Genetic influences on circulating retinol and its relationship to human health

Abstract

Retinol is a fat-soluble vitamin that plays an essential role in many biological processes throughout the human lifespan. Here, we perform the largest genome-wide association study (GWAS) of retinol to date in up to 22,274 participants. We identify eight common variant loci associated with retinol, as well as a rare-variant signal. An integrative gene prioritisation pipeline supports novel retinol-associated genes outside of the main retinol transport complex (RBP4:TTR) related to lipid biology, energy homoeostasis, and endocrine signalling. Genetic proxies of circulating retinol were then used to estimate causal relationships with almost 20,000 clinical phenotypes via a phenome-wide Mendelian randomisation study (MR-pheWAS). The MR-pheWAS suggests that retinol may exert causal effects on inflammation, adiposity, ocular measures, the microbiome, and MRI-derived brain phenotypes, amongst several others. Conversely, circulating retinol may be causally influenced by factors including lipids and serum creatinine. Finally, we demonstrate how a retinol polygenic score could identify individuals more likely to fall outside of the normative range of circulating retinol for a given age. In summary, this study provides a comprehensive evaluation of the genetics of circulating retinol, as well as revealing traits which should be prioritised for further investigation with respect to retinol related therapies or nutritional intervention.

Fig. 1. Overview of study design for the genome-wide meta-analysis of circulating retinol.

In this study, we used three input datasets – INTERVAL, METSIM, and ATBC/PLCO. Our primary meta-analysis was INTERVAL meta-analysed with METSIM. These two studies measured plasma retinol using the same platform and both had excellent genome-wide coverage that allowed the analysis of millions of overlapping common and rare variants. ATBC + PLCO was imputed using a much older imputation panel (HapMap2) and even after summary statistics imputation there were markedly fewer variants available. As a result, we performed a separate meta-analysis that included ATBC + PLCO, which had a larger sample size but fewer variants available genome-wide. This dual strategy was adopted to balance variant coverage (METSIM + INTERVAL), which improves discovery power, with the power afforded by increased sample size (METSIM + INTERVAL + ATBC + PLCO). Replication for the lead SNPs was then attempted in up 1635 participants in TwinsUK, with replication performed at each of the three-retinol measurement timepoints, respectively, for the twin pairs separated. Several post-GWAS analyses were then performed. These included: estimates of heritability and polygenicity, gene prioritisation, causal inference, signature mapping, polygenic score development for circulating retinol, as well as the integration of this retinol polygenic score with normative modelling derived deviations.

Fig. 2. Common variant influences on circulating retinol.

a Manhattan plot of the meta-analysis of common variants shared between the INTERVAL and METSIM cohorts (Stouffer’s sample size weighted meta-analysis). Variant-wise -log₁₀ P-values for association are plotted, with the dotted red line denoting genome-wide significance. The closest genic transcription start site is labelled for each lead SNP. Constituent GWAS performed using multiple-linear regression. b Manhattan plot, as above, for the larger sample size meta-analysis that includes the ATBC and PLCO cohorts, but with fewer variants available for meta-analysis. Constituent GWAS performed using multiple-linear regression. c Estimates of SNP heritability of retinol (h²), with the error bars denoting the standard errors of the estimates. The first two panels denote estimates using the BLD-LDAK model and the LDAK-thin model, respectively, both using LD tagging files from the Great British ancestry participants in the UK Biobank. The last panel estimates heritability using the LDSR model with LD from the 1000 genomes European participants. Estimates were for the METSIM + INTERVAL meta-analyses (Stouffer’s and IVW), as well as the larger meta-analysis including ATBC/PLCO. d Empirical Bayes’ estimation of non-null effects on retinol genome-wide, stratified by bins of ascendingly sorted LD score by magnitude. The LD score bins were different for each panel – 1000 bins, 1000 genomes European LD scores (top left); 5000 bins, 1000 genomes European LD scores (top right); 1000 bins, UKBB white British LD scores (bottom left); 5000 bins, white British LD scores (bottom right). Each point represents the proportion of non-null effect sizes for that bin, with the trendline estimated using a generalised additive model for the relationship between ascending LD score bin and the proportion of non-null effects.

Fig. 3. Representative subset of significant causal estimates of circulating retinol using the RBP4 instrumental variable (IV) subjected to colocalisation.

The effect size units are the beta per standard deviation (SD) increase in circulating retinol. The error bars denote 95% confidence intervals. Traits highlighted orange demonstrate strong evidence of a single shared causal variant with circulating retinol in the RBP4 region (PP_H4 > 0.8).

Fig. 4. Estimated causal effects of genetically predicted circulating retinol across the human clinical phenome.

a Prioritisation pipeline overview for retinol causal estimates [inverse-variance weighted estimator with multiplicative random effects (IVW-MRE)] that survive multiple-testing correction (FDR < 0.01). These estimates are then subjected to tests for heterogeneity and pleiotropy (Online Methods), with a tier then assigned based on how many of the five Mendelian randomisation (MR) methods applied are at least nominally statistically significant. In panels b and c, the left-hand plot denotes the Z-score (beta/SE) from the MR IVW-MRE estimates. Positive Z scores denote a positive IVW-MRE estimate of the effect of circulating retinol on that trait, and vice vera. The traits are coloured by their broad phenotypic category. The right-hand plot visualises the Z score using the IVW-MRE model verses that of the MR estimate using the RBP4 IV alone (Wald Ratio). The dotted lines approximately represent nominal statistical significance (P < 0.05). In panel b, just tier #2 traits are plotted (Online Methods), whilst panel c plots both tier #2 and tier #3 traits.

Fig. 5. Exploring the causal effects of continuous exposures on circulating retinol.

a Exposure traits that demonstrated a significant causal estimate (IVW-MRE) on circulating retinol after multiple-testing correction (FDR < 0.01). Traits are coloured relative to their broad phenotypic category. b Multivariable MR (MVMR) models investigating the effect of creatinine and major lipid species on circulating retinol. Each panel represents the results from a different MVMR model (each with different underlying assumptions (Online Methods). The exposure – retinol relationship plotted is conditional on the three other traits in the model.