Meta-analysis of sub-Saharan African studies provides insights into genetic architecture of lipid traits

Abstract

Genetic associations for lipid traits have identified hundreds of variants with clear differences across European, Asian and African studies. Based on a sub-Saharan-African GWAS for lipid traits in the population cross-sectional AWI-Gen cohort (N = 10,603) we report a novel LDL-C association in the GATB region (P-value=1.56 × 10^-8). Meta-analysis with four other African cohorts (N = 23,718) provides supporting evidence for the LDL-C association with the GATB/FHIP1A region and identifies a novel triglyceride association signal close to the FHIT gene (P-value =2.66 × 10^-8). Our data enable fine-mapping of several well-known lipid-trait loci including LDLR, PMFBP1 and LPA. The transferability of signals detected in two large global studies (GLGC and PAGE) consistently improves with an increase in the size of the African replication cohort. Polygenic risk score analysis shows increased predictive accuracy for LDL-C levels with the narrowing of genetic distance between the discovery dataset and our cohort. Novel discovery is enhanced with the inclusion of African data.

Fig. 1. Summary of the datasets and analyses.

For each lipid trait, Stage 1 of the GWAS involved a joint analysis of the full AWI-Gen dataset (10,603 participants from Eastern, Western and Southern Africa) and Stage 2 (N = 23,718) involved a meta-analysis of the Stage 1 results with the summary statistics from four African cohorts included in the Gurdasani et al. 2019 study (Uganda Genome Resource (UGR) study, the Africa-America Diabetes Mellitus (AADM) study, the Durban Diabetes Study (DDS), the Durban Case Control (DCC) study). Approximate geographic location and sample size of the cohorts are represented by position and size of the circles in the map. Each cohort is shown in a unique colour. Assessment of transferability of associations detected in the Global Lipid Genetics Consortium (GLGC) study and the Population Architecture using Genomics and Epidemiology (PAGE) study was performed in various replication sets. Predictability of polygenic risk score models based on Gurdasani et al. 2019, GLGC and PAGE was assessed in the AWI-Gen dataset. The map was created using R (https://www.r-project.org/).

Fig. 2. Genome-wide associations for LDL-C.

Miami plot showing summary data for Stage 1 GWAS (AWI-Gen, downward facing, N = 10,603) and Stage 2 GWAS (meta-analysis of AWI-Gen and four African cohorts, upward facing, N = 23,718). P-values (two-tailed, not adjusted for multiple comparisons, calculated using BOLT-LMM for Stage 1 GWAS and METASOFT for Stage 2 GWAS) are truncated at 10⁻²⁰ for clarity. The red horizontal lines show the genome-wide significance threshold (5 × 10⁻⁸) and SNPs with P-values below this threshold are shown in orange. The loci corresponding to the region showing novel association in Stage 1 and Stage 2 GWAS are indicated in red. Other possible novel loci that reached genome-wide significance only in the Stage 2 analysis are shown in purple. Known LDL-C associated regions that were represented by novel lead SNPs are shown in green. Loci represented by lead SNPs that are well-known for LDL-C associations across multiple studies, including ours, are shown in black.

Fig. 3. Fine mapping and novel association for LDL-C.

a Locuszoom (http://locuszoom.org/) plots showing LDL-C association around the LDLR region. The plot on top is based on the Prins et al. 2017 study (N = 9961). The 95% credible set in this study included over 40 SNPs (inferred using FINEMAP based results from CausalDB (http://mulinlab.org/causaldb/)). The bottom plot shows the same region in the AWI-Gen Stage 1 GWAS (N = 10,603). Here, the region is represented by a much narrower peak and the 95% credible set, inferred using FINEMAP, includes only two SNPs. b Locuszoom plot showing associations around the GATB region in the Stage 1 (top) (N = 10,603) and Stage 2 (bottom) (N = 23,718) analysis. Although, the lead SNP from Stage 1 GWAS, rs35804417 (pointed to by blue arrow), missed genome-wide significance, a set of SNPs, with the lead SNP rs6845395 (purple diamond), from the neighbouring FH1P1A gene was found to be significant in the Stage 2 GWAS.

Fig. 4. Transferability of previous lipid-trait associations to various African GWASs.

a LDL-C association signals detected in the Global Lipid Genetics Consortium (GLGC) study. b LDL-C association signals detected in Population Architecture using Genomics and Epidemiology Consortium (PAGE) study. The Y-axis shows the proportion of associated loci that are replicated in each of the individual African studies. The proportion of signals replicated at the genome-wide significance threshold P-value < 5 × 10⁻⁸ are shown in deep blue (F-GW), at P-value < 5 × 10⁻⁴ are shown in dark green (F_RepThr) and at the nominal threshold of P-value < 0.05 are shown in orange (F_NT). The signals from the GLGC study were partitioned on the basis of signal strength into Very Strong (P-value < 10⁻¹⁰⁰) indicated by the suffix “_VS” and a green background, Strong (10⁻²⁰  >  P-value > 10⁻¹⁰⁰) indicated by the suffix “_S” and a blue background, and moderate (5 × 10⁻⁸  >  P-value > 10⁻²⁰) indicated by the suffix “_M” and a grey background. For the PAGE study only two categories, Strong (S) (P-value < 10⁻²⁰) and Moderate (M) (5 × 10⁻⁸  >  P-value > 10⁻²⁰) were considered. The African replication datasets used in the analysis are – Stage 2 GWAS (Meta-analysis) (N = 23,718), Stage 1 GWAS (AWI-Gen) (N = 10,603), Uganda Genome Resource (UGR) study (N = 6407), Africa-America Diabetes Mellitus (AADM) study (N = 4116), Durban Diabetes study (DDS) (N = 1117) and Durban case control (DCC) study (N = 1475). Comparison of transferability of signals from c GLGC Consortium study, d PAGE Consortium study in the Stage 2 GWAS results for each of the four lipid traits are shown.

Fig. 5. Transferability of Polygenic Risk Score (PRS) Models derived from a sub-Saharan African, an European and a Multi-ancestry GWAS to the AWI-Gen dataset.

Plots showing additional variance explained (%R²) by each PRS for (a) LDL-C, (b) TC, (c) TG and (d) HDL-C in the AWI-Gen validation dataset (N = 7103). PRS based on sub-Saharan African discovery dataset (AFG, N = 13,115 individuals) is shown in blue, European (GLGC, N = 188,577 individuals) in yellow and Multi-ancestry (PAGE, N = 49,839 individuals) in red. Number of SNPs in each PRS is shown below and P-values (two-tailed estimates using PRSice2) are shown over the bars. All the PRS were significant for all lipid traits indicating transferability. However, the sub-Saharan African and Multi-ancestry PRS models had higher predictive accuracy compared to the European model. PRS stratification of (e) LDL-C, (f) TC, (g) TG and (h) HDL-C. Point range-plots comparing the difference in lipid-trait mean (mmol/L) of the upper PRS decile from the lowest, stratified by the discovery datasets are shown. The error bars show mean ± 95% confidence intervals. Additional details in Supplementary Data 8.

Fig. 6. Heterogeneity of effect size and minor allele frequencies of two LDL-C association signals in African datasets.

(a), (c) show effect size and MAF for rs7412 and (b), (d) show effect size and MAF for rs4788609. The GWASs compared include AWI-Gen West African (A_WEST) GWAS (N = 3763), AWI-Gen East African (A_EAST) GWAS (N = 1755), AWI-Gen South African (A_SOUTH) GWAS (N = 5085), Uganda Genome Resource (UGR) study (N = 6407), Africa-America Diabetes Mellitus (AADM) study (N = 4116), Durban Diabetes (DDS) (N = 1117) Study and Durban Case Control (DDC) Study (N = 1475). The West African populations are shown in orange, East African populations in green and Southern African populations in blue. AADM due to inclusion of both East and West African participants is shown in two colours. Error bars show effect sizes ± standard errors.