Private DNA Testing: What It Costs, What It Proves, and How to Keep the Data Yours

Oak
genomics privacy sequencing variant-calling dna-testing
An iridescent feathered reptile-bird inside a mirrored glass vault on a plinth, lit by one hard studio light against black.

A DNA test is private when you hold the primary files (FASTQ or CRAM), the lab does not retain an identified copy for research or commercial use, and no interpretation happens on a server you do not control. Very few consumer products meet all three conditions. Most sell you a report and keep the data. If you want privacy in the sense a security engineer would recognize, the shape of the purchase is: pay out of pocket at a CLIA lab, ask in writing for raw data delivery and destruction of the lab copy after a fixed retention window, and run annotation locally against downloaded databases. Cost ranges from about $100 for a genotyping array to roughly $400-$1,500 for 30x whole-genome sequencing at a consumer-facing sequencing vendor, up to several thousand for a clinical-grade genome with a reported interpretation.

What “private” has to mean technically

Three separate properties get collapsed into one word.

Data custody: who holds a copy, in what format, for how long, and under what terms. A consumer genotyping company that gives you a downloadable TSV of 650,000 genotypes still holds the array intensity data and your account identity. Deleting your account usually does not retract data already shared with research partners under the consent you clicked. Consent to broad, future, unspecified research use is the hard part of genomic privacy, because the data outlives the study it was collected for.12

Re-identification resistance: essentially zero, for any real genome. Your genome is an identifier. Stripping your name from a VCF does nothing. Perturbation and k-anonymity do not survive contact with high-density genotype data, which is why the ethics literature treats whole-genome data as categorically different from ordinary clinical results.3

Kin exposure: your genome discloses roughly half of each parent’s and each child’s variants and a quarter of each sibling’s. Uploading your file to a public genealogy database exposes relatives who never consented. This is the property people most often ignore, and it is the one with the largest blast radius.1

The three tiers, and what each can resolve

Genotyping array ($60-$150). Illumina Global Screening Array or similar. You get ~650k-900k pre-selected sites in a TSV or 23andMe-style text file, plus whatever imputation the vendor runs against a reference panel. Good for ancestry, common-variant polygenic scores, and pharmacogenomic star alleles that the chip probes. It cannot see anything not on the chip, which means most rare variants, nearly all indels, and structural variation. Imputed genotypes at low allele frequency are unreliable and should never be treated as calls.

Clinical panel or exome ($300-$2,000). Targeted capture, high depth (often 100x+ on target), curated interpretation by a lab that has standardized its pipeline and its variant classification process. This is where interpretive standards do a lot of quiet work: the value of a clinical exome report comes from the lab’s classification machinery as much as from the sequencing chemistry.4 The tradeoff is that you get a filtered answer to a specific question and usually not the underlying data in usable form.

Whole-genome sequencing, 30x short read ($400-$3,000 depending on vendor and whether interpretation is included). About 90-100 GB as a BAM, 45-60 GB as a CRAM referenced to GRCh38, 4-5 million variants in a single-sample VCF. This is the only tier that gives you a substrate you can re-analyze in five years against better annotations. Sequencing cost per genome has fallen by orders of magnitude since the ethical debates about the “$1000 genome” were first framed.56

We would buy the 30x genome and take the CRAM. Arrays are cheap enough to be a rounding error but they lock you into someone else’s choice of which 700,000 positions matter.

Where the errors live

“Can DNA test results be wrong” has a precise answer per variant class.

At 30x WGS in GIAB high-confidence regions, modern callers (DeepVariant, DRAGEN) reach SNV F1 above 99.5%. Indel F1 is lower, roughly 99%, and degrades sharply in homopolymers longer than about 8 bp. Outside high-confidence regions, all bets are off: segmental duplications, the MHC, CYP2D6, PMS2 (because of its pseudogene), SMN1/SMN2, and repeat expansions like FMR1 and C9orf72 are not reliably resolved by 30x short reads regardless of caller. If a report claims a call in those loci from short-read WGS without orthogonal confirmation, treat it as a hypothesis.

Consumer array raw data is worse than people assume for rare variants. A 2018 analysis in Genetics in Medicine of direct-to-consumer raw genotype files reanalyzed by a clinical lab found roughly 40% of the variants flagged in a set of samples were false positives. Arrays cluster genotypes across a population of samples, and clustering fails at sites where almost every sample is homozygous reference.

The second error class is interpretive, not technical. Sequence a healthy person and you find variants of uncertain significance in abundance, and uncertainty is the modal result rather than the exception.7 Participants in sequencing studies experience that uncertainty in several distinct ways, and it does not resolve cleanly into “positive” or “negative.”8 A correct call attached to a wrong or overstated clinical claim is still a wrong result. Anything that would change a medical decision needs confirmation in a clinical lab and a conversation with a genetic counselor or physician. Research-grade and consumer-grade calls are not diagnostic.

A setup that keeps the data yours

Take delivery of FASTQ or CRAM, not just a VCF. Verify before you delete the source:

samtools quickcheck -v *.cram
md5sum -c checksums.md5
samtools flagstat sample.cram
samtools coverage sample.cram | awk '{print $1, $6, $7}'

Expect mean depth ≥30x and ≥95% of the genome at ≥10x. If the vendor delivered BAM, convert it:

samtools view -T GRCh38_full_analysis_set_plus_decoy_hla.fa -C -o sample.cram sample.bam

CRAM is reference-dependent, so archive the exact reference FASTA and its index alongside it. Losing the reference makes the CRAM unreadable.

Call variants yourself if you want to control the pipeline:

run_deepvariant --model_type=WGS \
  --ref=GRCh38.fa --reads=sample.cram \
  --output_vcf=sample.vcf.gz --output_gvcf=sample.g.vcf.gz --num_shards=32

Annotate offline. VEP with a local cache means no coordinates leave your machine:

vep --offline --cache --assembly GRCh38 --fasta GRCh38.fa \
  --custom clinvar_20250xxx.vcf.gz,ClinVar,vcf,exact,0,CLNSIG,CLNREVSTAT \
  --af_gnomade --everything --vcf -i sample.vcf.gz -o sample.vep.vcf.gz

Filter ClinVar hits by review status, not by CLNSIG alone. criteria_provided,_multiple_submitters,_no_conflicts and above is the floor we use. Single-submitter pathogenic assertions are frequently stale.

Storage: keep the CRAM on an encrypted volume (cryptsetup luksFormat) and back it up encrypted at rest with age before it touches any cloud bucket. Do not upload raw files to public matching databases, free “upload your raw data” interpretation sites, or general-purpose LLM APIs. If you want an AI agent to reason over the data, run the model locally or pass it only derived summaries you have chosen deliberately.

Legal note: GINA restricts use of genetic information by health insurers and employers in the US. It does not cover life, disability, or long-term care insurance. Law enforcement access to consumer genealogy databases via familial matching is established practice. Also relevant to what any vendor can offer you: the patent situation around genome-wide sequence analysis has been contested for years, which shapes which labs report which genes.9 The broader point Collins made early on still holds, that the societal consequences of genomic data arrive faster than the frameworks governing them.10

Questions people also ask

How much does a DNA test cost at Walgreens? Shelf paternity kits run about $30 for the collection kit plus a lab fee of roughly $130-$160 when you mail the swabs in. Those are relationship tests only. They produce a probability of paternity, not sequence data, and the at-home versions are not admissible in court because there is no chain of custody.

How much does a private DNA test cost? For a relationship test, $150-$200 at-home or $300-$500 for a legally admissible AABB-accredited version. For your own genome, $400-$1,500 for 30x WGS with raw data delivery, more if clinical interpretation is included.

What is the cheapest way to get a DNA test done? A genotyping array during a holiday sale, around $40-$60. It is cheap because you are getting 0.02% of your genome at pre-chosen positions, and because the business model assumes long-term data retention.

Can you trust DNA test results? Trust the SNV calls from 30x WGS in high-confidence regions. Be skeptical of indels in repetitive sequence, of anything in CYP2D6 or the MHC from short reads, and of rare-variant calls from arrays. Trust the interpretation only to the extent the lab documents its classification process and evidence.4

How accurate is a DNA test? Accuracy is per-variant-class, not per-test. A relationship test computing a combined paternity index from 16-21 STR markers gives >99.99% probability of paternity or exclusion. A genome gives you 99.5%+ SNV accuracy in most of the genome and much less in about 2-5% of it. There is no single number.

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. Bartha Maria Knoppers. Consent to ‘personal’ genomics and privacy. The EMBO Reports, 2010. https://doi.org/10.1038/embor.2010.69 ↩ ↩2

  2. Jeantine E. Lunshof (Co-first author), Jason Bobe (Co-first author), John Aach, et al. Personal genomes in progress: from the Human Genome Project to the Personal Genome Project. Dialogues in Clinical Neuroscience, 2010. https://doi.org/10.31887/dcns.2010.12.1/jlunshof ↩

  3. Holly K. Tabor, Benjamin E. Berkman, Sara Chandros Hull, et al. Genomics really gets personal: How exome and whole genome sequencing challenge the ethical framework of human genetics research. American Journal of Medical Genetics Part A, 2011. https://doi.org/10.1002/ajmg.a.34357 ↩

  4. Stefan Timmermans. Trust in standards: Transitioning clinical exome sequencing from bench to bedside. Social Studies of Science, 2014. https://doi.org/10.1177/0306312714559323 ↩ ↩2

  5. John A. Robertson. The $1000 Genome: Ethical and Legal Issues in Whole Genome Sequencing of Individuals. The American Journal of Bioethics, 2003. https://doi.org/10.1162/152651603322874762 ↩

  6. Rachel L Goldfeder, Dennis P Wall, Muin J Khoury, et al. Human Genome Sequencing at the Population Scale: A Primer on High-Throughput DNA Sequencing and Analysis. American Journal of Epidemiology, 2017. https://doi.org/10.1093/aje/kww224 ↩

  7. Noralane M. Lindor, Stephen N. Thibodeau, Wylie Burke. Whole-Genome Sequencing in Healthy People. Mayo Clinic Proceedings, 2017. https://doi.org/10.1016/j.mayocp.2016.10.019 ↩

  8. Barbara B. Biesecker, William Klein, Katie L. Lewis, et al. How do research participants perceive “uncertainty” in genome sequencing?. Genetics in Medicine, 2014. https://doi.org/10.1038/gim.2014.57 ↩

  9. Robert Cook-Deegan, Subhashini Chandrasekharan. Patents and Genome-Wide DNA Sequence Analysis: Is it Safe to Go into the Human Genome?. Journal of Law, Medicine & Ethics, 2014. https://doi.org/10.1111/jlme.12161 ↩

  10. Francis S. Collins. Medical and Societal Consequences of the Human Genome Project. New England Journal of Medicine, 1999. https://doi.org/10.1056/nejm199907013410106 ↩