Predicting eye, hair, skin colour, and height

Pigmentation is one of the few complex traits where a small SNP panel achieves high prediction accuracy. The HIrisPlex-S system, developed at Erasmus MC, uses 36 SNPs and a multinomial logistic regression to predict eye colour, hair colour, and skin colour. Haeckel implements the same model with a per-SNP contribution tracker so you can see exactly which variants are pushing each prediction.

Why pigmentation is unusually predictable

Most of the variance in pigmentation is captured by a small set of large-effect variants in genes such as HERC2, OCA2, MC1R, SLC24A5, and SLC45A2. The variant rs12913832 in HERC2, in particular, is a single SNP that explains roughly 30% of the variance in eye colour. Compare this with intelligence or height, where the largest individual SNP explains less than a tenth of a percent of the variance, and it becomes clear why pigmentation prediction is the easiest case in genomic forensics.

Height

Height prediction uses a polygenic score derived from the GIANT consortium meta-analysis of more than five million people. The current ceiling is roughly 25% of variance explained for individuals of European ancestry, and somewhat less for other ancestries because of the cohort imbalance. Expect a 1-sigma prediction interval of about ±5 centimetres in adults.

A height prediction that disagrees sharply with your actual height usually has one of three explanations. Childhood nutrition or health may have suppressed your adult height below your genetic potential. Polygenic transferability may be poor for your ancestry, in which case the model is genuinely less reliable for you. Or you may be a perfectly normal individual sitting in the natural tail of the prediction interval, which is wider than people often realise.

Per-SNP contribution tracking

Every phenotype prediction the platform produces ships with a contributingSnps array that lists each SNP, its genotype in your file, the per-genotype effect on the prediction, and the literature reference for the effect estimate. You can therefore audit any prediction back to the variants that drove it, which is critical when a result surprises you. The single SNP rs12913832 in HERC2, for example, drives more than half of the eye-colour prediction's confidence; if your file is missing rs12913832 the prediction is suppressed even when other eye-colour SNPs are typed, because the missing-MC1R-style bias would otherwise dominate.

The skin pigmentation challenge

HIrisPlex-S skin prediction has high accuracy in European cohorts (where the model was trained) and degrades in African and East Asian cohorts where the SNP architecture differs. SLC24A5 dominates lightening signal in Europeans; SLC45A2 dominates in Iberians and North Africans; MC1R variants drive the red-haired phenotype but also influence skin response to UV; OCA2 variants behave differently in East Asians. The platform applies an ancestry-aware weighting to mitigate this, but the residual bias is real and we surface a confidence reduction when your inferred ancestry sits far from the model's training cohort.

Other phenotypes the platform does and does not predict

• Predicts: eye colour (blue/intermediate/brown), hair colour (red/blonde/brown/black), skin colour (very pale/pale/intermediate/dark/very dark), adult height in centimetres.
• Limited prediction (low confidence): freckling propensity, hair texture (straight/wavy/curly), male-pattern baldness onset, lactase persistence (which is a separate well-validated single-locus call).
• Does not predict: facial morphology, bone structure, weight, behavioural traits, intelligence (the polygenic score is reported separately on the PRS tab, not as a phenotype prediction), sexual orientation, political leanings, anything else that has been over-claimed by other consumer products.