Hidden heritability due to heterogeneity across seven populations

Abstract

Meta-analyses of genome-wide association studies (GWAS), which dominate genetic discovery are based on data from diverse historical time periods and populations. Genetic scores derived from GWAS explain only a fraction of the heritability estimates obtained from whole-genome studies on single populations, known as the 'hidden heritability' puzzle. Using seven sampling populations (N=35,062), we test whether hidden heritability is attributed to heterogeneity across sampling populations and time, showing that estimates are substantially smaller from across compared to within populations. We show that the hidden heritability varies substantially: from zero (height), to 20% for BMI, 37% for education, 40% for age at first birth and up to 75% for number of children. Simulations demonstrate that our results more likely reflect heterogeneity in phenotypic measurement or gene-environment interaction than genetic heterogeneity. These findings have substantial implications for genetic discovery, suggesting that large homogenous datasets are required for behavioural phenotypes and that gene-environment interaction may be a central challenge for genetic discovery.

Keywords: age at first birth; educational attainment; gene-environment interaction; hidden heritability; human reproduction; missing heritability.

Figure 1. Stacked Bar Charts of average between (σg2) and within (σgXp2,σgXc2σgXpXc2) variance explanation by common SNPs estimated for Height, BMI, education, age at first birth (AFB) and number of children (NEB) in four model specifications (G, G×P, G×C, G×P×C).

The best model (BM in white, in chart) is based on likelihood ratio tests comparing the full model with one constraining the respective variance component to 0; see Supplementary Table 6. ${\sigma }_g^2/{\sigma }_P^2$ = proportion of observed variance in the outcome associated with genetic variance across all environments, ${\sigma }_{gXp}^2/{\sigma }_P^2$ = proportion of observed variance in the outcomes associated with additional genetic variance within populations, ${\sigma }_{gXc}^2/{\sigma }_P^2$ = proportion of observed variance associated with additional genetic variance within demographic birth cohorts, ${\sigma }_{gXpXc}^2/{\sigma }_P^2$ = proportion of observed variance associated with additional genetic variance within populations and demographic birth cohorts. Models specifications G, G×P, G×C, G×P×C refer to the model specifications including the respective variance components as well as those of lower order – see Material and Methods. For detailed results see Supplementary Table 1-5.

Figure 2. Bar Charts of average % of hidden heritability due to heterogeneity (% of h²_SNP of the best fitting model which is not captured in standard GREML models) and of universal genetic effects (% of h²_SNP of the best fitting model which is effectively identical across the defined environments)

Figure 3. Trends in mean age at first birth of women indicating environmental changes across cohorts (1903-1970) from the US, UK, Sweden, the Netherlands, Estonia and Australia.

Trends in the mean age at first birth of women are based on aggregated data obtained from Human Fertility Database and the Human Fertility Collection (for details see Supplementary Note 3). For Estonia, from 1962 onwards, we used estimated age at first births based on women older than 40. For Australia, no official data was available and the trends have been estimated from the QIMR dataset, averaged for each decade.

Figure 4. Stacked Bar Charts of average between (σg2) and within (σgXp2) variance explanation by common SNPs estimated across 50 simulated phenotypes in two model specifications (standard GREML model and the gene-environment interaction model by study population (G×P) and for four simulated phenotypes.

Sim 1 with homogeneous SNP-based heritability 0.5 without gene-environment interaction, Sim 2 heterogeneous SNP-based heritability between 0.25-0.625 without gene environment interaction, Sim 3 with homogeneous SNP-based heritability 0.5 with gene-environment interaction (genetic correlation of 0.8 across populations) and Sim 4 with homogeneous SNP-based heritability 0.5 with gene-environment interaction (genetic correlation of 0.5 across populations). Individual model estimates are represented by black dots, individual ${\sigma }_g^2$ components in the G×P models in gray stripes.

Figure 5. Bar Charts of average % of hidden heritability due to heterogeneity (% of h²_SNP of the best fitting model which is not captured in standard GREML models) for Sim 1 including and excluding causal variants (Sim LD), for Sim 3 and 4. Individual estimates are represented by black dots.