What gnomAD v5 Will Change, and How to Work With v4 Today
gnomAD v5 has not been released. The gnomAD team has stated that it is planned for 2026 and that its defining change is the addition of All of Us Research Program whole genomes to the aggregate. That matters most for the genome half of the database. gnomAD v4.1 contains 730,947 exomes and 76,215 genomes, and All of Us has released a comparable order of magnitude of short-read genomes on its own. If those samples land in the aggregate, the genome-only callset roughly multiplies. That improves frequency estimates in introns, promoters, and untranslated regions where exomes give you nothing. Until then, v4.1 on GRCh38 is the correct reference, and the annotation pipeline you build for v4.1 will carry over to v5 with a swapped input file.
What has been announced, and what has not
The concrete public statements are narrow. A v5 release is anticipated in 2026, it includes All of Us data, and discussion of it happens on the gnomAD forum. Everything else is inference, so treat it as such.
Two inferences are safe. The build will be GRCh38, because v4 already is and All of Us genomes are aligned to GRCh38. And the release will be aggregate frequencies rather than individual-level data, because that is what gnomAD has always published and what the contributing consents permit.
One thing worth planning for is that All of Us is deliberately enriched for groups underrepresented in genomics. If those samples enter the aggregate, the ancestry composition of the genome callset shifts, and so do the per-group allele numbers. Any threshold you have tuned against v4 group frequencies should be re-derived rather than assumed.
What v4.1 already gives you
Before worrying about v5, get v4.1 wired in properly, because most of the value is already there. The sites VCFs are published per chromosome as bgzipped, tabix-indexed files. They are hosted on Google Cloud Storage, AWS Open Data, and Azure. There are separate exome and genome releases, plus a joint release that reports combined frequencies and joint filtering allele frequencies across both.
Use the joint release for allele frequency lookups when you can. It avoids the common error of quoting an exome frequency for a variant that is well covered in genomes and poorly covered in exomes, or the reverse.
Alongside the VCFs, three v4 products are worth pulling down separately. The constraint table gives per-transcript LOEUF, the upper bound of the observed-to-expected ratio for predicted loss-of-function variants, which is the standard way to ask whether a gene tolerates being broken. The structural variant callset covers deletions, duplications, and insertions that a short-read SNV pipeline will miss. The mitochondrial callset released in the v3 line reports heteroplasmy-aware frequencies for chrM, where nuclear-style allele frequency is the wrong model.
Annotating your own genome against gnomAD
For a single personal genome, the practical question is how to attach gnomAD frequencies to a few million variants in your own VCF without an afternoon of compute. We use echtvar for this and would recommend it over the alternatives.
echtvar encodes the gnomAD sites VCFs into a compressed archive keyed by position and allele, then annotates in a single streaming pass. Encoding is the slow step and you do it once:
echtvar encode gnomad.v4.1.joint.zip conf.json \
gnomad.joint.v4.1.sites.chr*.vcf.bgz
with a conf.json that names the fields you want carried over, for example AF_joint, AN_joint, and AF_grpmax_joint, each with an alias and an integer multiplier for fixed-point encoding. Then annotation of your own calls is fast:
echtvar anno -e gnomad.v4.1.joint.zip my.deepvariant.vcf.gz annotated.bcf
The alternatives have their places. bcftools annotate -a with the sites VCF works and needs no extra tooling, but it is slow across the whole genome and you have to be careful with multi-allelic records. Normalize first with bcftools norm -m-any -f GRCh38.fa so your alleles match gnomAD’s representation. VEP with the --custom flag gives you consequence prediction and frequencies in one run, which is what you want if you also need LOFTEE to flag low-confidence loss-of-function calls. vcfanno is a reasonable middle option if you are already annotating from several sources with a TOML config.
Whichever you choose, carry AN through. Do not carry AF alone. Allele number is your coverage proxy. A variant with AF = 0 and AN = 200 in a cohort of hundreds of thousands tells you nothing about rarity, and mistaking sparse coverage for absence is one of the recurring failure modes in population-database interpretation 1. A reasonable default is to ignore frequency at sites where AN is below half the cohort maximum for that chromosome.
The frequency thresholds that matter
Allele frequency is the single most useful filter you have, and gnomAD publishes the statistic designed for it. The filtering allele frequency is the upper bound of a 95% confidence interval on the frequency in the genetic group where the variant is most common. Using that upper bound rather than the point estimate is what keeps a variant seen twice in a small group from being dismissed as common 1.
The reverse error is more interesting and more common in practice. Variants reported in the literature as pathogenic often appear in gnomAD at frequencies far too high for the disease they are supposed to cause. Systematic review of reportedly pathogenic alleles for treatable pediatric conditions found very few at appreciable frequency in gnomAD, and much of what remained came apart under curation 2. Frequency is strong evidence against a high-penetrance claim, and it is worth applying to any variant your pipeline flags.
It does not settle causality on its own. A variant can stay absent or near-absent from gnomAD and still not be the cause of anything, and a disputed claim can survive reassessment against larger population data, as happened with the RP9 p.(H137L) allele in dominant retinitis pigmentosa 3. Where a variant bears on your own health, the interpretation belongs with a clinical geneticist working from a validated assay. Your annotation pipeline is not the place for it.
It also helps to know what gnomAD is a sample of. It excludes individuals with severe pediatric disease but is otherwise not a healthy cohort, which is why purpose-built reference sets of screened healthy elderly exist as a complement 4.
Where gnomAD will still be thin after v5
Adding All of Us improves ancestry representation without solving it. Population-specific resources keep surfacing common local variants that global aggregates estimate poorly, from deep whole-genome sequencing of Japanese cohorts 5 to clinical exome sets assembled specifically to prioritize variants in Indian patients 6. If your ancestry is underrepresented, expect a longer list of rare-looking variants that are ordinary where your grandparents came from.
The second gap is structural. For any given rare missense variant, the per-site population evidence is often a single count or zero, which carries almost no information. Methods that pool signal across homologous protein domains, as MetaDome does, give you a tolerance profile along the protein where site-level frequency is silent 7. This kind of evidence is complementary to gnomAD and does not improve when the cohort grows.
What to redo when v5 lands
Keep your annotation reproducible so the upgrade is cheap. Concretely, store your own calls as a normalized gVCF or VCF against GRCh38 and treat all gnomAD fields as derived. Re-encode the archive, re-run echtvar anno, and diff the two annotated outputs on the fields you filter with.
The diff is the useful artifact. Expect movement in three places: non-coding variants that had genome-only support, sites where AN rises above your coverage floor for the first time, and grpmax frequencies for groups that All of Us enlarges. Variants that cross your filtering allele frequency threshold in either direction are the ones to look at by hand.
Questions people also ask
When will gnomAD v5 be released? The gnomAD team has said it is anticipated in 2026, with All of Us data included. No date, sample count, or file manifest has been published. Build against v4.1 and plan for a swap.
Should I use v2 or v4? Use v4.1 unless you have a specific reason not to. v2 is on GRCh37 and much smaller. The main reason people still touch v2 is legacy constraint metrics or a region where v4 coverage is worse, and in that case lift over the variant rather than the whole callset.
What is the difference between the exome, genome, and joint releases? Exomes cover coding regions in far more people. Genomes cover the whole genome in fewer. The joint release combines both with consistent frequency and filtering allele frequency fields, and is the right default for lookups.
Does absence from gnomAD mean a variant is pathogenic? No. Most rare variants in any individual genome are absent from gnomAD and consequence-free. Absence removes one line of evidence against pathogenicity and supplies none in favor 1.
Footnotes
-
Sanna Gudmundsson, Moriel Singer‐Berk, Nicholas A. Watts, et al. Variant interpretation using population databases: Lessons from gnomAD. Human Mutation, 2021. https://doi.org/10.1002/humu.24309 ↩ ↩2 ↩3
-
Nina B. Gold, Steven M. Harrison, Jared H. Rowe, et al. Low frequency of treatable pediatric disease alleles in gnomAD: An opportunity for future genomic screening of newborns. Human Genetics and Genomics Advances, 2022. https://doi.org/10.1016/j.xhgg.2021.100059 ↩
-
Leon Chang, James A. Poulter, Andrew R. Webster, et al. RP9 revisited; RP9 p.(H137L) remains a likely cause of dominant splicing factor-Retinitis Pigmentosa. European Journal of Human Genetics, 2025. https://doi.org/10.1038/s41431-025-01964-0 ↩
-
Mark Pinese, Paul Lacaze, Emma M. Rath, et al. The Medical Genome Reference Bank contains whole genome and phenotype data of 2570 healthy elderly. Nature Communications, 2020. https://doi.org/10.1038/s41467-019-14079-0 ↩
-
Yukinori Okada, Yukihide Momozawa, Saori Sakaue, et al. Deep whole-genome sequencing reveals recent selection signatures linked to evolution and disease risk of Japanese. Nature Communications, 2018. https://doi.org/10.1038/s41467-018-03274-0 ↩
-
Neethukrishna Kausthubham, Anju Shukla, Neerja Gupta, et al. A data set of variants derived from 1455 clinical and research exomes is efficient in variant prioritization for early‐onset monogenic disorders in Indians. Human Mutation, 2021. https://doi.org/10.1002/humu.24172 ↩
-
Laurens Wiel, Coos Baakman, Daan Gilissen, et al. MetaDome: Pathogenicity analysis of genetic variants through aggregation of homologous human protein domains. Human Mutation, 2019. https://doi.org/10.1002/humu.23798 ↩
Pothos builds longitudinal molecular profiles of individuals: whole-genome sequencing, single-cell RNA sequencing, and plasma proteomics, delivered as a detailed report, every data file, and stored cells. Build your profile.