A pathogenic variant in your raw DNA data: why most rare chip calls are wrong

What a chip reads, and what it was built for

The DNA tests sold for ancestry and health use a genotyping chip, also called a SNP array. The chip carries probes for a fixed list of positions, several hundred thousand of them, and reports the two letters it finds at each one. The ten chip exports on our measured data page have between 563,320 and 955,958 rows^[2]. Everything between those positions goes unread; our guide to genome files explains how a sequenced genome differs.

Chips were designed for variants common enough to be carried by more than 1 person in 100, and for those they work very well^[1]. In recent years many chip designs, including those used by consumer companies, have added probes for rare variants that cause single-gene disorders^[1]. Those probes are where the trouble is, and they reach customers through the raw-data download.

23andMe's own health reports draw on a small set of these positions, each validated before it is reported. The raw file holds all the others as well, and the company's help pages say so:

Downloads are common all the same. In a survey of 1,137 consumer-test customers recruited through social media, 89 percent had downloaded their raw data, and 94 percent of those had run it through at least one third-party tool, most often GEDmatch or Promethease^[4].

Why rare calls go wrong

A chip does not read a letter directly. At each position it measures two signals, one for each possible letter, and software decides the genotype from where a sample's signals fall relative to everyone else's. People with two copies of one letter form one cloud, people with two copies of the other form a second, and carriers of one of each fall between them. Each sample gets the genotype of the cloud it lands nearest^{[1, 5]}.

This works when there are plenty of people in every cloud. UK Biobank ran its chips in batches of about 5,000 samples^[1], and a variant carried by 1 person in 10,000 has no carrier at all in most batches of that size. The software then has nothing to place the middle cloud by, and a point pushed off the main cloud by experimental noise can be read as a carrier. Weedon and colleagues, who measured the effect, describe telling a lone carrier apart from noise as extremely difficult, and note that smaller batches make it worse^[1].

The second half of the problem is arithmetic. Any one rare variant is absent from almost everyone, so nearly every reading of it should come back negative, and a chip with a specificity of 99.9 percent gets almost all of them right. But a chip reads thousands of rare positions at once. Across all of them, the small error rate produces more false positives than there are real carriers, and in the authors' words, a positive result for a very rare pathogenic variant is more likely to be wrong than right^[1].

Some errors are not random either. A probe that misbehaves can give the same wrong answer in sample after sample. In 21 Personal Genome Project volunteers whose 23andMe files Weedon's team compared with their genomes, one rare pathogenic variant in ABCC8 was called homozygous in 9 of the 21 people, wrongly every time^[1]. Cardiovascular genetics clinics described two men, seen separately, whose raw-data reports showed the same MYBPC3 variant, rs36211723; clinical testing found it in neither^[6].

The UK Biobank measurement

The largest test of this is Weedon and colleagues' 2021 study in the BMJ. They took 49,908 UK Biobank participants who had both chip data and exome sequencing, used the sequencing as the reference standard, and asked how often each chip call held up, splitting the results by how common the variant is. Frequency here is the share of chromosomes in the population that carry the variant^[1].

The fall-off is steep. Across 4,757 chip calls of one copy of a variant rarer than 1 in 100,000, sequencing confirmed 16 percent. In the 21 consumer-test volunteers, whose 23andMe chips were made by a different manufacturer to a different design, the confirmed share for variants rarer than 1 in 10,000 was 14 percent, and every one of the 47 rare pathogenic single-letter variants their chips reported was wrong^[1].

The columns answer different questions. Sensitivity is the share of real carriers the chip finds; specificity, the share of non-carriers it correctly leaves alone. The last column, the positive predictive value, is the one that matters to someone holding a positive result: of the people the chip called carriers, how many were. Specificity stays near 100 percent in every row, which is how a chip can look accurate on paper and still be wrong about most of its rare positive calls.

Pathogenic BRCA calls, counted in people

The study used BRCA1 and BRCA2 as its worked case, because a confirmed pathogenic variant there leads to extra cancer screening and, for some, preventive surgery. The two chips tested 1,139 pathogenic or likely pathogenic BRCA variants, 80 percent of them rarer than 1 in 10,000^[1].

In people rather than percentages: 889 participants had a chip call for a pathogenic BRCA variant, and sequencing found it in 37 of them. In the other direction, 107 participants carried a pathogenic BRCA variant at a position one of the chips tested, and the chips found 37 of those, the 34.6 percent sensitivity in the table. Another 137 carried one at a position neither chip tested at all^[1]. By our arithmetic on the study's counts, that is 37 of 244 known carriers found, about 15 percent.

The cancer records point the same way. Participants with a positive chip result had rates of breast, ovarian, prostate and pancreatic cancer close to those of age-matched controls, an odds ratio of 1.31 with a confidence interval from 0.99 to 1.71. Participants with a positive sequencing result had an odds ratio of 4.05, from 2.72 to 6.03^[1]. The chip-positive group looked like the general population because most of its members did not carry the variant.

Three chip files checked against the same people's genomes

Two of the people whose files are on our data page published both chip exports and a 30x whole genome to the Personal Genome Project^[7], which allows the check Weedon's team made, on chips from three other sources. We read every called single-letter position on three chip exports with the parser our chip report uses, moved each to the current reference assembly, and looked it up in the same person's genome. A chip call counts as confirmed when the genome carries the same letter, and as not in the genome when the genome reads the reference base there with at least 10 reads. Pathogenic means a single-letter variant that ClinVar's July 2026 release classifies as pathogenic or likely pathogenic^[8].

Ordinary calls held up almost without exception: at least 99.66 percent of the chips' non-reference calls were in the genome. The remaining fraction of a percent mixes chip errors with positions where the genome may be the one in error; we did not try to separate them. The pathogenic calls did not hold up. The AncestryDNA chip read 15,574 positions where ClinVar lists a pathogenic single-letter variant and reported the variant at 10 of them; the Genes for Good chip read 5,198 such positions and the Living DNA chip only 153. Across the first person's two chips there were 17 such calls at 14 distinct positions. The genome confirmed 4 and read the reference base at the other 10, at depths from 10 to 29 reads.

Frequency separated the two groups completely. In gnomAD v4.1, a public database of variant frequencies from sequenced people, the confirmed variants are carried on anywhere from 1 in 763 chromosomes to 69 percent of them. All ten contradicted variants are rarer than 1 in 5,000, and three have no carrier in gnomAD at all^[9]. Nine of the ten were calls of two copies, a genotype gnomAD records in no one for any of these variants. The second person's chip made a single pathogenic call, at a common variant, and the genome confirmed it.

AncestryDNA, 2018      rs386833618   10   126091596   G   G
Genes for Good         rs386833618   10   126091596   AA
30x genome, GRCh38     10   124401994   G   .   PASS   END=124403368;MinDP=23   GT:DP   0/0:23

This is one of the ten. The position holds a variant in OAT that ClinVar classifies as pathogenic for ornithine aminotransferase deficiency, on a two-star record^[8]. The condition, also called gyrate atrophy, is an inherited eye disease with vision loss that progresses from childhood. It is recessive, so it takes two pathogenic copies, and it is most common in Finland^[10]. The AncestryDNA chip read the reference letters here, G G. The Genes for Good chip, reading the same person's DNA, reported two copies of the variant, AA. The genome settles it: the reference base on both copies, with at least 23 reads across the stretch. In gnomAD, 30 of about 1.6 million chromosomes carry the variant and no one carries two copies^[9].

Two people and three chips are a small sample, and a genome is a strong reference standard but not a clinical test. The direction matches Weedon's: near-perfect at common positions, mostly wrong at rare pathogenic ones. Our account of the Prothrombin row shows the other side of chip reading, a common clinical variant that chips type well.

The second error: the label

A raw-data tool does two things: it takes the letters in the file as given, and it attaches a meaning to them from a database. Promethease, which 63 percent of the downloaders in the survey above had used^[4], describes itself as a literature retrieval system that connects a file of genotypes to the findings cited in SNPedia^[11]. Neither step checks whether the chip read the letters correctly, and the meaning is only as good as the entry behind it.

Ambry Genetics, a clinical laboratory, reviewed 49 patients sent to it between 2014 and 2016 to confirm a variant found in consumer raw data. Forty percent of the variants were not there^[12]. Separately, the laboratory reports that eight variants had been marked as increased risk by the raw data or by a third-party service and are classified as benign by Ambry and by several other laboratories. One, BRCA2 p.N372H, sits on about one chromosome in four in public population databases, far too common to cause a rare cancer syndrome^[12].

Our check found the same thing among the confirmed calls. Both people carry a variant in CDKN2B that 69 percent of the chromosomes in gnomAD carry^[9]. Its ClinVar record reads likely pathogenic and protective at once, with no stars: ClinVar grades the review behind each classification from zero to four stars, and zero means the submitter gave no assertion criteria^{[8, 13]}. The chips read this variant correctly; the label is what misleads. Our guide to carrier status covers another common misreading, one copy of a recessive variant taken as a personal risk.

What a company's own BRCA report covers

A vendor's validated report and its raw file are different things, and BRCA shows the difference. In 2018 the FDA authorized 23andMe's BRCA1/BRCA2 (Selected Variants) report, which then covered three variants most common in people of Ashkenazi Jewish descent. The agency noted that the test detected three of more than 1,000 known BRCA mutations, that those three are present in about 2 percent of Ashkenazi Jewish women and rarely in other groups, and that treatment decisions require confirmatory testing and genetic counselling^[14].

In October 2023 the report grew to 44 variants, which 23andMe says account for about 30 to 40 percent of cancer-related BRCA variants in people of African American, non-Ashkenazi European and Hispanic or Latino descent. The company reports that each variant showed more than 99 percent concordance with Sanger sequencing, and states that the report does not include all BRCA risk variants^[15]. Those 44 positions were validated one by one. A BRCA row elsewhere in the raw file has had the general quality review described in the quotation above^[3].

The well-known founder variants hold up. In the Ambry series, all 13 Ashkenazi founder BRCA variants sent for confirmation were confirmed, as were all four cystic fibrosis F508del calls, while eight other BRCA1 and BRCA2 calls were false positives^[12]. Our BRCA2 page covers what a confirmed pathogenic variant there means.

What confirmation looks like

Confirmation repeats the measurement in a clinical laboratory, on a fresh sample, with a method built to read rare variants. In the Ambry series it ran like this^[12].

The cardiology cases mentioned earlier show what rides on the answer. One of the two men with the MYBPC3 call had taken medical leave and given up cycling before the clinic saw him; clinical testing did not find the variant, and he was released from cardiac screening^[6]. For a positive result from a validated report, the FDA's release describes confirmatory testing and genetic counselling before any treatment decision^[14], and 23andMe's report text an independent test ordered by the person's own healthcare provider^[15].