Transferability of genetic loci and polygenic scores for cardiometabolic traits in British Pakistani and Bangladeshi individuals

Individuals with South Asian ancestry have a higher risk of heart disease than other groups but have been largely excluded from genetic research. Using data from 22,000 British Pakistani and Bangladeshi individuals with linked electronic health records from the Genes & Health cohort, we conducted genome-wide association studies of coronary artery disease and its key risk factors. Using power-adjusted transferability ratios, we found evidence for transferability for the majority of cardiometabolic loci powered to replicate. The performance of polygenic scores was high for lipids and blood pressure, but lower for BMI and coronary artery disease. Adding a polygenic score for coronary artery disease to clinical risk factors showed significant improvement in reclassification. In Mendelian randomisation using transferable loci as instruments, our findings were consistent with results in European-ancestry individuals. Taken together, trait-specific transferability of trait loci between populations is an important consideration with implications for risk prediction and causal inference.

Understanding the determinants of excess CAD burden in South Asian populations and improving prediction to enable preventive interventions to represent important public health priorities.
Common genetic variation is an important determinant of CAD and of upstream risk factors, such as blood pressure, lipids, and body mass index (BMI). The genetic component of disease risk can be harnessed to identify underlying disease genes and pathways, to estimate the unconfounded effects of risk factors by Mendelian randomisation, and to improve risk prediction through the application of polygenic scores (PGS). However, the genetic basis of CAD risk is not well characterised in South Asian ancestry populations because genome-wide association studies (GWAS) have been mostly limited to European ancestry populations 2 .
Fundamental questions remain about the extent to which the genetic determinants of cardiometabolic traits are shared by European and South Asian ancestry populations. These have important implications for translational applications of genetic data, such as causal inference with Mendelian randomisation which could prioritise different prevention strategies or drug targets between ancestries, and clinical risk prediction. Whilst the predictive performance of PGSs derived from European ancestry populations in other ancestry groups decreases with genetic distance [3][4][5][6] , the extent to which this attenuation is due to genetic drift (differences in linkage disequilibrium and allele frequency 7 ) versus heterogeneity of causal genetic effects remains unclear.
Furthermore, most previous large-scale studies assessed genetic risk prediction using data from a research setting. These findings may not generalise well to a real-world clinical setting. Firstly, clinical risk factors may be measured less comprehensively than in a research setting, affecting the performance of integrated risk models combining these factors with PGSs [8][9][10] . Secondly, there is evidence that the performance of PGSs may be modified by factors, such as educational attainment and socioeconomic status for which research studies are often not representative 11,12 . The robustness of PGSs applied to South Asian-ancestry individuals in a real-world healthcare system is largely unknown.
Here, we perform a comparative analysis of the genetics of CAD and upstream cardiometabolic traits in European and South Asian ancestry populations, using data from the Genes & Health (G&H) cohort 13 . G&H is a community-based cohort of British Pakistani and Bangladeshi individuals with linked electronic health record data (N = 22,490 individuals). This unique cohort represents an understudied and clinically vulnerable population with high levels of socioeconomic deprivation, and this is, to our knowledge, the first major genetic study focused on it. We apply new approaches to the transferability of genomic risk loci across populations, perform ancestryspecific and trans-ancestry Mendelian randomisation analysis, investigate the transportability of PGSs for CAD and its risk factors, and estimate the incremental improvement in CAD prediction when incorporating the CAD PGS into clinical risk tools.

Results
We conducted GWAS of CAD and key cardiometabolic traits in the G&H cohort which was the primary data resource for this study. In G&H, 4.9% (N = 1110) of the individuals had coronary artery disease (CAD), with the age of onset ranging from 17 to 97 years old (median 55). A quarter of the G&H participants were on active statin prescriptions, 23% on BP medications, 29% had high TC levels (>5 mmol/L), and 30% had high LDL-C levels (>3 mmol/L; Supplementary Data 6).
We used publicly available GWAS summary statistics derived from predominantly European ancestry individuals to compare the genetic architecture and assess whether reported GWAS loci are transferable to G&H (Fig. 1). We evaluated the performance of European ancestryderived PGSs in G&H and compared it to performance in European ancestry samples from eMERGE. Finally, we used Mendelian randomisation analysis to test the causal relationship between the cardiometabolic traits and CAD by comparing genetic instruments based on the GWAS data generated from European ancestry and from British South Asian ancestry individuals. The datasets that were used in each analysis are described in Supplementary Data 4.

Shared genetic architecture of cardiometabolic traits
We compared the genetic architecture of 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 from G&H, and European ancestry individuals from the electronic health record-based eMERGE cohort, since phenotypes had been ascertained in a similar way. All traits were found to have significant SNP heritability (h 2 = 0.04-0.20) in G&H, with estimates similar to those in eMERGE (Supplementary Data 7, Fig. 2a Fig. 2b). The only trait for which the genetic correlation differed nominally significantly from one was BMI (r g = 0.85 [95% CI: 0.71-0.99], p = 0.02, not adjusted for multiple comparisons).

High transferability of cardiometabolic loci
We assessed whether published trait-associated genomic loci identified in predominantly European ancestry populations were shared by the British Pakistani and Bangladeshi populations 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 a 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. We report the observed number divided by the expected number of loci and call this new approach the power-adjusted transferability (PAT) ratio. Across most traits examined, the observed number of transferable loci closely matched the loci we expected (Table 1 and Supplementary Data 8). 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, yielding a PAT ratio of 1.05. The PAT ratio for CAD was only 0.62, with the number of observed transferable loci (13%) lower than the expected number (21%), although this difference was only marginally significant (binomial p-value = 0.05; one-sided and not adjusted for the multiple comparisons). To explore whether this was likely to be due to ancestry differences or other factors, we also calculated the PAT ratio in eMERGE, and observed a similarly low PAT ratio for CAD (0.69, binomial p-value = 6 × 10 −4 ) ( Table 1 and Supplementary Data 9).
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 50 kb of the locus was significant at p < 1 × 10 −3 (Fig. S4). These nine loci were all associated with lipid traits: EVI5, NBEAL1, GPAM, CETP, STAB1, TTC39B, SH2B3, ACP2, and NECAP2 (Supplementary Data 10). Of these loci, CETP, which has been reported to be associated with both HDL-C and LDL-C levels in European ancestry samples, was strongly associated with HDL-C in G&H (p = 7.08 × 10 −56 ), but not with LDL-C levels (p = 0.23) (Fig. 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 14 . 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 European ancestry samples as the reference. Colocalisation methods can estimate the likelihood of causal variant sharing without the need to identify the specific causal variant. 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 . Medication data are not available in eMERGE so the highest measurements for LDL-C, SBP, and DBP are used. Error bars represent 95% confidence intervals in both plots. b Genetic correlations were estimated using Popcorn based on GWAS summary statistics generated from G&H and European-ancestry individuals from UK Biobank. Red indicates that the genetic correlation is nominally significantly lower than 1 (p-value = 0.02 for BMI; two-sided and not adjusted for multiple comparisons). Medication-adjusted lipid and blood pressure levels are used. For r g estimates of 1 (TG and DBP), the method cannot derive confidence intervals. Sample sizes of GWAS for cardiometabolic traits in G&H are in Table 1.

Variable performances of polygenic scores
To assess the performance of PGSs for cardiometabolic traits derived from European ancestry populations in British Pakistani and Bangladeshi individuals, we compared predictive performance in G&H to that in European ancestry individuals from eMERGE. We quantified predictive accuracy using the "incremental AUC" statistic for CAD and the "incremental R 2 " statistic for continuous risk factor traits; these are the gain in AUC or R 2 when adding the PGS to the regression of phenotype on the baseline covariates (sex, age, and genetic PCs). We first evaluated previously published PGSs from the PGS Catalog (Supplementary Data 5). The CAD PGSs that proved to have the best performance in G&H and eMERGE were two different scores optimised in South Asian 15 and European ancestry samples 16 , respectively. PGSs for all phenotypes assessed were significant predictors of their target trait in G&H (Fig. 3). For prediction in G&H, the incremental R 2 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 G&H versus eMERGE, determined by the ratio of incremental AUC or R 2 , 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%; Supplementary Data 5). Amongst the cardiometabolic traits, 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 British Pakistani and Bangladeshi individuals were more likely to be treated with statins 17 . Incremental R 2 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 (Supplementary Data 5), although the heritability for statin-adjusted and unadjusted LDL-C was not significantly different (Supplementary Data 7; one-sided p-value from z test = 0.34).
We explored the factors that may impact the relative accuracy of PGSs. We considered the effect on the relative accuracy of the transancestry genetic correlation, ratio of heritability estimates in G&H versus eMERGE, as well as the product of the previous two terms. However, none of them showed a significant association with the relative PGS performance (Fig. S7).
To assess whether the performance of PGS based on European ancestry GWAS could be improved in British Pakistani and Bangladeshi samples, we next constructed PGS using the clumping and p-value thresholding (C + T) method and optimised them separately within G&H and eMERGE using 10-fold cross-validation. The numbers of SNPs in the most frequently selected best C + T PGSs are similar between eMERGE and G&H, and PGSs for lipids contained fewer SNPs (194-454) than other traits (>20,000; Supplementary Data 12, Fig. S8). 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 (British Pakistani and Bangladeshi and primarily European ancestry, respectively; Fig. S9).
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 18 and using PRS-CSx 19 , both integrating the European ancestry GWAS and that from UKBB South Asian ancestry samples. The improvement in accuracy was modest (0.3-10.5%) (Fig. S10). This may be due to the low sample sizes in the UKBB South Asian ancestry GWASs.
Modest improvement in CAD risk prediction by adding PGS to clinical risk score A CAD PGS derived from European ancestry GWAS summary statistics and tuned in South Asian ancestry individuals from UKBB 15 (PGS000296 in the PGS Catalog), showed the highest predictive accuracy in British Pakistani and Bangladeshi 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; Supplementary Data 5). 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; Fig. S11). 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; Fig. S12, Supplementary Data 13). Consistent with previous findings in European ancestry individuals 8 , the CAD PGS was uncorrelated with QRISK3 (Pearson's correlation coefficient r = −0.0056 and p-value = 0.62). The integrated score combining QRISK3 and the CAD PGS had a non-significant improvement in the C-index (0.853, 95% CI: 0.838-0.867) but a 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 20 . The integrated score reclassified 3.2% of the population as high risk and 2.5% as low risk (Supplementary Data 13). This improvement was mostly driven by the enhanced identification of CAD cases in people between 25 and 54 years of age (NRI in cases being 7.0% vs. NRI in controls being −1.2%), and of controls in people between 55 and 84 of age (NRI in cases being 0.0% vs. NRI in controls being 6.8%) (Fig. 4, Supplementary Data 13). 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 (Supplementary Data 13). Using continuous NRI, the integrated score showed significant improvement (27.0%; 95% CI: 17.7%-36.2%) and similar trends in age groups (Fig. S13, Supplementary Data 13). To assess the potential effects of missingness of QRISK3 variables (Fig. S3), we included additional data for HDL-C and TC that were measured more recently, which were not used in the above standard method (Supplementary Methods). The new QRISK3 score was more accurate with the C-index increased to 0.851, but we still

Estimated causal effects of CAD risk factors largely consistent
We carried out two-sample Mendelian randomisation analyses to estimate potentially causal effects of the risk factors on CAD in G&H and compared findings with European ancestry samples from eMERGE. For G&H, we used transferable loci as genetic instruments to benefit from the precision of largely European ancestry discovery GWAS whilst ensuring only valid instruments are used. In eMERGE, estimates of causal effects for BMI, BP, and lipids, except TG, were statistically significant (Fig. 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 associated with increased risk of CAD 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. We also compared different strategies for instrument selection in G&H, such as using all loci associated at genome-wide significance in European ancestry GWAS for the risk factors (Fig. S14). When following the standard approach of using an independent ancestry-matched sample (UKBB South Asian ancestry) to derive the instruments, an insufficient number of genome-wide significant instruments (p < 5 × 10 −8 ) were identified (Fig. S15). 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. However, the association of BMI with CAD was significant only for transferable loci (Fig. S14).
We found evidence of heterogeneity between MR estimates based on Cochran's Q statistic for DBP when using the trait loci from European ancestry GWAS as instruments (p-value = 0.04), LDL-C when using the UKBB South Asian ancestry-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 Mendelian randomisation model (Supplementary Data 14). We also carried out a multivariable MR for the lipid biomarkers to adjust for potential horizontal pleiotropy. Effect estimates remained highly consistent with those observed in the univariable MR (Supplementary Data 15).

Discussion
We conducted the first study to systematically assess the transferability of genetic loci and PGSs for cardiometabolic traits in individuals of South Asian descent with real-world clinical data, using~22,000 individuals from the G&H cohort. For lipids and blood pressure, we found evidence that causal genetic variants at published loci are widely shared with European ancestry populations. The prediction accuracy of PGSs derived from European ancestry GWASs for these traits was similar between G&H and European ancestry samples. However, the predictive performance of BMI and CAD PGS was reduced by 22% and 58%, respectively (for the PGS Catalog scores) and CAD also had fewer transferable loci. A CAD PGS optimised for South Asian ancestry individuals nonetheless yielded an appreciable improvement in risk reclassification (categorical NRI = 3.9%; 95% CI: 0.9-7.0%) when combined with the QRISK3 clinical risk score.
Other genetic studies of CAD and related traits that have evaluated reproducibility of loci in South Asian ancestry populations have either been limited by small sample sizes or have restricted their comparisons to the index SNP identified in the GWAS, which does not take LD or statistical power for replication into account [21][22][23] . A recent study compared genetic determinants of >200 lipid metabolites in 5000 South Asians from Pakistan and 13,000 Europeans and found high overlap in the detected associations 24 . Using a new method, our paper goes further by empirically demonstrating that, in most cases where loci do not replicate, it is due to the lack of power. These findings suggest that, in large part, the genes and pathways that influence the tested cardiometabolic traits are shared between these ancestrally divergent populations. One surprising finding was that the major  LDL-C locus at CETP was not associated with this biomarker in G&H but exhibited pleiotropic effects on HDL-C. This is consistent with the observation from a recent study in ancestrally diverse individuals 25 . Abnormalities in CETP are linked to accelerated atherosclerosis and might play an important role in increasing risk in South Asian ancestry individuals 26 .
For benchmarking we also assessed the transferability of genetic loci in European-ancestry participants from eMERGE. For HDL-C, triglycerides, and blood pressure, the PAT ratio was close to one (Table 1). For LDL-C, we observed a low PAT ratio (0.60, one-sided binomial p-value = 1.6 × 10 −5 ), probably because statin-adjust data were used in the discovery GWAS but lipid-lowering medication data were not available in eMERGE. For CAD, eMERGE had more cases (6815 vs. 1110), and we were able to replicate 31 loci (9 in G&H). Nevertheless, this is significantly lower than what would be expected given the replication power for eMERGE. In fact, the PAT ratio (PAT = 0.69, onesided binomial p-value = 6 × 10 −4 ) was similar to G&H (PAT = 0.62, onesided binomial p-value = 0.05). The lower PAT ratio for CAD in both G&H and eMERGE may indicate that it is a more complex outcome than measurable continuous biomarkers and the definition of CAD is potentially affected by cohort-specific factors such as how the diagnosis is coded in different health care systems. Procedure codes were not available in eMERGE and thus were not considered in defining CAD cases, which adds another source of uncertainty to the comparison.
BMI had the lowest proportion of transferable loci with shared causal variants as well as lower accuracy of the PGS in G&H relative to samples of European descent in eMERGE and a genetic correlation nominally significantly lower than one. South Asian ancestry individuals are known to have higher visceral fat at the same BMI compared to European ancestry individuals in Western countries 27,28 . Consistent with this, the estimated causal effect of BMI on CAD was significant only when using the transferable loci as instruments in the Mendelian randomisation analysis. Visceral adiposity is a strong risk factor for cardiometabolic diseases, independent of total fat mass; these findings warrant further study and may suggest that BMI may not be an optimal biomarker of adiposity in South Asian ancestry individuals 29 .
We observed variable performances of PGS developed in European ancestry and applied in British Pakistani and Bangladeshi individuals for the cardiometabolic traits that were investigated in this work, with relative accuracy in G&H versus eMERGE ranging from 131% for DBP to 42% for CAD. Consistent with previous studies 30, 31 , PGSs for HDL-C and triglycerides had similar predictive accuracy between the two ancestry groups. It is perhaps not surprising to observe a reduction in the accuracy of the CAD PGS in BPBs, given that it did not transfer well even within European-ancestry subpopulations 32 . Medication data were not available in eMERGE, we thus compared the PGS accuracies using the unadjusted, "highest ever" measurements for LDL-C and blood pressure. In G&H where medication data were available, PGSs for SBP and DBP showed similar performance in predicting the adjusted and unadjusted values (Supplementary Data 5). Unlike HDL-C and TG which showed high relative accuracy, PGS for LDL-C showed lower accuracy in G&H than eMERGE, but we were limited in drawing any conclusion because we did not correct for statin usage and PGS showed higher accuracy in G&H when using statinadjusted LDL-C.
We explored the factors that may impact the relative accuracy of PGSs. Based on a recently proposed theory, relative accuracy is proportional to the product of the trans-ancestry genetic correlation and the ratio of heritability estimates 7 . Neither the trans-ancestry genetic correlation nor the heritability of the trait was associated with the relative PGS performance. This may be because the theory was derived for PGSs based on genome-wide significant SNPs (whereas our PGSs include many SNPs with less significant p-values), and because the relative accuracy also depends on differences in allele frequencies and LD patterns at these SNPs between populations, which we have not factored in and may differ between traits.
Several groups have shown improvements in PGS performance in diverse ancestry groups when incorporating summary statistics from ancestry-matched samples 18,19,33 . Incorporating UKBB South Asian ancestry GWAS data in meta-PGSs proposed by Marquez-Luna et al. 18,33 and using PRS-CSx 19 did not show a large improvement in G&H. A likely reason is the limited sample size of the South Asian ancestry samples in UKBB for some of the traits. Larger samples of South Asian ancestry individuals are needed to examine if ancestrymatched GWAS data can improve prediction accuracy over and above what would be expected from the increased sample size. The increased value of increasing European-ancestry samples versus diversifying ancestries in GWAS will depend on the extent to which the causal variants are shared. For traits for which the causal variants are shared, there is more to be gained from more powerful European ancestry GWASs, even without adding samples of the target ancestry. However, increasing diversity in GWASs will greatly improve the resolution of fine-mapping and the power to identify the causal variants by leveraging the LD differences across ancestries 31,34 .
We assessed the clinical value of the PGS for CAD on top of the traditional clinical risk factors captured in the QRISK3 algorithm. Similar work has been done previously in research cohorts [8][9][10]35 ; our study represents an important addition since it captures the noise with which QRISK3 is actually measured within a real-world clinical setting (as opposed to using comprehensive measures taken for research purposes), which may affect the performance of integrated risk models combining these factors with PGSs. We note that only about 4% of the~8 million individuals used for developing QRISK3 were of South Asian ancestry 36 , and the weights for each conventional risk factor might not be optimal for South Asian ancestry individuals. QRISK3 was developed to predict cardiovascular disease (CVD), which is a composite outcome of CAD and stroke. However, our analysis focused on CAD, which is an important component of CVD and the main focus in GWASs and genetic prediction studies. The PGS for CAD developed by Wang et al. showed robust association with CAD in G&H, with a similar OR per SD in PGS (1.63, 95% CI: 1.51-1.76) as in their study (1.60, 95% CI: 1.32-1.94) 15 . The integrated score combining PGS and QRISK3 showed significant reclassification improvement against QRISK3 alone (NRI 3.9% (95% CI: 0.9-7.0%)). Previous studies in UKBB European ancestry samples reported similar improvement, with NRI estimates of 3.5% (95% CI: 2.4-4.5%) 8 and 3.7% (95% CI: 3.0-4.4%) 35 in two different analyses using CAD as the outcome. However, these NRI estimates are probably affected by using UKBB samples that are healthier than the general UK population without recalibrating risk to a primary care setting 9 . In G&H, the PGS improved the identification of high-risk individuals in people younger than 55 years as well as low-risk individuals in people older than 55 years, both of which are important in a clinical setting. We anticipate that, like European ancestry individuals 8,9,35 , the British Pakistani and Bangladeshi community (and potentially other South Asian ancestry populations) would also benefit from the use of integrating PGS in primary prevention settings.
Mendelian randomisation has emerged as a powerful tool to explore the potential causal effects of risk factors on disease outcomes. Statistical power can be the limiting factor when extending these analyses to ancestrally diverse populations because independent ancestry-matched GWAS for risk factors of interest may not be sufficiently large. To increase power to estimate the MR causal effects of risk factor traits on CAD in British Pakistanis and Bangladeshis, we used genetic instruments derived from large European ancestry GWAS. Some of the loci may be invalid instruments for other populations. However, restricting the published loci to the ones that were transferable in this population successfully addressed this issue for BMI and shows promise as a new approach for trans-ancestry Mendelian randomisation. An assumption that requires further study is whether the effect sizes of transferable loci are the same for each ancestry group. Future research should also further investigate the impact of pleiotropy on the causal effect estimate for HDL-C.
Our study has several limitations. Firstly, due to the limited sample size in each age-by-sex subgroup, we could not recalibrate risk prediction models in G&H to what would be expected in an unbiased primary care setting 9 . Secondly, for the comparisons of results in G&H with other cohorts, it should be noted that each of the cohorts examined here is unique. We selected eMERGE, which is also based on electronic health records, for comparison with G&H. Although UK Biobank would have been a better match in terms of country, we were unable to use it for comparison because it was one of the studies included in the published GWAS meta-analyses for the cardiometabolic traits that formed the source of locus discovery as well as the PRS training data. Differences in ascertainment (including the age distribution) and clinical measurements within different cohorts and healthcare systems may have impacted the genetic associations. Different genotyping arrays and imputation panels of different sizes and ancestries were used in G&H and eMERGE, which might lead to potential bias in the comparisons of the two cohorts. The lack of medication data in eMERGE meant we were limited in the conclusions we could draw from comparisons of heritability, transferability (PAT ratio), and PGS performance for LDL-C and blood pressure data. G&H is enriched for young people (median age 40 years old), given that the median age of onset for CAD is 55, some young individuals in this cohort will develop CAD beyond the observation period, which might lead to the underestimation of accuracy and clinical value of the CAD PGS. Finally, our transferability analyses evaluated whether there is evidence for a directionally consistent association in G&H but we did not assess heterogeneity in effect sizes which would also impact genetic risk prediction.
In conclusion, our work provides the first comprehensive assessment of the transferability of cardiometabolic loci to a non-European ancestry population and its impact on two key applications of genetics, causal inference and risk prediction. Our protocol and our new approach for transferability can serve as methodological standards in this developing field. We have shown high transferability of GWAS loci across several cardiometabolic traits between European ancestry and British Pakistani and Bangladeshi populations. The performance of PGSs is trait-specific. Our results suggested there would be clinical value in adding PGS to conventional risk factors in the prediction of CAD in primary care settings to improve the more efficient use of preventive interventions, such as lipid-lowering medications. Our investigation contributes to the increasing representation of individuals of diverse ancestry and varying socio-economic status in research studies, which we hope will help to decrease health disparities.

Genes & Health cohort
Genes & Health (G&H) is a community-based cohort of British Pakistani and Bangladeshi individuals recruited primarily in East London 13 . All participants have consented for lifelong electronic health record 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 (interquartile range [IQR] = 16) and 43 (IQR = 19) years for female and male participants, respectively (Fig. 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 .

Quality control and imputation of genotype data from Genes & Health
We used the 2020 February data release which contained 28,022 individuals genotyped on the Illumina Infinium Global Screening Array v3 with additional multi-disease variants. Of these, 22,490 (80%) individuals had linkage to primary or secondary care data, of which 56.5% were female. Quality control of genotype data was performed using Illumina's GenomeStudio and plink v1.9. We removed variants with low call rate (<0.99), rare variants with minor allele frequency (MAF) < 1%, and variants that failed the Hardy-Weinberg test (p < 1 × 10 −6 ) in a subset of samples with low level of autozygosity (Supplementary Methods). 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 (Fig. S2).
We used the Michigan Imputation Server 37 to perform imputation with the GenomeAsia pilot reference panel 38 , 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 European ancestry samples from the eMERGE cohort (henceforth eMERGE), a consortium of US medical research institutions, to carry out comparisons with 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 40 . We projected eMERGE participants onto the PC space generated from the 1000 Genomes project phase 3 dataset and applied Uniform Manifold Approximation and Projection (UMAP) 41  We calculated the QRISK3 10-year predicted risk for CAD 36 in G&H using the R package "QRISK3" v0.3.0 43 . We used data available up until 1 January 2010 to calculate QRISK3 (Supplementary Methods). Definitions of variables in the QRISK3 algorithm are shown in Supplementary Data 3, following 8 .
Phenotype data in eMERGE were downloaded from dbGaP (phs001584.v1.p1, phs000888.v1.p1, and phs001584.v2.p2; Supplementary Methods). 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 heritability estimates and performance of PGSs with G&H in order to minimise the effects of medications.

Genome-wide association analyses in Genes & Health
GWAS was performed with SAIGE 44 and adjusted for age, age 2 , sex and the first twenty principal components. For total cholesterol and LDL-C, adjustments were made for use of statins as described above.

Heritability and trans-ancestry genetic correlations
Datasets that were used in analyses are provided in Supplementary Data 4. We used GCTA to estimate SNP heritability in unrelated individuals from G&H and eMERGE, correcting for age, sex, and first 10 genetic PCs 45 . For CAD, we estimated SNP heritability on the liability scale using 6.7% as the prevalence estimate in the US 46 , 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 European ancestry 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 European ancestry individuals while accounting for differences in LD structure (Supplementary Methods) 47 . A two-sided pvalue < 0.05 indicated that the genetic correlation was significantly different from one.

Assessment of transferability of trait loci
Previous studies that evaluated the reproducibility of GWAS loci in South Asian individuals did not formally account for differences in power or LD patterns [21][22][23] . We assessed whether published traitassociated loci were reproducible in G&H (Supplementary Data 4), i.e. whether a locus affects the same trait in both populations, regardless of effect sizes. Credible sets for trait loci were generated and consisted of lead (independent) variant plus proxy SNPs (r 2 ≥ 0.8) within a 50 kb window (based on the European ancestry 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 two-sided 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). The statistical power to observe an association of a given locus in G&H was calculated using alpha = 0.05, the effect size estimate for the lead variant from the European ancestry discovery GWAS, and the allele frequency of the variant and sample size in G&H (Supplementary Methods). For SBP and DBP, the raw measurements were used in the discovery GWAS, we thus calculated power with effect size estimates in UK Biobank European-ancestry individuals by Neale's group where normalised blood pressure values were used. The power estimates were summed up across published loci for a given trait to give an estimate of the number of loci expected to be significantly associated in G&H. This is the expected number if all loci are transferable and account for the statistical power for replication. We calculated the power-adjusted transferability (PAT) ratio by dividing the observed number of loci with p < 0.05 amongst the published loci in G&H over the expected number. To our knowledge, this is a novel approach for assessing the reproducibility of GWAS findings.
We also highlighted published trait loci that we deemed to be 'non-transferable' despite sufficient statistical power: they contained at least one variant in the credible set with >80% power for replication and yet none of the variants in the credible set had p < 0.05 and no variant within 50 kb of locus had p < 1 × 10 −3 in G&H. LocusZoom (http://locuszoom.org/) was used to create regional association plots.

Trans-ancestry colocalisation
We used the Trans-ancestry colocalisation method (TAColoc) (https:// github.com/KarolineKuchenbaecker/TEColoc) 30 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 European ancestry individuals (Supplementary Methods).

Construction of polygenic scores
We evaluated the performance of PGSs in G&H and eMERGE. We first assessed PGSs that were previously constructed (mostly optimised in European ancestry samples) from the PGS Catalog 48 . 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. When multiple PGSs were available in the PGS Catalog, we reported the best score per trait. The details of the scores are in Supplementary Data 5.
We also calculated PGSs using the clumping and p-value thresholding method (C + T) and optimised PGSs in G&H and eMERGE separately using 10-fold cross-validation (Supplementary Methods). Lastly, in G&H we calculated meta-PGSs proposed by Marquez-Luna et al. 18 and PGS using the PRS-CSx method 19 that incorporated GWAS summary data from the panUKBB South Asian-ancestry individuals (Supplementary Methods).

Assessment of PGS accuracy and clinical performance
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, age 2 , 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 R 2 when adding PGS as an additional predictor, or incremental R 2 , was calculated as the difference between the R 2 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 1000 times to estimate the 95% confidence intervals for incremental R 2 and incremental AUC. We estimated the effect size per SD of PGS from the full model. Effect size for quintiles, and for the top 10% versus middle 40-60% was reported as well. Relative accuracy was calculated as the ratio of incremental AUC (or incremental R 2 for continuous traits) in G&H to that in eMERGE.
QRISK3 scores were calculated for 8112 unrelated individuals as described in Supplementary Methods (420 CAD cases and 7702 controls). We followed Riveros-Mckay et al. 8 to integrate QRISK3 scores with the PGS for CAD developed by Wang et al. 15 . 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 + sex, (2) PGS + age at assessment + sex, (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): NRI = Pðup|caseÞ À Pðdown|caseÞ + Pðdown|noncaseÞ À Pðup|noncaseÞ 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 age-by-sex subgroups. Bootstrap resampling (1000 times) was used to estimate confidence intervals for NRI. All reported p-values are two-sided.

Mendelian randomisation analysis
We modelled liability to CAD as our outcome within a univariable twosample Mendelian randomisation (MR) 49 framework using the cardiometabolic traits (BMI, SBP, DBP, LDL-C, HDL-C, TG) as exposures. To identify genetic instruments for the exposure, we explored three alternative approaches: (a) published loci significant at p < 5 × 10 −8 in the original European ancestry GWAS; (b) transferable loci defined as described in the transferability section of the methods, taking the effect size from the original European ancestry GWAS; and (c) loci significant at p < 5 × 10 −8 in the South Asian ancestry group of the Pan-UKBB GWAS, LD-clumped to an r 2 < 0.2 with a LD window of 50 kb, based on South Asian 1000 Genomes project LD reference. Where insufficient numbers of genome-wide significant instruments were identified, we used a more permissive p-value threshold of p < 5 × 10 −5 for instrument selection in UKBB South Asian. The primary Mendelian randomisation 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 package v0.5.5 50 . For comparison, a two-sample Mendelian randomisation approach was also performed using summary data for CAD from eMERGE and established loci significant at p < 5 × 10 −8 in the original European ancestry GWAS. We also undertook several sensitivity analyses. In brief, we evaluated the Egger intercept to assess directional pleiotropy and Cochran's Q statistic 51 as an indicator of heterogeneity. Mendelian randomisation analysis using MR pleiotropy residual sum and outliers methods (MR-PRESSO) 52 , weighted median 53 and weighted mode methods 54 models were additionally performed in the presence of heterogeneity. To investigate the individual direct effect of HDL-C, LDL-C, and TG on the risk of CAD and simultaneously account for horizontal pleiotropy, we replicated the analysis in a multivariable MR (MVMR) 55 setting with the TwoSampleMR R package. Genetic instruments for the HDL-C, LDL-C, and TG joint exposure were selected from samples with European ancestry in UKBB, if associated with at least one of the three lipids (p < 5 × 10 −8 ). The instruments were filtered based on MAF > 0.005 and LD-clumped to an r 2 < 0.01 with a window of 50 kb, based on the EUR 1000 Genome project LD reference, using plink2. All reported p-values are two-sided.

Reporting summary
Further information on research design is available in the Nature Research Reporting Summary linked to this article.

Data availability
Genes & Heath imputed genotype data (GRCh 37) have been deposited in EGA under study accession number: EGAS00001005373 (https:// ega-archive.org/datasets/EGAD00001007815). The electronic health records from Genes & Health are available under restricted access for bona fide research; researchers wishing to access them should apply to the G&H Executive (www.genesandhealth.org/research/scientistsusing-genes-health-scientific-research). GWAS summary statistics generated in Genes & Health are available at www.genesandhealth.org/ research/scientific-data-downloads. The transferable loci generated in this study are provided in the Supplementary Data file. Publicly available GWAS summary statistics that were used in this study (Supplementary Data 4) are available via the CARDIoGRAMplusC4D Consortium (http://www.cardiogramplusc4d.org), GIANT (https:// portals.broadinstitute.org/collaboration/giant/index.php/Main_Page), GLGC (http://csg.sph.umich.edu/willer/public/lipids2017/), and GWAS Atlas (https://atlas.ctglab.nl/traitDB/). SNPs and the weights for polygenic risk scores are available in the PGS Catalog (www.pgscatalog.org) and score IDs are provided in Supplementary Data 5.

Code availability
Code used to assess the transferability of genetic loci and calculate the PAT ratio is available at https://github.com/Nsallah1/GH_Manuscript. TAColoc used to perform trans-ancestry colocalisation analysis is available at https://github.com/KarolineKuchenbaecker/TEColoc. Code used to assess the accuracy and clinical utility of polygenic risk scores is available at https://github.com/QinqinHuang/GnH28k_ polygenic_scores 56 . All additional software is available online.