Integrating polygenic risk scores in the prediction of type 2 diabetes risk and subtypes in British Pakistanis and Bangladeshis: A population-based cohort study.

BACKGROUND: Type 2 diabetes (T2D) is highly prevalent in British South Asians, yet they are underrepresented in research. Genes & Health (G&H) is a large, population study of British Pakistanis and Bangladeshis (BPB) comprising genomic and routine health data. We assessed the extent to which genetic risk for T2D is shared between BPB and European populations (EUR). We then investigated whether the integration of a polygenic risk score (PRS) for T2D with an existing risk tool (QDiabetes) could improve prediction of incident disease and the characterisation of disease subtypes. METHODS AND
FINDINGS: In this observational cohort study, we assessed whether common genetic loci associated with T2D in EUR individuals were replicated in 22,490 BPB individuals in G&H. We replicated fewer loci in G&H (n = 76/338, 22%) than would be expected given power if all EUR-ascertained loci were transferable (n = 101, 30%; p = 0.001). Of the 27 transferable loci that were powered to interrogate this, only 9 showed evidence of shared causal variants. We constructed a T2D PRS and combined it with a clinical risk instrument (QDiabetes) in a novel, integrated risk tool (IRT) to assess risk of incident diabetes. To assess model performance, we compared categorical net reclassification index (NRI) versus QDiabetes alone. In 13,648 patients free from T2D followed up for 10 years, NRI was 3.2% for IRT versus QDiabetes (95% confidence interval (CI): 2.0% to 4.4%). IRT performed best in reclassification of individuals aged less than 40 years deemed low risk by QDiabetes alone (NRI 5.6%, 95% CI 3.6% to 7.6%), who tended to be free from comorbidities and slim. After adjustment for QDiabetes score, PRS was independently associated with progression to T2D after gestational diabetes (hazard ratio (HR) per SD of PRS 1.23, 95% CI 1.05 to 1.42, p = 0.028). Using cluster analysis of clinical features at diabetes diagnosis, we replicated previously reported disease subgroups, including Mild Age-Related, Mild Obesity-related, and Insulin-Resistant Diabetes, and showed that PRS distribution differs between subgroups (p = 0.002). Integrating PRS in this cluster analysis revealed a Probable Severe Insulin Deficient Diabetes (pSIDD) subgroup, despite the absence of clinical measures of insulin secretion or resistance. We also observed differences in rates of progression to micro- and macrovascular complications between subgroups after adjustment for confounders. Study limitations include the absence of an external replication cohort and the potential biases arising from missing or incorrect routine health data.
CONCLUSIONS: Our analysis of the transferability of T2D loci between EUR and BPB indicates the need for larger, multiancestry studies to better characterise the genetic contribution to disease and its varied aetiology. We show that a T2D PRS optimised for this high-risk BPB population has potential clinical application in BPB, improving the identification of T2D risk (especially in the young) on top of an established clinical risk algorithm and aiding identification of subgroups at diagnosis, which may help future efforts to stratify care and treatment of the disease.

Introduction

People of South Asian origin are disproportionately affected by type 2 diabetes (T2D) and tend to develop the condition at younger ages and at a lower body mass index than European ancestry individuals [1]. Despite this, they are underrepresented in studies assessing the genetic aetiology of the disease. Furthermore, to our knowledge, there has been no systematic assessment of the extent to which genetic risk loci identified in European ancestry individuals can be transferred into South Asians. Given the known pathophysiological differences between these populations [2,3], such an assessment is important for understanding the extent to which the aetiology of the disease varies between them.

Characterisation of the genetic aetiology of T2D, using genome-wide association studies, has allowed the development of polygenic risk scores (PRS) to aid the individualised clinical prediction of common complex diseases [4-10]. To date, PRS have been developed and tested predominantly in highly selected white European populations (EUR) with bias towards healthy and older people [10,11]. For coronary artery disease, integration of PRS with clinical risk tools has been shown to enhance the prediction of incident disease, in multiple ancestral groups [5,6,9,12], which may help target preventative care. Similarly, there is considerable potential to use PRS to improve the prediction of T2D risk [11], and, to date, their integration with well-validated clinical risk instruments already in routine use, such as QDiabetes, has not yet been evaluated. Building enhanced risk tools that combine PRS with tools such as QDiabetes could offer significant opportunities for clinical benefit including enhanced individualised screening and preventive measures such as referral to diabetes prevention programmes [13]. There is a particular need to investigate PRS in understudied groups who are at high risk of T2D, including people of South Asian ancestry [11,14] and women with a history of gestational diabetes mellitus (GDM) [15]. PRS may also enhance the characterisation of T2D “subgroups,” a recent area of significant research and clinical interest due to their potential to capture important heterogeneous features at diabetes diagnosis associated with common aetiological disease pathways that may predict future diabetes complications and treatment responsiveness [2,16-18].

In this study, we aimed to develop and evaluate a T2D PRS in British Bangladeshis and Pakistanis (BPBs) enrolled in the Genes & Health (G&H) programme [19]. This real-world, community-based cohort (n = 48,144) combines genetic data with rich, lifelong electronic health record data and comprises a minority ethnic group living predominantly in socioeconomically deprived circumstances, otherwise underrepresented in health research [19]. We therefore aimed, firstly, to examine the transferability of T2D genetic loci already ascertained in individuals of European ancestry (EUR) to BPBs, taking into account power and differences in linkage disequilibrium, and to optimise a PRS for this population. Secondly, we tested whether the PRS enhanced the prediction of incident T2D when integrated with the commonly used clinical risk score, QDiabetes [20]. Thirdly, we sought to investigate whether the PRS alone might predict the progression to T2D from gestational diabetes in BPB women, as has been observed in European and Southeast Asian populations [21,22]. Finally, we explored whether the PRS might predict clinically heterogeneous T2D subgroups at diagnosis, which are increasingly well characterised [2,16-18,23].

Methods

Study population

G&H recruits BPB people aged 16 years and above, predominantly in community and primary care settings. We used the 2020 data release [24], which comprises electronic health record (EHR) data and genotype data from the Illumina Infinium Global Screening Array V3 Chip on 22,490 participants [19]. Descriptions of quality control and imputation of genotype data are provided in . We applied specific inclusion and exclusion criteria to the G&H population for each analysis, described below and summarised in . All analyses were planned at study outset, with the exception of the derivation of a second clustering model (Integrated Model), the rationale for which is given below.

Summary of analyses in study.

FPG, fasting plasma glucose; GWAS, genome-wide association study; G&H, Genes & Health; HDL, high-density lipoprotein; PRS, polygenic risk score.

Ascertainment of conditions using clinical coding

T2D, GDM, and associated complications were ascertained using clinical codes extracted from primary and secondary care EHR, presented in full in , developed using widely used clinical coding resources [25]. We excluded individuals with clinical codes of conditions causing secondary or monogenic diabetes.

Transfer of previously identified GWAS loci to the study population

We obtained previously identified genetic loci associated with T2D from a genome-wide association study (GWAS) conducted in people of European ancestry (EUR) by Vujkovic and colleagues. [26]. We performed a GWAS of T2D in G&H and assessed if previously identified loci were reproducible in G&H at p-value <0.05 using previously established methods () [9]. We assessed 338 T2D loci that had variants well imputed in G&H. The expected power for replication was estimated assuming the same effect size as in the EUR discovery sample and accounting for the allele frequency and sample size in G&H. Even though the same locus and target gene may affect T2D risk in both populations, it is possible that this is caused by distinct causal variants in the region. Therefore, trans-ancestry colocalisation analysis was performed using TEColoc to assess whether a transferable locus shared the same causal variant between BPB and EUR populations, using UK Biobank (UKBB) EUR individuals for the latter [27]. We used overlapping variants in a 50-kb window and assessed 27 transferable loci in which ≥10% of the variants from UKBB were well imputed in G&H and vice versa. We might not be able to detect shared causal variants due to the low coverage for some loci, but it is not likely to cause false positives. We also applied the Popcorn algorithm to estimate the trans-ancestry genetic correlation between BPB and UKBB EUR populations [28]. Further methodological details are given in .

Polygenic risk score construction

We applied previously constructed scores from the PGS Catalog [29], principally developed in EUR populations, to participants in G&H. These were compared to PRS optimised within G&H using the clumping and p-value thresholding (C+T) method implemented in PRSice2 v2.2.11 [30,31] based on the largest published T2D multiancestry GWAS to date [26] (), downloaded from dbGaP (phs001672.v4). While the EUR-specific GWAS was used for assessing the transferability of loci as mentioned above, we constructed the PRS using the more powerful multiancestry GWAS, which contained 9,004 cases and 12,066 controls with Pakistani ancestry. We also tried using a South Asian–specific GWAS [32] to construct a PRS and combining the PRS from the South Asian–specific GWAS with the one from the multiancestry GWAS using the method from Marquez-Luna and colleagues [33] (). Neither of the scores showed better performance; thus, we used the C+T PRS constructed from the multiancestry GWAS. For the C+T score, LD estimated from EUR samples (N = 503) from the 1,000 Genomes project was used for clumping (r2 = 0.1). We calculated multiple scores using various p-value thresholds. We excluded one sample from each pair of 2nd degree relatives identified using the kinship coefficient from KING v2.3.4 [34] and assessed the accuracy of PRS in unrelated individuals. We selected the PRS with the highest predictive accuracy in an independent set of samples that were not included in the QDiabetes analysis described below, i.e., prevalent cases with onset before the assessment date, and controls who did not meet inclusion criteria for longitudinal analysis (aged 25 to 84 years with no prior history of T2D as of 1 May 2010, no recorded value of HbA1c >48 mmol/mol or FPG >7.0 mmol/L; see for characteristics of cases and non-cases). The area under the receiver operating characteristic curve (AUC) was estimated with the R package “pROC.” Predictive accuracy of PRS was quantified by incremental AUC, or the gain in AUC when adding PRS to the reference model, which accounted for participant age, gender, and 10 genetic principal components (PCs). We also calculated the incremental pseudo R2 on the liability scale [35]. We used 13.6% as the prevalence estimate for T2D in South Asian ancestry individuals in the UK, the 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 4 east London boroughs. The C+T PRS with p-value threshold < 1 × 10−5 (2,877 SNPs; ) showed the highest predictive accuracy among all the scores we constructed and was thus used in downstream analysis (). A scaled PRS following the normal distribution with mean and median of 0 and standard deviation (SD) of 1 was constructed across BPB participants after regressing out the first 10 PCs derived as described in , allowing direct comparison between ancestry groups [36]. We used additional PRSs associated with T2D and its aetiology that are previously described in the literature by Mansour Aly and colleagues [18] and Udler and colleagues [37] ().

Development of an integrated risk tool to predict incident type 2 diabetes

We assessed whether PRS could enhance 10-year risk prediction of T2D compared to QDiabetes, a validated, EHR-based risk prediction tool commonly used in UK primary care to estimate an individual’s 10-year risk of developing T2D [20]. There are 3 QDiabetes models. Model A provides estimates based on age, ethnicity, family history, comorbidities, and prescribed medications. Model B (which has the highest clinical predictive value) uses the same variables as model A, plus fasting plasma glucose (FPG), while model C is composed of model A plus HbA1c. Analysis was performed in 13,648 individuals aged 25 to 84 years with no prior history of T2D as of 1 May 2010, no recorded value of HbA1c >48 mmol/mol or FPG >7.0 mmol/L (), using established inclusion criteria [20]. QDiabetes scores were calculated on the assessment date of 1 May 2010 for each participant on the basis of available clinical data required to run the model using the R package “QDiabetes” [38]. Numbers available for analysis were model A (n = 13,648), model B (n = 4,334), and model C (n = 864). We applied multiple imputation to fill in missing data for body mass index (BMI) and the Townsend Deprivation Index using the R package “MICE” [39].

We combined the PRS and QDiabetes score, which were not significantly correlated, to build an integrated risk tool (IRT) to estimate risk of developing T2D over the next 10 years, as previously described [12]. For each individual, QDiabetes score was converted to an odds ratio (OR) and multiplied by the odds calculated from the individual’s PRS given their QDiabetes score, calculated from a logistic regression model incorporating PRS, QDiabetes, and an interaction between them, and training it separately on male and female participants. Ten-year risk of disease was classified as high (>10%) or low (<10%) in line with [12]. We assessed relative model performance using the concordance index (C-Index) and categorical net reclassification index (NRI). C-Indices were calculated from Cox proportional hazard models using the R package “survival.” NRI was calculated as NRI = P(up|case) − P(down|case) + P(down|noncase) − P(up|noncase), i.e., the sum of (1) the proportion of individuals who subsequently developed T2D correctly reclassified as high risk, minus the proportion of individuals who subsequently developed T2D incorrectly reclassified as low risk; and (2) the proportion of individuals who did not develop T2D correctly reclassified as low risk, minus the proportion of individuals who did not develop T2D incorrectly reclassified as high risk. Categorical NRI performance was assessed in the entire analysed sample, plus in age-by-sex subgroups, with 40 years chosen as the threshold for high/low age group comparison to reflect (1) a mean age at T2D diagnosis of 49.1 years and median of 47.8 years in our cohort; (2) a mean age at study entry of 39.1 years; (3) the fact that routine NHS health checks that include T2D screening are offered at age 40; and (4) previous use of age 40 as a cutoff discriminating early from late-onset diabetes in the literature [40]. We observed that findings were robust to alteration of this threshold across a broad age range (). NRI confidence intervals were calculated using bootstrapping (number of iterations = 1,000). The characteristics of reclassified individuals were compared with descriptive statistics. In addition to the categorical NRI, at the request of a peer reviewer, we also calculated continuous NRI using the R package “PredictABEL.” P(up|case) and P(down|case) in the above equation now indicate the proportion of cases that had higher or lower risk estimates using the IRT than QDiabetes, respectively.

Prediction of type 2 diabetes after gestational diabetes using PRS

In women with a history of GDM, we compared characteristics (including T2D PRS) of those who subsequently did (n = 127) and did not (n = 175) develop T2D. We compared characteristics between the 2 groups using multivariate logistic regression models that included risk factors used to construct the QDiabetes score. Details of a power calculation for this analysis are available in The association between PRS and T2D in women with a history of GDM was assessed in Cox proportional hazard models controlling for (1) QDiabetes score or (2) unadjusted clinical risk factors used to construct the QDiabetes score.

Identification of type 2 diabetes subgroups and their association with PRS

We applied data-driven clustering techniques to define subgroups of individuals with distinct characteristics in 2 separate models using latent class analysis (Stata V16.0) based on the methodology used by [2,16,17]. The optimal number of clusters was selected on the basis of information criteria elbow plots () and was found to be 5 for all model iterations. The first model (Clinical Model) was based on 5 clinical variables at the time of T2D diagnosis: age, BMI, HbA1c, serum triglycerides, and serum high-density lipoprotein (HDL). Analysis was restricted to 4,266 participants for whom values for these variables were available within 1 year of T2D diagnosis. Differences in PRS between clusters were compared using analysis of variance (ANOVA) (RStudio V1.1.413). Having observed differing PRS distribution between clusters derived using clinical variables, we next explored whether inclusion of PRS as a clustering covariate could further delineate the aetiological processes underlying cluster membership. In the second model (Integrated Model), we repeated clustering using age, BMI, HbA1c, serum triglycerides, and PRS as covariates in 5,905 individuals with available data within 12 months of diagnosis. HDL was omitted from the Integrated Model as its inclusion did not alter cluster characteristics but did reduce sample size. Differences in previously reported glycaemic trait PRSs including fasting glucose, beta cell function, insulin sensitivity index, and corrected insulin response were compared between each cluster and nondiabetic controls (n = 10,841) using Bonferroni-corrected unpaired one-tailed t tests [18,37]. Further details are presented in

Research ethics and reporting guidelines

G&H operates under approval from the National Research Ethics Committee (London and Southeast), and the Health Research Authority (reference 13/LO/124); Queen Mary University of London is the Sponsor. Written informed consent is obtained from all study volunteers, and it allows analysis of health and genetic data and publication of results. This study is reported according to the Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD) statement (checklist available in ).

Results

As of 1 May 2020 (the end of the study period), 7,599 individuals with an EHR diagnosis of T2D were enrolled in G&H, followed up for a mean of 9.69 years after diagnosis. 52.8% of these individuals were in the most deprived Index of Multiple Deprivation quintile in the UK. Mean age at time of diagnosis was 46.2 years. A total of 1,205 individuals (15.9%) developed macrovascular complications between T2D diagnosis and the end of the study period; 2,300 (30.3%) developed microvascular complications.

Genetics of type 2 diabetes in BPB versus European individuals

We first investigated the extent to which genetic risk for T2D was shared between G&H (BPB) and EUR individuals. The GWAS in G&H identified 3 significant associations at p-value < 5 × 10−8 (, ), and they were all previously identified [26]. Four out of the 6 genetic loci identified in the South Asian–specific GWAS by Kooner and colleagues [32] were replicated in G&H (). The trans-ancestry genetic correlation (i.e., the correlation of causal-variant effect sizes) between G&H and UKBB EUR individuals was significantly lower than 1 (rg = 0.68, standard error = 0.15, p-value (for the null hypothesis that rg = 1) = 0.03). Among the 338 genetic loci identified in EUR populations that had variants well imputed in G&H, we observed significant evidence of transferability for 76 (22.5%) loci (among them 13 were significant at p-value < 0.05/338; ), which was lower than would be expected (30.0%) accounting for differences in power (one-sided binomial p-value = 0.001). This suggested that a large proportion (75%) of the loci ascertained in EUR populations, which were well powered to replicate in G&H did so. We did not observe any loci that were well powered (power to replicate >0.8) but not transferable in G&H. The evidence of transferability was consistent with other cardiometabolic traits (Observed/Expected = 0.62 for coronary artery disease and 1.0 to 1.2 for BMI, lipids, and blood pressure) that were reported in the same cohort [9].

To assess whether causal variants were shared between EUR and BPB populations for the transferable loci, we applied trans-ancestry colocalisation analysis. Of 27 replicated loci assessed, 9 (33%) had significant evidence of shared causal variants (). For example, we observed shared causal variants at the TCF7L2 locus, one of the known loci with the strongest association with T2D, and at the KCNJ11 locus, which is the target gene for drugs such as Glyburide [41,42]. The proportion of transferable loci with shared causal variants for T2D was lower than for triglycerides (56%; binomial p-value = 0.015) and total cholesterol (61%; p = 0.003) and similar to HDL (48%), LDL (47%), and BMI (26%) (binomial p-value >0.05 for all) in the same cohort [9].

Performance of an integrated risk tool to predict incident type 2 diabetes

The characteristics of participants included in incident T2D risk prediction analyses are shown in . We constructed a PRS for T2D using the clumping and p-value thresholding method, based on multiancestry GWAS meta-analysis [26]. This had an OR per SD of 1.57 (95% confidence interval [CI]: 1.50 to 1.65), an incremental AUC of 0.032 (95% CI: 0.026 to 0.039), and an incremental R2 on the liability scale of 4.6% (95% CI: 3.7 to 5.6) in G&H. There was no correlation between PRS and QDiabetes scores (Pearson’s coefficients −0.03, 0.08, 0.13 for QDiabetes models A, B, and C, respectively; associated p-values 0.31, 0.18, 0.16) (). PRS was weakly correlated with fasting glucose (Pearson’s coefficient 0.11, p < 0.001) and HbA1c (Pearson’s coefficient 0.07, p = 0.048) (). Compared to QDiabetes alone, the IRT combining QDiabetes model A (multiple clinical risk factors) with PRS improved the 10-year prediction of T2D as assessed by categorical NRI: NRI 3.22% (95% CI: 2.00% to 4.38%) in 13,648 individuals (). The IRT C-index, a goodness-of-fit metric approximating the area under the receiver–operator curve, was superior to QDiabetes model A for participants aged less than 40 years (p = 0.002) (), but not for people aged over 40 years. Enhancement of T2D prediction with the IRT persisted but was attenuated when QDiabetes model B (model A plus FPG) was used: categorical NRI 0.80% (95% CI 0.21% to 1.42%). The IRT did not improve on the QDiabetes score with model C (model A plus HbA1c): categorical NRI 0.20% (95% CI: −0.09% to 0.44%) (). Higher estimates of model enhancement were observed with continuous NRI (): Model A 28.1% (95% CI 23.9% to 32.1%; Model B 23.8% (95% CI 17.2% to 30.6%); Model C 19.4% (95% CI 4.92% to 33.9%). There were no observed differences in C-Index between QDiabetes and IRT in models B and C ().

Performance of the IRT using NRI.

NRI and 95% CIs comparing 10-year prediction of T2D between QDiabetes models A (n = 13,648), B (n = 4,334), and C (n = 864) and IRTs combining each QDiabetes model with T2D PRS. NRI is presented overall and for age-by-sex subgroups. The NRI for individuals who subsequently develop T2D (cases) and who do not develop T2D (non-cases) are presented alongside the NRI for cases and non-cases combined (all individuals). CI, confidence interval; IRT, integrated risk tool; NRI, net reclassification index; PRS, polygenic risk score; T2D, type 2 diabetes.

Performance of the IRT using C-Index.

Comparison of C-Index calculated from Cox proportional hazard models for QDiabetes models A, B, and C and IRTs combining each QDiabetes model with individual T2D PRS. C-index is presented for all participants and for age-by-sex subgroups. Data are presented as mean with 95% CIs. CI, confidence interval; C-Index, concordance index; IRT, integrated risk tool; PRS, polygenic risk score; T2D, type 2 diabetes.

Across all IRT models (for QDiabetes model A, B, and C), NRI was higher in participants aged less than 40 years who subsequently developed T2D (cases) than those who did not (non-cases), implying improved ability to correctly classify younger individuals as high risk. The converse pattern was seen in individuals aged greater than 40 years, implying enhanced ability to correctly classify older individuals as low risk. The clinical features of individuals whose risk was reclassified are presented in The number of individuals reclassified in each model are presented by age band and sex in

In women who had a history of GDM, T2D PRS was higher in those who had subsequently developed T2D, compared to those who had not (mean value (SD) 0.408 (0.93) versus 0.140 (0.98); p = 0.02) (). Among individuals without T2D (non-cases), women with a history of GDM displayed a higher mean PRS than age- and BMI-matched females without GDM and compared to males (p = 0.003 and p = 0.001, respectively) (). In multivariate survival analysis restricted to women with a history of GDM, T2D PRS was independently associated with development of T2D after adjustment for (1) QDiabetes score model A (hazard ratio (HR) per SD of PRS 1.23, 95% CI 1.05 to 1.42; p = 0.028) and (2) established risk factors for T2D (HR 1.37, 95% CI 1.17 to 1.57; p = 0.002) (). Similar findings were observed in age- and BMI-matched male and female (without a history of GDM) controls (), indicating that the utility of the T2D PRS for risk prediction does not appear to differ between women with versus without GDM.

Mean T2D PRS in women with GDM.

Mean T2D PRS in women with GDM who subsequently did and did not develop T2D, compared to age- and BMI-matched control groups of females without GDM, and males. Data are presented as group mean with 95% CIs according to sex-by-GDM subgroup, stratified by subsequent T2D status. BMI, body mass index; CI, confidence interval; GDM, gestational diabetes mellitus; PRS, polygenic risk score; T2D, type 2 diabetes.

Type 2 diabetes subgroup identification using clustering

In our Clinical Model, we identified T2D subgroups by clustering age, BMI, HbA1c, HDL, and triglycerides at the time of diagnosis (). We replicated some previously identified clusters: Mild Obesity-related Diabetes (MOD), Mild Age-Related Diabetes (MARD), Severe Insulin-Resistant Diabetes (SIRD), and Mild Diabetes (MD), and used their previous nomenclature [2,16,17,43]. In the absence of clinical measures of insulin secretion such as C-peptide, we were unable to fully delineate the previously well-replicated Severe Insulin-Deficient Diabetes (SIDD) cluster in the Clinical Model. However, our MD cluster may include individuals who are insulin deficient as it is characterised by low serum triglycerides (1.21 (0.43) mmol/L) in addition to high HDL (2.07 (0.25) mmol/L), and our Clinically Undifferentiated Diabetes (CUD) cluster likely also contains people with SIDD as its membership is characterised by high HbA1c and low serum triglycerides. There were no statistically significant differences in rates of progression to micro- or macrovascular complications between clusters identified in the Clinical Model. However, we found that the PRS differed significantly between clusters (ANOVA, p = 0.002) and that CUD was characterised by the highest observed PRS (0.34 (0.97)). PRS was lowest in MOD and MARD, with mean PRS being 0.07 (1.02) and 0.23 (0.97), respectively.

Cluster analysis using clinical variables at T2D diagnosis with and without PRS.

Radar plots showing cluster mean values of the variables used to derive 5 clusters of individuals with T2D at the time of diagnosis in 2 separate models. In panel A (Clinical Model), latent class analysis was performed using age, BMI, HbA1c, serum triglycerides, and HDL. In panel B (integrated model), clustering was repeated using T2D PRS in place of HDL. The centre and edge of each polygon represent the minimum and maximum mean cluster values for each variable, respectively. All polygons represent the same scale; all scales are linear. Mean values for each variable within each cluster are represented by coloured dots. For example, mean BMI is highest in cluster 1. BMI, body mass index; HbA1c, hemoglobin A1c; HDL, high-density lipoprotein; PRS, polygenic risk score; T2D, type 2 diabetes.

Given these results, we next explored whether inclusion of PRS as a covariate in the clustering model could further delineate the subgroups by revealing their likely aetiology (henceforth termed “Integrated Model”). We included PRS in the latent class analysis along with age, BMI, HbA1c, and triglycerides (). We identified broadly similar clusters, although with some additional differentiation not seen in the Clinical Model: MOD and MARD were still identified, but in the latter, mean age was notably higher than in the Clinical Model (61.6 (7.6) versus 50.9 (12.2) years). In this analysis, 2 clusters representing insulin-resistant diabetes were identified, differentiated by both the severity of hypertriglyceridaemia and PRS, and therefore, we call these SIRD and Insulin-Resistant Diabetes (IRD). The MD cluster was not identified in this analysis. Instead, this integrated analysis identified a cluster associated with a high PRS (mean PRS 0.48 (0.95)), young age of onset (41.6 (7.6) years), low BMI (26.7 (3.3) kg/m2), and low triglycerides (1.66 (0.74) mmol/L). Given that our PRS is constructed predominantly with variants that are known to have strong effect sizes on T2D () and that these loci are well known [44] to represent defects in beta-cell function (including TCF7L2, KCNJ11, HNF4A, CDKAL1, MTNR1B, SLC30A8, and IGF2BP2), we call this cluster Probable Severe Insulin-Deficient Diabetes (pSIDD). Mean differences were compared between nondiabetic controls and each cluster for previously identified glycaemic trait PRSs [18,37] (). Compared to nondiabetic controls, the pSIDD cluster had higher fasting glucose PRS (mean difference 0.298, one-sided p < 0.001) and the beta cell PRS (mean difference 0.302, one-sided p < 0.001). Additionally, we saw increasing odds of allocation to pSIDD with increasing quintiles of fasting glucose PRS (OR for top versus bottom PRS quintile 1.15 (95% CI 1.09 to 1.20) and beta cell PRS (OR for top versus bottom PRS quintile 1.16 (95% CI 1.10 to 1.21)), and decreasing odds of allocation to pSIDD with increasing corrected insulin response PRS (OR for top versus bottom PRS quintile 0.87 (95% CI 0.83 to 0.91). These associations of pSIDD with glycaemic traits are in keeping with previous reports of SIDD in the literature [18]. The beta cell PRS was also significantly higher in SIRD (mean difference 0.339, one-sided p = 0.02) and IRD (mean difference 0.239, one-sided p < 0.001) than in controls, suggesting the possibility of combined insulin resistant and deficient aetiology underlying these clusters, as described in other studies of South Asians [2].

In Cox proportional hazard models (), macrovascular complication rates were higher in MARD and MOD (p < 0.001 and p = 0.009, respectively); microvascular rates were highest in IRD and MARD (p < 0.001 and p = 0.002). In age-adjusted Cox proportional hazard models, rates of progression to micro- and macrovascular complications in the MARD cluster were not significantly different to other clusters, but results for MOD and IRD were unchanged.

Discussion

In this study, we harness the power of a large population-based study of BPB people with linked health and genomic data to improve the prediction of T2D. For the first time in a real-world study population, we show that use of PRS enhances prediction of incident T2D on top of an established and clinically validated risk prediction tool, QDiabetes. We also show that among BPB women with a history of GDM, PRS is associated with T2D development. Additionally, we show that our PRS is variably associated with T2D subgroups and can itself distinguish a subgroup that is undifferentiated by clinical features.

To our knowledge, this is the first study to systematically assess the transferability of genetic loci associated with T2D in EUR individuals into BPB individuals. Previous studies have assessed directional consistency of genetic effects [45] or heterogeneity in allelic ORs [46] across ancestry groups but, to our knowledge, have not explicitly asked how many genetic loci one might expect to replicate in a sample of a given ancestry given the power and linkage disequilibrium differences. We replicated a lower percentage of previously identified GWAS loci (22.5%) in G&H than would be expected (30.0%) after accounting for power, consistent with the only moderate trans-ancestry genetic correlation between the 2 populations. That fewer loci were replicated than expected might also be due to ancestry-specific gene–environment interactions and the Winner’s Curse phenomenon, an ascertainment bias in which the effect sizes of previously reported loci were likely overestimated, which could lower the chances of replication [47]. Among the genomic loci, which did transfer between EUR and BPB individuals and were powered to interrogate this, only 33% showed evidence of shared causal variants. Hence, despite the considerable overlap in genetic risk between BPB and EUR individuals, our results should still motivate larger studies of T2D genetics in different ancestry groups [48] in order to better characterise the genetic contribution to disease and ancestry-specific aetiological features. In the future, the findings of such studies may point to different effects of risk factors via mendelian randomisation, or different potential drug targets across populations.

Our IRT that combines QDiabetes scores with PRS modestly improved risk classification in a population at high risk of T2D. Specifically, the IRT improved NRI compared to QDiabetes models A (3.22%) and B (0.80%). The IRT did not improve NRI compared to model C, which may reflect reduced power in model C due to lower sample size (n = 864) than the other 2 models (n = 13,648 for model A; n = 4,344 for model B). Alternatively, the lower NRI for models B and C could imply that the benefit of integrating genetic information with risk models already incorporating metrics of hyperglycaemia is more limited than models without hyperglycaemic metrics (Model A). This hypothesis is supported by the association of HbA1c and fasting glucose with PRS (correlation coefficients 0.11 and 0.07, respectively), although the strength of the association is weak. This interpretation would be in line with previous findings reported in EUR individuals by Lyssenko and colleagues [49] and Meigs and colleagues [50], although as these studies were performed prior to the availability of any large GWAS data reassessment with an updated PRS is of benefit, as is its reporting in BPB individuals. Hippisley-Cox and Coupland [20] describe model B (incorporating FPG) as the best-fitting model in 11.5 million (multiethnic, but dominantly Europeans with only approximately 4.5% of South Asian ancestry) individuals in the QResearch database (C-Index 0.89 for women, 0.87 for men for model B, compared to 0.83 for women and 0.81 for men in model A), but we observed lower C-Index values for QDiabetes alone in our study (). The lower C-Index values for comparable models in our study could be also due to differing age, and deprivation distributions between the QResearch database and G&H participants, and the lack of ethnic diversity in our model, since ethnicity is strongly weighted in the QDiabetes risk estimation algorithms and explains a large proportion of variance in T2D risk.

We observed positive NRIs with the IRT driven principally by enhanced reclassification of younger individuals from low to high risk and improved downgrading of older individuals from high to low risk. In the context of a growing burden of T2D in increasingly resource-constrained health systems, such reclassification would support more effective resource allocation, with intensification of preventative care pathways for those at highest risk (e.g., early referral to T2D prevention programmes [13]) and relaxation of those who are identified to be at reduced risk. We found that individuals being reclassified as higher risk tended to be young (mean age 36.0 years), free from comorbidities, and relatively slim (mean BMI 25.9 kg/m2; ). According to estimates by the International Diabetes Federation, half of cases of T2D remain undiagnosed [51]. Using our IRT to identify these individuals, who would likely otherwise have been considered healthy, as high risk could offer significant opportunities to identify and manage T2D early and prevent subsequent morbidity and mortality [52]. This finding is particularly important given the observation that early-onset T2D is associated with rapid progression to vascular complications [53].

We showed that in BPB women with a history of GDM, a T2D PRS was associated with development of T2D. This finding persisted after adjustment for established clinical risk factors and the QDiabetes risk score. Our findings are in keeping with prior reports of an association between a T2D PRS and development of T2D in white [21] and Southeast Asian [22] women with a history of GDM. While previous studies have applied alternative T2D PRSs to women of South Asian origin to identify those at higher risk of GDM [54], none to date have explored their association with T2D development in women with GDM. Although we observed similar associations between PRS and T2D development in age- and BMI-matched male and female controls, the clinical utility of this finding in women with a history of GDM is clearer: this patient group is at extremely high risk of T2D, where improvements to follow-up and screening processes achieved by individualisation could be valuable. This is particularly clinically important in BPB women given their increased risk of developing T2D relative to other ethnic groups, and the globally low uptake of postpartum diabetes screening in non-white individuals [55]. Among individuals without T2D (non-cases), women with a history of GDM displayed a higher mean PRS than age- and BMI-matched females without GDM and compared to males. This could be because these individuals have increased propensity to develop T2D in the future, but have not yet developed it, or because they have undiagnosed T2D, for example, through missed screening tests postdelivery.

We used data-driven clustering approaches in 2 separate models to explore how PRS might associate with the T2D subgroups that are increasingly recognised as a route to developing stratified diabetes care. Despite the absence of biomarkers such as auto-antibodies and C-peptide, which are seldom measured in the routine primary care datasets we had available, we were able to reproduce previously described diabetes subgroups [2,16,17] in a previously uninvestigated population of BPBs. We showed heterogeneous distributions of a T2D PRS across clusters in a model whose membership was defined by clinical measures. PRS was lowest in the MARD cluster, concordant with our findings that it performed best in predicting the risk of T2D in people aged under 40 years, and the MOD cluster, suggesting that other polygenic influences (e.g., on body weight) may be more important. Our SIRD cluster was characterised by markedly raised serum triglycerides and is worthy of further exploration in studies of rare genetic variants. In the absence of clinical measures of insulin secretion such as C-peptide, we were unable to fully delineate the previously well-replicated SIDD cluster in the Clinical Model. However, our MD cluster may include individuals who are insulin deficient as it is characterised by low serum triglycerides (1.21 (0.43) mmol/L) in addition to high HDL (2.07 (0.25) mmol/L), and our CUD cluster likely also contains people with SIDD as its membership is characterised by high HbA1c and low serum triglycerides.

When we incorporated the PRS in an integrated cluster model with clinical features, we observed additional delineation of subgroups than our previous model, with emergence of clear clusters representing IRD and SIRD. This enhanced delineation was possible without biochemical measures of insulin secretion and resistance or diabetes autoantibody data. As such measures are rarely performed in routine primary care for people with diabetes (due to cost, performance, and interpretative challenges), these findings suggest PRS could be a pragmatic tool to aid clinical care if cheap genotyping chips become widely used in health systems. Interestingly, we were also able to delineate a probable SIDD cluster, well described in other studies of South Asians [23,56], on the basis of high polygenic susceptibility that was underpinned by gene variants associated with insulin secretion, in combination with supportive clinical features including high HbA1c, and low BMI and serum triglycerides. These findings are supported by tendencies towards higher scores on previously reported beta cell PRS [37] and fasting glucose PRS [18]. This analysis showed that the diabetes subgroups identified using the integrated cluster model were associated with differential rates of progression to complications, which were not apparent in the clusters found by the Clinical Model, implying that addition of PRS to data-driven identification of T2D subgroups could provide an unexplored clinical tool to risk-stratify populations and target care.

This study has several limitations, including our need to impute missing data not present in health records and the lack of a replication cohort to externally validate findings. Our analysis of the progression to T2D after GDM was limited by the low uptake of postpartum diabetes screening that may have resulted in underdiagnosis of T2D and ascertainment bias. Across all of our analyses, it is likely that some individuals coded as having T2D actually have type 1 diabetes or rare monogenic forms of diabetes, although the absolute number of these and impact on overall findings is likely to be very small. It is also likely that undiagnosed T2D may be present in some of our controls, and this is expected to lead to a higher mean PRS in that group, but attenuation of any effect sizes of analyses investigating its predictive ability. Misclassification of type 1 and type 2 diabetes may have occurred in this real-world data set, although the small number of individuals diagnosed with T2D managed with insulin alone (n = 14) suggests the effect of this on our findings would be minimal. In the absence of C-peptide and auto-antibody data in our primary care data set, estimating misclassification rates is complex and beyond the scope of this study. While miscoding of T2D may also have impacted results, previous research has suggested this occurs in fewer than 2% of individuals [57], and we would similarly expect the effect of this on results to be small.

In conclusion, our T2D PRS, optimised in British South Asians, modestly enhances prediction of incident disease when combined with an established clinical risk tool compared to using the clinical tool alone, and particularly in young people. Additionally, the PRS has value in predicting the onset of T2D in a specific high-risk group, women affected by gestational diabetes. These findings could aid the personalisation of care of people at risk of T2D. The PRS also helps elucidate aetiologically different diabetes subgroups at diagnosis, in the absence of insulin secretion/resistance measures, and these differ in their association with future complications. The value of PRS in this context may assist effective stratification of care in the future. Our work provides important insight into the genetic risk in an ethnic group underrepresented in research but disproportionately affected by T2D, and has significant potential to be translated into clinical practice.

Additional methods.

(DOCX)

Click here for additional data file.

TRIPOD checklist.

(DOCX)

Click here for additional data file.

PRSs constructed using the clumping and p-value thresholding (C+T) method or from the PGS Catalog.

AUC, area under the curve; GWAS, genome-wide association study; OR, odds ratio; PRS, polygenic risk score; SD, standard deviation.

(DOCX)

Click here for additional data file.

Effect of altering age cutoff for older and younger participants on NRI in risk of incident T2D analysis.

NRI, net reclassification index; T2D, type 2 diabetes.

(DOCX)

Click here for additional data file.

Elbow plot to identify optimal number of clusters in latent class analysis using BIC and AIC.

AIC, Akaike information criteria; BIC, Bayesian information criteria.

(DOCX)

Click here for additional data file.

Manhattan plot and Q-Q plot for the GWAS in G&H.

G&H, Genes & Health; GWAS, genome-wide association study.

(DOCX)

Click here for additional data file.

Distribution of previously reported glycaemic trait PRSs across clusters in the Integrated Model.

BMI, body mass index; CI, confidence interval; CIR, corrected insulin response; FINS, fasting insulin; HbA1c, hemoglobin A1c; IRD, Insulin-Resistant Diabetes; ISI, insulin sensitivity index; MARD, Mild Age-Related Diabetes; MOD, Mild Obesity-related Diabetes; PRS, polygenic risk score; pSIDD, Probable Severe Insulin-Deficient Diabetes; SIRD, Severe Insulin-Resistant Diabetes; T2D, type 2 diabetes.

(DOCX)

Click here for additional data file.

Cox proportional hazard models to show association between cluster membership and development of macro- and microvascular complications of T2D.

CUPS, Clinically Undifferentiated High Polygenic Susceptibility Diabetes; IRD, Insulin-Resistant Diabetes; MARD, Mild Age-Related Diabetes; MOD, Mild Obesity-related Diabetes; SIRD, Severe Insulin-Resistant Diabetes; T2D, type 2 diabetes.

(DOCX)

Click here for additional data file.

Clinical codes used to define conditions.

(XLSX)

Click here for additional data file.

Characteristics of individuals used to construct T2D PRS in G&H.

G&H, Genes & Health; PRS, polygenic risk score; T2D, type 2 diabetes.

(XLSX)

Click here for additional data file.

List of variants used in the T2D PRS for G&H.

G&H, Genes & Health; PRS, polygenic risk score; T2D, type 2 diabetes.

(XLSX)

Click here for additional data file.

Three genome-wide significant variants associated with T2D in G&H.

G&H, Genes & Health; T2D, type 2 diabetes.

(XLSX)

Click here for additional data file.

Six genome-wide significant variants associated with T2D that were reported by Kooner and colleagues [32].

T2D, type 2 diabetes.

(XLSX)

Click here for additional data file.

T2D-associated loci transferable between Europeans and BPB individuals in G&H.

BPB, British Pakistani and Bangladeshi; G&H, Genes & Health; T2D, type 2 diabetes.

(XLSX)

Click here for additional data file.

Results of the trans-ancestry colocalisation analysis to assess sharing of causal variants for transferable loci.

(XLSX)

Click here for additional data file.

Association between T2D PRS (T2D PRS) and selected variables included in analyses.

PRS, polygenic risk score; T2D, type 2 diabetes.

(XLSX)

Click here for additional data file.

C-Index, categorical NRI, continuous NRI, and associated 95% CIs.

CI, confidence interval; C-Index, concordance index; NRI, net reclassification index.

(XLSX)

Click here for additional data file.

Characteristics of participants after reclassification by IRT versus QDiabetes Model A alone.

IRT, integrated risk tool.

(XLSX)

Click here for additional data file.

Number of participants reclassified as high (>10%) and low (<10%) risk of developing T2D over 10 years of follow-up by the IRT compared to QDiabetes Models A, B, and C.

IRT, integrated risk tool; T2D, type 2 diabetes.

(XLSX)

Click here for additional data file.

Cox proportional hazard models describing association between T2D PRS and development of T2D in 302 women with a history of GDM.

GDM, gestational diabetes mellitus; PRS, polygenic risk score; T2D, type 2 diabetes.

(XLSX)

Click here for additional data file.

Distribution of previously reported glycaemic trait PRS and wGRS across clusters in the Integrated Model.

PRS, polygenic risk score; wGRS, weighted genetic risk scores.

(XLSX)

Click here for additional data file.

26 Aug 2021

Dear Dr Finer,

Thank you for submitting your manuscript entitled "Harnessing the power of polygenic risk scores to predict type 2 diabetes and its subtypes in a high-risk population of British Pakistanis and Bangladeshis in a routine healthcare setting" for consideration by PLOS Medicine.

Your manuscript has now been evaluated by the PLOS Medicine editorial staff as well as by an academic editor with relevant expertise and I am writing to let you know that we would like to send your submission out for external peer review.

However, before we can send your manuscript to reviewers, we need you to complete your submission by providing the metadata that is required for full assessment. To this end, please login to Editorial Manager where you will find the paper in the 'Submissions Needing Revisions' folder on your homepage. Please click 'Revise Submission' from the Action Links and complete all additional questions in the submission questionnaire.

Please re-submit your manuscript within two working days, i.e. by Aug 30 2021 11:59PM.

Login to Editorial Manager here: https://www.editorialmanager.com/pmedicine

Once your full submission is complete, your paper will undergo a series of checks in preparation for peer review. Once your manuscript has passed all checks it will be sent out for review.

Feel free to email us at plosmedicine@plos.org if you have any queries relating to your submission.

Kind regards,

Callam Davidson

Associate Editor

PLOS Medicine

3 Nov 2021

Dear Dr. Finer,

Thank you very much for submitting your manuscript "Harnessing the power of polygenic risk scores to predict type 2 diabetes and its subtypes in a high-risk population of British Pakistanis and Bangladeshis in a routine healthcare setting" (PMEDICINE-D-21-03668R1) for consideration at PLOS Medicine.

Your paper was evaluated by an associate editor and discussed among all the editors here. It was also discussed with an academic editor with relevant expertise, and sent to independent reviewers, including a statistical reviewer. The reviews are appended at the bottom of this email and any accompanying reviewer attachments can be seen via the link below:

[LINK]

In light of these reviews, I am afraid that we will not be able to accept the manuscript for publication in the journal in its current form, but we would like to consider a revised version that addresses the reviewers' and editors' comments. Obviously we cannot make any decision about publication until we have seen the revised manuscript and your response, and we plan to seek re-review by one or more of the reviewers.

In revising the manuscript for further consideration, your revisions should address the specific points made by each reviewer and the editors. Please also check the guidelines for revised papers at http://journals.plos.org/plosmedicine/s/revising-your-manuscript for any that apply to your paper. In your rebuttal letter you should indicate your response to the reviewers' and editors' comments, the changes you have made in the manuscript, and include either an excerpt of the revised text or the location (eg: page and line number) where each change can be found. Please submit a clean version of the paper as the main article file; a version with changes marked should be uploaded as a marked up manuscript.

In addition, we request that you upload any figures associated with your paper as individual TIF or EPS files with 300dpi resolution at resubmission; please read our figure guidelines for more information on our requirements: http://journals.plos.org/plosmedicine/s/figures. While revising your submission, please upload your figure files to the PACE digital diagnostic tool, https://pacev2.apexcovantage.com/. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email us at PLOSMedicine@plos.org.

We hope to receive your revised manuscript by Nov 24 2021 11:59PM. Please email us (plosmedicine@plos.org) if you have any questions or concerns.

***Please note while forming your response, if your article is accepted, you may have the opportunity to make the peer review history publicly available. The record will include editor decision letters (with reviews) and your responses to reviewer comments. If eligible, we will contact you to opt in or out.***

We ask every co-author listed on the manuscript to fill in a contributing author statement, making sure to declare all competing interests. If any of the co-authors have not filled in the statement, we will remind them to do so when the paper is revised. If all statements are not completed in a timely fashion this could hold up the re-review process. If new competing interests are declared later in the revision process, this may also hold up the submission. Should there be a problem getting one of your co-authors to fill in a statement we will be in contact. YOU MUST NOT ADD OR REMOVE AUTHORS UNLESS YOU HAVE ALERTED THE EDITOR HANDLING THE MANUSCRIPT TO THE CHANGE AND THEY SPECIFICALLY HAVE AGREED TO IT. You can see our competing interests policy here: http://journals.plos.org/plosmedicine/s/competing-interests.

Please use the following link to submit the revised manuscript:

https://www.editorialmanager.com/pmedicine/

Your article can be found in the "Submissions Needing Revision" folder.

To enhance the reproducibility of your results, we recommend that you deposit your laboratory protocols in protocols.io, where a protocol can be assigned its own identifier (DOI) such that it can be cited independently in the future. Additionally, PLOS ONE offers an option to publish peer-reviewed clinical study protocols. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols

Please ensure that the paper adheres to the PLOS Data Availability Policy (see http://journals.plos.org/plosmedicine/s/data-availability), which requires that all data underlying the study's findings be provided in a repository or as Supporting Information. For data residing with a third party, authors are required to provide instructions with contact information for obtaining the data. PLOS journals do not allow statements supported by "data not shown" or "unpublished results." For such statements, authors must provide supporting data or cite public sources that include it.

We look forward to receiving your revised manuscript.

Sincerely,

Callam Davidson,

PLOS Medicine

plosmedicine.org

-----------------------------------------------------------

Requests from the editors:

Please revise your title according to PLOS Medicine's style. Your title must be nondeclarative and not a question. It should begin with main concept if possible. "Effect of" should be used only if causality can be inferred, i.e., for an RCT. Please place the study design ("A randomized controlled trial," "A retrospective study," "A modelling study," etc.) in the subtitle (ie, after a colon).

Abstract Background: Provide the context of why the study is important. The final sentence should clearly state the study question.

Abstract Methods and Findings:

* Please include the study design, years during which the study took place, and main outcome measures.

* Please quantify the main results (with 95% CIs and p values).

* Please include the important dependent variables that are adjusted for in the analyses.

* In the last sentence of the Abstract Methods and Findings section, please describe the main limitation(s) of the study's methodology.

Abstract Conclusions:

* Please address the study implications without overreaching what can be concluded from the data; the phrase "In this study, we observed ..." may be useful.

* Please avoid vague statements such as "these results have major implications for policy/clinical care". Mention only specific implications substantiated by the results.

Please add line numbering in the margins of the document to facilitate the review process.

Author summary bullet points 2 and 3: Please temper assertions of primacy (“No study has systematically…”) by adding ‘to our knowledge’ or similar. Please check throughout (other instances I noted were in the first and second paragraphs of the discussion).

Please remove the final line of your introduction (“Our findings have major clinical relevance to a wider UK and global population of people of south Asian ancestry at risk of type 2 diabetes”). I would suggest toning this statement down and relocating it to the discussion.

Please remove funding information from the acknowledgements, in the event of publication this information is published as metadata based on your responses to the submission form questions.

For similar reasons, please also remove the conflict of interest and contributions statements from the end of the main text and ensure this information is captured as part of the submission form.

Please add ‘[preprint]’ after medRxiv for references 9, 13, and 41.

Please ensure you define all abbreviations used in your tables.

In Figures 1-3, please show the axis beginning at zero. If this is not possible, please show a break in the axis.

At this stage, we ask that you include a short, non-technical Author Summary of your research to make findings accessible to a wide audience that includes both scientists and non-scientists. The Author Summary should immediately follow the Abstract in your revised manuscript. This text is subject to editorial change and should be distinct from the scientific abstract. Please see our author guidelines for more information: https://journals.plos.org/plosmedicine/s/revising-your-manuscript#loc-author-summary

Did your study have a prospective protocol or analysis plan? Please state this (either way) early in the Methods section.

a) If a prospective analysis plan (from your funding proposal, IRB or other ethics committee submission, study protocol, or other planning document written before analyzing the data) was used in designing the study, please include the relevant prospectively written document with your revised manuscript as a Supporting Information file to be published alongside your study, and cite it in the Methods section. A legend for this file should be included at the end of your manuscript.

b) If no such document exists, please make sure that the Methods section transparently describes when analyses were planned, and when/why any data-driven changes to analyses took place.

c) In either case, changes in the analysis-- including those made in response to peer review comments-- should be identified as such in the Methods section of the paper, with rationale.

Please consider adding an additional figure providing an overview of the study design (see Figure 1 in https://journals.plos.org/plosmedicine/article?id=10.1371/journal.pmed.1003498 for an example). This will aid comprehension for the non-specialist reader.

Please ensure that the study is reported according to the TRIPOD guideline, and include the completed TRIPOD checklist as Supporting Information. Please add the following statement, or similar, to the Methods: "This study is reported as per the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) guideline (S1 Checklist)."

The TRIPOD guideline can be found here: https://www.equator-network.org/reporting-guidelines/tripod-statement/

When completing the checklist, please use section and paragraph numbers, rather than page numbers.

Comments from the Academic Editor:

1. The diagnosis of diabetes was based on EMR (no OGTT was performed to screen for diabetes). This will have led to a low incidence of diabetes, and will have been largely affected by economic status.

2. The genetic risk score was calculated based on the findings from European-ancestry individuals, I would suggest to use a South Asian genetic risk score or simply perform a replication for all reported loci for T2D with samples from South Asian individuals first.

Comments from the reviewers:

Reviewer #1: The authors perform a wide range of analysis to explore the genetic architecture of type 2 diabetes (T2D) in British Pakistanis and Bangladeshis (BPB). First the authors perform a genome-wide association study (GWAS) and contrast results with a large GWAS of European ancestry individuals. Secondly, the authors construct a polygenic risk score specific for BPB and check added predictive performance over the existing clinical score QDiabetes.

My comments focus on the methodological details of these two main analysis steps.

1. Transfer of previously-identified GWAS loci to the study population

a) Please provide more details (e.g. citation, formulas or code) how the "expected power for replication" was calculated.

b) How did you calculate the expected proportion of genetic loci one might expect to replicate in a sample of a given ancestry given the power and linkage disequilibrium differences? Please provide more details (e.g. citation, formulas or code).

c) 76 loci are classified as "transferrable". Yet only 27 loci are analysed for colocalisation using TEColoc. Please explain why 49 loci are not considered. Is this due to the filter that ≥10% overlapping variants should be present? If so, please explain why there is so little overlap between Vujkovic et al European-ancestry GWAS and the BPB G&H GWAS.

d) Regarding imputation, in the Supplementary File 1 the number of SNPs used to construct PRS (N=9,527,863) is provided, but not the number for GWAS or colocalization analysis. Please provide these missing two numbers of SNPs as well. Why did you use different set of SNPs for different analysis?

e) Please provide more details on the TEColoc or the JILM method and why you selected this method in contrast to the more commonly used colocalisation method by Giambartolomei et al.

2. Development of an integrated risk tool to predict incident type 2 diabetes

a) The authors show two ways to calculate the categorical net reclassification index (NRI):

(1) the proportion of individuals who subsequently developed type 2 diabetes correctly reclassified as high risk, minus the proportion of individuals who subsequently developed type 2 diabetes incorrectly reclassified as low risk;

(2) the proportion of individuals who did not develop type 2 diabetes correctly reclassified as low risk, minus the proportion of individuals who did not develop type 2 diabetes incorrectly reclassified as high risk.

Please be more precise how the proportions are defined, what exactly is the numerator and the denominator of the proportion?

b) Which of the two definitions was used for Figure 1 and in the Results?

c) Figure 1 shows the NRI for all subjects and cases and non-cases. How can the "proportion of individuals who subsequently developed type 2 diabetes" be calculated for non-cases?

d) There is a big distinction in the pattern of NRI for subjects younger or older than 40. How robust are the results with respect to this cut-off?

Minor:

- Page 5: "with descriptive statistics. ." Please remove the second full period.

Reviewer #2: In this paper, Hodgson et al. construct polygenic risk scores and integrate them with clinical models to enhance the prediction of type 2 diabetes (T2D) in UK Pakistani and Bangladeshi individuals in a large population study. They first assess the transferability of known T2D genetic risk loci (derived from predominantly European populations) to their cohort and replicated significantly fewer loci than predicted. They then constructed a number of polygenic scores using variants identified from a multi-ancestry GWAS and adjusted/tuned the scores for Pakistani and Bangladeshi individuals in 2249 cases and 4443 controls, selecting the one with the highest odds for prediction. This score was then integrated with three versions of a clinical model (QDiabetes) to predict onset of T2D over 10 years in 13,642 individuals free from diabetes at baseline. The authors show that incorporation of the PRS with the QDiabetes provided better discrimination of risk of progression to T2D, especially in those developing T2D <40 years. The PRS also was noted to be higher in women with a history of gestational diabetes who progressed to T2D versus non-progressors. Finally, they incorporated the PRS into a cluster analysis of baseline routine clinical characteristics, first largely replicating clusters defined previously in white populations, but also defining a 'clinically undifferentiated' cluster, associated with higher PRS but with no relationship to risk of progression to microvascular and macrovascular complications.

This study has several strengths:

1. The analysis was conducted in a large population-based cohort of a hitherto understudied ethnic group; UK Pakistani and Bangladeshi individuals, who are at a disproportionately higher risk of developing T2D even within the south Asian ethnic group and at relatively lean body mass. It would be reasonable to postulate that this underlying risk is, in part, mediated by genetic predisposition.

2. Longitudinal data linkage of participants to their electronic health record further provides a very powerful methodology in which genetic data can be associated with clinical outcomes. There are very few longitudinal cohorts of this size with linked data.

3. This study uses multimodal approaches to decipher phenotype-genotype associations, including cluster analyses to identify heterogeneity in T2D sub-phenotypes, which is relatively novel in this ethnic group.

On the other hand, there are some issues that must be addressed to enhance the relevance and focus of the paper:

1. The authors chose to apply polygenic scores derived in predominantly white European ethnicity to their Pakistani and Bangladeshi population and 'fine tuned' the effect size of variants to enable the best prediction of T2D. If their only goal had been to examine the transferrability of European PRS to South Asians, this may have been justified; however, their stated goal in the first sentence of the Discussion (also the more clinically relevant one) is to improve the prediction of T2D in South Asians via genetic means, and in that case the genetic predictive tool should include variants discovered in this population. One alternative, somewhat hindered by sample size, would have been to leverage the GWAS performed on cases and controls in this specific G&H cohort, and incorporate those variants into the score. Arguably, higher rates of consanguinity in this ethnicity have the potential to uncover novel loci. Nevertheless, at a minimum, the variants discovered by Kooner et al. (Nature Genetics 2011) should have also been included, as the trans-ancestry GWAS used to identify T2D loci has a very small number of South Asian individuals. The decision not to include South Asian-specific variants should be discussed or justified.

2. In this vein, the GWAS conducted in G&H should be at least briefly presented, including the standard Manhattan and Q-Q plots and a table of the genome-wide significant associations, and placed in the context of the prior Kooner et al. GWAS.

3. One other major limitation is the absence of C-peptide or fasting insulin in this cohort, which pre-empts the ability to generate clusters analogous to those described by Ahlqvist et al. (ref. 19). Since this is an essential marker to denote beta-cell function, the authors are unable to reconstruct the "severe insulin deficient diabetes" (SIDD) cluster originally reported by Ahlqvist et al. Indeed, it is curious that what they term a novel "clinically undifferentiated" cluster has the strongest association with the T2D PRS. This is very telling: a close look at Supplementary Table 4 reveals that the PRS constructed here is largely composed of the variants that have the strongest effect sizes on T2D (i.e. those that were originally discovered in Europeans but replicate in the current smaller sample). Therefore they are the ones that emerged from the earliest European GWAS, which by design favored loci that caused defects in beta-cell function (e.g. TCF7L2, MTNR1B, SLC30A8, KCNJ11, HNF4A, CDKAL1, IGF2BP2; see Pubmed ID 18504548 for an in-depth discussion of how the early T2D GWAS yielded beta-cell genes). Thus it is quite possible that the "clinically undifferentiated" cluster might indeed correspond to a cluster of severe, genetically driven beta-cell dysfunction, analogous to the SIDD cluster of Ahlqvist et al. As it turns out, a genetic exploration of the Ahlqvist clusters already available in preprint form (https://www.medrxiv.org/content/10.1101/2020.09.29.20203935v1) does demonstrate the strongest association of the T2D PRS with the SIDD cluster.

4. The authors do not discuss a clear limitation that the time to follow-up of 10 years does not conclusively distinguish cases from controls, in that some participants classified as controls will eventually develop diabetes. Therefore, scores in the control group, given the high rates of T2D in this population are likely to be higher than true diabetes-free controls. Indeed, the mean age of participants in the dataset used to construct scores was 30.8 years (supp Table 2). Could they not have chosen 'older' diabetes free controls? Further the controls used to construct polygenic scores are described as having 'insufficient data to be included in the study' - it is not clear what is meant by this.

5. The authors acknowledge that there may be miscoding of T2D due to misclassification of diabetes as a limitation, but to identify cases, it appears they relied only on the first appearance of the diagnostic code for T2D in the electronic health record. Could they confirm whether this was verified for example, through confirmation of a repeat diagnostic code at a subsequent encounter, prescription data and if not, could they discuss how likely this is to have been an issue in real-world electronic health records.

6. What might the effect of a highly consanguineous population be on generation and efficacy of a PRS. The authors say they selected 'unrelated' cases and controls - however, should not a PRS in this specific population take into account high proportions of relatedness? Could the genomic data be used to exclude a given degree of relatedness, measured as "identity by descent" in available software such as PLINK?

7. There are likely to be a number of reasons why fewer loci were replicated including differences in minor allele frequency, linkage disequilibrium, gene-environment interactions or the simple statistical phenomenon of the "winner's curse" (an overestimate of the effect size reported in the initial study) - a more extensive discussion of this might have been useful.

8. For the GDM analysis, the sample sizes used to compare PRS between progressors and non-progressors are small. Could the authors verify there was sufficient power based on expected effect sizes from previous studies?

9. Models B and (less so) C that incorporate glycemic measures (glucose and HbA1c, respectively), attenuate the ability of the PRS is predicting T2D onset, suggesting that these biochemical measures probably identify those at risk of T2D. This undermines the utility of the PRS, and confirms previous findings from multiple studies that polygenic prediction of T2D does not add much to clinical predictors if glycemia is included in the instrument (see, e.g. Meigs et al. and Lyssenko et al., both in NEJM 2008). The Lyssenko paper did show that genetic prediction was most useful in younger individuals or those with the longest follow-up, indicating that clinical predictors may not manifest themselves fully until later in life or closer to disease onset, a finding that is consistent with the current study. Since there was no relationship between PRS and incidence of complications, the counter-argument of genetic prediction lacking clinical utility could perhaps be discussed more critically?

10. There are multiple statistical comparisons (e.g. all the SNPs taken forward into this cohort), but there were no adjustments to the p-values for multiple comparisons, as a nominal p-value of 0.05 was taken as evidence of replication. While this is sensible when one is trying to estimate how many SNPs will reach that alpha level over the predicted number based on effect size and sample size, this should be acknowledged and explained.

11. The cluster analysis seems 'tacked on' rather than being integral to the core work of PRS transferrability. As such, there is no discussion of the fact that the PRS is associated with different clusters in this analysis but few significant associations with SNPs were found in the original analysis by Ahlqvist et al. The more extensive genetic evaluation undertaken by that group and published in the preprint cited above provides a potential blueprint of what could be done here with the GWAS data obtained in G&H. There should have been some reference to the recent Indian study that undertook clustering and identified a novel insulin-deficient but non-autoimmune cluster (Anjana et al. Pubmed ID 33527813).

12. This is a small point, but the Ahlqvist et al. study preceded and provided the reason for the subsequent Dennis et al. study - since they are cited at the same place in the manuscript, they should be listed in chronological order.

Reviewer #3: The study by Hodgson et al aims at developing a PRS to estimate the 10-year risk of developing T2D and to identify T2D subtypes in patients of South Asian ancestry living in the UK. The authors tested the transferability to South Asians, of genomic variants associated to T2D in people of European descent and then tested the predictive ability of a more South Asian-specific PRS that they combined with the clinical score QDiabetes to estimate the risk of developing T2D in women with gestational diabetes and in men and women, younger or older than 40 years of age.

Strengths:

The manuscript is well written, and results are clear.

The research question is important. People of South Asian origin are more affected by T2D. This occurs at a younger age and at lower BMI. The question that there may be a genetically-distinct T2D subtype in this population is sound. Most of PRS have been developed with and for populations of European descent. The transferability of data from Europeans to people of South Asian origin is unknown.

The authors have access to a large set of clinical and genomic data including electronic health records and follow-up of British Pakistani and Bangladeshi (BPB).

The analyses are appropriate and carefully done.

The PRS may be useful for women with history of GDM. Women who subsequently developed T2D had a higher PRS than those who did not. Women with GDM had also higher PRS than age- and BMI-matched controls without GDM and higher than men. This clinically important observation could have been investigated further.

Weaknesses.

The construction of the PRS is not detailed enough. The number of SNPs that have been included and their effect size should be stated. How many of the European transferable SNPs have been included in the PRS?

Some of the conclusions are not well supported by the results. For instance, the predictive performance of the PRS over that of the QDiabetes score is very modest. Are the differences in AUC (Table S5) significant after the addition of the PRS? They seem to be significant only for Model A. The incremental effect of the PRS is eliminated by the addition of fasting glucose (Model B) or HbA1c (Model C) to the clinical score.

Same comment could be made for the NRI that is significant only for Model A which has the lowest predictive performance.

Data-driven clustering: T2D cluster driven by the PRS is the largest in term of number of individuals. However, the progression of microvascular and macrovascular outcomes in individuals of this cluster does not differ from the other clusters. The only cluster that differs between macro and microvascular complications is the MARD one. This could be explained by the fact that age has a different impact on the progression of macrovascular and microvascular outcomes.

The cluster analyses could be done also individuals of European descent and the radar plots compared between the two ethnic groups.

Minor:

The title should be more specific to the findings.

Any attachments provided with reviews can be seen via the following link:

[LINK]

13 Jan 2022

Submitted filename: Response to Editors and Reviewers.docx

Click here for additional data file.

8 Feb 2022

Dear Dr. Finer,

Thank you very much for submitting your revised manuscript "Integrating polygenic risk scores in the prediction of type 2 diabetes risk and subtypes: an analysis of genomic and routine primary care health data in a population-based study of British Pakistanis and Bangladeshis" (PMEDICINE-D-21-03668R2) for consideration at PLOS Medicine.

Your paper was sent back to independent reviewers and the reviews are appended at the bottom of this email. Any accompanying reviewer attachments can be seen via the link below:

[LINK]

As there are remaining concerns to be addressed, we will not be able to accept the manuscript for publication in the journal in its current form. We would however like to consider a revised version that addresses the reviewers' and editors' comments. As stated previously, we cannot make any decision about publication until we have seen the revised manuscript and your response, and we plan to seek re-review by one or more of the reviewers.

In revising the manuscript for further consideration, your revisions should address the specific points made by each reviewer and the editors. Please also check the guidelines for revised papers at http://journals.plos.org/plosmedicine/s/revising-your-manuscript for any that apply to your paper. In your rebuttal letter you should indicate your response to the reviewers' and editors' comments, the changes you have made in the manuscript, and include either an excerpt of the revised text or the location (eg: page and line number) where each change can be found. Please submit a clean version of the paper as the main article file; a version with changes marked should be uploaded as a marked up manuscript.

We hope to receive your revised manuscript by Mar 01 2022 11:59PM. Please email us (plosmedicine@plos.org) if you have any questions or concerns.

Please use the following link to submit the revised manuscript:

https://www.editorialmanager.com/pmedicine/

Your article can be found in the "Submissions Needing Revision" folder.

We look forward to receiving your revised manuscript and please don't hesitate to get in touch if you have any questions.

Sincerely,

Callam Davidson,

Associate Editor

PLOS Medicine

plosmedicine.org

-----------------------------------------------------------

Requests from the editors:

The current title is slightly too long. Please update to 'Integrating polygenic risk scores in the prediction of type 2 diabetes risk and subtypes in British Pakistanis and Bangladeshis: a population-based cohort study', or similar.

Thanks for providing a completed TRIPOD checklist - please update the checklist to use section and paragraph numbers as opposed to page numbers (as these change during the revision process).

Please also ensure the TRIPOD checklist is cited in the methods (e.g. 'This study is reported according to the Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD) statement (S1 checklist)'.

Comments from the reviewers:

Reviewer #2: The authors have been very responsive to the review process and the manuscript is significantly improved. Particularly I appreciate the new designation of the pSIDD cluster. There is a small number of minor outstanding issues that should be relatively straightforward to resolve.

1. The choice of SNPs to construct the most informative polygenic score: thank you for the clarification, and it does make sense to use the multiethnic resource that includes more samples than Kooner et al. I still think that some mention of the rationale used and elegantly outlined in the response to the reviewers should find its way into the manuscript, if anything to allay the concern of the informed reader. This could be done in the Supplement or as a couple of sentences in the Methods.

2. Multiple testing correction for SNPs taken forward: I agree that for a replication of a single variant, a P-value <0.05 is sufficient. When one is testing replication of multiple individual variants, the number of hypotheses is as large as the number of uncorrelated variants taken forward: thus strictly speaking it does make sense to use a Bonferroni correction, and this is done routinely. You could simply state that the variants passed the nominal P-value, but that experiment-wide significance would be set at 0.05/n.

3. As minor edits, in Supplementary Table 4, the locus "CDKAL1" has a typo ("CKDAL1"). Gene names should be italicized throughout, including Supplementary Tables

Reviewer #3: The modifications made to the revised version of the manuscript clearly add to the quality of this paper, and most of my concerns from my previous review have been addressed adequately.

Reviewer #4: The authors explore the genetic determinants of Type 2 Diabetes (T2D) in people of British Pakistani and Bangladeshi (BPB) ancestry. Using genetic data from the Genes & Health project, they first explore whether SNPs associated with T2D in European populations are also associated with T2D in their BPB sample. They then construct a polygenic risk score (PRS) for T2D and show that it can be useful for predicting which individuals will develop T2D alongside an already established clinical tool (QDiabetes).

I would like to commend the authors for working with a previously understudied population group, and I agree with them that further genetic research involving such groups is needed.

As a statistician, I am perhaps unable to fully scrutinise the authors' applied results, but I will attempt to provide some comments on their use of appropriate methodology to obtain these results. Since I was not involved in the previous round of review, some of my comments will be complementary to those previously raised by other reviewers, while other comments will be different.

Here are my comments on the manuscript.

1) The authors used a European LD panel (European 1000G samples) for their LD clustering and thresholding analysis. Was it not possible to use a south Asian panel instead?

2) In their supplement, the authors present a formula according to which they calculated the proportion of SNPs from the European GWAS that would be expected to replicate in the G+H GWAS. First of all, the formula seems wrong; for example, in a case-control study with n_cases = n_controls, one would have 1 - phi = 0 and hence only a 0.05 probability to replicate a SNP regardless of its beta coefficient in the European GWAS. Please clarify (could it be that phi = n_cases / n instead?). Moreover, the authors should provide an explanation about why the formula does what they are trying to achieve, or a reference to support its use. In general, I found it a bit surprising that 72% of these GWAS-significant variants should be expected to not associate with T2D even at a nominal 0.05 level in the authors' genetic sample.

3) In the authors' colocalization analyses, the overlap of genetic variants between the European and G+H GWAS in each region was as low as 10%. Therefore, it is quite likely that in some regions the truly causal variant(s) may not have been genotyped in one of the two samples (or even in none of the two samples). Would that be a problem for the TEColoc method used by the authors?

4) What was the explanatory power (e.g. in terms of R^2 values) of the diabetes PRS constructed by the authors? Also, did the PRS associate with any of the other variables included in their analysis (BMI, HDL, TG, HBA1c, glucose)? Following from a previous comment by another reviewer, any associations with glucose and HBA1c would be particularly interesting to report, since they would explain why the PRS offered little benefit over QDiabetes when using models B or C which incorporated these traits.

5) To assess the predictive accuracy of their PRS, the authors use the net reclassification index (NRI), which amounts to first binarizing each method's predictions by splitting individuals to high-risk and low-risk groups and then comparing the proportions of cases in each group. Wouldn't it make more sense to not binarize and instead compare each method's predictive accuracy using the actual estimated probabilities? E.g. something like Sum( |Pr(case) - Case.status| ) where Case.status = 1 for cases and 0 for controls, and the summation extends over all individuals.

6) One possible limitation of the study is the potential for its results to be affected by selection bias. For example, since BPB individuals were only recruited at three centres across the UK (East London, Bradford and Manchester), some of their baseline characteristics may differ to those of other such populations. The construction of the PRS was based on an even more selective sample, since individuals included in the QDiabetes assessment sample were excluded from the PRS construction. In addition, genetic studies conducted on individuals with a pre-established disease are known to suffer from a form of selection bias called index event bias; this may have been present in the authors' analysis on T2D incidence in women who had previously developed gestational diabetes.

7) As mentioned by the authors, one of the main reasons for studying T2D in BPB individuals is the increased frequency of diabetes in this population group. Can any of the authors' findings help to explain why T2D is more common in BPB individuals?

8) In their conclusions, the authors state that the addition of their PRS to data-driven identification of T2D subgroups can improve the classification of individuals into T2D subgroups. It is perhaps worth clarifying that this seems to be the case only if information on other relevant variables is missing. In general, the results of a statistical clustering algorithm depend on what variables the clustering was based on. Presumably, the authors' PRS was only needed to identify and characterize the pSIDD cluster because information on fasting glucose and beta cell function-related traits was not available in their dataset, and the PRS would be redundant if such information was available.

Any attachments provided with reviews can be seen via the following link:

[LINK]

Submitted filename: Manuscript Number PMEDICINE.docx

Click here for additional data file.

1 Mar 2022

Submitted filename: PLoS_Medicine_Reviewers_Comments_2022Feb.docx

Click here for additional data file.

24 Mar 2022

Dear Dr. Finer,

Thank you very much for re-submitting your manuscript "Integrating polygenic risk scores in the prediction of type 2 diabetes risk and subtypes in British Pakistanis and Bangladeshis: a population-based cohort study" (PMEDICINE-D-21-03668R3) for review by PLOS Medicine.

I have discussed the paper with my colleagues and the academic editor and it was also seen again by two reviewers. I am pleased to say that provided the remaining editorial and production issues are dealt with we are planning to accept the paper for publication in the journal.

The remaining issues that need to be addressed are listed at the end of this email. Any accompanying reviewer attachments can be seen via the link below. Please take these into account before resubmitting your manuscript:

[LINK]

***Please note while forming your response, if your article is accepted, you may have the opportunity to make the peer review history publicly available. The record will include editor decision letters (with reviews) and your responses to reviewer comments. If eligible, we will contact you to opt in or out.***

In revising the manuscript for further consideration here, please ensure you address the specific points made by each reviewer and the editors. In your rebuttal letter you should indicate your response to the reviewers' and editors' comments and the changes you have made in the manuscript. Please submit a clean version of the paper as the main article file. A version with changes marked must also be uploaded as a marked up manuscript file.

Please also check the guidelines for revised papers at http://journals.plos.org/plosmedicine/s/revising-your-manuscript for any that apply to your paper. If you haven't already, we ask that you provide a short, non-technical Author Summary of your research to make findings accessible to a wide audience that includes both scientists and non-scientists. The Author Summary should immediately follow the Abstract in your revised manuscript. This text is subject to editorial change and should be distinct from the scientific abstract.

We hope to receive your revised manuscript within 1 week. Please email us (plosmedicine@plos.org) if you have any questions or concerns.

We ask every co-author listed on the manuscript to fill in a contributing author statement. If any of the co-authors have not filled in the statement, we will remind them to do so when the paper is revised. If all statements are not completed in a timely fashion this could hold up the re-review process. Should there be a problem getting one of your co-authors to fill in a statement we will be in contact. YOU MUST NOT ADD OR REMOVE AUTHORS UNLESS YOU HAVE ALERTED THE EDITOR HANDLING THE MANUSCRIPT TO THE CHANGE AND THEY SPECIFICALLY HAVE AGREED TO IT.

Please ensure that the paper adheres to the PLOS Data Availability Policy (see http://journals.plos.org/plosmedicine/s/data-availability), which requires that all data underlying the study's findings be provided in a repository or as Supporting Information. For data residing with a third party, authors are required to provide instructions with contact information for obtaining the data. PLOS journals do not allow statements supported by "data not shown" or "unpublished results." For such statements, authors must provide supporting data or cite public sources that include it.

To enhance the reproducibility of your results, we recommend that you deposit your laboratory protocols in protocols.io, where a protocol can be assigned its own identifier (DOI) such that it can be cited independently in the future. Additionally, PLOS ONE offers an option to publish peer-reviewed clinical study protocols. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols

Please review your reference list to ensure that it is complete and correct. If you have cited papers that have been retracted, please include the rationale for doing so in the manuscript text, or remove these references and replace them with relevant current references. Any changes to the reference list should be mentioned in the rebuttal letter that accompanies your revised manuscript.

Please note, when your manuscript is accepted, an uncorrected proof of your manuscript will be published online ahead of the final version, unless you've already opted out via the online submission form. If, for any reason, you do not want an earlier version of your manuscript published online or are unsure if you have already indicated as such, please let the journal staff know immediately at plosmedicine@plos.org.

If you have any questions in the meantime, please contact me or the journal staff on plosmedicine@plos.org.

We look forward to receiving the revised manuscript by Mar 31 2022 11:59PM.

Sincerely,

Callam Davidson,

Associate Editor

PLOS Medicine

plosmedicine.org

------------------------------------------------------------

Requests from Editors:

Please trim your Author Summary, aiming for 2-3 single sentence bullet points per question (I think you can provide slightly less detail for the ‘Why was this study done?’ question). Please include the headline numbers from the study, such as the sample size and key findings, under the ‘What did the researchers do and find?’ question.

Please define any abbreviations used in your Tables and Figures in the associated legend.

Your line numbering should be continuous throughout the document rather than beginning again on each page.

Page 5, Line 30: Please cite Table 1 in the Results rather than the Methods section.

Please consider using a different colour scheme for your Figures and Supplementary Figures to aid those with red/green colour blindness.

The ordering of your Supporting Figures is non-sequential (Figure S4 is cited before S2 and S3). Please rearrange the figures such that they are cited sequentially.

Please define the 95% CI bars in the legend of Figure 3.

Page 7, Line 47: Should this be Figure 3B-C rather than A-B?

In your Supplementary Tables, please report p values as P<0.001.

It appears the x-axis label in Figure 4 has been cut off.

Page 8, Lines 23-25: This feels like Discussion material (speculative) – please consider relocating.

Page 8, Lines 45-50: Similar to the above, this feels like Discussion material. Please keep to reporting findings in the Results and add interpretation only in the Discussion.

I could not find reference to Figure S5 in the main text.

Page 10, Line 9: ‘To our knowledge’.

Comments from Reviewers:

Reviewer #2: The authors ought to be commended for their thorough and responsive approach to the review process. I have no further comments.

Reviewer #4: I am happy with the authors' response to my previous comments and would like to recommend that their manuscript is accepted for publication.

Any attachments provided with reviews can be seen via the following link:

[LINK]

6 Apr 2022

Dear Dr Finer,

On behalf of my colleagues and the Academic Editor, Dr Weiping Jia, I am pleased to inform you that we have agreed to publish your manuscript "Integrating polygenic risk scores in the prediction of type 2 diabetes risk and subtypes in British Pakistanis and Bangladeshis: a population-based cohort study" (PMEDICINE-D-21-03668R4) in PLOS Medicine.

Before your manuscript can be formally accepted you will need to complete some formatting changes, which you will receive in a follow up email. Please be aware that it may take several days for you to receive this email; during this time no action is required by you. Once you have received these formatting requests, please note that your manuscript will not be scheduled for publication until you have made the required changes.

When making the formatting changes, please also ensure the institutional affiliation of author NS is up to date (the competing interests statement suggests it may not be).

In the meantime, please log into Editorial Manager at http://www.editorialmanager.com/pmedicine/, click the "Update My Information" link at the top of the page, and update your user information to ensure an efficient production process.

PRESS

We frequently collaborate with press offices. If your institution or institutions have a press office, please notify them about your upcoming paper at this point, to enable them to help maximise its impact. If the press office is planning to promote your findings, we would be grateful if they could coordinate with medicinepress@plos.org. If you have not yet opted out of the early version process, we ask that you notify us immediately of any press plans so that we may do so on your behalf.

We also ask that you take this opportunity to read our Embargo Policy regarding the discussion, promotion and media coverage of work that is yet to be published by PLOS. As your manuscript is not yet published, it is bound by the conditions of our Embargo Policy. Please be aware that this policy is in place both to ensure that any press coverage of your article is fully substantiated and to provide a direct link between such coverage and the published work. For full details of our Embargo Policy, please visit http://www.plos.org/about/media-inquiries/embargo-policy/.

To enhance the reproducibility of your results, we recommend that you deposit your laboratory protocols in protocols.io, where a protocol can be assigned its own identifier (DOI) such that it can be cited independently in the future. Additionally, PLOS ONE offers an option to publish peer-reviewed clinical study protocols. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols

Thank you again for submitting to PLOS Medicine. We look forward to publishing your paper.

Sincerely,

Callam Davidson

Associate Editor

PLOS Medicine

Table 1

Characteristics of participants included in analysis with each QDiabetes model.

	Model A (N = 13,648)	Model B (N = 4,334)	Model C (N = 864)	p-value
% Developing T2D	18.2	27.0	33.3	<0.001
T2D PRS	−0.046 (0.99)	−0.080 (0.99)	−0.062 (1.01)	0.210
Age (Years)	35.4 (8.8)	42.0 (11.1)	45.2 (12.3)	<0.001
% Female	49.8	59.2	64.9	<0.001
% Bangladeshi	69.9	73.9	58.0	<0.001
% Family history of diabetes	35.9	40.2	33.9	<0.001
Index of Multiple Deprivation (2015) score	7.4 (2.2)	7.8 (1.9)	7.2 (2.4)	<0.001
BMI (kg/m2)	25.8 (4.6)	26.704 (4.624)	27.7 (4.7)	<0.001
HbA1c (mmol/mol)	39.1 (4.3)	NA	38.0 (4.2)	0.271
Fasting glucose (mmol/L)	4.9 (0.7)	4.8 (0.7)	NA	0.154
HDL (mmol/L)	1.2 (0.3)	1.2 (0.3)	1.2 (0.3)	<0.001
Triglycerides (mmol/L)	1.9 (1.3)	1.8 (1.2)	1.7 (1.2)	0.077
Preexisting Conditions
Gestational diabetes (% females)	4.1	5.2	10.7	<0.001
Polycystic ovarian syndrome (% females)	4.7	7.3	6.6	<0.001
Cardiovascular disease (%)	2.5	4.7	9.7	<0.001
Hypertension (%)	6.1	14.2	13.7	<0.001

QDiabetes Model A is calculated in individuals without HbA1c or FPG; Model B by adding FPG to Model A; and Model C by adding HbA1c to Model A. Values show the mean with SD in brackets, unless otherwise indicated. Means were compared with ANOVA.

ANOVA, analysis of variance; BMI, body mass index; FPG, fasting plasma glucose; HbA1c, hemoglobin A1c; HDL, high-density lipoprotein; PRS, polygenic risk score; SD, standard deviation; T2D, type 2 diabetes.

Table 2

Clinical and Integrated cluster analyses.

AClinical Model		MOD (N = 211)	MARD (N = 861)	SIRD (N = 32)	MD (N = 95)	CUD (N = 3,067)	p-value
	Age (Years)	44.5 (11.4)	50.9 (12.2)	43.1 (10.9)	49.3 (12.3)	45.2 (10.6)	<0.001
	BMI (kg/m 2 )	40.2 (3.6)	27.6 (4.0)	28.2 (3.9)	27.1 (5.2)	27.5 (3.7)	<0.001
	HbA1c (mmol/mol)	57.5 (15.8)	53.4 (13.4)	71.9 (26.9)	55.9 (17.63)	59.0 (16.4)	<0.001
	Serum triglycerides (mmol/L)	1.72 (0.73)	1.39 (0.63)	12.2 (3.95)	1.21 (0.43)	1.97 (1.08)	<0.001
	HDL (mmol/L)	1.17 (0.24)	1.45 (0.15)	0.79 (0.21)	2.07 (0.20)	0.99 (0.17)	<0.001
	PRS	0.07 (1.02)	0.23 (0.97)	0.29 (1.14)	0.26 (1.09)	0.34 (0.97)	0.002
BIntegrated Model		MOD (N = 556)	MARD (N = 1180)	SIRD (N = 37)	IRD (N = 579)	pSIDD (N = 3,489)	p-value
	Age (Years)	44.8 (9.6)	61.6 (7.6)	43.7 (11.3)	42.8 (9.9)	41.6 (7.6)	<0.001
	BMI (kg/m 2 )	37.2 (3.9)	27.5 (3.4)	28.1 (3.7)	27.8 (3.8)	26.7 (3.3)	<0.001
	HbA1c (mmol/mol)	60.0 (17.1)	53.1 (11.6)	72.8 (26.8)	63.3 (19.1)	60.0 (16.5)	<0.001
	Serum triglycerides (mmol/L)	1.77 (0.73)	1.62 (0.70)	13.0 (3.63)	4.97 (1.05)	1.66 (0.74)	<0.001
	PRS	0.16 (0.91)	−0.10 (0.92)	0.41 (0.96)	0.25 (1.02)	0.48 (0.95)	<0.001

Distribution of clustering variables and T2D PRS across 5 data-driven clusters of individuals with T2D at the time of diagnosis in a clinical (panel A) and integrated clinical and genetic (panel B) model. PRS was not used as a clustering variable in the clinical model; it was used as a clustering variable in the integrated model. In the clinical model, the distribution of PRS between groups was compared after cluster allocation. Data are presented as mean (standard deviation). Mean values were compared using ANOVA.

ANOVA, analysis of variance; BMI, body mass index; CUD, Clinically Undifferentiated Diabetes; HbA1c, hemoglobin A1c; HDL, high-density lipoprotein; IRD, Insulin-Resistant Diabetes; MARD, Mild Age-Related Diabetes; MD, Mild Diabetes; MOD, Mild Obesity-related Diabetes; PRS, polygenic risk score; pSIDD, Probable Severe Insulin-Deficient Diabetes; SIRD, Severe Insulin-Resistant Diabetes.