Celiac genes in your DNA file: what DQ2 and DQ8 can and can't rule out

The immune molecule behind celiac disease

HLA-DQ is a molecule on the surface of immune cells. It holds short fragments of protein and shows them to T cells, which decide whether to mount a response. It is built from two chains: an alpha chain made by the HLA-DQA1 gene and a beta chain made by HLA-DQB1^[3]. Both genes come in many versions, and different pairs hold different fragments.

In celiac disease two of these pairs matter, DQ2 and DQ8. In the gut, an enzyme called tissue transglutaminase modifies fragments of gluten, and DQ2 and DQ8 bind the modified fragments far more tightly than the originals. The gluten-specific T cells found in the gut of people with celiac disease recognise gluten presented by these two molecules^[4]. The authors of the first crystal structure of DQ2 holding a gluten fragment, published in 2004, explain the link by DQ2's unusual ability to hold the proline-rich gluten fragments that survive digestion^[5].

DQ2.5 is the main risk type, and most people with celiac disease carry it^[4]. Together, DQ2.5 and DQ8 are found in almost 95 percent of patients^[3]. DQ2.2 matters less on its own. In the risk model of Monsuur and colleagues, it counts mainly in combination: beside DQ2.5, or beside DQ7, a type whose alpha chain can pair with DQ2.2's beta chain to make DQ2.5^[3].

A positive result is common

Because the risk types are common, a gene test finds one in a large share of everyone tested. The British Society of Gastroenterology's guideline gives the frequency of DQ2 as anywhere from 0 to 40 percent of a population, and of DQ8 from 0 to 20 percent, depending on the country, and calls the positive predictive value of the test very low^[4]. The US National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK) puts it more simply: most people with these gene variants do not develop celiac disease^[6].

Applied to a thousand people, those rates mean roughly 300 carriers, of whom about 9 develop celiac disease. Nine in a thousand is close to the 1 in 100 that MedlinePlus gives for the whole population, so the figures agree with each other: nearly every case falls inside the carrier group, and 97 of every 100 carriers never become one.

Two cautions about that picture. The 3 percent is an average over a population, and it moves with how common celiac disease is where people live: a 2018 meta-analysis found prevalence ranging from 0.4 percent in South America to 0.8 percent in Europe and Oceania^[7]. And it is nobody's personal figure. MedlinePlus notes that other contributors, environmental factors and changes in other genes among them, also influence who develops the disease^[2], and no HLA result says which carriers will be among the few.

A negative is the informative result, within limits

Turn the arithmetic around. If almost everyone with celiac disease carries DQ2 or DQ8, someone who carries neither is very unlikely to have it. That is the use clinical guidelines describe: the BSG puts the diagnostic value of HLA typing in its high negative predictive value^[4], and NIDDK writes it as one sentence.

Very unlikely is not impossible, and the size of the remainder has been measured. The European Genetics Cluster on Celiac Disease typed 1,008 European patients and found 61 who carried neither a complete DQ2 nor DQ8. Of those 61, 57 carried one half of DQ2, either its alpha chain or its beta chain. The remaining four carried neither DQ2 nor DQ8, nor any half of DQ2^[8].

Those numbers describe a test that reads the genes completely. A test that reads two tag markers in place of the genes has a wider blind spot, and part of it comes from the way the two chains of DQ2.5 can be inherited.

One molecule, two copies of chromosome 6

Everyone carries two copies of chromosome 6, one from each parent, and each copy has its own DQA1 and DQB1. A cell makes alpha and beta chains from both copies, and an alpha chain from one copy can pair with a beta chain from the other. In 1989 Sollid and colleagues typed 94 children with celiac disease and found that all but one appeared to carry the same pair of DQA1 and DQB1 genes, the pair now called DQ2.5: either both on the same copy of chromosome 6, the arrangement called cis, or one on each copy, called trans^[9]. Most patients carry the cis form on the DR3-DQ2 haplotype; a smaller group builds the same molecule in trans from a DR5 and a DR7 haplotype^[4].

That is how a two-marker test can miss DQ2.5 without misreading a single letter. The DQ2.5 tag travels with the DR3-DQ2.5 haplotype, the copy of chromosome 6 that carries both genes together^[3]. In the trans arrangement the alpha chain comes from a DQ7 haplotype and the beta chain from a DQ2.2 haplotype, and neither copy carries the tag. The alpha chains differ too, DQA1*05:05 on DQ7 against DQA1*05:01 on DQ2.5, but only by one or a few DNA letters, and they are thought to work the same way^[3]. To capture this case, Monsuur and colleagues needed a separate marker for DQ7 and three for DQ2.2^[3].

What 23andMe and other chip files read: two markers

23andMe's celiac report reads two positions: rs2187668 in HLA-DQA1, which tags DQ2.5, and rs7454108 near HLA-DQB1, which tags DQ8. The company adds that the variants are best studied in people of European descent^[10]. The report was one of the ten genetic health risk reports the US Food and Drug Administration authorised for 23andMe in April 2017^[11]. These are the same two markers our report reads from a chip file.

They are good markers. In the original validation, the DQ2.5 tag had a sensitivity of 1.000 and a specificity of 0.999 for the DQ2.5 haplotype, and the DQ8 tag 0.991 and 0.996^[3]. What they measure is narrow. That study needed six markers to cover the risk types carried by more than 95 percent of patients^[3]; 23andMe's report reads two of the six. Neither of the two sees DQ2.2, a DQ2.5 built in trans, or a lone half of DQ2.

A file reading CC at rs2187668 and TT at rs7454108 holds the chip version of a negative, with the blind spots described above. That assumes both rows are there. We checked the ten chip exports measured on our data page, public files from the Personal Genome Project, for the two rows^[14].

Seven of the ten carry both rows under their rsIDs. A file missing one can still show a positive in the row it has. It cannot support a negative, because the missing marker could have been the positive one, so when the row it has shows nothing, our report says the markers were not called. The same goes for the Genes for Good export: our reader looks for rs7454108 by name, and that file lists the DQ8 position under a coordinate name. Chip versions change from year to year, so a file from another year can differ from the one we checked.

Why a whole-genome VCF can't give a negative

A sequenced genome covers both positions, so it ought to give the same answer as a chip. For the DQ8 marker it does. For the DQ2.5 marker it does not, and public frequency data show it. The 1000 Genomes Project's 2015 release, built from low-coverage sequencing, exome sequencing and genotyping arrays^[15], counted the DQ2.5 tag on 389 of 5,008 chromosomes, 7.8 percent^[16]. In 2022 the project sequenced the same 2,504 samples again, together with 698 relatives, at 30x; the new data were aligned to GRCh38, where the 2015 release had used GRCh37^[17]. That release counts the tag on 44 of 6,404 chromosomes: 0.69 percent^[16].

gnomAD shows a second sign that reads go missing at this spot: it has a called genotype at the DQ2.5 position for 127,560 alleles, against 152,280 at the DQ8 position^{[18, 19]}. We think the cause is where reads from this region end up, though we have not proven it. The main GRCh38 sequence of this stretch of chromosome 6 comes from a DR15 haplotype, which carries no DQ2.5^[21]. GRCh38 also contains alternative sequences of the region^[22], and two of them, called COX and QBL, are DR3-DQ2.5 haplotypes^[21] whose reference letter at rs2187668 is the tag's T^[12]. Reads from a DQ2.5 chromosome match those alternatives better than the main sequence, and an aligner can place them there, out of sight of a variant caller reading chromosome 6.

Our report therefore reads a VCF in one direction only. It reports a DQ2.5 or DQ8 positive from any VCF or gVCF and never issues a negative from one; when neither marker is found, the celiac section says they were not called. A gVCF reference block at rs2187668 is no better than silence here, because the reads that carried the T may never have been placed at that position. The general case, why a missing VCF line is not a reference call, has its own article.

Typing the genes from aligned reads

The way around both problems, the tags' blind spots and the variant caller's, is to read the genes themselves. An HLA typing tool compares the reads from DQA1 and DQB1 with the IPD-IMGT/HLA Database, the reference catalogue of known HLA allele sequences, and reports which two versions of each gene are present^{[23, 24]}. That needs the aligned reads, a BAM or CRAM, because a VCF no longer holds them. Once both genes are typed, the cis and trans question answers itself: a DQA1*05 alpha chain and a DQB1*02 beta chain make DQ2.5 whether they sit on one copy of chromosome 6 or on two.

Our Deep Read add-on does this with T1K, a published typing tool^[24], against release 3.65.0 of that catalogue^[23]. It reports DQ2.5 in cis or trans, DQ8, DQ2.2 and a lone half of DQ2 as separate results, and it returns a negative only when both genes were typed. We checked it against HG002, a Genome in a Bottle reference genome whose HLA types have been published^[22]. From short reads it typed both genes correctly, DQ8 included. From PacBio HiFi long reads of the same genome it returned the wrong alleles for both genes, and would have reported neither DQ2 nor DQ8 for a genome that carries DQ8. Deep Read does not accept long-read files for that reason, and it types these genes only from short-read files aligned to GRCh38. The result is a research-grade reading of consumer sequencing, not clinical HLA typing.

Finland measured some of the highest rates

Some of the field's clearest numbers come from Finland. Mäki and colleagues tested blood that 3,654 Finnish schoolchildren had given in 1994, none of whom had a celiac diagnosis at the time. Biopsy-proven celiac disease turned out to affect at least 1 in 99 of them, and every antibody-positive child but two carried DQ2 or DQ8^[25]. In two samples of Finnish adults taken about twenty years apart, Lohi and colleagues found total prevalence rising from 1.05 percent in 1978–80 to 1.99 percent in 2000–01, a rise that better detection could not explain^[26].

The tag markers have been checked in Finns as well. A Finnish, Hungarian and Italian replication found that the six-marker method predicted the risk haplotypes with sensitivity and specificity from 95 to 100 percent^[27]. The two markers in consumer files belong to that set of six. Their blind spots come from which types they look for, so they are the same in every population.

What each file can and can't say

Which column applies depends on the file you hold. Our free file check reads the first lines of a file in your browser and names its format without uploading anything.

A gene result is not a celiac test

Celiac disease is diagnosed with other tests. Antibody blood tests, such as the tissue transglutaminase test, are used to screen for it, and in most patients a biopsy of the small intestine confirms the diagnosis^[1]. NIDDK lists genetic testing among the tests that can help rule celiac disease in or out, and adds that having DQ2 or DQ8 alone does not mean a person has it^[6].

The two kinds of test also depend on different things. Antibody tests and biopsies look for the effects of the immune response to gluten, and NIDDK notes that a gluten-free diet can affect their results^[6]. A gene result does not depend on diet: the DNA is the same whatever a person eats. That difference is why the BSG guideline describes HLA typing as a test that still applies to someone who began a gluten-free diet without having been tested first^[4].