All of Us Genomics: What the Data Is, How to Access It, and What Participants Get Back

All of Us is an NIH research program that has sequenced hundreds of thousands of whole genomes from a deliberately diverse US cohort and makes the de-identified data available to approved researchers through a cloud-hosted Researcher Workbench. It is funded by the NIH (authorized under the 21st Century Cures Act of 2016, with the program growing out of the Precision Medicine Initiative announced in 2015) and began national enrollment in May 2018. Two things trip people up. First, if you join as a participant you do not get your CRAM or your VCF back. You get a curated set of reports covering hereditary disease risk, pharmacogenomics, ancestry, and a handful of traits. Second, if you get access as a researcher, you cannot download individual-level genomic data at all. You compute on it inside Google Cloud, in their environment, on your billing account.

What the program sequences

The core product is short-read whole-genome sequencing (the program abbreviates it srWGS). The libraries are PCR-free and run on Illumina NovaSeq at 150 bp paired-end. The target is mean coverage around 30x, with a quality bar requiring the large majority of callable bases at 20x or better. This is the same reversible-terminator chemistry that made population-scale human sequencing tractable in the first place ¹. Alignment is to GRCh38, including alt contigs, with per-sample gVCFs joint-called into a single cohort callset.

Alongside WGS, a large fraction of participants have Illumina Global Diversity Array genotypes (roughly 1.9M markers, chosen for coverage across non-European ancestries rather than tuned to European linkage disequilibrium). The array data exists partly for cost reasons and partly because it was generated earlier in the program’s history. For most analyses the WGS callset supersedes it. The array data is still useful for sample QC cross-checks and for participants who have array data but no genome.

The cohort composition is the part that matters scientifically. Reference panels and GWAS catalogs have been overwhelmingly European-ancestry for two decades, a bias that traces back to how the first dense SNP maps were built ². A majority of All of Us participants come from communities historically underrepresented in biomedical research, which changes what allele frequency spectra and polygenic scores look like when you recompute them here.

The callsets you will work with

Each data release (v7, v8, and onward) ships several derived products at different granularities. Pick the wrong one and you will either waste money or silently drop the variants you care about.

• Hail VariantDataset (VDS). The full joint callset, stored sparsely with reference blocks so that homozygous-reference regions are run-length encoded instead of materialized per sample. This is the complete data: every variant, every sample, plus reference confidence. It is the only correct starting point for rare-variant work.
• ACAF threshold callset. A filtered, dense Hail MatrixTable (also distributed as VCF and PLINK bed/bim/fam) restricted to variants above a population allele frequency or allele count threshold, roughly MAF > 1% or AC > 100 in any computed ancestry group. This is what you want for GWAS and for anything where you need a dense genotype matrix without paying to densify the VDS. It is the wrong input for burden tests, because the variants a burden test is built on are exactly the ones this callset removed.
• Exome-region and ClinVar subsets. Prefiltered to coding intervals or to ClinVar-annotated positions. Fast, cheap, and adequate for a lot of exploratory work.
• CRAM and gVCF per sample. Available in controlled-tier storage. You need these if you are doing anything alignment-dependent, such as structural variants, mitochondrial heteroplasmy, repeat expansions or custom realignment.

In the Workbench, paths are exposed as environment variables rather than hardcoded (WGS_VDS_PATH, WGS_ACAF_THRESHOLD_MULTI_HAIL_PATH, WGS_EXOME_SPLIT_HAIL_PATH, and similar), and the exact names drift between releases. Read the release notes before you copy a notebook from a previous version.

The standard VDS pattern is: read, filter intervals first, then densify.

import hail as hl, os
vds = hl.vds.read_vds(os.getenv('WGS_VDS_PATH'))
vds = hl.vds.filter_intervals(
    vds, [hl.parse_locus_interval('chr17:43044295-43125483',
                                  reference_genome='GRCh38')])
mt = hl.vds.to_dense_mt(vds)
mt = hl.split_multi_hts(mt)

Calling to_dense_mt() before narrowing the intervals is the single most common and most expensive mistake. You are asking a Dataproc cluster to materialize a matrix with hundreds of thousands of columns across three billion rows, and the bill lands on your account.

QC artifacts you should read before you compute anything

Each release publishes auxiliary tables that encode decisions you would otherwise have to make yourself and get wrong:

• Flagged samples: contamination estimates (freemix), call rate, coverage, and sex-ploidy outliers. These are not removed from the callset. Filtering is your job.
• Relatedness: KING-robust kinship estimates with a flagged-related-samples list, typically pruned at a kinship threshold around 0.1 (second-degree). The cohort contains genuine relatives, so ignoring this inflates association statistics.
• Genetic ancestry: predicted ancestry labels and principal components computed against a reference panel. Use the provided PCs as covariates rather than recomputing them on a subset, unless you have a specific reason.
• Sex at birth versus genetic sex ploidy. These are separate fields and they are not always concordant. Handle the discordance explicitly instead of coercing it.

Access tiers and what you can and cannot do

The Registered Tier gives you survey responses, EHR-derived data, and physical measurements. The Controlled Tier is where genomic data lives. It additionally requires that you be affiliated with an institution that has signed a Data Use and Registration Agreement, that you complete the responsible-conduct training, and that you verify your identity. Everything runs in a Jupyter or RStudio environment on Google Cloud inside the program’s perimeter.

Egress of individual-level data is blocked. You can export aggregate results and figures. You cannot pull a VCF to your laptop. New accounts get a modest amount of free cloud credit (on the order of a few hundred dollars), after which you attach your own billing account and pay for the Dataproc clusters and persistent disks you spin up. Budget accordingly: a genome-wide dense operation over the full cohort is a real expense. Such a run is not a warm-up.

What a participant receives

Participants can opt into health-related DNA results. The program returns them under an FDA investigational device exemption, with the reported variants confirmed in a CLIA-certified laboratory before anything goes out ³. In practice that means two report types. The first is hereditary disease risk across the ACMG secondary-findings gene set. That set covers BRCA1, BRCA2, Lynch syndrome genes and familial hypercholesterolemia genes, along with cardiomyopathy and arrhythmia genes and the rest. The second is pharmacogenomic star-allele calls for a small number of well-characterized genes. Results that indicate elevated risk are delivered with genetic counseling, which is the right design. A positive finding in one of these genes is a clinical matter and belongs in a conversation with a clinician and a genetic counselor. A spreadsheet is not the place for it.

You do not receive your CRAM, your gVCF, or your array intensities. All of Us is a research cohort that returns a bounded, clinically validated slice. If your goal is to hold and analyze your own molecular data, the program is not the mechanism.

That boundary is a considered one. Whole-genome sequencing generates findings the participant did not ask about, with uncertain penetrance and implications for relatives who never consented, and the literature on consent and incidental findings has argued for years that returning everything by default is not obviously the ethical choice ⁴. Reasonable people disagree. We think a technically fluent adult should be able to get the raw files, and we also think the program’s conservatism about returning unvalidated variant calls is defensible.

Questions people also ask

Who funds All of Us? The National Institutes of Health, with funding authorized in part by the 21st Century Cures Act of 2016. It is administered by the NIH Office of the Director, and the work is distributed across genome centers, biobanking, and data and research centers under NIH contracts and cooperative agreements.

Who started the All of Us Research Program? It began as the Precision Medicine Initiative Cohort Program, announced by President Obama in his 2015 State of the Union address, and was built out by the NIH under then-director Francis Collins. It was renamed All of Us and opened national enrollment on May 6, 2018.

Does All of Us do whole-exome sequencing? No. The program generates short-read whole-genome sequencing. There are exome-region subsets of the WGS callset for convenience, which are derived by interval filtering rather than by capture-based exome sequencing, so coverage characteristics differ from a true WES experiment.

Can I download All of Us genetic data? Individual-level genomic data cannot leave the Researcher Workbench. You analyze it in place on Google Cloud and export aggregate results only.

Can I get my own raw DNA data as a participant? No. Participants who opt in receive curated hereditary disease risk, pharmacogenomic, ancestry, and trait reports. Raw sequence files are not returned.

Footnotes

1. David Bentley, Shankar Balasubramanian, Harold Swerdlow, et al. Accurate whole human genome sequencing using reversible terminator chemistry. Nature, 2008. https://doi.org/10.1038/nature07517 ↩

2. The International SNP Map Working Group, Cold Spring Harbor Laboratories:, Ravi Sachidanandam, et al. A map of human genome sequence variation containing 1.42 million single nucleotide polymorphisms. Nature, 2001. https://doi.org/10.1038/35057149 ↩

3. Eric Venner, Donna M. Muzny, Joshua D. Smith, et al. Whole-genome sequencing as an investigational device for return of hereditary disease risk and pharmacogenomic results as part of the All of Us Research Program. Genome Medicine, 2022. https://doi.org/10.1186/s13073-022-01031-z ↩

4. Wim Pinxten, Heidi Howard. Ethical issues raised by whole genome sequencing. Best Practice & Research Clinical Gastroenterology, 2014. https://doi.org/10.1016/j.bpg.2014.02.004 ↩