A study in scarlet: MC1R as the main predictor of red hair and exemplar of the flip-flop effect

Abstract

Genetic variation in melanocortin-1 receptor (MC1R) is a known contributor to disease-free red hair in humans. Three loss-of-function single-nucleotide variants (rs1805007, rs1805008 and rs1805009) have been established as strongly correlated with red hair. The contribution of other loss-of-function MC1R variants (in particular rs1805005, rs2228479 and rs885479) and the extent to which other genetic loci are involved in red hair colour is less well understood. Here, we used the UK Biobank cohort to capture a comprehensive list of MC1R variants contributing to red hair colour. We report a correlation with red hair for both strong-effect variants (rs1805007, rs1805008 and rs1805009) and weak-effect variants (rs1805005, rs2228479 and rs885479) and show that their coefficients differ by two orders of magnitude. On the haplotype level, both strong- and weak-effect variants contribute to the red hair phenotype, but when considered individually, weak-effect variants show a reverse, negative association with red hair. The reversal of association direction in the single-variant analysis is facilitated by a distinguishing structure of MC1R, in which loss-of-function variants are never found to co-occur on the same haplotype. The other previously reported hair colour genes' variants do not substantially improve the MC1R red hair colour predictive model. Our best model for predicting red versus other hair colours yields an unparalleled area under the receiver operating characteristic of 0.96 using only MC1R variants. In summary, we present a comprehensive statistically derived characterization of the role of MC1R variants in red hair colour and offer a powerful, economical and parsimonious model that achieves unsurpassed performance.

Figure 1

Previously reported variants in MC1R associated with red hair. The colour-coded circles in the figure correspond to the wild-type amino acid residues that are changed in the variants at the positions and to the resulting residues as specified here. High-penetrance R variants are in yellow: D84E (rs1805006), R142H (rs11547464), R151C (rs1805007), I155T (rs1110400), R160W (rs1805008) and D294H (rs1805009), and low-penetrance r variants are in green: V60L (rs1805005), V92M (rs2228479) and R163Q (rs885479). Frameshift variants N29insA (rs312262906), 179insC (rs555179612) and Y152OCH (rs201326893) are in red. Image courtesy of GPCRdb.org.

Figure 2

LD plot for 12 MC1R variants. Metric coding: value in diamonds = D′; colour scheme: r-squared (white = 0, red = 1 [here none]).

Figure 3

Interaction between r allele count and R allele count. Association for r allele count with red hair given an invariant R allele count in tabular (A) and graphical (B) format.

Figure 4

AUROC curves for all pairwise hair colour comparison GLMs with MC1R variants as predictors.

Figure 5

AUROC curves for (A) mRMR and (B) LASSO models using 10 top MC1R genetic variants and 10 top genetic variants from each gene for mRMR and 100 top genetic variants from each gene for LASSO. Red versus other hair colour. For both mRMR and LASSO, the model performance for all genes is statistically significantly different from the model using only MC1R variants. Despite the 10 iterations of 10-fold cross-validation to obtain an estimate of mean ROC performance, error bars for the 95% confidence interval are based on a standard error of the mean assuming a sample size of 10 rather than 100 due to lack of test set independence in folds between cross-validation iterations.

Figure 6

Red versus other hair colour prediction using LASSO models with 10 MC1R and 100 top mRMR-ranked variants from 1000 randomly selected genes. All the genes shown fall within 2. ASIP, OCA2, IRF4 and HERC2 (not shown) have AUROC values 0.970, 0.965, 0.965 and 0.965, respectively, and are the only genes whose variants improve predictive performance above and beyond MC1R variants. The variants of these four genes and two other genes outperform MC1R-alone models with a statistically significant difference (t-test P-values: ASIP, <1e-16; HERC2, <1e16; OCA2, <1e-16; IRF4, <1e-16; POMC, 5.7e-10; SLC45A2, 8.3e-3).