Transferability of genetic risk scores in African populations

Abstract

The poor transferability of genetic risk scores (GRSs) derived from European ancestry data in diverse populations is a cause of concern. We set out to evaluate whether GRSs derived from data of African American individuals and multiancestry data perform better in sub-Saharan Africa (SSA) compared to European ancestry-derived scores. Using summary statistics from the Million Veteran Program (MVP), we showed that GRSs derived from data of African American individuals enhance polygenic prediction of lipid traits in SSA compared to European and multiancestry scores. However, our GRS prediction varied greatly within SSA between the South African Zulu (low-density lipoprotein cholesterol (LDL-C), R² = 8.14%) and Ugandan cohorts (LDL-C, R² = 0.026%). We postulate that differences in the genetic and environmental factors between these population groups might lead to the poor transferability of GRSs within SSA. More effort is required to optimize polygenic prediction in Africa.

Fig. 1. Performance of GRSs for lipid traits in the South African Zulu cohort using the MVP GWAS summary statistic results of various ancestry populations, including individuals of African American, European and multiethnic ancestry populations.

a, Violin plots showing GRSs that explained the highest proportion of variance (R²) for lipids derived from African American (AFR), European (EUR) and multiancestry (MEA) populations. b, GRSs in deciles compared to the first decile. The y axis shows the mean, and the x axis is the GRSs in deciles. The points show mean, and error bars represent standard errors of the mean. All South African Zulu cohorts (n = 2,598) were used in this analysis.

Fig. 2. GRSs of individuals of African ancestry with dyslipidemia.

a, Map of Africa showing sample collection points in Kyamulibwa in Kalungu district, Uganda and Durban, Kwazulu-natal province, South Africa. b, Bar plot showing comparative performance of polygenic prediction of TC using the same GRS comprising 286 SNPs, which was developed in Ugandan cohort (n = 6,407) and then replicated in the South African Zulu cohort (n = 2,598). The y axis is the prediction accuracy (R²), and the x axis is the number of SNPs in the GRS for TC used. c, Correlation coefficients between African American-derived GRSs and serum lipid levels in the Ugandan cohort. d, Scatter plot for the correlation of the same minor allele frequencies (MAF) between the South African Zulu and Ugandan cohorts (R = Pearson correlation, one-sided test). PC, principal component. e, Scatter plot for the principal component analysis of the 1000 Genomes Project reference populations with the South African Zulu and Ugandan cohorts (GBR, British; MSL, Mende; UGR, Uganda genome resource; Zulu, South African Zulu; YRI, Yoruba).

Extended Data Fig. 1. Proportion of variance of TC explained by GRS in the South African Zulu samples using GRS derived from GWAS.

(a) African American ancestry, (a) European ancestry and (a) multiethnic ancestry. The bars represent GRS calculated for subsets of markers at different p-value thresholds. The best GRS in red color was selected based on having the highest proportion of the variance (R2) for the trait in linear models adjusted for age, sex and principal components.

Extended Data Fig. 2. Correlation coefficients between GRS and serum lipid levels.

(a) African American derived GRS in the South African Zulu dataset. (b) European derived GRS in the South African Zulu. (c) African American derived GRS in the Ugandan cohort. (d) European derived GRS in the Ugandan cohort. The correlation coefficients r2 are given with colors corresponding to the direction and strength of r2. The r2 on the diagonal represents the strength of correlation of a GRS with its target lipid trait. The off-diagonal r2 represents the strength of correlation of a GRS with other lipid traits.

Extended Data Fig. 3. Box plots showing the distribution of age, BMI and lipid traits among the Ugandan and South African Zulu cohorts.

The horizontal line is the median values, error bars are 25th and 75th percentiles. Extreme values are maximum and minimum for respective traits. Data analysis were performed in all Ugandan (n = 6,407) and South African Zulu (n = 2,598) cohorts.

Extended Data Fig. 4. The discriminative power of our polygenic risk score or GRS to successfully identify the individuals of African ancestry with dyslipidaemia.

(a) Distribution of total cholesterol (TC) among South African Zulus. The top 10% deciles were named “cases,” and the lower deciles were designated as “controls.” (b) The area under the curve in South African Zulu.