A saturated map of common genetic variants associated with human height

Abstract

Common single-nucleotide polymorphisms (SNPs) are predicted to collectively explain 40-50% of phenotypic variation in human height, but identifying the specific variants and associated regions requires huge sample sizes¹. Here, using data from a genome-wide association study of 5.4 million individuals of diverse ancestries, we show that 12,111 independent SNPs that are significantly associated with height account for nearly all of the common SNP-based heritability. These SNPs are clustered within 7,209 non-overlapping genomic segments with a mean size of around 90 kb, covering about 21% of the genome. The density of independent associations varies across the genome and the regions of increased density are enriched for biologically relevant genes. In out-of-sample estimation and prediction, the 12,111 SNPs (or all SNPs in the HapMap 3 panel²) account for 40% (45%) of phenotypic variance in populations of European ancestry but only around 10-20% (14-24%) in populations of other ancestries. Effect sizes, associated regions and gene prioritization are similar across ancestries, indicating that reduced prediction accuracy is likely to be explained by linkage disequilibrium and differences in allele frequency within associated regions. Finally, we show that the relevant biological pathways are detectable with smaller sample sizes than are needed to implicate causal genes and variants. Overall, this study provides a comprehensive map of specific genomic regions that contain the vast majority of common height-associated variants. Although this map is saturated for populations of European ancestry, further research is needed to achieve equivalent saturation in other ancestries.

Fig. 1. Relationship between frequency and estimated effect sizes of minor alleles.

Each dot represents one of the 12,111 quasi-independent GWS SNPs that were identified in our cross-ancestry GWAS meta-analysis. Data underlying this figure are available in Supplementary Table 10. SNP effect estimates (y axis) are expressed in height standard deviation (s.d.) per minor allele as defined in our cross-ancestry GWAS meta-analysis. SNPs were stratified in five classes according to their P value (P) of association. We show two curves representing the theoretical relationship between frequency and expected magnitude of SNP effect detectable at P < 5 × 10⁻⁸ with a statistical power of 90%. Statistical power was assessed under two experimental designs with sample sizes equal to n = 0.5 million and n = 5 million. Source data

Fig. 2. Brisbane plot showing the genomic density of independent genetic associations with height.

Each dot represents one of the 12,111 quasi-independent GWS (P < 5 × 10⁻⁸) height-associated SNPs identified using approximate COJO analyses of our cross-ancestry GWAS meta-analysis. Data underlying this figure are available in Supplementary Table 10. GWS SNPs with the largest density on each chromosome were annotated with the closest gene. We highlight 24 of 12,111 associations that are mainly contributed by groups of non-European ancestry (3 from African ancestries, 10 from Hispanic ethnicities or ancestries and 11 from East Asian ancestries). The full list of height-associated SNPs detected in groups of non-European ancestry and independent of associations detected in European ancestry GWASs is reported in Supplementary Table 9. Signal density was calculated for each associated SNP as the number of other independent associations within 100 kb. A density of 1 means that a GWS COJO SNP shares its location with another independent GWS COJO SNP within less than 100 kb. The mean signal density across the genome is 2 and the median signal density is 1 (s.e. 0.14 and 0.0, respectively). The s.e. values were calculated using a leave-one-chromosome-out jackknife approach (LOCO-S.E.). SNPs that did not reach genome-wide significance  are not represented on the figure. Source data

Fig. 3. Variance of height explained by HM3 SNPs within GWS loci.

a, Stratified SNP-based heritability ($h_{\mathrm{SNP}}^2$) estimates obtained after partitioning the genome into SNPs within 35 kb of a GWS SNP ('GWS loci' label) versus SNPs that are more than 35 kb away from any GWS SNP. Analyses were performed in samples of five different ancestries or ethnic groups: European (EUR: meta-analysis of UK Biobank (UKB) + Lifelines study), African (AFR: meta-analysis of UKB + PAGE study), East Asian (EAS: meta-analysis of UKB + China Kadoorie Biobank), South Asian (SAS: UKB) and Hispanic (HIS: PAGE). Error bars represent standard errors. b, More than 90% of $h_{\mathrm{SNP}}^2$ in all ancestries is explained by SNPs within GWS loci identified in this study. The cumulative length of non-overlapping GWS loci is around 647 Mb; that is, around 21% of the genome, assuming a genome length of around 3,039 Mb (ref. ). The proportion of HM3 SNPs in GWS loci is around 27%. Source data

Fig. 4. Accuracy of PGSs within families and across ancestries.

Prediction accuracy (R²) was measured as the squared correlation between PGS and actual height adjusted for age, sex and 10 genetic principal components. a, Accuracy of PGSs assessed in participants of five different ancestry groups: European (EUR) from the UKB (n = 14,587) and the Lifelines Biobank (n = 14,058); South Asian (SAS; n = 9,257) from UKB; East Asian (EAS; n = 2,246) from UKB; Hispanic (HIS; n = 5,798) from the PAGE study; and admixed African (AFR) from UKB (n = 6,911) and PAGE (n = 8,238). PGSs used for prediction, in a, are based on GWS SNPs or around 1.1 million HM3 SNPs. When using all HapMap 3 SNPs, SNP effects were calculated using the SBayesC method (Methods), whereas PGSs based on GWS SNPs used joint SNP effects estimated using the COJO method (Methods). Both SBayesC and COJO were applied to (1) our cross-ancestry meta-analysis (turquoise bar); (2) our EUR meta-analysis (yellow bar); and (3) each ancestry-specific meta-analysis (red bar). b, Squared correlation of height between EUR participants in UKB and their first-degree relatives, and the accuracy of a predictor combining PGS (denoted PGS_GWS, as based on GWS SNPs) and familial information. The accuracies of PGS_GWS and PGS_HM3 shown in b are the average of the respective accuracies of these PGSs in EUR participants from UKB and the Lifelines Biobank as shown in a. Sibling correlation was calculated in 17,492 independent EUR sibling pairs from the UKB and parent–offspring correlations in 981 EUR unrelated trios (that is, two parents and one child) from the UKB. PA, parental average. Source data

Extended Data Fig. 1. Broad ancestries composition.

Geographical mapping and ancestries composition of 281 studies meta-analysed in this study. Various analyses were performed including (1) dectection of height-associated SNPs (Genetic discoveries box), (2) quantification of the genomic distribution of height-associated loci (Genomic distribution box), (3) assessement of the performances of polygenic predictors of height (Polygenic prediction box), and (4) assessment of the relationship between GWAS sample size and discoveries (Saturation of discovery from GWAS box).

Extended Data Fig. 2. Colocalization of height-associated signals across ancestries.

Proportion (y-axis) of GWS SNPs identified in our GWAS meta-analyses of non-European (non-EUR: African – AFR; East Asian – EAS; South Asian – SAS; Hispanic – HIS) ancestry/ethnicity participants thar are located within a certain distance (x-axis) of GWS SNPs identified in our GWAS meta-analysis of EUR participants only. Source data

Extended Data Fig. 3. Replication of marginal associations in the EBB.

a, Each dot represents one the 12,111 SNPs detected in our trans-ancestry meta-analysis. The x-axis represents the expected statistical power to replicate each association (P<0.05/9,473 = 5.3×10⁻⁶; where 9,473 is the number of associations reaching marginal genome-wide significance in our discovery trans-ancestry GWAS and with a minor allele frequency>1% in the EBB sample). The y-axis represents the -log₁₀ of the association p-value in the EBB multiplied by the product of signs of estimated SNP effects in the discovery and in the EBB. Horizontal dotted line represents replication at P<0.001 and the vertical dotted line indicates 80% of statistical power. SNPs highlighted in green have an expected statistical power for replication >80%. One outlier (rs11100870), highlighted in red, does not replicate in the EBB sample. b, Proportion (P) of SNPs with a sign-consistent estimated effect between discovery GWAS (N~5.3M) and EBB. Expected proportions (E[P]) are calculated using equation (2) in the Methods. Error bars are defined as 1.96×$\sqrt{P(1-P)/m}$, where m is the number of SNPs in the corresponding MAF interval. Source data

Extended Data Fig. 4. Enrichment of genes containing pathogenic mutations that cause extreme height or abnormal skeletal growth syndromes near hotspots of GWS SNPs.

Four hundred and sixty-two (462) autosomal genes were curated from the Online Mendelian Inheritance in Man (OMIM) database. a, Red arrow indicates the observed enrichment statistic (OR = 2.5-fold) measuring the odds ratio of the presence of an OMIM gene within 100 kb of a GWS SNPs with a density > 1. The blue histogram represents the distribution of enrichment statistics from 1,000 random genes matched, which length distribution matches that of the OMIM genes. b, Enrichment of OMIM genes near high density GWS SNPs. High density is defined by on the x-axis by the minimum number of other independent GWS SNPs detected within 100 kb. Source data

Extended Data Fig. 5. Haplotypic analysis at the ACAN locus.

a, Distribution of estimated haplotype effects from 14,117 haplotypes covering a 100 kb long genomic region near the ACAN gene (hg19 genomic coordinates: chr15:89,307,521-89,407,521). b, Quantile-quantile plot of associations between these 14,117 haplotypes and height. c, Distribution of the variance explained by each of the 14,117 haplotypes. d, Mean signals density (y-axis) across simulated data where 1 causal SNP within the locus explains between 0.5% and 5% (x-axis) of trait variance. Causal variants were sampled from a pool of 13 SNPs with a 1.4×10⁻⁵ < MAF < 1% genotyped in 291,683 unrelated EUR participants of the UKB, with no missing values at these 13 SNPs. Standard errors were calculated as the standard deviation (s.d.) of signal density across 100 simulation replicates. GCTA-COJO analyses to identify independent signals were performed using a subset of 10,000 unrelated EUR participants of the UKB to mimic the large discrepancy between the size of the discovery GWAS and that of the LD reference used in our real data analyses. e, Proportion of VNTR length explained by 25 GWS SNPs identified near ACAN in 4 ancestries (European: EUR; South Asian: SAS; East Asian: EAS; African: AFR). f, Proportion of height variance explained in a sample of EUR UK Biobank participants by various sets of polymorphisms at the ACAN locus. rs3817428 and rs34949187 are two missense variants and rs7176941 is an intronic variant with high posterior causal probability identified in ref. . In e and f, error bars represent standard error (s.e.). Source data

Extended Data Fig. 6. Variance of height explained by common SNPs within 35 kb of GWS SNPs.

Stratified SNP-based heritability ($h_{\mathrm{SNP}}^2$) estimates were obtained from a partition of the genome into two sets of 1000 Genomes imputed SNPs with a minor allele frequency (MAF) >1%: (1) SNPs within +/− 35 kb of GWS (GWS loci) vs. all other SNPs. Analyses were performed in samples of five different ancestry groups: European (EUR; UK Biobank only), African (AFR), East Asian (EAS) and South Asian (SAS) as described in the legend of Fig. 3. Estimates from stratified analyses were compared with SNP-based heritability estimates obtained from analysing HM3 SNPs only (dotted horizontal violet bar). Source data

Extended Data Fig. 7. Accuracy of PGSs derived from joint effects of SNPs ascertained at various significance thresholds.

The six panels show on their y-axes the prediction accuracy (R²) of multiple PGS across five target samples. The ancestry group and size of each target sample is indicated in the panel title. The top-left panel shows the averaged prediction accuracy in two European ancestry (EUR) target samples from the UK Biobank (UKB) and Lifelines Biobank (LLB). The other panels show prediction accuracies in individual target samples of African ancestry (AFR) from UKB and the PAGE study, East Asian ancestry (EAS) and South Asian ancestry (SAS) ancestry from the UKB and Hispanic ethnicity from the PAGE study. Each panel is divided in four columns representing the four significance levels used to ascertain SNPs using the GCTA-COJO algorithm. GCTA-COJO was applied to each ancestry-group specific GWAS meta-analysis with an ancestry-match linkage disequilibrium (LD) reference. We used genotypes from 50,000 (vs 350,000 for results reported in the main text) unrelated EUR participants as LD reference to run GCTA-COJO on the EUR- and the cross-ancestry GWAS meta-analysis. For the other ancestry groups, we used genotypes from 10,636 AFR individuals, 5,875 EAS individuals, 4,883 HIS individuals and 9,448 SAS individuals as LD reference (as described in Methods). Error bars are standard error (s.e.). The number of SNPs used in each PGS is indicated (in white) within each bar. Source data

Extended Data Fig. 8. Enrichment of height-associated genes identified at various GWAS sample sizes within 20 clusters of gene sets representing broad categories of biological pathways.

Gene-set enrichment was performed with MAGMA and DEPICT across seven GWAS with increasing sample sizes. Samples used (Lango Allen et al. (2010), n = 0.13M; Wood et al. (2014), n = 0.24M; Yengo et al. (2018), n = 0.7M; GIANT-EUR (no 23andMe), n = 1.63M; 23andMe-EUR, n = 2.5M; European-ancestry meta-analysis, n = 4.08M; and cross-ancestry meta-analysis, n = 5.31M) are described in Tables 1–2. The degree of enrichment of gene sets (MAGMA, DEPICT) of known skeletal growth disorder genes catalogued in the Online Mendelian Inheritance in Man (OMIM) database among 20 clusters of gene sets (see Methods section in Supplementary Note 5) is indicated by the blue-red colour scale. Enrichment for MAGMA and DEPICT was defined to be the number of prioritized gene sets (top 10% of gene sets) in each cluster divided by the 10% of the number of gene sets in the cluster. Enrichment for OMIM was defined to be the number of OMIM genes in a gene set (Z > 1.96) divided by the size of the gene set divided by the proportion of all genes in OMIM, then averaged across the cluster. Significant enrichment (compared to shuffled prioritization of gene sets or genes) is marked with *. Source data

Extended Data Fig. 9. Annotation-level saturation of GWAS discoveries as a function of sample size.

Increase in sample size from ~4 million to ~5 million is achieved by including ~1 million participants of non-European ancestry. a, Number of annotations showing a significant heritability enrichment as function the function of the sample size of the GWAS used to estimate these enrichment. Heritability enrichment was detected using a stratified LD score regression (LDSC) analysis of 97 genomic annotations included in the “baseline + LD” model from Gazal et al. b, Correlation between Z-scores measuring the statistical significance of heritability enrichments of 97 annotations (each dot is an annotation) in our largest GWAS (x-axis) as compared to down-sampled GWAS (y-axis). Sample size is denoted by the colour-code. c, Distribution of estimated enrichment statistics for 21 annotations found significantly enriched (P < 0.05/97) in at least 6 of the 7 GWAS analysed here. LoF-i genes: Loss of function intolerant genes; TSS: Transcription Start Sites; DGF: Digital genomic footprint; TFBS: Transcription Factor Binding Sites; DHS: DNAse I hypersensitive sites; GERP (NS): GERP++ score (number of substitutions). Source data

Extended Data Fig. 10. Partitioning of low-frequency SNP-based heritability within GWS loci.

Panels b–d represent partitioned SNP-based heritability estimates from three samples (EBB: Estonian Biobank; UKB: UK Biobank; LLB: Lifelines Biobank) of unrelated European ancestry individuals independent of our discovery GWAS. a, Partitioned SNP-based heritability estimates obtained from an inverse-variance weighted meta-analysis of estimates shown in b–d. SNPs were partitioned into four classes according to their minor allele frequency (MAF: 0.1% < MAF < 1% vs. MAF > 1%) and their position within versus outside GWS loci. The SNP-based heritability contributed by SNPs within GWS loci is denoted $h_{\mathrm{GWS}}^2$, and that contributed by SNPs outside these loci is denoted $h_{\mathrm{other}}^2$. These results are further discussed in Supplementary Note 6. Source data