What the Million Veteran Program’s 459,777-Person DNA Dataset Shows About Ancestry and Data Quality

The Million Veteran Program DNA dataset quality control took an initial tranche of 485,856 people and produced a high-quality set of 459,777 unique individuals with 668,418 genetic markers. Nearly 30% of those people have non-European ancestry. MVP's substantial ancestral diversity surpasses that of other large biobanks. A genotype from this cohort becomes more interpretable when read next to a VA electronic health record and the population it was called in.

Million Veteran Program DNA dataset quality control, ancestry and diversity

The MVP is run by the Department of Veterans Affairs. At the time of this paper, blood samples had been collected from over 800,000 enrolled participants. The genotyping was array-based, using the MVP 1.0 custom Axiom array designed as a single assay for a multi-ethnic cohort. Array-based means the chip tests a fixed set of positions across the genome, not every base. That set has to be chosen so it works across different ancestry groups.

A genotype means more when it is read next to a health record and the population it was called in.

The 459,777-person dataset is large, but the diversity is the harder part. A variant that is common in one population can be rare or absent in another. If a study is mostly one ancestry, an association can reflect population history instead of biology. The MVP paper tested this by replicating established genetic associations with height in European Americans and African Americans. Seeing the same established signal in both groups is a sign the data and population context are working.

Why nearly 30% non-European ancestry changes the result

Ancestry is not a label you can ignore. It changes the allele frequency at thousands of markers. It changes how likely a rare variant is to appear by chance. When researchers run a genome-wide association study, they adjust for ancestry because otherwise a variant that tracks with ancestry will look associated with any health outcome that also tracks with ancestry.

In the MVP dataset, the abstract reports that non-European individuals make up nearly 30%. That is enough to ask whether a finding replicates across groups, not just within one. The height association result is one clear example in the paper. Height is well studied, highly heritable, and has well-established associated markers. Failure to reproduce established associations would raise QC concerns.

What QC actually filters in 485,856 samples

QC in a genotype study is not a single pass. It is a series of filters, and the numbers that survive are the practical summary. The abstract does not break down the exact thresholds. The categories below are standard examples of genotype QC filters, not a paper-specific list taken from the abstract. The full paper supports the specific MVP thresholds and any additional checks.

• Sample call rate: how many of the array's markers were successfully read for a person.
• Marker call rate: how often a specific marker was read across all people.
• Ancestry-aware checks: whether a marker behaves consistently across populations.
• Unique sample count: removing duplicate or unexpected related samples.

The starting tranche was 485,856 individuals. The final high-quality set was 459,777 unique individuals. The difference comes from samples that did not meet the QC bar. The paper does not break down every reason in the abstract. On the marker side, 668,418 markers passed QC. That count includes common variants and rare variants. Rare variants are harder to call because fewer people carry them. A noisy probe can look like a rare allele. QC has to separate true rare variation from assay noise.

What a genotype looks like in real data

In a QC report, you would not see a single good or bad genotype. You would see a sample call rate, a marker call rate, and plots that show whether samples cluster by population. In a cohort this size, one bad sample is a tiny fraction. One bad marker is more dangerous because it can be systematically misread in cases or controls. Marker-level QC matters more than raw sample size. Size alone does not fix a probe that fails in one ancestry group.

The marker count matters because it is the set that survived the QC filters and could be trusted for downstream analysis.

Common mistake: a DNA kit result is not a fixed readout

People often treat a genotype from a DNA kit as a stable, self-contained fact. It is not. The same DNA tested on a different array may be called at a different set of markers. A result can change after imputation, after ancestry adjustment, or when moved between genome builds. The MVP dataset is high quality because both the array design and the QC were tuned to a diverse cohort.

The genotype becomes more useful next to the person's health record. In MVP, that means VA electronic health records: diagnoses, lab values, pharmacy data, and follow-up visits. Without context, a genetic marker is just a position in a file.

Can I apply this to a raw 23andMe or AncestryDNA file?

Sort of. A consumer raw DNA file is usually a subset of the same kind of array data, but it has not been through MVP-level unified QC. The file contains the markers called in your sample, not a clinical interpretation. You can look at the markers you have and see whether they overlap with known variants, but that is not the same as MVP-level QC. For a personal file, you would still need to check which markers and genome build were used.

What the evidence does not tell you

This paper describes the genotyping array design and quality control for the MVP 1.0 release. It is not a clinical test. It is not whole-genome sequencing for every participant, though MVP was doing sequencing and other assays in parallel. The data were made available to approved MVP researchers. Further releases were expected as recruitment continued. The exact QC thresholds and the full ancestry breakdown are in the full paper, not the abstract. For personal decisions, treat a research cohort result as evidence, not a diagnosis.

HealthOS call-to-action: if you have a raw DNA file, see https://healthosx.com/dna.