What DNA Upload Sites Can Tell You, and What They Cannot

Oak
dna genomics arrays imputation bioinformatics
A lab instrument scans a huge dark grid of tiny wells where only scattered points glow bright and others glow faintly.

If you are asking which genetic site is better than Ancestry, the answer is that the site matters much less than the file you give it. Every consumer upload service (GEDmatch, Genomelink, Sequencing.com, MyHeritage, FamilyTreeDNA) receives the same thing: a text file of roughly 600,000 to 700,000 genotype calls from a SNP array, which samples about 0.02 percent of your 3.1 billion base pairs at positions chosen for population genetics rather than for your biology. GEDmatch remains the best of them for genealogy because its one-to-many matching and segment triangulation tools work across databases. For anything beyond relative matching and admixture estimates, the limiting factor is the array, and the fix is sequencing rather than a different website. If you have a clinical question, the path runs through a clinician who can order a validated test, not through an upload site.

What is inside the file you upload

The 23andMe export is a tab-delimited text file with a comment header and four columns: rsid, chromosome, position, genotype. Positions are on GRCh37 (hg19). Genotypes are given as two characters on the plus strand of the reference, -- for no-calls, and I/D for the handful of insertion and deletion probes, which carry no information about what was inserted or deleted. The v5 chip is built on the Illumina Global Screening Array, so most markers are common variants with a minor allele frequency above one percent, plus a curated set of pharmacogenetic and ancestry-informative sites.

The AncestryDNA export has five columns instead, splitting the call into allele1 and allele2, and uses 0 rather than - for missing data. It is also on GRCh37. MyHeritage and FamilyTreeDNA emit CSV variants of the same idea. None of these files contain read data, quality scores, or coverage, so there is no way to re-examine a surprising call. You get one genotype per probe and a manufacturer’s claim about accuracy.

Two things follow. First, a variant absent from the file is not absent from you: it was simply never assayed. Second, the concordance between any two upload sites for the same file will be near perfect, because they are all reading the same 600,000 calls. Differences in their reports come from reference panels and modeling choices, not from new measurement.

What the sites do with it, and how to do the same thing yourself

Three analyses dominate: relative matching, admixture estimation, and imputation-backed trait reports. All three are reproducible locally, and running them yourself tells you how much of the output is data and how much is model.

Start by converting to VCF against an explicit reference so that build and strand are pinned down:

bcftools convert --tsv2vcf genome.txt \
  -f human_g1k_v37.fasta -s ME -Oz -o me.b37.vcf.gz
bcftools index me.b37.vcf.gz
bcftools +fixref me.b37.vcf.gz -Ob -o me.fixed.bcf -- -f human_g1k_v37.fasta -m flip

+fixref reports how many sites had reference and alternate alleles swapped or needed a strand flip. If that count is large, your input file was not on the build you assumed. Lift to GRCh38 only when your downstream reference panel requires it, using CrossMap with the UCSC chain file, and expect to lose a few thousand sites that fail to map uniquely.

For quality control and population structure, plink 1.9 reads the 23andMe format directly:

plink --23file genome.txt ME ME --snps-only --make-bed --out me
plink --bfile me --missing --het --check-sex --out me.qc

A no-call rate above roughly two percent, or a heterozygosity rate far from the cohort norm, means the array run was marginal and every downstream report inherits the noise. For ancestry, merge with the 1000 Genomes phase 3 genotypes on the intersecting sites, prune for linkage disequilibrium with --indep-pairwise 50 5 0.2, and project yourself onto the reference principal components. The 1000 Genomes integrated map is the panel almost every consumer admixture engine ultimately depends on, directly or through its successors 1. Your position on PC1 and PC2 will be stable and informative at continental scale. The percentages in a commercial pie chart are model output conditioned on which populations the vendor chose to include, which is why two sites give you different numbers from identical genotypes. Fine-scale claims about ancient populations rest on a small number of sequenced ancient genomes and shift as new ones are published 2.

Imputation is where an array file gains the most. Submit your VCF to the TOPMed or Michigan Imputation Server, which phases with Eagle and imputes with Minimac4 against a panel of roughly a hundred thousand sequenced genomes. You will get back dosages at hundreds of millions of sites, each with an estimated squared correlation, R2, in the INFO field. Filter hard:

bcftools view -i 'INFO/R2>0.8' chr1.dose.vcf.gz -Oz -o chr1.filt.vcf.gz

Common variants impute well. Rare variants, meaning minor allele frequency below about half a percent, impute poorly, and those are exactly the variants that carry large effects on protein function. Imputed data is appropriate for polygenic scores and population comparisons. It is not a substitute for sequencing a specific gene, and no reputable pipeline treats an imputed genotype as a clinical result.

Where arrays fail

The failure modes are structural, not fixable by better software. Arrays genotype only preselected sites, so a typical genome’s millions of variant positions, including thousands predicted to alter protein sequence, are mostly unobserved 1. Structural variation is invisible: copy number changes, inversions, and repeat expansions are not on the chip. Mobile element content varies substantially between individuals, and characterizing it requires sequencing reads rather than probe intensities 3. Mitochondrial heteroplasmy, which is a quantitative measure of the fraction of mitochondrial genomes carrying a variant, cannot be represented by a diploid genotype call at all.

There is also a probe design problem at rare sites. A probe optimized for a common polymorphism can misbehave when the true underlying sequence is unusual, and because the array reports a hard call with no supporting reads, a false positive looks identical to a true one. This is the reason consumer arrays repeatedly produce spurious calls in well-studied disease genes, and the reason that any finding from an upload report needs confirmation by a clinician-ordered, validated assay before it means anything about you.

Even the well-behaved cases illustrate how thin the data is. APOE, mapped to chromosome 19 four decades ago, is usually captured by two probes at rs429358 and rs7412, which resolve the common isoforms and nothing else about the locus 4. Reports about Neanderthal-derived variation are similarly built from a modest set of tagged sites, though the underlying biology is real enough: introgressed haplotypes show measurable effects on lipid catabolism in Europeans 5.

What we would do instead

We would sequence once, deeply, and keep the primary data. Thirty-fold whole-genome sequencing gives you a CRAM or BAM you can revisit, a gVCF that distinguishes “reference” from “not covered”, and access to indels and structural variants that arrays cannot see. Keep the alignment, not only the VCF, because reanalysis against an improved reference or a new variant caller is then a compute job rather than a new blood draw. The physical and genetic maps that made coordinate-based reference genomes possible took years of consortium work to build 6 7, and the practical payoff is that your reads stay interpretable as those resources improve.

Sequence also tells you almost nothing about state. Your genome is fixed, and static variant interpretation is why array reports feel thin even when correct. The measurements that change on the timescale of weeks are transcript abundance, plasma protein levels, standard blood chemistry, and glucose dynamics. Study designs that combine genotype with tissue-level molecular readouts, as cancer genome projects did by pairing sequence with expression and methylation, recover mechanisms that genotype alone cannot 8. The same logic applies to one person measured repeatedly.

Questions people also ask

Is there a site better than Ancestry? For genealogy, GEDmatch gives you more analytical control over the same genotypes, including one-to-one segment comparison and multiple admixture calculators you can compare against each other. For biology, no consumer site is better, because they all read the same array file, and the improvement comes from changing the input to whole-genome sequence.

What are the most popular DNA sites? By database size, AncestryDNA and 23andMe dominate direct testing, with MyHeritage and FamilyTreeDNA next, and GEDmatch serving as the main cross-vendor upload and matching hub. Genomelink and Sequencing.com sit on top of uploads and sell interpretation layers rather than measurement.

Is 23andMe or Ancestry more popular? AncestryDNA has the larger genotyped database and therefore finds more relatives, while 23andMe historically shipped more health-oriented reports. For raw data quality the two are comparable, since both use Illumina arrays of similar density on GRCh37.

Can I get a clinical answer from a raw data upload? No. Upload reports are unvalidated interpretations of unvalidated calls, and a variant that appears in one should be confirmed by a clinician-ordered diagnostic test before it informs any decision. Family-based sequencing studies show how much careful design goes into attributing a phenotype to a locus even with full sequence in hand 9.

Oak builds longitudinal molecular profiles of individuals: whole-genome sequencing, RNA sequencing, proteomics, blood biomarkers, and continuous glucose data, integrated into one model of you. Build your profile.

Footnotes

  1. The 1000 Genomes Project Consortium. An integrated map of genetic variation from 1,092 human genomes. Nature, 2012. https://doi.org/10.1038/nature11632 ↩ ↩2

  2. Maanasa Raghavan, Pontus Skoglund, Kelly E. Graf, et al. Upper Palaeolithic Siberian genome reveals dual ancestry of Native Americans. Nature, 2013. https://doi.org/10.1038/nature12736 ↩

  3. Adam D. Ewing, Haig H. Kazazian. High-throughput sequencing reveals extensive variation in human-specific L1 content in individual human genomes. Genome Research, 2010. https://doi.org/10.1101/gr.106419.110 ↩

  4. H K Das, J McPherson, G A Bruns, et al. Isolation, characterization, and mapping to chromosome 19 of the human apolipoprotein E gene. Journal of Biological Chemistry, 1985. https://doi.org/10.1016/s0021-9258(18)88963-3 ↩

  5. Ekaterina E. Khrameeva, Katarzyna Bozek, Liu He, et al. Neanderthal ancestry drives evolution of lipid catabolism in contemporary Europeans. Nature Communications, 2014. https://doi.org/10.1038/ncomms4584 ↩

  6. P.S Harper. Mapping and sequencing the human genome. Endeavour, 1989. https://doi.org/10.1016/0160-9327(89)90035-5 ↩

  7. Maynard Olson, Leroy Hood, Charles Cantor, et al. A Common Language for Physical Mapping of the Human Genome. Science, 1989. https://doi.org/10.1126/science.2781285 ↩

  8. The Cancer Genome Atlas Research Network. Comprehensive genomic characterization defines human glioblastoma genes and core pathways. Nature, 2008. https://doi.org/10.1038/nature07385 ↩

  9. Nandita Mukhopadhyay, Madison Bishop, Michael Mortillo, et al. Whole genome sequencing of orofacial cleft trios from the Gabriella Miller Kids First Pediatric Research Consortium identifies a new locus on chromosome 21. Human Genetics, 2019. https://doi.org/10.1007/s00439-019-02099-1 ↩