last row is the failure in the opposite direction, and it is why the filtering allele frequency is the right instrument rather than a raw frequency with a lower cut-off. Note also that gnomAD reports genetic ancestry groups, not races or ethnicities; the mapping between a patient's self-reported background and a database grouping is itself approximate, and least reliable for admixed individuals.

The admixture problem underneath

The group labels are a convenience, and treating them as homogeneous is its own source of error. Applying local ancestry inference to over 27 million variants in gnomAD's Admixed American (n = 7,612) and African/African American (n = 20,250) groups, one analysis found that 78.5% and 85.1% of variants respectively showed at least a two-fold difference in ancestry-specific allele frequencies within the group⁹.

So even the group-level frequency an admixed patient is compared against is an average over components that differ substantially. The label is a coarse summary of a continuum — which is the same structural point the polygenic-score piece made about population labels, arriving here from the filtering side rather than the prediction side.

What to do about it

1. Use the filtering allele frequency, not the raw AF. It is already in gnomAD, it accounts for sample size, and it is the single highest-value change available here⁶⁷.
2. Set the threshold from the disease, not from habit. Prevalence, heterogeneity and penetrance give a maximum credible population allele frequency; a generic 1% is an order of magnitude too permissive for many conditions⁸.
3. Record the allele number behind every filtering decision. “Absent from gnomAD” means something different at AN = 1,000,000 than at AN = 4,000, and only one of those is evidence.
4. Check the patient's likely ancestry group against the database depth for that group before trusting a negative filter result. If the relevant denominator is small, the filter did less work than it appears to have done.
5. Treat “no frequency evidence” as unmeasured, not as benign or as suspicious. It is the absence of a measurement, and reporting it as either is an inference the data did not supply.
6. Flag candidates that survived only because the relevant population is thinly sequenced — they warrant an explicit note in the report rather than equal footing with the rest of the list.
7. Consult population-specific resources where they exist, and treat their absence for a given population as a known limitation of the analysis rather than a silent one.
8. Record the database version with the result. A variant unfilterable in one release becomes filterable in the next when the relevant samples arrive — over a million variants crossed that line between gnomAD releases¹ — so this is a strong candidate for periodic reanalysis.
9. Say what the filter did in the report. A methods line giving the threshold, the database version and the ancestry group used converts an invisible assumption into a stated one.

The shape of the error

The recurring structure in this series is a correct measurement carrying an inference it does not support. This case has an unusual property: the inference fails differently for different patients, using the identical code path.

Nothing in the pipeline varies. The threshold is the same, the lookup is the same, the database is the same file. What varies is how much information that file contains about the person in front of you — and the output format has no field for that. A candidate list does not record which entries survived because they are genuinely rare and which survived because nobody like the patient has been sequenced in quantity.

The encouraging part is that this is the most tractable problem the series has covered. The statistical instrument exists and is already computed. The databases are growing, and the growth is measurably doing the work — over a million variants became filterable for non-European patients in a single release. The gap is closing.

Until it closes, a rarity filter is not a statement about the population. It is a statement about the sample, and the sample was not assembled with every patient in mind.

References

1. Analysis of >800,000 diverse sequenced humans in gnomAD improves clinical interpretation and provides insight into gene function. Genetics in Medicine Open 2024 (conference abstract; gnomAD v4 release figures). Author list not fully captured — verify before publication. gimopen.org — S2949-7744(24)00621-6
2. Introducing gnomAD v4's enhanced data. Golden Helix (vendor blog) — cited only for the UK Biobank contribution figure and its effect on European proportion; a primary source would be stronger. goldenhelix.com — gnomAD v4 curation
3. Manrai AK, Funke BH, Rehm HL, et al. Genetic misdiagnoses and the potential for health disparities. New England Journal of Medicine 2016;375(7):655–665. nejm.org/doi/full/10.1056/NEJMsa1507092
4. Ancestry-informative markers and variants of uncertain significance on hereditary cancer panels. Frontiers in Oncology 2026. pmc.ncbi.nlm.nih.gov/articles/PMC13167527
5. Variant classification of hereditary cancer genes is affected by genomic underrepresentation of admixed populations. Molecular Genetics and Genomics 2025. link.springer.com — s00438-025-02295-x
6. Filtering allele frequency — gnomAD browser documentation. gnomad.broadinstitute.org/help/faf
7. Whiffin N, Minikel E, Walsh R, et al. Using high-resolution variant frequencies to empower clinical genome interpretation. Genetics in Medicine 2017;19(10):1151–1158. nature.com/articles/gim201726
8. Overview of specifications to the ACMG/AMP variant interpretation guidelines (ClinGen; BA1/BS1 thresholds, FAF caveats, maximum credible population allele frequency). pmc.ncbi.nlm.nih.gov/articles/PMC6885382
9. Improved allele frequencies in gnomAD through local ancestry inference. Nature Communications 2025. nature.com/articles/s41467-025-63340-2

Zetobit, LLC · Bioinformatics consulting · Lexington, KY · zetobit.com