Transferability of genetic loci and polygenic scores for cardiometabolic traits in British Pakistanis and Bangladeshis

Genes & Health cohort

Genes & Health (G&H) is a community-based cohort of BPB individuals recruited primarily in East London ¹³. All participants have consented for lifelong EHR access and genetic analysis. The study was approved by the London South East NRES Committee of the Health Research Authority (14/LO/1240). 97.4% of participants in G&H are in the lowest two quintiles of the Index of Multiple Deprivation in the UK. About two thirds are British Bangladeshi and the remainder British Pakistani. The median age at recruitment was 37 and 43 years for female and male participants, respectively (Figure S1). The cohort is broadly representative of the background population with regard to age, but slightly over-sampled females and those with medical problems since two-thirds of people were recruited in healthcare settings such as GP surgeries ¹³. We used the 2020 February data release which contained 28,022 individuals genotyped on the Illumina Infinium Global Screening Array v3 chip (with the additional multi-disease variants). Of these, 22,490 (80%) individuals had linkage to primary or secondary care data, of which 56.5% were female. Having identified related individuals (second degree or closer; kinship coefficient >0.0884) using KING v2.2.4¹⁴, we performed principal component analysis (PCA) in unrelated samples, and projected the remainder onto the same PC space using smartpca from EIGENSOFT v7.2.1¹⁵.

Quality control and imputation of genotype data from Genes & Health

Quality control of genotype data was performed using Illumina’s GenomeStudio and plink v1.9. We first removed variants with cluster separation scores <0.57, Gentrain score <0.7, excess of heterozygotes >0.03, or ChiTest 100 (Hardy-Weinberg test) <0.6 in GenomeStudio, as well as variants that were included on the array in order to tag specific structural variants. We removed samples with low call rate (<0.995 for male samples and <0.992 for female samples across all 637,829 variants including those on Y chromosome for males) and those that failed gender checks. When there were duplicate samples, we retained the sample with the highest call rate. Using plink, we further removed variants with low call rate (<0.99), and the variant with the lowest call rate amongst duplicate variant pairs. We excluded rare variants with minor allele frequency (MAF) <1%. The high levels of autozygosity in this cohort can cause variants to fail Hardy-Weinberg equilibrium test. We thus removed variants that failed the Hardy-Weinberg test (p<1×10⁻⁶) in a subset of samples with low level of autozygosity. To define these ‘low-autozygosity’ individuals, we pruned SNPs (LD r² >0.8) and called runs of homozygosity (RoHs) using plink1.9 with default parameters, then took the 64% of the individuals who had a fraction of the genome in RoHs <0.5%. We excluded individuals who did not have Bangladeshi or Pakistani ancestry (further than +/- 3 standard deviations [SD] from the mean of PC1 for the individuals who self-reported as coming from that group), and those who self-reported as coming from other ethnic groups or who did not report this information (Figure S2).

We used the Michigan Imputation Server¹⁶ to perform imputation with the GenomeAsia pilot reference panel¹⁷, imputing from 336,133 autosomal, biallelic SNPs with matched alleles. Eagle v2.4 and Minimac v4 were used for phasing and imputation, respectively. We excluded SNPs with imputation INFO score <0.3 or MAF <0.1%, which left 9,527,863 autosomal SNPs.

Quality control and imputation of genotype data from eMERGE

We used EUR samples from the eMERGE cohort (henceforth eMERGE), a consortium of US medical research institutions, to compare with BPB individuals from G&H. Network Phase III data (N=61,377) were downloaded from dbGaP (Accession number: phs001584.v1.p1). Quality control of genotype data and imputation to the Human Reference Consortium (HRC) reference panel have been described previously ¹⁸. To identify EUR samples, we performed PCA in samples from the 1000 Genomes project phase 3 dataset, and projected eMERGE participants onto the same PC space using smartpca from EIGENSOFT v7.2.1¹⁵. For PCA, we restricted to LD-pruned common SNPs (MAF ≥1%) with imputation INFO score ≥0.98 in eMERGE. We identified samples that were clustered together with the EUR samples from the 1000 Genomes project using a dimension reduction method, Uniform Manifold Approximation and Projection (UMAP), applied to the first 20 PCs, performed using the R package “umap” v0.2.6.0¹⁹. Self-reported Hispanic or Latino, African, Asian, American Indian or Alaska Native individuals were excluded. This resulted in 43,877 EUR individuals available for the comparison with G&H. Well-imputed (INFO ≥0.3) bi-allelic SNPs with MAF ≥0.1% (N=11,625,805) were retained for downstream analysis.

Phenotype and covariate definitions from electronic health-record data in Genes & Health

Of the 22,490 genotyped G&H individuals with EHR data, 20,830 had primary care data available through the Discovery Data Service²⁰ which includes clinical observations as well as current and historic diagnoses (coded using READ version 2 codes, and recently converted to SNOMED CT codes using standard mapping protocols²¹). 17,226 had diagnosis and procedure codes (ICD10 and OPCS4 codes, respectively) extracted from the UK’s largest secondary care health provider, Barts Health NHS Trust.

Coronary artery disease (CAD) cases and controls were defined using the same ICD10 and OPCS4 codes as Khera et al. ²² (Table S1). We defined CAD cases as those with myocardial infarction or coronary revascularization in either primary and secondary care data. We excluded individuals with angina, chronic ischemic heart disease, aneurysm or atherosclerotic cardiovascular disease from the control sample ²³. Since procedure codes were not available in eMERGE, we performed a sensitivity analysis in G&H to investigate the effects of excluding OPCS4 codes in CAD ascertainment. For this, we defined CAD solely using ICD10 codes in individuals with secondary care data, ignoring OPCS codes and primary care data; we excluded individuals without secondary care data for this analysis.

We used median adult height and weight measurements within the past 5 years to calculate BMI. For lipids, we took the latest adult measurements and corrected for statin usage if lipid levels were measured between the start and end date of any statin prescriptions. No adjustment was made on HDL cholesterol (HDL-C) or triglycerides. Adjustment of lipids followed the procedure in Liu et al. ²⁴, as follows. To correct for statin usage, total cholesterol (TC) was replaced by TC/0.8. LDL cholesterol (LDL-C) levels were calculated using the Friedewald equation, and statin-adjusted LDL-C was recalculated using adjusted TC levels as follows: corrected LDL-C = uncorrected LDL-C + 0.2*adjusted TC. LDL-C/0.7 was used for 32 individuals for whom we couldn’t find a TC measurement on the same date. Rank-based inverse normal transformation was applied to the lipid levels.

We extracted the latest systolic blood pressure (SBP) and diastolic blood pressure (DBP) measurements and adjusted for blood pressure medication use by adding 15 and 10 mmHg to SBP and DBP, respectively, if the measurement coincided with any prescription date ²⁵. Sample sizes are shown in Table 1 (all individuals) and Table S2 (unrelated).

To calculate a standard clinical risk score to compare with the PGS, we calculated the QRISK3 10-year predicted risk for CAD ²⁶ in G&H using the R package “QRISK3” v0.3.0 ²⁷. QRISK3 was calculated based on the data available up until 1 January, 2010, which is about 10 years prior to the latest data extraction. We excluded about one third of CAD cases whose diagnosis was made earlier than this assessment date (prevalent cases) and used incident cases who developed CAD later. Follow-up varied for cases and was fixed at 10 years for controls. We used clinical data that were extracted earlier than the assessment date (1 January 2010) to calculate QRISK3. The QRISK3 algorithm has variables that indicate whether a patient has a variety of other diseases, and these were defined using the codes shown in Table S3, following ¹⁰. Medication use (hypertension treatment, corticosteroid, and atypical antipsychotic medication) was defined as two or more prescriptions, with the most recent one having been issued within 28 days prior to the assessment. We used the most recent measurements taken prior to the assessment date, and kept individuals with at least three non-missing measurements out of four (height, weight, SBP, and TC). Pattern of missingness is shown in Figure S3. Townsend index was not available in G&H, so we used the mean value (3.307) of the lowest two quintiles from the 2011 census data in the UK ²⁸. HDL-C levels were all measured later than 2010 in G&H, so for TC/HDL-C ratio, we used 3.905 and 4.882 (averages calculated using later data) for females and males, respectively. To deal with missing data, we applied multiple imputation which accounts for sex, age, and genetically-defined ancestry (Bangladeshi versus Pakistani; identified using PCA-UMAP), using the R package “mice” v3.13.0 to impute height, weight, SBP, SD of SBP measurements within 2 years, and smoking status.

Phenotype definitions from electronic health-record data in eMERGE

Phenotype data in eMERGE were downloaded from dbGaP (phs001584.v1.p1, phs000888.v1.p1, phs001584.v2.p2). Individuals younger than 16 years old were excluded. BMI was provided and we took the median value from adult measurements. Lipid and blood pressure measurements were taken from dataset phs000888.v1.p1. Data on medications affecting lipid and BP measurements were not available, so the highest measurements for LDL, TC, SBP, and DBP were used when comparing PGSs with G&H in order to minimise the effects of medications. CAD was ascertained using ICD9/10 codes which were available in the updated eMERGE Phase III dataset (phs001584.v2.p2). Coronary artery disease (CAD) cases and controls were defined based on secondary care ICD10 codes as described above for G&H (Table S1).

Genome-wide association analyses in Genes & Health

GWAS was performed with SAIGE ²⁹ and adjusted for age, age², sex and the first twenty principal components. For total cholesterol and LDL-C, adjustments were made for use of statins as described above. We followed the QC procedure in ³⁰ (EasyQC package) with the following exclusion criteria for variants: monomorphic variants, missing / invalid estimates, allele mismatch and allele frequency difference of >0.2 with reference panel, imputation INFO score <0.7 (<0.9 for downstream analysis i.e. correlation and colocalisation), MAF <0.005 (<0.01 for downstream analysis i.e. correlation and colocalisation).

Heritability and trans-ancestry correlations

Datasets that were used in analyses are provided in Table S4. We used GCTA to estimate SNP heritability in G&H and eMERGE³¹. We excluded one sample in each pair of 3^rd-degree relatives (kinship coefficient >0.0442 calculated using KING v2.2.4¹⁴). We used SNPs with INFO >0.9 and MAF >0.01 in each cohort separately. We also calculated SNP heritability using the intersection of these SNP sets in both cohorts. For CAD, we estimated SNP heritability on the liability scale using 6.7% as the prevalence estimate in the US ³², and 3.33% for the UK background population from which G&H is sampled, defined as all people from South Asian ethnicities (N=255,066 aged ≥20 years) registered with a primary health physician/GP in four east London boroughs.

For the genetic correlation analyses, we used GWAS summary statistics generated in EUR individuals from UK Biobank (UKBB), since we needed a larger sample size of ancestrally homogeneous individuals than is available through eMERGE to obtain accurate estimates. We used Popcorn (https://github.com/brielin/Popcorn) to estimate the trans-ancestry genetic correlations between G&H and UKBB EUR individuals while accounting for differences in LD structure ³³ (i.e. the correlation of causal-variant effect sizes across the genome at SNPs common to both populations). Variant LD scores were estimated for ancestry-matched 1000 Genomes v3 data for each study combination (i.e. SAS-EUR). The estimation of LD scores failed for chromosome 6 for some groups, so we left out the major histocompatibility complex (MHC) region (positions 28,477,797 to 33,448,354) from chromosome 6 from all comparisons. Variants with INFO score <0.9 or MAF <0.01 were excluded. A p-value <0.05 indicated that the genetic correlation was significantly less than 1 i.e. r_g <1.0.

Assessment of transferability of established loci

Previous studies that evaluated reproducibility of GWAS loci in SAS individuals did not formally account for differences in power or LD patterns ^34–36. We assessed whether established trait-associated loci were reproducible in G&H by performing a lookup of loci identified in non-SAS ancestry GWAS (Table S4). Credible sets for established loci were generated and consisted of lead (independent) variant plus proxy SNPs (r² >=0.8) within a 50kb window (based on the EUR 1000 Genomes data) of the sentinel variant and with p-value <100 × p_sentinel. The locus was defined as being ‘transferable’ if at least one variant from the credible set was associated at p <0.05 with the relevant trait in G&H, and the direction of effect matched in both datasets. For loci harbouring multiple signals, we only kept the most strongly associated variant (i.e. smallest p-value). Expected power for replication was calculated using alpha=0.05, the effect size estimated in the EUR GWAS, and the allele frequency of the variant and sample size in G&H. The power of lead variants per locus was summed up and divided by the number of loci to give an estimate of the number of expected significant loci per trait, which was compared with the observed number of such loci; to our knowledge, this is a novel approach for assessing reproducibility of GWAS findings. Loci were only deemed to be ‘non-transferable’ if they contained at least one variant in the credible set with >80% power and yet none of the variants in the credible set had p <0.05 and no variant within 50kb of locus had p <1×10⁻³ in G&H. LocusZoom (http://locuszoom.org/) was to create regional association plots.

Trans-ancestry colocalisation

We used the Trans-ethnic colocalisation method (TEColoc) (https://github.com/KarolineKuchenbaecker/TEColoc)³⁷ which tests whether a specific locus has the same causal variant in two groups with different ancestry, and applied it to G&H and UKBB EUR individuals. This method adopts the joint likelihood mapping (JLIM) statistic developed by Chun and colleagues³⁸ that estimates the posterior probabilities for colocalisation between GWAS signals and compares them to probabilities of distinct causal variants while explicitly accounting for LD structure. For this, LD scores were estimated using a subset of samples from the 1000 Genomes Project v3 that had matching ancestry to all Europeans for UK Biobank. For G&H we used raw genotype data and LD was estimated directly for these samples. JLIM assumes only one causal variant within a region in each study. We therefore used small windows of 50Kb for each known locus to minimise the risk of interference from additional association signals. Distinct causal variants were defined by separation in LD space by r² ≥0.8 from each other. We excluded loci where the overlap between UKBB and G&H was <10 SNPs and the proportion of well-imputed SNPs overlapping between cohorts (SNP coverage) was <10%; this left no loci to consider for CAD, SBP and DBP. We used a significance threshold of p <0.05 to determine evidence of sharing. LocusZoom (http://locuszoom.org/) was to create regional association plots.

Construction of polygenic risk scores

We evaluated the performance of PGSs in G&H and eMERGE. We first assessed PGSs that were previously constructed (mostly optimised in EUR samples) from the PGS Catalog ³⁹. We restricted to 7,353,388 bi-allelic SNPs that had INFO ≥0.3 and MAF ≥0.1% in both eMERGE and G&H. Variant information in existing PGS was harmonised to GRCh37 using dbSNP mappings from Ensembl Variation and liftover. We calculated PGSs as weighted sums of imputed allele dosages using plink2.0 --score function. There were often multiple PGSs that were previously developed from different studies available for each trait, and below we report the one that had the highest accuracy in each cohort. The best PGS (defined as described in the next section of the Methods) for BMI was derived from GWAS conducted in primarily EUR samples and optimised in EUR individuals, and those for lipids and BP contained genome-wide significant variants identified in EUR GWASs. We selected different PGSs for CAD in eMERGE and G&H, with the former optimised in EUR individuals and the latter in SAS individuals; in both cases these were based on GWAS conducted in primarily EUR samples. The details of each PGS are in Table S5.

Next we calculated PGSs using the clumping and p-value thresholding method (C+T) and optimised PGSs in G&H and eMERGE separately. We used GWAS summary data from primarily EUR samples (Table S4). We used LD estimated using EUR samples (N=503) from the 1000 Genomes project for clumping using PRSice2 v2.2.11⁴⁰. We calculated multiple scores using combinations of various LD r² thresholds (0.1, 0.2, 0.5, 0.8) and p-value thresholds (5×10⁻⁸, 1×10⁻⁷, 5×10⁻⁷, 1×10⁻⁶, 5×10⁻⁶, 1×10⁻⁵, 5×10⁻⁵, 1×10⁻⁴, 5×10⁻⁴, 0.001, 0.005, 0.01, 0.02, 0.05, 0.08, 0.1, 0.2, 0.3, 0.4, 0.5, 0.8, 1) for each trait, and reported the PGS with the best predictive performance within each target cohort.

Lastly, we calculated meta-PGSs proposed by Marquez-Luna et al.⁴¹ that incorporate GWAS summary data from the target populations. We downloaded GWAS summary data that were generated in SAS samples of the UKBB from the Pan-UK Biobank website (https://pan.ukbb.broadinstitute.org), and constructed scores (PGS_SAS) using the C+T method described above and using SAS samples from the 1000 Genomes project for the LD reference. We combined the scores derived from EUR GWASs (PGS_EUR) and PGS_SAS in linear regression to construct meta-PGSs.

Assessment of PGS accuracy and clinical performance

We excluded one sample in each pair of 2^nd-degree relatives (kinship coefficient >0.0884 calculated using KING v2.2.4¹⁴). Individuals with the highest number of relatives (and controls, if the trait is binary) were removed first. Sample sizes for each trait are in Table S2. Quantitative traits were inverse normal transformed. Age at recruitment was used as a covariate for analysis of disease status, and age at measurement for analysis of quantitative traits. PGSs were standardised to a mean of 0 and SD of 1. We fitted the following two models: (1) the full model which had PGS and covariates namely sex, age, age², and the first 10 genetic PCs, and (2) the reference model which accounted for the covariates only. For continuous risk factors, linear regression was fitted, and the gain in R² when adding PGS as an additional predictor, or incremental R², was calculated as the difference between the R² of the full model and the reference model. Logistic regression was used to assess the associations between PGSs and CAD. The area under the receiver operating characteristic curve (AUC) was estimated for both models with the R package “pROC” v1.16.2 and incremental AUC was calculated similarly. We performed bootstrap resampling of individuals 1,000 times to estimate the 95% confidence intervals for incremental R² and incremental AUC. The best PGS per trait was the one with the highest incremental R² for continuous risk factors and the one with the highest incremental AUC for CAD. We estimated the effect size (or odds ratio for binary traits) per SD of PGS from the full model. Effect size or odds ratio for quintiles, and for top 10% versus middle 40-60% were reported as well. Relative accuracy was calculated as the ratio of incremental AUC (or incremental R² for continuous traits) in G&H to that in eMERGE.

QRISK3 scores were calculated for 8,112 unrelated individuals as described above (420 CAD cases and 7,702 controls). To integrate QRISK3 scores with PGS for CAD, we followed Riveros-Mckay et al.¹⁰ and calculated an integrated score by multiplying the odds converted from the QRISK3 score with the odds ratio given an individual’s PGS, where the odds ratio per SD of PGS was estimated using a logistic regression in which QRISK3 and their interaction were accounted for. The logistic regression was performed in males and females separately. We used the most accurate PGS for CAD in SAS from the PGS Catalog, which was developed by Wang et al.⁴²; this score was derived from EUR GWAS using LDpred and tuned in SAS individuals in UKBB. We regressed out 10 PCs from the PGS, and used the scaled residuals in the Cox regression analysis. Cox regression was performed using the R package “survival” v3.2-7. The concordance indices (C-indices) of the following models were compared: (1) age at assessment + gender, (2) PGS + age at assessment + gender, (3) QRISK3, and (4) the integrated score. We calculated the continuous net reclassification index (NRI) and categorical NRI (using 10% as the threshold to classify high-risk individuals) for the integrated score compared to QRISK3 alone. NRI was calculated as the sum of NRI for cases and NRI for controls (noncases):

For continuous NRI, P(up|case) and P(down|case) indicate the proportions of cases that had higher or lower risk estimates using the integrated score, respectively. For categorical NRI, P(up|case) indicates the proportions of cases that were reclassified as high-risk individuals (i.e. with <10% risk by QRISK3 but >10% by the integrated scores). We calculated NRI in two age groups (25–54 versus 55–84 years old at baseline, chosen since the average age of onset in this cohort was 55.3 years old), as well as in sex-by-gender subgroups. Bootstrap resampling (1,000 times) was used to estimate confidence intervals for NRI.

Mendelian randomisation analysis

We modelled liability to CAD as our outcome within a two-sample Mendelian randomisation⁴³ (MR) framework using the risk factors (BMI, SBP, DBP, LDL-C, HDL-C, TG) as exposures. To identify genetic instruments for the exposure, we explored three alternative approaches: (a) established loci significant at p<5×10⁻⁸ in the original EUR GWAS; (b) transferable loci defined as described above, taking the effect size from the original EUR GWAS; and (c) loci significant at p<5×10⁻⁸ in the SAS ancestry group of the Pan-UKBB GWAS, LD-clumped to an r²<0.2 with a LD window of 50kb, based on SAS 1000 Genomes project LD reference. Where insufficient genome-wide significant instruments were identified, we used a more permissive p-value threshold of p<5×10⁻⁵ for instrument selection in UKBB SAS. The primary MR analysis was performed using, as outcome, summary association data from the G&H CAD GWAS performed as described above, using the inverse-variance weighted method under a random effect model, implemented with the TwoSampleMR R package⁴⁴. For comparison, a two-sample MR approach was also performed using summary data for CAD from eMERGE and established loci significant at p<5×10⁻⁸ in the original EUR GWAS. We also undertook several sensitivity analyses. In brief, we evaluated the MR-Egger intercept to assess directional pleiotropy and Cochran’s Q statistic⁴⁵ as an indicator of heterogeneity. MR analysis using weighted median⁴⁶ and weighted methods⁴⁷ models were additionally performed in the presence of heterogeneity.

Shared genetic architecture of cardiometabolic traits

We compared the genetic architecture of coronary artery disease (CAD) and upstream risk factors, namely HDL-C, LDL-C, triglycerides (TG), total cholesterol (TC), systolic and diastolic blood pressure (SBP & DBP), between British Pakistanis and Bangladeshis (BPB) from G&H, and European-ancestry populations (EUR) (Figure 1). We used EUR individuals from the EHR-based eMERGE cohort to estimate heritability, since phenotypes had been ascertained in a similar way to G&H (i.e. EHRs). All traits were found to have significant SNP heritability (h²= 0.03–0.23) in G&H, with estimates similar to those in eMERGE (Table S7, Figure 2A), except for LDL-C, SBP and DBP which had significantly lower values in G&H than eMERGE (e.g. for LDL-C, h² was 0.21 [95% CI: 0.17–0.25] in eMERGE and 0.06 [95% CI: 0.02–0.10] in G&H). Conclusions were unchanged when restricting the heritability estimates to the same set of well-imputed SNPs in both cohorts (Table S7). We observed high genetic correlations between G&H and EUR from UKBB for all traits, with the lowest value seen for SBP (r_g=0.71 [95% CI: 0.36–1.06], p=0.09; Figure 2B). The only trait for which the genetic correlation differed significantly from one was BMI (r_g=0.85 [95% CI: 0.71–0.99], p=0.02).

• Download figure
• Open in new tab

Figure 1. Summary of study design, research questions and analyses conducted.

The coloured boxes indicate input data. Within the white boxes, black text indicates the analyses we used to address the questions in blue. BPB: British Pakistanis and Bangladeshi ancestry; EUR: European ancestry; SAS: South Asian ancestry; CAD: coronary artery disease; BMI: body mass index; SNP: single nucleotide polymorphism; GWAS: genome-wide association study; MR: Mendelian randomisation; PGS: polygenic score; UKBB: UK Biobank. Datasets and discovery GWAS that were used in each analysis are provided in Table S4.

• Download figure
• Open in new tab

Figure 2. SNP heritability and trans-ancestry genetic correlations for cardiometabolic traits.

A. SNP heritability was estimated using GCTA in G&H (orange) and eMERGE (grey) for cardiometabolic traits. Dashed line indicates statin-adjusted LDL cholesterol levels in G&H which are not available in eMERGE. Error bars represent 95% confidence intervals in both plots. B. Genetic correlations were estimated using GWAS summary statistics generated from G&H and European-ancestry individuals from UK Biobank. Red indicates that the genetic correlation is significantly lower than 1 (p-value = 0.02 for BMI). BMI: body-mass index; BPB: British Pakistani ancestry; CAD: coronary artery disease; EUR: European ancestry; LDL-C: low-density lipoprotein cholesterol; G&H: Genes & Health; HDL-C: high-density lipoprotein cholesterol; SBP: systolic blood pressure; DBP: diastolic blood pressure.

High transferability of cardiometabolic loci

We assessed whether published trait-associated genomic loci identified in predominantly EUR populations were shared by the BPB population represented by G&H. To account for differences in LD patterns, our assessment of transferability was based on the credible sets of variants per locus, likely to contain the causal variant, rather than the sentinel variants alone. Low numbers of transferable loci may be due to limited statistical power rather than lack of causal variant sharing. Therefore, we compared the number of observed transferable loci with the number expected given the sample size and allele frequency in G&H if all causal variants were shared. The number of expected transferable loci varied widely between traits (e.g. we expected to be able to detect significant associations for 56% of HDL-C loci but only for 18% of SBP loci), highlighting the importance of accounting for power when assessing transferability. Across most traits examined, the observed number of transferable loci closely matched the loci we expected (Table 1 and Table S8). For example, for BMI we expected to be able to find evidence for transferability for 20% of loci and we did indeed observe transferability for 21% of loci. However, the exception was CAD for which the number of observed transferable loci (13%) was below the expected number (21%), although this difference was only marginally significant (binomial p-value = 0.05).

We also assessed whether there were any specific loci that were not transferable despite being well powered to observe an association (power >80%). Out of a total of 184 well-powered loci tested across all traits, only nine were non-transferable; that is, no variant in the credible set was significant at p<0.05 and no variant within 50kb of locus was significant at p<1×10⁻³ (Figure S4). These nine loci were all associated with lipid traits: EVI5, NBEAL1, GPAM, CETP, STAB1, TTC39B, SH2B3, ACP2 and NECAP2 (Table 2). Of these loci, CETP, which was previously associated with LDL-C levels in Europeans (established variant in Europeans -rs7499892), was strongly associated with HDL-C in G&H (p=7.08×10⁻⁵⁶), but not with LDL-C levels (p=0.23) (Figure S5) despite having >80% power for replication.

Even when there are associations in the same region in two ancestry groups, it is possible that they are driven by different causal variants, as previously seen ⁴⁹. To assess the extent of sharing of causal variants between ancestries at previously reported loci with evidence of transferability, we applied trans-ancestry colocalisation for G&H with UKBB EUR samples as the reference. We found evidence for the most extensive sharing of causal variants for transferable lipid loci: total cholesterol (61% of loci had significant colocalisation), followed by TG (56%), HDL-C (48%) and LDL-C (47%) (Table 1). For BMI we found evidence for sharing of causal variants for only 26% of transferable loci assessed (Table 1 and Table S9). Causal variants in major lipid loci such as PCSK9 were among variants that were consistently not shared (p_JLIM>0.05) between the two populations (Figure S6 and Table S9).

Variable transferability of polygenic scores

Polygenic scores (PGSs) for CAD have been shown to have predictive value over risk scores based on clinical factors alone ^{10,11,42,50–54}. To assess the transferability of PGSs for cardiometabolic traits derived from EUR populations into BPB individuals, we compared predictive performance in G&H to that in EUR individuals from eMERGE. We quantified predictive accuracy using the “incremental AUC” statistic for CAD and the “incremental R²” statistic for continuous risk factor traits; these are the gain in AUC or R² when adding the PGS to the regression of phenotype on the baseline covariates (sex, age and genetic PCs).

We first evaluated the previously published PGSs from the PGS Catalog (Table S5). The PGSs for risk factors were developed using data from primarily EUR individuals, and the CAD PGSs that proved to have the best performance in G&H and eMERGE were two different scores optimised in SAS ⁴² and EUR samples ⁵³, respectively. PGSs for all traits assessed were significant predictors in G&H (Table S5, Figure 3A). For prediction in G&H, the incremental R² for BP was low (∼1.8%), but it was higher for lipids and BMI, ranging from 3.9% to 6.7%. Relative accuracy of PGS in eMERGE and G&H, determined by the ratio of incremental AUC or R², was close to 1 for HDL-C, TG, SBP and DBP, and lower for CAD (42%, 95% CI: 30%–59%) and BMI (78%, 95% CI: 68%–88%; Figure 3B). Amongst the risk factors, prediction of LDL-C had the lowest relative accuracy (66%, 95% CI: 53%–79%), probably due to the fact that we did not adjust for statin usage since medication data were not available in eMERGE, and BPB individuals were more likely to be treated with statins ⁵⁵. Incremental R² for the PGS for LDL-C increased from 3.9% (3.3%–4.5%) to 6.2% (5.3%– 7.1%) when using statin-adjusted LDL-C in G&H (Table S5, Figure 3A), although the heritability was not significantly different (Figure 2A). In a sensitivity analysis, the relative accuracy of the CAD PRS in eMERGE versus G&H was consistent when defining CAD based on diagnostic codes only, rather than with the inclusion of procedure codes in the G&H definition (Table S5).

• Download figure
• Open in new tab

Figure 3. Comparison of the predictive accuracy of polygenic scores in people of British Pakistani and Bangladeshi versus European ancestry.

A. Predictive accuracy of polygenic scores (PGSs) for cardiometabolic traits in British Pakistani and Bangladeshi (BPB) individuals from G&H (purple) and European-ancestry (EUR) individuals from eMERGE (green). Incremental AUC was calculated for coronary artery disease (CAD), and incremental R² was calculated for its continuous risk factors. Error bars indicate 95% confidence intervals (CIs) estimated by bootstrap resampling of samples. The highest measurements for low-density lipoprotein cholesterol (LDL-C), systolic blood pressure (SBP), and diastolic blood pressure (DBP) are compared between eMERGE and G&H, and statin-adjusted LDL-C data are also shown for G&H. B. Relative accuracy of PGSs (i.e. the ratio of incremental AUC for CAD or incremental R² for risk factors estimated in G&H to that in eMERGE) for PGS Catalog scores (red) and C+T scores (blue). Error bars represent 95% bootstrap CIs. Panel C and D show the effect sizes of PGSs from the PGS Catalog. C. The odds ratio per standard deviation (SD) of PGS is shown for CAD on the left panel, and the differences in phenotypic SD per SD of PGS are shown for quantitative traits on the right panel. D. The odds ratio for CAD comparing the four quintiles to the middle quintile (40–60%) is shown on the left panel. Quintiles are determined in controls. The differences in phenotypic SD compared to the reference quintile are shown on the right panel. Error bars show 95% CIs estimated using the standard error in C and D.

To assess whether the performance of PGS based on EUR GWAS could be improved in BPB, we next constructed PGS using the clumping and P-value thresholding (C+T) method and optimised them separately within G&H and eMERGE. The numbers of SNPs in the best C+T PGSs are similar between eMERGE and G&H, and PGSs for lipids contained fewer SNPs (194 to 454) than other traits (>20,000; Table S10, Figure S7). C+T PGSs and PGSs from the PGS Catalog showed similar performance in G&H across traits, although they were optimised in different ancestry populations (BPB and primarily EUR, respectively; Figure 3A). For BMI, triglycerides and HDL-C, we observed slightly larger differences in predictive accuracies between G&H and eMERGE for C+T PGSs than observed with the PGS Catalog scores (Figure 3B).

We then assessed whether PGS methods that account for ancestry differences improved predictive accuracy in G&H. PGSs were constructed using a meta-score strategy ⁴¹, combining the EUR-derived PGS (described above) and that from UKBB SAS samples. The improvement in accuracy was modest (5–11%) (Figure S8). This may be due to the low sample sizes in the UKBB SAS GWASs.

Modest improvement in CAD risk prediction by adding PGS to clinical risk score

A CAD PGS derived from EUR GWAS summary statistics and tuned in SAS individuals from UKBB⁴² (PGS000296 in the PGS Catalog), showed the highest predictive accuracy in BPB individuals in G&H. This score had an OR per SD of 1.63 (95% CI: 1.51–1.76) and incremental AUC of 0.009 (95% CI: 0.006–0.012; Table S5). Individuals in the top quintile of PGS were predicted to have a 2.2-fold increase (95% CI: 1.78–2.76) in disease risk relative to the middle quintile (quintiles were determined in controls; Figure 3D). We investigated the additional predictive power of PGS on top of established clinical risk factors for CAD, and the net reclassification improvement (NRI) achieved by adding the PGS to a clinical risk score.

To calculate the clinical risk score, we used the QRISK3 algorithm to estimate 10-year risk of cardiovascular disease at a baseline time point, selected so that the participants in G&H had about 10 years of follow-up. QRISK3 was a strong predictor of CAD events and had a concordance index (C-index) of 0.843 (95% CI: 0.828–0.858; Figure 4A, Table S11). Consistent with previous findings in EUR individuals¹⁰, the CAD PGS was uncorrelated with QRISK3 (Pearson’s correlation coefficient r=-0.0056 and p-value=0.62). We followed Riveros-Mckay et al.¹⁰ to construct an integrated score combining QRISK3 and the CAD PGS. The integrated score had a non-significant improvement in the C-index (0.853, 95% CI: 0.838– 0.867). However, compared with QRISK3 alone, the integrated score showed significant improvement in reclassification (categorical NRI: 3.9%; 95% CI: 0.9%–7.0%) using a 10-year risk threshold of 10% based on the threshold for preventive intervention with statin treatment recommended by National Institute for Health and Care Excellence ⁵⁶. The integrated score reclassified 3.2% of the population as high risk and 2.5% as low risk (Table S11). This improvement was mostly driven by the enhanced identification of CAD cases in people at 25– 54 years old (NRI in cases being 7.0% versus NRI in controls being -1.2%), and of controls in people at 55–84 years old (NRI in cases being 0.0% versus NRI in controls being 6.8%) (Figure 4B, Table S11). The QRISK3 classified most (91.4%) of the individuals at 55–84 years old as high risk. Using the integrated score, 7.6% of the individuals older than 55 years were down-classified from high to low risk (Table S11). Using continuous NRI, the integrated score showed significant improvement (27.0%; 95% CI: 17.7%–36.2%) and similar trends in age groups (Figure S9, Table S11).

• Download figure
• Open in new tab

Figure 4. Model discrimination and net reclassification index for coronary artery disease with addition of a polygenic score to QRISK3.

A. The concordance index (C-index) of QRISK3 (red) and an integrated score that combines QRISK3 and a polygenic score (PGS) for CAD (blue) in all British Pakistani and Bangladeshi individuals from G&H as well as in age-by-gender subgroups. The error bars represent 95% confidence intervals (CIs) estimated using the standard error. B. Categorical net reclassification index (NRI) for the integrated score compared to QRISK3 in all samples as well as in age-by-gender subgroups. NRI in cases (red) and controls (blue) are also shown. The error bars indicate 95% CIs estimated using the bootstrap method.

Causal effects of CAD risk factors largely consistent across ancestries

We carried out two-sample Mendelian randomisation (MR) analyses to assess the causal effects of the risk factors on CAD in G&H and compared findings with EUR samples from eMERGE. For G&H, we used transferable loci as genetic instruments to benefit from the precision of large EUR discovery GWAS whilst ensuring only valid instruments are used. In eMERGE, causal effects for BMI, BP and lipids, except TG, were statistically significant (Figure 5). Consistent with this, we found that higher BMI (OR=1.73, p-value=0.01), higher LDL-C (OR=1.55 p-value=4×10⁻⁴) and lower HDL-C levels (OR=0.75, p-value=8×10⁻³) were causally associated with increased risk of CAD in G&H. The OR for LDL-C was larger than the one in eMERGE (OR=1.15) although with overlapping confidence intervals (CI: 1.03-1.29 in eMERGE, CI: 1.22-2.00 in G&H). The effects for SBP and DBP were not statistically significant in G&H. However, both had relatively small numbers of loci as instruments and confidence intervals of the effect estimates were wide.

• Download figure
• Open in new tab

Figure 5. Estimates of causal effects of risk factors on coronary artery disease in European (eMERGE) and British South Asian (G&H) ancestry individuals.

Two-sample Mendelian randomisation (MR) estimates for the causal effects are presented based on genetic instrument variables identified from EUR discovery GWAS for each risk factor. All independent genome-wide significant loci were used as instruments for eMERGE and only the transferable loci for G&H. Effect estimates are presented as odds ratios with 95% confidence intervals per standard deviation increase in the reported unit of the trait: triglycerides (TG), systolic blood pressure (SBP), low-density lipoprotein cholesterol (LDL-C), high-density lipoprotein cholesterol (HDL-C), diastolic blood pressure (DBP), body mass index (BMI). The p-value (P) and number of single nucleotide polymorphism instruments (N SNPs) included in the MR analysis are shown for each exposure.

We also assessed different strategies for instrument selection in G&H, such as using all loci associated at genome-wide significance in EUR GWAS for the risk factors (Figure S11). When following the standard approach of using an independent ancestry-matched sample (UKBB SAS) to derive the instruments, an insufficient number of genome-wide significant instruments (p<5×10⁻⁸) were identified (Figure S10). To address this, we also tested a less stringent p-value threshold (p<5×10⁻⁵) for selecting instruments. For the lipid biomarkers, the results were consistent regardless of which loci were chosen as instruments (Figure S11). However, the association of BMI with CAD was significant only for transferable loci (Figure S11).

We found evidence of heterogeneity between causal estimates based on Cochran’s Q statistic for DBP when using the established loci as instruments (p-value=0.04), LDL-C when using the UKBB SAS-ascertained loci (p-value=0.02) and HDL-C for transferable loci (p-value=1×10⁻³). However, the results of the weighted median and weighted mode models were consistent with those obtained by the inverse-variance weighted MR model (Table S12).