Development and application of type 1 diabetes polygenic scores across diverse populations

Abstract

Type 1 diabetes is highly influenced by genetic risk factors, especially variation in the HLA genes. Polygenic scores can strongly predict type 1 diabetes risk by integrating variants across the genome. Most polygenic scores perform best in European-ancestry populations, due to low representation of other ancestry groups in existing studies, as well as substantial variation in HLA alleles across the globe. However, recent multi-ancestry genome-wide association studies have broadened our understanding of type 1 diabetes genetic risk, and advanced computational techniques can accurately capture HLA alleles at high resolution. This article will review novel approaches to develop type 1 diabetes polygenic scores that demonstrate high predictive power across diverse populations.

Graphical Abstract

Introduction

Type 1 diabetes is a complex disease associated with multiple genetic and environmental risk factors. Polygenic scores, which integrate the risk conferred by multiple variants across the genome, are highly predictive of type 1 diabetes risk. Current type 1 diabetes polygenic scores have an area under the receiving operating characteristic curve (ROC AUC) approaching 0.9 or greater, whether comparing individuals with type 1 diabetes against healthy control individuals (1,2) or against those with type 2 diabetes (3), MODY (4) or neonatal diabetes (4).

However, polygenic score performance is highest in ancestry-matched populations. Most genome-wide association studies (GWAS) have been performed in European populations; polygenic score performance declines when applied to other populations, which may exacerbate underlying health disparities (5). Recent GWAS have expanded the number of known type 1 diabetes genetic loci (6,7), including in multi-ancestry cohorts (8–10). In addition, new methods have been developed to impute HLA alleles (11), compute genome-wide polygenic scores (12), and integrate polygenic scores across ancestries (13,14). Various screening programmes now use polygenic scores to predict type 1 diabetes risk in healthy newborns (e.g. EarlyCheck (15) in USA, Type1Screen (16) in Australia). Therefore, it is imperative to apply polygenic scores that accurately predict type 1 diabetes risk in diverse settings.

In this article, we will discuss methods to apply type 1 diabetes polygenic scores across populations, focusing on developments since the publication of a recent review (17). Whereas other articles have reviewed general principles of multi-ancestry polygenic scores (14), this article will focus on issues specific to type 1 diabetes, including the heterogeneity of type 1 diabetes across populations and the critical importance of HLA alleles. We will review prior type 1 diabetes polygenic scores and discuss three recent scores that aim to predict type 1 diabetes risk across ancestries. Finally, we will discuss future directions to further optimise multi-ancestry risk prediction.

Heterogeneity of type 1 diabetes across populations

The prevalence, incidence and clinical presentation of type 1 diabetes exhibit significant heterogeneity across different ethnic, geographic and genetic backgrounds (Fig. 1). Here, we use the term population as a surrogate for ancestry, ethnicity or race, which are all used in the literature to describe human populations, although their definitions differ. Type 1 diabetes heterogeneity across populations reflects differences in the genetic architecture, environmental triggers and sociodemographic characteristics (16). Individuals of European ancestry are more likely to have multiple autoantibodies and a classic autoimmune type 1 diabetes phenotype. African and African-Caribbean populations show greater heterogeneity, characterised by single-GADA positivity and higher rates of antibody-negative diabetes and ketosis-prone diabetes. The rate of autoantibody-negative type 1 diabetes varies by ethnicity; about 5–15% of individuals with type 1 diabetes lack detectable autoantibodies in predominantly European samples (18) but this proportion is higher in non-European populations, particularly in African and Asian populations. These atypical type 1 diabetes presentations suggest mixed autoimmune and non-autoimmune mechanisms in type 1 diabetes aetiology and have important clinical implications, since the risk of misclassification as type 2 diabetes is higher in these populations.

Fig. 1.

 Prevalence and incidence of type 1 diabetes per age group across the globe. The map shows the prevalence of type 1 diabetes per 100,000 people in each country (98), as well as the incidence per age group in each continent in 2017 (31). This figure is available as part of a downloadable slideset

Type 1 diabetes incidence

Over the past several decades, the incidence of type 1 diabetes has been increasing globally, particularly in countries with high baseline rates. Diabetes Atlas reports that type 1 diabetes incidence has been rising by 3–4% annually, notably in younger children (19). Population-specific trends in incidence may be due to differences in diagnostic capabilities plus true changes in the disease burden, likely driven by changing environmental exposures and their interaction with genetics.

While White European populations tend to be at greater risk for type 1 diabetes than other populations (20), with a prevalence of around 4 in 1000, differences within European-ancestry populations have been described. According to Diabetes Atlas (19), Scandinavian countries and Gulf states (Kuwait, Saudi Arabia and Qatar) have reported the highest incidence of type 1 diabetes worldwide. In Europe, Finland is the country with the highest annual incidence of approximately 60 per 100,000 children (21). Other Scandinavian countries also report high incidence rates, with Sweden having an incidence of around 40–50 per 100,000 children (22). In the USA, the incidence of type 1 diabetes is estimated to be around 24 per 100,000 children annually (23) but this varies between different states and communities. The number of youth in the USA with type 1 diabetes is projected to increase in the next 30 years by 2.5-fold in non-Hispanic White populations and up to 6.6-fold in Hispanic populations (24). Canada has incidence rates comparable with those of the USA (25). The incidence of type 1 diabetes in Kuwait is one of the highest in the world, and is reported to have doubled in the past 30 years (26).

In East Asia, type 1 diabetes is less common but its rates are increasing. Japan has an incidence rate of around 4–5 per 100,000 children (27). In China, the incidence rate of type 1 diabetes is rising but it remains lower than in Western countries (28). In Australia, the incidence rate of type 1 diabetes in children is around 30 per 100,000 annually, with a slightly higher rate in the state of New South Wales (29). The incidence of type 1 diabetes is lower in low- and middle-income countries but it is on the rise, especially in urban areas, possibly due to changes in lifestyle, diet and healthcare infrastructure (23).

Type 1 diabetes clinical presentation

While type 1 diabetes can present at any age and its incidence is increasing across all ancestries and age groups, the age at peak incidence varies across populations (Fig. 1) (30,31). In European populations, ~40% of incident cases manifest in childhood or adolescence (30), with a peak incidence at around 10–14 years of age (32). The higher cumulative incidence of type 1 diabetes in adults (~60% of incident cases per year) can be due to the longer at-risk time in adulthood, even if incidence peaks in childhood. Within individuals of European ancestry, the age of onset can be influenced by genetic factors, with higher type 1 diabetes polygenic score predicting younger age at onset and heterozygosity of certain HLA-DR/DQ and class I genes associated with younger ages at diagnosis (33,34). However, in non-European populations, there is evidence of a later age of onset and more atypical presentations. In China, a tendency for onset in adulthood has been described (35), with 65% of incident cases described in people older than 20 years old, although the prevalence of childhood-onset disease is lower than in Europeans (28). Non-Hispanic Black children in the USA are older at diagnosis than non-Hispanic White Children (36).

The presentation of type 1 diabetes varies across populations not only in terms of age of onset but also in disease presentation. Black and Hispanic children in the USA have higher BMI and are more likely to present with diabetic ketoacidosis than non-Hispanic White children (36). This can be explained by differences in healthcare access, socioeconomic status and pathophysiology (insulin resistance). The rates of latent autoimmune diabetes in adults (LADA), an autoantibody-positive type 1 diabetes that presents more insidiously in adulthood and can be often misdiagnosed and initially treated as type 2 diabetes, vary substantially across ancestries (37). In Europeans, the rates of LADA vary among 2–12% of all cases of diabetes (38). Whether LADA is a true separate entity or a misclassification of type 1 diabetes remains debatable (39) and its genetic similarities and differences from type 1 diabetes have been explored in a single small European GWAS (40). In Asian and Hispanic populations, a higher prevalence of LADA has been reported (37). In China, Japan and India, atypical presentations are frequent, and known as fulminant type 1 diabetes. This is characterised by absence of autoantibodies at diagnosis with possible later seroconversion, acute onset and intense insulitis (41) also affecting the exocrine pancreas (42). Autoantibody-negative type 1 diabetes wherein the affected individual is prone to ketosis and diabetic ketoacidosis (referred to as ketosis-prone diabetes), was first described in individuals of African or Asian ancestry and has distinct features such as intermittent insulin dependence and no HLA associations (43). Recently, a form of autoantibody-negative youth-onset type 1 diabetes was described in sub-Saharan Africa, involving individuals with clinically diagnosed type 1 diabetes (young-onset, insulin-treated) but low type 1 diabetes genetic risk, suggesting a non-autoimmune aetiology (44). These results imply that type 1 diabetes polygenic scores derived from HLA associations in European populations can still identify individuals with true autoimmunity vs those with a novel non-autoimmune form in non-European populations.

Notably, in the majority of GWAS in both European and non-European populations, from which these polygenic scores are developed, autoantibody status is not confirmed. This is very relevant to large population biobanks that do not always have biomarker-validated definitions of type 1 diabetes, imposing trade-offs (sample size vs accuracy of type 1 diabetes diagnosis), which can affect the accuracy of the polygenic scores developed using their data.

To accommodate this shortfall, it is critical to prioritise the development of well-characterised type 1 diabetes case cohorts across all ethnicities whose disease onset, progression and autoantibody status are systematically documented and cannot be confused with type 2 diabetes. Such efforts are essential for refining polygenic risk scores, ensuring equitable risk prediction and ultimately reducing disparities in type 1 diabetes diagnosis and management.

Genetic variability and environmental influences

One of the major contributors to the heterogeneity in incidence and presentation of type 1 diabetes is genetic diversity. Notably, HLA-DR3 and HLA-DR4 have been associated with an increased risk of type 1 diabetes but the frequency of these alleles varies significantly between populations (17). European populations, including those of Scandinavian and northern European descent, have a higher frequency of HLA-DR3/DR4 haplotypes compared with Asian and African populations (17). Middle Eastern populations exhibit different risk alleles and gene–environment interactions that may influence the development of type 1 diabetes (45). Furthermore, certain non-HLA genes, such as those involved in immune response regulation, also contribute to disease heterogeneity. Genetic variants influencing immune tolerance and beta cell survival also vary across populations, further underscoring the complexity of type 1 diabetes pathogenesis (34).

Beyond genetics, environmental factors play a crucial role in the heterogeneity in the pathogenesis and presentation of type 1 diabetes across populations. Since genetic drift (i.e. the change in allele frequencies over generations) is a slow evolutionary phenomenon, gene–environment interactions might explain population-specific trends in type 1 diabetes incidence, given a rapidly changing environment worldwide. Environmental factors contributing to type 1 diabetes pathogenesis include viral infections, dietary habits (affecting the gut microbiome (46)) and socioeconomic conditions, each of which differs widely between populations (47). The hygiene hypothesis, which suggests that reduced childhood infections lead to a higher incidence of autoimmune diseases, may also play a role in geographical differences in type 1 diabetes incidence and presentation (48). Enteroviral infections are more commonly associated with type 1 diabetes in high-income countries compared with low-income countries (49). In rural North-West Ethiopia, compared with industrialised regions, type 1 diabetes presents with phenotypic differences, including single GADA positivity, later presentation and specific HLA haplotypes conferring disease risk or protection (50). Interestingly, fulminant diabetes described in Asian populations is often connected with acute viral infection and exhibits unique genetic characteristics such as association with HLA-DRB1*04:05-DQB1*04:01 (51). Integrating ancestry-specific HLA haplotypes into polygenic scores can ensure representation of atypical forms of type 1 diabetes that might reflect environmental exposures. Furthermore, genetically similar populations can have different type 1 diabetes incidence in different geographical locations (52–54), further highlighting the gene–environment interplay in type 1 diabetes pathophysiology.

In summary, the interplay between genetics and environment remains a critical area of research to better understand ancestry-specific differences in the aetiology of type 1 diabetes. Environmental triggers of type 1 diabetes vary across the globe and may explain differences in disease incidence and phenotype. This directly impacts the utility of polygenic scores from diverse ancestries, and validation and calibration in local populations is needed. Ideally, well-phenotyped individuals representing the full diversity of ancestries and geographies are needed to generate diversity-informed polygenic scores, and not simply multi-ethnic samples from a limited geographic region.

Importance of MHC region in type 1 diabetes genetic risk

Genetic predisposition to type 1 diabetes is strongly influenced by variation in HLA genes in the MHC on chromosome 6, which encodes proteins responsible for antigen presentation. HLA class II alleles, particularly within the HLA-DRB1-DQA1-DQB1 locus, account for a large proportion of type 1 diabetes susceptibility and represent some of the strongest known genetic risk factors, a finding established as early as the 1970s (55). Classical HLA alleles explain approximately half of the total genetic risk for type 1 diabetes in many populations, exceeding the contribution of individual non-HLA loci (56); in most other common complex diseases, risk is typically distributed across many loci of small effect. These observations highlight the importance of accurately capturing HLA variation when developing and applying polygenic scores, especially in ancestrally diverse populations.

Directly measuring HLA alleles in large studies is expensive and time-consuming; thus, most studies rely on imputation methods that predict HLA alleles from nearby SNPs. Early work in type 1 diabetes demonstrated that specific SNPs and small SNP panels could effectively tag high-risk HLA haplotypes, enabling scalable genetic risk assessment without direct HLA typing (57,58). These studies established the foundation for subsequent HLA imputation approaches by showing that local linkage disequilibrium patterns could be leveraged to infer classical HLA variation. In simple terms, certain SNPs tend to be inherited together with specific HLA alleles because they are located close to each other on the chromosome. As a result, the presence of a particular SNP can act as a proxy, or ‘tag’, for an underlying HLA allele even if that allele is not directly measured. One widely used approach is implemented in the HLA-TAPAS pipeline (11,59), which uses high-resolution, multi-ancestry reference panels to impute HLA alleles and amino acid polymorphisms across populations. The HLA-TAPAS panel is available through the Michigan Imputation Server (https://imputationserver.sph.umich.edu/index.html#!), enabling imputation at four-digit resolution using reference haplotypes from tens of thousands of individuals with whole genome sequencing (WGS), spanning five global populations. Accurate HLA imputation is critical for developing and evaluating trans-ancestry polygenic scores (60). Type 1 diabetes risk alleles and haplotypes differ in frequency and linkage disequilibrium (LD) patterns across populations (9). In other words, the correlation between a SNP and the HLA allele it tags can vary between populations, meaning that SNPs that reliably predict HLA genotypes in one ancestry may perform poorly in another. Multi-ancestry reference panels and imputation frameworks reduce bias inherent in single-ancestry resources, improving comparability of HLA-inclusive scores across cohorts.

In practice, the choice of HLA imputation approach depends on the type of genetic data available. For genotype array data, commonly used tools such as HLA-TAPAS and SNP2HLA leverage reference panels to impute classical alleles and amino acid polymorphisms from dense SNP information. For WGS data, HLA alleles can be inferred either through imputation-based frameworks or by using dedicated HLA typing algorithms, such as HLA*LA (61) or HLA-HD (62), that directly analyse sequence reads. Large-scale resources such as the UK Biobank and All of Us Research Program provide access to imputed or sequence-based genetic data, although accurate HLA calling from WGS remains computationally intensive and may require specialised pipelines and quality-control procedures.

While methods such as HLA-TAPAS leverage dense reference panels and LD structure, newer machine-learning approaches address challenges in trans-ancestry HLA imputation. For example, DEEP*HLA, a convolutional deep learning model, has been shown to improve imputation accuracy, particularly for low-frequency and rare alleles that are difficult to resolve with traditional statistical models (63). Reported gains are generally modest for common alleles but more substantial for rare variants and in cross-ancestry settings, where complex or long-range haplotypic patterns are not well captured by standard models. However, despite these improvements, overall accuracy remains highly dependent on the size and ancestry composition of the reference panel. Deep learning methods can capture complex, non-linear relationships in SNP data beyond local LD structure and may enhance performance across diverse populations. Nevertheless, these advanced models are often computationally intensive and require greater technical expertise, introducing a trade-off between incremental gains in imputation accuracy and ease of implementation. For large-scale or clinical applications, this trade-off is particularly relevant when considering reproducibility and regulatory requirements (e.g. CLIA certification).

A key factor in applying polygenic risk scores is the type of genetic data available. In practice, there is a trade-off between smaller SNP panels and more comprehensive genotyping or sequencing approaches. Targeted SNP panels, including those used in some type 1 diabetes genetic risk scores, are cost-efficient and easy to implement but rely on a limited set of tagging variants and may not fully capture HLA diversity, particularly across ancestries. In contrast, genome-wide arrays or WGS enable HLA imputation using dense SNP data, which can improve the accuracy and resolution of inferred HLA alleles. However, more comprehensive approaches come at higher cost, and the incremental benefit compared with simpler genotype-based strategies remains unclear (64). Therefore, selecting an appropriate genotyping and imputation strategy is critical for achieving robust and equitable polygenic score performance. To provide an overview of available approaches, the main types of genetic data used for HLA typing and polygenic risk scoring are summarised in Table 1.

Table 1.

Overview of genetic data types for HLA typing and polygenic scoring

Data type	Approach	Resolution	Advantages	Limitations	Example software packages
Serological HLA typing	Antibody-based	Low	Historically important Simple	Low resolution	NA
Targeted HLA typing (PCR/probe-based)	Direct typing of HLA loci	High	Accurate allele-level calls	Costly Limited scalability
Genotype arrays	Analyse common SNPs + perform imputation	Moderate–high	Cost-efficient Scalable Widely available	Poor coverage of highly complex genetic regions Limited accuracy in diverse ancestries Depends on reference panels	HLA-TAPAS DEEP*HLA (11,63)
Custom SNP panels	Select SNPs that tag HLA alleles + perform imputation	Low–moderate	Very cost-effective Easy to implement	Limited resolution May miss rare or ancestry-specific variants
Whole exome sequencing (WES)	Sequence-based + infer HLA alleles	Moderate	Captures coding variation	Incomplete HLA coverage	HLA*LA HLA-HD (61,62)
WGS	Sequence-based + infer HLA alleles	High	Comprehensive Highest resolution	Expensive Computationally intensive

These considerations are reflected in how type 1 diabetes polygenic scores incorporate HLA variation. For example, type 1 diabetes GRS2 includes 67 SNPs that capture both HLA risk and protective haplotypes and has shown strong performance in distinguishing type 1 diabetes from other forms of diabetes in European populations (1). This approach relies on tag SNPs and its performance may be reduced in non-European populations unless ancestry-specific adjustments or additional HLA imputation are applied.

Accurately capturing HLA variation is essential for developing reliable type 1 diabetes polygenic scores across diverse populations. The use of multi-ancestry HLA reference panels and newer imputation methods represents important progress toward more equitable risk prediction, which means improving the transferability and performance of genetic risk scores across individuals of different ancestries, thereby reducing biases introduced by historically Eurocentric datasets. However, most multi-ancestry polygenic score studies have been conducted in high-income settings and important gaps remain in their evaluation and implementation in low- and middle-income countries. In this broader context, it is important to account for differences in cost, infrastructure and clinical feasibility that may limit the implementation of genetic risk prediction in resource-limited settings. At the same time, achieving an appropriate balance between accuracy, computational demands and clinical feasibility remains a key challenge for the broader implementation of trans-ancestry type 1 diabetes risk scores.

Development of type 1 diabetes polygenic scores

The development of early polygenic scores for type 1 diabetes originated from GWAS conducted primarily in populations of European ancestry. These foundational efforts relied on cohorts such as the Type 1 Diabetes Genetics Consortium (T1DGC (65)) and the Wellcome Trust Case Control Consortium (66). Early models were constrained by the limited size of these cohorts and typically incorporated the main HLA haplotypes alongside a limited set of non-HLA variants. Due to the strong influence of HLA predictors, these early scores (3,67,68) demonstrated an impressive predictive performance in European populations, as quantified by ROC AUC (Text box). With a ROC AUC above 0.85, they often exceeded the predictive performance achieved by polygenic scores for many non-autoimmune diseases, which tend to be between 0.6 and 0.8 (69). In 2019, type 1 diabetes GRS2 was published (1). Compared with previous scores, GRS2 accounts for interactions between HLA haplotypes. This new approach has better performance than other polygenic scores, with a ROC AUC of 0.92 when validated in an external dataset.

When GRS2 is applied to non-European populations, however, the discriminative performance declines substantially. Despite this reduction, it is important to note that the performance of GRS2 in non-European ancestry groups remains higher than the performance observed for most other diseases. GRS2 performed similarly in Hispanic, White and Black individuals in the SEARCH for Diabetes in Youth population-based cohort (70). HLA-DR3/DR4 status is the major discriminator of type 1 diabetes risk but discrepancies in risk allele frequency and ORs can decrease the performance of GRS2, as seen in the South Asian population (ROC AUC 0.83–0.84) (71,72). The loss of performance is particularly pronounced in populations of East Asian ancestry (ROC AUC 0.77 (73)), reflecting differences in genetic diversity, allele frequencies and LD structure relative to European cohorts.

Moreover, the distribution of raw polygenic scores differs markedly across ancestries (70,74,75). This can influence other aspects of score performance such as calibration, which is not captured by ROC AUC. For example, the distribution of GRS2 values tends to be lower in the African population compared with white Europeans and admixed Americans. Consequently, a given score threshold corresponds to different sensitivity and specificity profiles in different populations (Fig. 2a). Furthermore, a raw score value that confers high absolute disease risk in one population may correspond to a lower disease risk in another population, which can lead to miscalibration (Fig. 2b). Calibration can adjust polygenic score models to ensure that the predicted disease risk is similar across ancestries (Fig. 2c, d). More sophisticated tools can calibrate scores using multiple covariates, such as age and sex (76). However, this process may suppress true differences between populations. For instance, lower GRS2 values in African populations may reflect lower frequency of high-risk HLA haplotypes, thereby leading to lower overall type 1 diabetes risk.

Fig. 2.

 Discrimination and calibration of GRS2. (a) Discrimination of GRS2 in MGB Biobank. Each density plot shows the distribution of type 1 diabetes polygenic scores within African or European ancestry, separated into type 1 diabetes cases (blue) and controls (red). The dotted line corresponds to the 90th percentile in the European population. Using the dotted line as a fixed threshold to classify cases and controls, the sensitivity of GRS2 is higher in European ancestry, whereas the specificity is higher in African ancestry. (b) Calibration of GRS2 in MGB Biobank. The scatter plot displays the observed type 1 diabetes risk in MGB Biobank (with 95% CIs) vs the predicted type 1 diabetes risk for each quintile of GRS2, separated by African ancestry (green circles) or European ancestry (red triangles). The dashed line represents the line of equality. The model is well-calibrated in European ancestry but the observed type 1 diabetes risk is higher than the predicted risk in African ancestry. To build the calibration plot, a logistic regression model was used to predict type 1 diabetes risk at the median value of each GRS2 quintile. The model was trained on individuals with European ancestry in MGB Biobank. (c, d) GRS2 scores were calibrated by adding 1.5 points to individuals with African ancestry. After calibration, the sensitivity and specificity are similar across populations (c), and the observed type 1 diabetes risk is closer to the predicted risk in both African and European ancestry (d). AFR, African; EUR, European; T1D, type 1 diabetes. This figure is available as part of a downloadable slideset

The potential for miscalibration is especially problematic for screening and risk stratification, where consistent clinical decision thresholds are required. Two initial approaches have partially addressed the challenge of the limited transferability of polygenic scores derived from European populations. The first approach uses population-specific models, prompting the development of ancestry-tailored scores for African (77) and East Asian populations (73). While these ancestry-specific models improve predictive accuracy within their target populations, they also expose a critical limitation: population-specific polygenic scores have reduced applicability in admixed populations, where individual genetic ancestry is heterogeneous. This challenge is particularly important in increasingly diverse societies, where admixed individuals represent a growing proportion of the population. Another successful approach consists of modifying GRS2 by including SNPs that are specific to a population (73,78) or are common across populations (2). However, the improved performance in multi-ancestry settings can come at a cost for European populations (2) (ROC AUC 0.895 vs 0.92 for GRS2).

Novel type 1 diabetes polygenic scores

Recently, three new polygenic scores were developed that seek to improve type 1 diabetes risk prediction across populations (60,75,79) (Table 2). All three scores use variant weights from recent multi-ancestry GWAS to ensure broad transferability across diverse populations (Fig. 3). However, the three novel polygenic scores use different methods to model genetic risk. GRS_HLA focuses on genetic risk conferred by approximately 40 HLA alleles or SNPs. In contrast, T1D MAPS and TA-PS model genetic risk conferred across the genome, including HLA and non-HLA alleles. For the HLA region, TA-PS uses the same approach as GRS2x (accounting for interactions between 14 HLA-DR-DQ haplotypes), whereas T1D MAPS uses 38 HLA-DRB1-DQA1-DQB1 haplotypes and does not account for interactions. For non-HLA alleles, both TA-PS and T1D MAPS model the risk conferred by over one million variants across the genome, including variants that are not significantly associated with type 1 diabetes. TA-PS uses PRS-CSx (13) to integrate GWAS summary statistics from diverse populations, but there are many approaches to combine ancestry-specific summary statistics (14).

Table 2.

Comparison of GRS2 vs novel multi-ancestry polygenic scores

Cohort description	GRS2 (1)	GRSHLA (60)a	MAPS (79)	TA-PS (75)
Development cohort	T1DGC	T1DGC	MGB Biobank	Montreal T1D Cohort
No. of T1D cases	6670	16,104 14,491 EUR 1613 Non-EUR	372 335 EUR 37 Non-EUR	2416 2193 EUR 223 Non-EUR
No. of controls	9416	25,262 21,914 EUR 3348 Non-EUR	60,319 53,487 EUR 6832 Non-EUR	2241 1908 EUR 333 Non-EUR
Populations included	EUR	AFR, AMR, EUR, FIN	AFR, AMR, EUR	AFR, AMR, EAS, EUR, SAS
Maximum age at T1D diagnosis	16 years	No maximum age	No maximum age	No maximum age, but the majority are childhood-onset
Testing cohort	UKBB	T1DGC–WGS	All of Us	Multipleb
No. of T1D cases	387	510 211 EUR 299 Non-EUR	86 61 EUR 25 Non-EUR	3230 1791 EUR 1439 Non-EUR
No. of controls	11,885	6342 2580 EUR 3762 Non-EUR	183,250 107,380 EUR 75,870 Non-EUR	521,411 459,710 EUR 61,701 Non-EUR
Populations included	EUR	AFR, AMR, EAS, EUR	AFR, AMR, EUR	AFR, AMR, EAS, EUR
Maximum age at T1D diagnosis	20 years	No maximum age	30 years	25 years (UKBB only)
HLA score
Methodology	HLA DR-DQ haplotypes, including interaction terms	Conditionally independent HLA SNPs and HLA alleles	HLA DR-DQ haplotypes	HLA DR-DQ haplotypes, including interaction terms (same as GRS2)
No. of variants	14 HLA DR-DQ SNPs with interaction terms plus 21 other HLA variants	5 AFR 3 AMR 38 EUR 6 FIN 36 All	38 HLA DR-DQ haplotypes	14 HLA DR-DQ SNPs with interaction terms plus 21 other HLA variants
Non-HLA score
Source GWAS	T1DGC (65)	Robertson et al (7)	Chiou et al (6)	Chiou et al (6) Sakaue et al (10) Verma et al (8) Michalek et al (9)
Methodology	Candidate genes independently associated with T1D	Conditionally independent SNPs	Construct genome-wide polygenic score with PRS-CS (12)	Construct polygenic score across populations with PRS-CSx (13)
No. of variants	32	78	1,067,951	1,116,345
Overall score
Methodology	Unweighted combination of HLA + non-HLA score	Conditionally independent HLA + non-HLA SNPs	Weighted combination of HLA + non-HLA score	Weighted combination of HLA DR-DQ haplotypes, their interactions and the non-HLA score
ROC AUC in development cohort	EUR: 0.93	AFR: 0.88 AMR: 0.85 EUR: 0.91 FIN: 0.84	EUR: 0.88 Non-EUR: 0.89	EUR: 0.9 Non-EUR: 0.87 Multi-ancestry: 0.89
ROC AUC in testing cohort	EUR: 0.92	Multi-ancestry: 0.81	EUR: 0.89 Non-EUR: 0.90	N/A

^aData for GRS_HLA refers to GRS_HLA score derived from all ancestries based on HLA SNPs

^bTesting cohorts for TA-PS: All of Us; UKBB; CHOP-CAG (Children’s Hospital of Philadelphia Center for Applied Genomics); GRACE (Genetic Risk Assessment for Chinese Eaglet-T1D)

AFR, African; AMR, admixed American; EAS, East Asian; EUR, European; FIN, Finnish; MGB Biobank, Mass General Brigham Biobank; SAS, South Asian; T1D, type 1 diabetes; UKBB, UK Biobank

Fig. 3.

 Overview of the development and testing of multi-ancestry type 1 diabetes polygenic scores. This figure outlines steps to develop and validate type 1 diabetes polygenic scores across multiple ancestries, following reporting standards for polygenic scores (99). T1D, type 1 diabetes. This figure is available as part of a downloadable slideset

The three novel scores demonstrate high discriminatory power across ancestry groups. For example, the ROC AUC was 0.81–0.91 for GRS_HLA, 0.88–0.90 for T1D MAPS, and 0.87–0.90 for TA-PS. For each score, applying a single threshold to discriminate between cases and controls yields reasonably high specificity and sensitivity across populations (ESM Fig. 1). Notably, within European ancestry, each novel score has similar or better performance compared with GRS2, demonstrating that multi-ancestry approaches can improve score performance in diverse settings without reducing performance in more homogeneous cohorts. Additional work should replicate these findings in external datasets, including cohorts representing diverse populations. Future studies should also assess whether these scores can accurately predict absolute disease risk in all populations without further ancestry-specific calibration.

Although multi-ancestry scores offer clear advantages in diverse settings, there are some trade-offs compared with ancestry-specific scores. A given variant may confer disease risk in one population and disease protection in another population; however, a multi-ancestry score would not account for these varying effects. One potential solution is to use different weights for each variant depending on the genetic ancestry of the target individual. However, such approaches would be computationally intensive and the resulting scores could be challenging to interpret.

Other features of multi-ancestry polygenic scores may complicate implementation. For instance, GRS2x uses variants from commonly used reference panels (e.g. 1000 Genomes). In contrast, GRS_HLA, TA-PS and T1D MAPS require dedicated imputation of the MHC region using an HLA-specific reference panel, such as HLA-TAPAS (11). Genotype imputation is feasible in a research setting, especially if imputed data are already available in large biobanks, but it may pose issues in a clinical setting. Likewise, it is more expensive to genotype over one million variants for T1D MAPS and TA-PS, compared with only 67 variants for GRS2x. Nevertheless, genotyping costs have dramatically declined over recent years and genotype arrays are becoming more common in clinical use.

Future avenues: how can we further optimise trans-ancestry type 1 diabetes prediction?

Although recent trans-ancestry type 1 diabetes polygenic scores have improved risk stratification in non-European populations, substantial biological and methodological limitations remain. Further gains will require deeper integration of functional genomics, intermediate molecular phenotypes and clinical heterogeneity.

Functionally informed polygenic scores

One limitation of conventional polygenic scores is their reliance on marginal GWAS effect estimates without prioritising causal variants. While the function of HLA alleles is well-known, SNPs outside the HLA region may tag different functional variants or exhibit allele flips across populations (14). Incorporating fine-mapped credible sets and tissue-specific regulatory annotations can therefore improve biological resolution and trans-ancestry portability. Annotation-informed models such as LDpred-funct enhance performance by prioritising regulatory variants (80). OmniPRS demonstrates that incorporating functional prior probabilities into Bayesian models can improve polygenic prediction across traits (81).

Integration of pancreatic islet regulatory landscapes and immune-cell chromatin accessibility is particularly relevant. Sex-stratified GWAS and sex-specific polygenic scores in European-ancestry datasets suggest that sex-related differences in type 1 diabetes biology may modestly improve prediction beyond pooled models (82). At present, sex-specific and regulatory annotation-informed approaches are not routinely incorporated into clinical type 1 diabetes polygenic scores and should be viewed as promising research directions rather than established implementation-ready models. Pathway-specific polygenic scores aggregating genetic risk within immune and beta cell regulatory pathways may further enhance interpretability. Such biologically stratified polygenic scores also enable mechanistic disease subtyping, distinguishing immune-dominant vs beta cell-vulnerability-driven pathways of type 1 diabetes risk, with direct implications for precision prevention and trial stratification.

Moving from genetic risk to multi-omics integrated risk prediction

Inherited genetic variation captures only part of type 1 diabetes liability. Integrating polygenic scores with proteomic, metabolomic and transcriptomic profiles improves disease gene prioritisation and prediction across complex traits (83). For type 1 diabetes, inflammatory proteins, beta cell stress peptides, immune activation markers and metabolomic signatures of dysglycaemia provide dynamic intermediate layers that enable longitudinal risk assessment. Systems-level approaches such as OmicKriging anticipated this framework (84), and modern high-throughput omics now make such strategies feasible. Effective integration will likely require artificial intelligence-based representation learning (multimodal data integration models) to jointly embed genetic and molecular variation.

Meta-PRS

Meta-PRS combines GWAS of biologically proximal endophenotypes (e.g. C-peptide, islet autoantibody phenotypes, immune-cell composition and T cell activation states). These intermediate traits may capture components of disease biology that are obscured in case–control GWAS alone. Integrative polygenic score frameworks have outperformed disease-only scores in other complex traits (85), and novel methods can fine-tune polygenic scores using GWAS summary statistics (86). In type 1 diabetes, this approach remains at an early stage and will require better-powered ancestry-diverse studies of these endophenotypes before it can be translated into routine risk prediction. Integrating biomarkers such as C-peptide and/or autoantibodies in discovery GWAS studies might be particularly relevant in non-European ancestry populations where misclassification is more common. Establishing an optimal representation of larger sample size with simple definitions and smaller samples with stricter, ancestry-informed definitions can decrease the noise in the discovery GWAS.

Cross-age optimisation and life-course modelling

Age at onset varies across ancestries and reflects differences in immune kinetics, beta cell reserve and exposures. Joint training across age strata may improve prediction in populations with higher burdens of later-onset disease. Life-course models incorporating time-varying biomarkers further refine individualised risk trajectories. Autoantibody status defines presymptomatic staging and progression risk (87), while dysglycaemia and C-peptide decline refine short-term risk (88).

Gene–environment interaction, family history and global data sharing

Environmental modifiers of genetic risk vary across populations. Large-scale gene–environment interaction studies may explain ancestry-related polygenic score heterogeneity. Additionally, because of the important environmental differences in populations across the globe and the gene–environment interactions in type 1 diabetes aetiology, it is crucial to calibrate the newly developed polygenic score in locally collected samples from diverse ancestries and not just in multi-ethnic samples collected in the same geographical region.

Family history remains a strong non-genomic indicator of type 1 diabetes risk (89), and integrated risk models that combine polygenic scores with age, sex, BMI or other metabolic measures, family history, and especially islet autoantibodies are likely to outperform single-component approaches (90,91). When family history is incompletely recorded, risk models can treat it as a partially observed variable and estimate its contribution using available pedigree structure or correlated clinical information. Continued optimisation depends on international collaboration and harmonised phenotyping. The Global Biobank Meta-analysis Initiative illustrates the value of coordinated multi-ancestry GWAS (92).

Admixed ancestry and local ancestry-aware modelling

Admixed populations pose major challenges due to heterogeneous LD, particularly in the HLA region. Local ancestry-aware models that apply ancestry-matched effect size improve prediction in admixed individuals (93). Improved imputation and ancestry-weighted association methods also enhance performance (94). Future models must be trained and validated directly in admixed populations.

Machine learning for non-linear biological complexity

Machine learning methods enable modelling of non-linear genetic architectures, epistasis and ancestry-dependent effects but risk overfitting without biological constraints. In type 1 diabetes, where risk reflects interactions among HLA alleles, immune loci, beta cell networks, age, sex and environment, biologically constrained machine learning architectures, such as pathway-aware neural networks and ancestry-aware models, offer improved performance and interpretability. Ensemble-based models have shown consistent gains across traits (95).

Conclusion

Novel methods have enhanced the ability of polygenic scores to predict type 1 diabetes risk across diverse populations. Recent scores still show somewhat lower predictive power in many non-European populations than in European populations. Closing this gap will require more inclusive and flexible approaches to score development and more explicit recognition that type 1 diabetes liability is shaped by age, environmental exposures and their interaction with genetic susceptibility. Heritability differs between childhood-onset and adult-onset disease, with a larger environmental contribution in adult-onset type 1 diabetes (96). Polygenic scores can also contribute to risk stratification before clinical disease, including prediction of progression from islet autoimmunity to overt type 1 diabetes in at-risk individuals (97). These observations support population-specific calibration and careful interpretation of genetic risk estimates.

Ultimately, clinical translation is likely to require large, well-phenotyped datasets with broader ancestral representation, shared standards for score development and external validation, improved calibration across populations, and implementation studies in clinical settings. Trans-ancestry polygenic scores should be regarded as tools that require local validation and population- and sex-specific calibration. In the near term, progress is likely to depend as much on stronger cohorts, more consistent phenotyping and rigorous evaluation as on further increases in model complexity. The clinical and translational readership of Diabetologia is well positioned to contribute to the collaborative phenotyping, cohort harmonisation and prospective validation efforts needed to move type 1 diabetes polygenic scores across diverse populations towards equitable clinical use.

Assessment of polygenic score performance.

Polygenic scores should be validated in external testing cohorts that are independent of the training cohorts used to develop the score

The following measures can be used to assess score performance, in the presence or absence of covariates (e.g. age, sex, clinical biomarkers)

Sensitivity

What proportion of type 1 diabetes cases are correctly identified as cases?

Specificity

What proportion of controls are correctly identified as controls?

Positive predictive value

What proportion of predicted cases truly have type 1 diabetes?

Negative predictive value

What proportion of predicted controls truly do not have type 1 diabetes?

Discrimination

How well does the polygenic score separate type 1 diabetes cases from controls?

This can be quantified with a ROC AUC, which compares the sensitivity and specificity at various values of the polygenic score

ROC AUC values range from 0.5 to 1, where 0.5 indicates random chance (no discrimination) and 1 indicates perfect discrimination

Calibration

How well does the predicted disease risk generated from the model align with the true disease risk observed in the population?

Abbreviations

GWAS

Genome-wide association study

LADA

Latent autoimmune diabetes in adults

LD

Linkage disequilibrium

ROC AUC

Area under the receiver operating characteristic curve

T1DGC

Type 1 Diabetes Genetics Consortium

WGS

Whole genome sequencing