Learn Explainer
Eye colour genes: the HERC2 switch, and why green and hazel are hard to predict
Blue versus brown is one of the few looks a DNA file can call with some confidence, and most of that confidence comes from a single letter near a pigment gene on chromosome 15. Green and hazel are another matter: the largest study so far found 61 regions of the genome involved, and prediction tools still miss most intermediate eyes.
Key takeaways
- The strongest eye-colour variant, rs12913832, lies in HERC2 but acts on its neighbour OCA2: the G allele weakens an enhancer loop onto the OCA2 promoter, so the iris makes less melanin[2].
- Six variants, the IrisPlex test, told blue from brown with an AUC of 0.96 for each across 3,840 Europeans from seven countries[13].
- Green, hazel and other intermediate colours are predicted far less well: an AUC of 0.74 in one Dutch study[10], and in a Norwegian test set two standard tools predicted no intermediate eyes at all[7].
- The largest study so far, almost 195,000 people, found 124 independent signals in 61 regions of the genome, 50 of them new[1].
- In gnomAD, 78.2 percent of Finnish genomes carry two copies of the blue-associated G allele, the highest share of the genetic-ancestry groups the database reports[9].
Eye colour depends mostly on how much melanin the iris holds: brown irises carry more of it, and blue or green ones very little[1]. The biggest single genetic influence is one position, rs12913832, inside the HERC2 gene. It does nothing to HERC2 itself. It sits in an enhancer that switches on the neighbouring pigment gene OCA2, and the G allele weakens that switch[2]. In one of the two 2008 studies that found it, this one position predicted eye colour better than the OCA2 haplotype the same group had found before[3].
A switch inside the wrong gene
Eye colour is still taught as a one-gene trait, brown dominant over blue[4]. In a study of Queensland twins, modelling put about three quarters of the variation in eye colour down to one dominant genetic factor, and a genome scan of the same twins found a single strong signal on chromosome 15, next to OCA2[4]. Association studies had pointed at OCA2 as the main contributor[5]. OCA2 makes the P protein of the pigment cells, and two non-working copies cause oculocutaneous albinism type 2[6]. In 2008 two groups found the variant itself.
Hans Eiberg's group followed blue eyes through a large Danish family and narrowed the locus to 166,000 bases inside HERC2. Two variants there were perfectly associated with blue and brown eyes in their samples; one, rs12913832, sat in a highly conserved stretch of intron 86, about 21,000 bases upstream of the OCA2 promoter[5]. All 155 blue-eyed Danes they typed shared one haplotype around it, as did seven blue-eyed people from Turkey and Jordan, so they proposed a single founder mutation behind blue eyes[5].
Richard Sturm's group in Brisbane reached the same position from the other direction. They screened 92 more variants in 300 to 3,000 Europeans and found that rs12913832 on its own predicted eye colour better than the best OCA2 haplotype they had published before, with an R² of 0.68 in an ordinal regression[3]. They also found that a coding variant in OCA2, rs1800407, modifies how strongly the HERC2 variant shows.
How the switch works came four years later. In darkly pigmented human melanocytes, Mieke Visser and colleagues found three transcription factors bound to the enhancer and a chromatin loop joining it to the OCA2 promoter, with OCA2 expressed at a higher level. In lightly pigmented melanocytes carrying the other allele, the loop, the binding and the expression were all reduced[2].
One warning about the letters. HERC2 and OCA2 lie on the reverse strand of chromosome 15[8], so papers written from the gene's side, Visser's among them, call the brown-associated allele T and the blue-associated one C. Genome files and gnomAD use the forward strand, A and G[9], and so do this article and our report.
Blue irises have no blue pigment
There is no blue pigment in the eye; a blue iris is a structural colour, made by the way its layers scatter light, the effect that also colours the sky[4]. The rest of the colour comes from how much melanin sits in the pigment layer at the back and in the cells of the stroma in front, from the ratio of the two kinds of melanin, dark eumelanin and reddish pheomelanin, and from that scattering[1]. Brown irises hold more melanin and a higher share of eumelanin. Blue and green ones hold very little of either, with relatively more pheomelanin[1].
So a variant that turns down a pigment gene can make an eye blue, and colour runs as a range from little melanin to a lot. The studies below still sort eyes into three boxes: blue, intermediate and brown, where intermediate covers green, hazel and mixed shades. People grade that middle box inconsistently, which makes it a hard target for any prediction[10].
IrisPlex: six variants for blue and brown
Forensic geneticists want to describe an unknown person's appearance from a trace of DNA when a standard profile matches no one, and IrisPlex was one of the first tests built for it[11]. Susan Walsh, Manfred Kayser and colleagues built IrisPlex from the six variants that had best predicted blue and brown eyes in a study of 6,168 Dutch Europeans, with prevalence-adjusted accuracies above 90 percent for those two colours[11]. The assay gave complete profiles from 31 picograms of DNA, about six cells' worth.
| Gene | Variant | Read by our report |
|---|---|---|
| HERC2 | rs12913832 | Yes, on both eye-colour cards |
| OCA2 | rs1800407 | No |
| SLC24A4 | rs12896399 | No |
| SLC45A2 | rs16891982 | Yes, on the tanning-response card, not for eye colour |
| TYR | rs1393350 | No |
| IRF4 | rs12203592 | Yes, on the freckling cards, not for eye colour |
Source: The six variants as listed by the IrisPlex authors[11, 12]; the last column is our report's trait content as of October 2026.
The model was then tested well beyond the Netherlands. In the European Eye Study, 3,840 people from seven sites had their DNA typed and their irises photographed. Counting only predictions the model made with a probability above 0.7, it named blue or brown correctly 94 percent of the time on average, between 91 and 98 percent depending on the site[13].
0.96
AUC for blue eyes and, separately, for brown eyes, across 3,840 Europeans from seven sites[13]
94%
of blue-or-brown predictions above the 0.7 probability threshold that were correct, 91 to 98% by site[13]
The AUC is the chance that a randomly chosen blue-eyed person gets a higher blue score than someone without blue eyes; 0.5 is a coin toss and 1 is perfect. For blue against brown, six variants come close to the ceiling.
Green and hazel slip through
The same research shows where the method stops working. In 5,951 Dutch participants of the Rotterdam Study, a model of 17 predictors, age, sex and 11 variants among them, scored an AUC of 0.92 for blue and 0.93 for brown, and 0.74 for intermediate eyes[10]. Split into five categories instead of three, green and mixed eyes fell to 0.66[10].
In practice the tools rarely commit to the middle. Of 523 Norwegians whose irises were photographed and scored, 24 percent had intermediate eyes, yet neither rs12913832 alone nor the public IrisPlex web tool predicted intermediate for anyone; 69 and 72 percent of those people were predicted blue[7]. In an independent set of 849 samples from a Polish study, the original IrisPlex model's sensitivity for intermediate eyes was zero; machine-learning models found up to 39 percent of them[14].
Even blue and brown have exceptions. About 3 percent of Europeans with the G/G genotype have brown eyes, and about 5 percent of Europeans have eyes quite different from what rs12913832 would lead anyone to expect[15]. In a Norwegian sample, 43 of 166 people with A/A or A/G, the brown-associated genotypes, did not have brown eyes; seven rarer variants in the OCA2 and HERC2 region could account for 37 of those 43[16].
61 regions, and room for more
Until 2021 about ten genes had been tied to eye colour by genome-wide studies[1]. A 2021 consortium then pooled up to 192,986 people of European ancestry from ten populations, most of them 23andMe customers in the discovery stage, and found 124 independent associations in 61 regions of the genome, 50 of them new. Some point at melanin genes; others at genes for the structure of the iris[1]. A smaller analysis of 1,636 people of Han Chinese and Indian ancestry found the same genetics behind variation in brown eyes, with smaller effects[1].
The study's 112 independent autosomal variants accounted for practically all of the liability for blue against brown eyes in the TwinsUK cohort, but only 38.5 percent of the liability for intermediate eyes. Across the whole three-category scale they accounted for 53.2 percent of the variation[1]. The rest may sit in variants not yet found, in rarer ones like those the Norwegian study turned up, and in the vagueness of the categories themselves.
Many traits are spread over small effects with nothing large on top, which is why a polygenic score usually explains a modest share of its trait. Eye colour has one very large effect above the many small ones. That is why blue and brown are predictable at all, and why green and hazel are not yet.
Finland has the highest share of G/G genomes
The prevalence of blue eyes rises with latitude across Europe[1], and a 2022 review lecture puts the highest frequency of blue eyes around the Baltic Sea, citing an earlier compilation[4]. We could not read the surveys behind it, and eye colour is graded so differently from study to study that country percentages compare poorly. A genotype compares exactly.
In Finnish genomes the G allele has a frequency of 88.5 percent, against 76.4 percent in other European genomes[9]. The 1000 Genomes samples inside gnomAD show the north-to-south slope within Europe, though each group holds only about a hundred people.
| Group | People | G allele frequency | G/G genotype |
|---|---|---|---|
| Finnish, gnomAD | 5,308 | 88.5% | 78.2% |
| Non-Finnish European, gnomAD | 33,991 | 76.4% | 59.3% |
| Finnish in Finland, 1000 Genomes | 98 | 91.3% | 82.7% |
| British in England and Scotland, 1000 Genomes | 87 | 81.0% | 66.7% |
| Toscani in Italy, 1000 Genomes | 103 | 40.8% | 14.6% |
| Iberian in Spain, 1000 Genomes | 104 | 32.7% | 9.6% |
| All gnomAD genomes | 76,108 | 48.7% | 34.3% |
Source: gnomAD v4.1 genomes, variant 15-28120472-A-G, including the 1000 Genomes subsets; the G/G shares are our arithmetic from the homozygote counts[9].
A genotype share is not an eye-colour share. G/G goes with blue eyes most of the time but not always, and A/G with brown or intermediate eyes most of the time[15, 16]. The closest Nordic eye-colour measurement we found is the Norwegian one above: of 523 people, 56 percent had blue eyes, 24 percent intermediate and 20 percent brown, graded from photographs[7]. Why the G allele became so common in the north is a separate question. Eiberg's group proposed a single founder[5], and Simcoe and colleagues name migration, sexual selection and possibly natural selection as the likely causes of the north-south slope[1].
What our report shows, file by file
The report reads one eye-colour variant, rs12913832, on two cards. The Traits card, Eye colour (blue vs brown), maps A/A to Brown-eye associated, A/G to Intermediate and G/G to Blue-eye associated; its Intermediate means the genotype in between, and its text says most people with A/G have brown or hazel eyes. The Fringe card, Blue or brown eyes (HERC2), calls the same three genotypes Brown-leaning, Brown-leaning (carrier of blue) and Blue-leaning. Neither reads the other IrisPlex variants or gives a probability for any colour.
| File | When the file has a call | When it has none |
|---|---|---|
| Chip export | Read by rsID, on either strand | Shown as not read, among the traits the file could not resolve, with no outcome |
| Plain VCF | Read by rsID or by position, 15:28,120,472 on GRCh38; GRCh37 files are lifted over first | Read as A/A, Brown-eye associated, with a note that this is an inference |
| gVCF | As for a plain VCF | As for a plain VCF, unless the reference blocks cover at least 90% of the HERC2 region at a depth of 10 or more; then A/A is shown as examined |
| BAM or CRAM alone | Not read: Deep Read genotypes a fixed set of medication-related and health positions, and rs12913832 is not one of them | Not read |
Source: The report's Traits and Fringe modules as of October 2026; the gVCF rule is the one our coverage guide describes.
The empty case matters here because the reference genome carries A, the brown-associated letter, at this position. A plain VCF lists only differences from the reference, so a person with A/A has no line and a person with a G has one. When a file is silent, the report reads brown and says it inferred it. On a gVCF the examined label rests on the whole HERC2 region, about 221,000 bases, having been read: strong evidence about the region, not proof about this base.
VCF
#CHROM POS ID REF ALT QUAL FILTER FORMAT SAMPLE
15 28120472 rs12913832 A G . PASS GT 1/1We found the cost of that rule on our own pipeline. The 1000 Genomes reference sample NA12878 is G/G at rs12913832[17]. Before a fix on 6 October 2026, the trait module matched calls by rsID, and DRAGEN and GATK gVCFs write a dot in the ID column. Run on the public gVCF of NA12878, the real G/G call went unmatched, good coverage of HERC2 marked the assumed A/A as examined, and the card said brown-eye associated. Calls are now matched by position as well as rsID. Our article on missing VCF records goes through other ways silence was misread.
The optional couple outlook on our family planning page, marked as just for fun, reads the same marker through the Traits table for the four equally likely combinations of two parents' copies, with a note that eye colour is shaped by more genes than this one. Traits sit in the report's well-supported tier and the Fringe in its speculative one, as our guide to confidence tiers explains.
What Aimosti would (and wouldn't) show you
The report reads rs12913832 from a chip export, a VCF or a gVCF and shows it twice: on a Traits card, Eye colour (blue vs brown), and on a Fringe card, Blue or brown eyes (HERC2). Each gives the genotype and the outcome most associated with it. No other eye-colour variant is read, and the report gives no colour probability and no green or hazel prediction. A BAM or CRAM on its own does not produce these cards.
What we won't claim
We report one well-replicated variant and what studies of whole populations found. We won't predict anyone's exact eye colour, read eye colour as evidence of who is related to whom, or treat a genotype as a description of a person.
Bottom line. One letter near OCA2 settles most of the blue-or-brown question, which is why a single-marker card is worth showing. Everything between blue and brown is spread over dozens of other regions that current prediction tools read poorly. The G allele is commoner in Finland than in any other group gnomAD reports, so for most Finnish files the card's answer will be the blue-associated one.
Questions people ask
Can DNA predict eye colour?
For blue against brown, fairly well: the six-variant IrisPlex test reached an AUC of 0.96 for each across 3,840 Europeans[13]. For green, hazel and other intermediate colours, much less well, with AUCs around 0.74 and many intermediate eyes predicted as blue[7, 10].
What does rs12913832 G/G mean?
Two copies of the allele associated with blue eyes. It weakens an enhancer that switches on the pigment gene OCA2[2], and most people with it have blue eyes, though about 3 percent of Europeans with G/G have brown ones[15]. In gnomAD it is the genotype of 78.2 percent of Finnish genomes[9].
Can two blue-eyed parents have a brown-eyed child?
Rarely, but yes. G/G people usually have blue eyes, yet some have brown ones, which a 2020 study linked to variants in TYR, TYRP1 and SLC24A4[15]. A one-marker Punnett square does not capture that.
Does a 23andMe or AncestryDNA file show my eye colour genotype?
If the chip carried rs12913832 and made a call there, the file holds the genotype, and our report reads it on both eye-colour cards. If the row is missing or a no-call, the Traits card shows it as not read rather than guessing. The free file check names what kind of file you have.
References
- Simcoe M, Valdes A, Liu F, et al. Genome-wide association study in almost 195,000 individuals identifies 50 previously unidentified genetic loci for eye color. Science Advances, 2021. doi:10.1126/sciadv.abd1239
- Visser M, Kayser M, Palstra RJ. HERC2 rs12913832 modulates human pigmentation by attenuating chromatin-loop formation between a long-range enhancer and the OCA2 promoter. Genome Research, 2012. doi:10.1101/gr.128652.111
- Sturm RA, Duffy DL, Zhao ZZ, et al. A single SNP in an evolutionary conserved region within intron 86 of the HERC2 gene determines human blue-brown eye color. American Journal of Human Genetics, 2008. doi:10.1016/j.ajhg.2007.11.005
- Mackey DA. What colour are your eyes? Teaching the genetics of eye colour and colour vision. Edridge Green Lecture, RCOphth Annual Congress, Glasgow, May 2019. Eye, 2022. doi:10.1038/s41433-021-01749-x
- Eiberg H, Troelsen J, Nielsen M, et al. Blue eye color in humans may be caused by a perfectly associated founder mutation in a regulatory element located within the HERC2 gene inhibiting OCA2 expression. Human Genetics, 2008. doi:10.1007/s00439-007-0460-x
- OCA2 gene. MedlinePlus Genetics, US National Library of Medicine.
- Meyer OS, Salvo NM, Kjærbye A, et al. Prediction of eye colour in Scandinavians using the EyeColour 11 (EC11) SNP set. Genes, 2021. doi:10.3390/genes12060821 Model data set of 523 Norwegians: 293 blue (56%), 123 intermediate (24%), 107 brown (20%) by PIE-score.
- Gene: HERC2 (ENSG00000128731) and OCA2 (ENSG00000104044), GRCh38: both on the reverse strand of chromosome 15. Ensembl.
- gnomAD v4.1 variant 15-28120472-A-G (rs12913832), genomes. Genome Aggregation Database (gnomAD). Read 2026-10-10. Finnish: G allele count 9,400 of 10,616, 4,153 homozygotes; non-Finnish European: 51,909 of 67,982, 20,169 homozygotes; all genomes: 74,054 of 152,216.
- Liu F, Wollstein A, Hysi PG, et al. Digital quantification of human eye color highlights genetic association of three new loci. PLoS Genetics, 2010. doi:10.1371/journal.pgen.1000934
- Walsh S, Liu F, Ballantyne KN, van Oven M, Lao O, Kayser M. IrisPlex: a sensitive DNA tool for accurate prediction of blue and brown eye colour in the absence of ancestry information. Forensic Science International: Genetics, 2011. doi:10.1016/j.fsigen.2010.02.004
- Wollstein A, Walsh S, Liu F, et al. Novel quantitative pigmentation phenotyping enhances genetic association, epistasis, and prediction of human eye colour. Scientific Reports, 2017. doi:10.1038/srep43359
- Walsh S, Wollstein A, Liu F, et al. DNA-based eye colour prediction across Europe with the IrisPlex system. Forensic Science International: Genetics, 2012. doi:10.1016/j.fsigen.2011.07.009
- Kukla-Bartoszek M, Teisseyre P, Pośpiech E, et al. Searching for improvements in predicting human eye colour from DNA. International Journal of Legal Medicine, 2021. doi:10.1007/s00414-021-02645-5
- Meyer OS, Lunn MMB, Garcia SL, et al. Association between brown eye colour in rs12913832:GG individuals and SNPs in TYR, TYRP1, and SLC24A4. PLoS ONE, 2020. doi:10.1371/journal.pone.0239131
- Salvo NM, Andersen JD, Janssen K, et al. Association between variants in the OCA2-HERC2 region and blue eye colour in HERC2 rs12913832 AA and AG individuals. Genes, 2023. doi:10.3390/genes14030698
- rs12913832: sample genotypes, 1000 Genomes phase 3. Ensembl. NA12878: G|G (read through the Ensembl REST API, 2026-10-10).
Last reviewed . Every number on this page links to the source it comes from; if one of them has moved, tell us.