What to Do With Your 23andMe Raw Data

23andMe tests you on a DNA chip (a genotyping array) rather than sequencing your whole genome. The chip reads a fixed set of positions: about 640,000 on the current v5 chip (2017 onward), roughly 570,000 on v4 (2013–2017), and roughly 960,000 on v3 (2010–2013). That is a small slice of the 3 billion positions in your genome, but a well-chosen one: most markers are common variants that also stand in for the stretch of DNA around them.

The raw data is a plain text file with a long comment header, then one line per marker and four tab-separated columns:

# rsid	chromosome	position	genotype
rs4988235	2	136608646	AG
rs12913832	15	28365618	GG
rs429358	19	45411941	TT
rs7412	19	45412079	CC
rs1801133	1	11856378	--

• rsidThe variant's public ID in the dbSNP database, such as rs4988235. You can look these up in dbSNP, ClinVar, or SNPedia. IDs that start with "i" are 23andMe's own internal markers, often custom probes for specific variants.
• chromosome and positionWhere the marker sits, using the GRCh37 (hg19) reference genome. This matters when you compare against papers or tools that use the newer GRCh38, where positions differ.
• genotypeYour two copies, one from each parent. They are unphased, so AG doesn't say which parent gave you which letter. "--" means the chip couldn't make a confident call; a typical file has calls at 98–99% of markers. Men have a single letter at most X and Y markers.

Chips are very good at common variants, the ones carried by at least a few percent of people. That covers more than most people expect:

• Polygenic risk for common diseasesHeart disease, type 2 diabetes, breast and prostate cancer, and most other common conditions are shaped by thousands of common variants, each with a small effect. A polygenic risk score adds them up and compares your total with a reference population. The top of the distribution can matter: in one large study, about 8% of people had at least three times the average risk of coronary artery disease based on their score alone (Khera et al., Nature Genetics, 2018).
• APOE and other well-studied single variantsA handful of individual common variants matter on their own. Two markers, rs429358 and rs7412, together give your APOE type; carrying one or two ε4 copies substantially raises Alzheimer's risk. Others include HFE C282Y (rs1800562, hereditary hemochromatosis) and Factor V Leiden (rs6025, blood clots). Chips call these reliably because they are common, but anything with health consequences is worth confirming clinically. If you would rather not know your APOE type, avoid tools that report it.
• How you respond to medicationsCommon variants in genes like CYP2C19 and SLCO1B1 affect how your body processes many medications, including some antidepressants, clopidogrel, and statins. Genomisaur doesn't cover drug response, but our sister site Gene2Rx does: it reads the same raw data file and checks how your genes may affect 120 medications, with results based on clinical guidelines such as CPIC's. Get a pharmacogenetics report on Gene2Rx
• Physical and everyday traitsLactase persistence (rs4988235), blue versus brown eyes (rs12913832 in HERC2), wet or dry earwax (rs17822931 in ABCC11), and the alcohol flush reaction (rs671 in ALDH2) are each driven mostly by one well-mapped variant. Traits shaped by many genes, like height, are better estimated with polygenic scores.
• Relatives who tested elsewhereGEDmatch, FamilyTreeDNA, and MyHeritage accept 23andMe files for relative matching, which reaches people who never tested with 23andMe.

Imputation stretches the file further. Because nearby variants tend to be inherited together, statistical imputation against a large reference panel can fill in millions of markers the chip never measured. Genomisaur imputes array files to about 30 million variants before scoring, so each score can use far more of its variants than the chip covers directly. Imputation is accurate for common variants and much less reliable for rare ones.

DNA chips are designed around common variation. They struggle with rare variants, and rare variants are where most high-impact disease mutations live.

• Rare variant calls are often wrongWhen a clinical lab re-tested variants that people had found in their consumer raw data, 40% turned out to be false positives (Tandy-Connor et al., Genetics in Medicine, 2018). In UK Biobank chip data, calls for the rarest variants were confirmed by sequencing only about 16% of the time (Weedon et al., BMJ, 2021). If any tool flags a rare disease-causing variant in your file, treat it as a reason to get a clinical test, not as a result.
• A clean result doesn't rule much outMore than a thousand disease-causing variants are known in BRCA1 and BRCA2. The best-known consumer BRCA report covers three of them, which are mainly found in people of Ashkenazi Jewish ancestry. Not seeing a variant in your raw data says little about whether you carry a different one.
• Some kinds of variation are invisibleChips cannot reliably detect deleted or duplicated genes (such as the SMN1 deletions behind spinal muscular atrophy carrier status), repeat expansions (Huntington's disease, fragile X syndrome), or most larger structural changes.
• Scores travel poorly across ancestriesMost genetic studies have been done in people of European ancestry, so polygenic scores predict less well in other groups. One widely cited analysis found accuracy 2 to 4.5 times lower in East Asian and African ancestry groups (Martin et al., Nature Genetics, 2019). Genomisaur's percentiles are calibrated against a European-ancestry reference, and its reports say so.

The file opens in any text editor, and at a few hundred thousand rows it fits in a spreadsheet. Search for an rsID and you will find your genotype. A few things trip people up:

• StrandConsumer files report genotypes on the forward (plus) strand of the reference genome. Papers and SNPedia sometimes describe the same variant on the opposite strand, where A pairs with T and C pairs with G, so a "risk allele T" in a paper can appear as A in your file. Check the orientation before concluding anything.
• One SNP rarely means muchMost variants you will read about shift risk by a few percent. Without the thousands of other variants and a reference population to compare against, a single "risk" genotype is usually noise.
• Popular variants with weak evidenceMTHFR C677T (rs1801133) is the classic example. It is heavily marketed, but the American College of Medical Genetics and Genomics recommends against testing for it because it has little clinical value.
• Missing callsA marker the chip could not read means no data, not a normal result.

Services that accept raw data do quite different things. Choose by what you want to learn.

Before uploading anywhere, check three things: what method sits behind the health results (published polygenic scores, single-SNP lookups, or something proprietary), whether you can delete your data, and whether the company sells or shares it. Prices change often; confirm on each site.

Your raw data identifies you, and partly identifies your relatives. Once a copy sits on another company's server, you depend on that company's security and deletion policy.

In the US, the Genetic Information Nondiscrimination Act (GINA) bars health insurers and most employers from using your genetic information. It does not cover life, disability, or long-term care insurance.

Genealogy databases are a special case: some allow law-enforcement searches, usually depending on your settings. Check each site's options before you upload.

23andMe itself changed hands in 2025, when TTAM Research Institute, a nonprofit led by co-founder Anne Wojcicki, bought the company out of bankruptcy. If you are deciding whether to keep your account, download your raw data first. Deleting the account deletes the file, and 23andMe cannot send it to you afterwards.