Most polygenic scores (PS) for type 1 diabetes were developed using European (EUR) ancestry datasets, limiting performance in non-European (non-EUR) populations. We evaluated novel type 1 diabetes PS across diverse populations to assess whether newer models improve prediction in underrepresented populations. Seven type 1 diabetes PS (T1D GRS2, T1D GRS2′, T1D GRS HLA , T1GRS, TA-PS, TA-PS (S), T1D MAPS) were evaluated in All of Us and Mass General Brigham Biobank. Predictive performance was assessed using the area under the receiver operating characteristic curve (AUC). T1D GRS2 and T1D MAPS showed the highest AUC in All of Us, while other scores demonstrated comparatively lower performance. These highest-performing scores were then evaluated in a meta-analysis across three additional biobanks (Genomic Health Initiative at Endeavor Health, Penn Medicine BioBank, Geisinger MyCode). Among 2,782 individuals with type 1 diabetes and 546,577 controls, T1D MAPS showed comparable performance to T1D GRS2 in EUR populations but significantly improved discrimination in non-EUR populations (meta-analysis ΔAUC=0.049, p =1.7×10 -7 ). Overall, T1D MAPS improves prediction in non-EUR populations, highlighting the importance of multi-ancestry approaches for equitable genetic risk prediction. These findings could inform precision medicine in diverse populations.

Introduction and Objective: Type 1 diabetes (T1D) is characterized by autoimmune destruction of pancreatic beta cells. However, the clinical presentation of T1D is heterogeneous, with variation in age of onset, rate of beta-cell loss, and predisposition to other autoimmune conditions. In this study, we generated T1D genetic clusters and analyzed whether these clusters could identify distinct mechanisms in T1D pathogenesis. Methods: We assembled summary statistics from recent genome-wide association studies for T1D and related traits, including glycemic indices, cardiometabolic phenotypes, and immune cell markers. We applied Bayesian nonnegative matrix factorization (bNMF) to generate soft clusters that link T1D loci with associated traits. We calculated cluster-specific partitioned polygenic scores (pPS) in the Mass General Brigham Biobank, and we used regression models to analyze the association between pPS and various clinical phenotypes. Results: After removing highly correlated loci and traits, we analyzed 205 variants associated with T1D at genome-wide significance (including 79 major histocompatibility complex [MHC] variants) and 91 T1D-related traits. We identified eight clusters, which were distinguished by variation in immune cell markers and clinical phenotypes. For example, one cluster was associated with HLA-DRB1*03:01-DQA1*05:01-DQB1*02:01, was characterized by decreased lymphocyte count, and conferred increased risk of celiac disease (odds ratio [OR] 1.46 per standard deviation of pPS, P = 1.2x10-33). Another cluster was associated with multiple HLA variants, was characterized by increased expression of CD20, and conferred decreased risk of celiac disease (OR 0.75, P = 1.4x10-19). Outside of the MHC, two clusters were associated with altered measures of beta-cell function, such as HOMA-B; one cluster was linked to variation near GLIS3, while the other was linked to INS. Conclusion: Genetic clustering methods provide insight into the heterogeneity of T1D. Further studies should validate the clinical implications of these genetic clusters. Disclosure A.J. Deutsch: None. K. Smith: None. J.C. Florez: Research Support; Current; Novo Nordisk. Consultant; Current; Alveus Therapeutics. M. Udler: Advisory Panel; Ended; Novo Nordisk. Research Support; Current; Novo Nordisk. Funding NIH/NIDDK (K23 DK140643)

Precision medicine for complex diseases uses individual-level characteristics to improve prediction of risk, therapeutic response and prognosis. Many precision medicine studies leverage existing data types and analytic methods to reveal new insights; however, beyond oncology, there has been limited success in translating precision medicine research for complex diseases into clinical practice. Thus, there is a need to identify areas for improvement, particularly in translation-oriented analytical methods and study designs. In this perspective article, we outline five fundamental tenets to enhance the efficient clinical translation of precision medicine research. These tenets focus on addressing (1) heterogeneity in risk, response and prognosis; (2) signal robustness; (3) structured statistical benchmarking against key performance indicators; (4) precision trial designs; and (5) risks and benefits to individuals and society. Our intention is to promote clinically meaningful, reproducible, scalable and equitable health outcomes through precision medicine, beyond those possible through contemporary approaches.

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.

Objective: Clinical heterogeneity in youth-onset type 2 diabetes is less understood than that of adult-onset type 2 diabetes. We performed phenotypic clustering of youth-onset type 2 diabetes to determine if clusters provided clinical utility. Research Design and Methods: We performed data-driven clustering in a diverse subset of autoantibody-negative, clinician-diagnosed type 2 diabetes before age 20 years in the TODAY [n=525] and SEARCH [n=333] studies. Participants were clustered using: 1) similar variables as previously described in adults and 2) novel routinely available clinical variables. We assessed the effectiveness of the clusters, as well as that of simple clinical measures, to predict treatment response in the TODAY clinical trial. Results: There were three youth-onset type 2 diabetes clusters: 1) Youth-onset insulin-deficient diabetes (YIDD-T2), 2) Youth-onset insulin-resistant diabetes (YIRD-T2), and 3) Intermediate youth-onset diabetes (IYOD-T2). These clusters had differential responses to therapies and risk of treatment failure in the TODAY study, with those in the YIDD-T2 cluster experiencing the highest rate of treatment failure, regardless of treatment arm. YIDD-T2 also had high rates of type 2 diabetes complications. We then generated three novel clusters, with also different rates of treatment failure, using variables available in routine clinical practice. Compared to both clustering methods, simple clinical measures performed comparably or better at predicting treatment response and complications. Conclusion: Youth-onset type 2 diabetes can be characterized into reproducible clusters that demonstrate differential response to treatments and risk of complications. Nevertheless, cluster membership did not add clinical utility beyond simple clinical measures for predicting outcomes.

Introduction and Objective: Type 2 diabetes (T2D) is polygenic and heterogeneous. Clustering of genome-wide association study (GWAS) signals has identified variant clusters underlying T2D trait heterogeneity but has not resolved them to genes and pathways. We tested whether clustering diabetes-associated SNPs by gene and functional annotation could resolve T2D genetic signals to pathways. Methods: We developed PIGEAN, a Bayesian method inferring SNP-to-gene/annotation associations from GWAS summary statistics, and EAGGL, a soft-clustering method using non-negative matrix factorization on PIGEAN outputs. We applied both to meta-analyzed T2D GWAS SNPs to define pathway clusters (factors). Factor polygenic scores (FPS) were made from each factor to test associations with T2D traits in 4 models, with or without T2D polygenic adjustment and using joint or individual factors, in AMP-T2D Knowledge Portal GWAS summary statistics and individual-level data from the United Kingdom Biobank. Results: PIGEAN identified 149 genes with >50% likelihood of T2D involvement by annotation; 67 were factor-specific. EAGGL clustered these into 11 annotation-derived factors. These factors overlapped T2D SNP-trait clusters, but aggregated liver clusters, resolved additional beta cell and lipid clusters, and introduced 2 novel insulin-response clusters. FPS trait associations were concordant, notably CRP with Adipocyte Hypertrophy p = 2.0 × 10-3, NAFLD with Hepatic Steatosis p = 7.0 × 10-3, LPA with Non-Leptin Adipokine Signaling p = 9.4 × 10-4, BMI with Leptin Signaling p = 3.3 × 10-3 and Obesity p < 10-300, and ALT with Adipose Failure p = 6.2 × 10-5. Of PIGEAN-identified genes, 58% (86) were novel findings, including PDE8B. Conclusion: Annotation-informed clustering identified biologically interpretable factors spanning pancreatic, adipose, hepatic, and mixed dysmetabolic pathways. This approach refines our understanding of T2D heterogeneity by mapping signals to genes and pathways. Disclosure S.D. Gage: None. K. Smith: None. A.J. Deutsch: None. M. Udler: Advisory Panel; Ended; Novo Nordisk. Research Support; Current; Novo Nordisk. J. Flannick: None. Funding NIDDK (U54 DK118612)

Polygenic scores strongly predict type 1 diabetes risk, but most scores were developed in European-ancestry populations. In this study, we leveraged recent multi-ancestry genome-wide association studies to create a type 1 diabetes multi-ancestry polygenic score (T1D MAPS). We trained the score in the Mass General Brigham (MGB) Biobank (372 individuals with type 1 diabetes) and tested the score in the All of Us program (86 individuals with type 1 diabetes). We evaluated the area under the receiver operating characteristic curve (AUC), and we compared the AUC to two published single-ancestry scores: T1D GRS2EUR and T1D GRSAFR. We also developed an updated score (T1D MAPS2) that combines T1D GRS2EUR and T1D MAPS. Among individuals with non-European ancestry, the AUC of T1D MAPS was 0.90, significantly higher than T1D GRS2EUR (0.82) and T1D GRSAFR (0.82). Among individuals with European ancestry, the AUC of T1D MAPS was slightly lower than T1D GRS2EUR (0.89 vs. 0.91). However, T1D MAPS2 performed equivalently to T1D GRS2EUR in European ancestry (0.91 vs. 0.91) and performed better in non-European ancestry (0.90 vs. 0.82). Overall, these findings advance the accuracy of type 1 diabetes genetic risk prediction across diverse populations.

Background:Accurately classifying pediatric diabetes can be challenging for providers, and misclassification can result in suboptimal care. In recent years, type 1 diabetes (T1D) polygenic scores, which quantify one's genetic risk for T1D based on T1D risk allele burden, have been developed with good discriminating capacity between T1D and not-T1D. These tools have the potential to improve significantly diagnostic provider accuracy if used in clinic. Methods:We applied T1D polygenic scores to a group of pediatric patients (n=1846) with genetic data available in the Boston Children's Hospital PrecisionLink Biobank, including 96 individuals diagnosed with T1D. Results:Patients with a clinical diagnosis of T1D had higher T1D polygenic scores compared to controls (Wilcoxon rank-sum P<0.0001). Sixty-nine of the 74 individuals with diabetes and a T1D polygenic score exceeding an externally validated cutoff for distinguishing T1D from not-T1D were confirmed to have T1D. There were multiple cases where T1D polygenic scores would have clinical utility. An elevated T1D polygenic score suggested T1D in a pancreatic autoantibody (PAA)- negative individual with negative MODY genetic testing and a phenotype matching T1D. A low T1D polygenic score accurately indicated atypical diabetes in an individual found to have HNF1B-MODY. One individual had positive PAA, but the provider noted that the patient may not have classic T1D, as later suggested by a low T1D polygenic score. Conclusion:T1D polygenic scores already have clinical utility to aid in the accurate diagnosis of pediatric diabetes. Efforts are now needed to advance their use in clinical practice.

Type 1 diabetes (T1D) polygenic risk scores (PRS) are effective tools for discriminating T1D from other diabetes types and predicting T1D risk, with applications in screening and intervention trials. A previously published T1D Genetic Risk Score 2 (GRS2) is widely adopted, but challenges in standardization and accessibility have hindered broader clinical and research utility. To address this, we introduce GRS2x, a standardized and cross-compatible method for accurate T1D PRS calculation, demonstrating genotyping and reference panel independent performance across diverse datasets. GRS2x as a unified approach facilitates accessible and portable measurement of T1D polygenic risk.

Diabetes mellitus is a heterogeneous condition with substantial clinical variability across global populations. The standard classification of type 1 and type 2 diabetes is primarily based on phenotypic characteristics in European-ancestry populations. However, diabetes exhibits diverse phenotypes in other populations, including a varied relationship between diabetes risk and body mass index. These differences may be partly attributable to genetic variation among populations. Understanding and leveraging this genetic variation can aid in the development of precision medicine approaches to diabetes diagnosis and treatment, ultimately helping to reduce health care disparities in diabetes among minoritized populations.

Introduction and Objective: Clustering type 2 diabetes (T2D) genetic loci based on phenotype associations is a powerful approach for uncovering disease mechanisms. Here, we compared two recent efforts from the T2D Global Genomics Initiative (TGGI) and Mass General Hospital (MGH) groups, which used distinct input matrices and clustering methods (k-means vs Bayesian non-negative matrix factorization, bNMF). Our goal was to identify robust and biologically meaningful clusters that transcend methodological differences. Methods: To compare the published clusters, we assessed: i) cluster-specific partitioned polygenic risk scores (pPS) in individuals from the Mass General Brigham Biobank (MGBB, N=64k), ii) ) variant overlap, iii) colocalization analyses, and iv) a cross-clustering analysis, where each clustering method was applied to the opposing study’s matrix. Results: In MGBB, we identified four cluster pairs with correlated pPS (MGH/TGGI): Beta Cell 2/Beta Cell PI+; Lipodystrophy 1/Lipodystrophy; Obesity/Obesity, Cholesterol/Liver Lipid Metabolism (R>0.4, p<1e-200). Phenotypic associations with cluster pPS reinforced these findings, such as both MGH/TGGI Lipodystrophy clusters with increased triglycerides and decreased BMI and HDL. These same cluster pairs had strong overlap of shared variants (Fisher’s p< 1e-10). In colocalization analyses, loci that overlapped between matrices and belonged to one of the above four cluster pairs were enriched for quantitative trait loci signals in metabolites, proteins, and GWAS traits (Chi-square p<0.002). When bNMF was applied to the TGGI matrix, >90% of variants from the four robust clusters still clustered together (excluding those that went unclustered). Conclusion: Our analysis demonstrates that while clustering results are influenced by inputs and methods, certain T2D genetic clusters emerge consistently, indicating robust biological signatures. These findings support cluster reproducibility and highlight key clusters for further investigation into T2D pathophysiology. K. Smith: None. A.J. Deutsch: None. J.E. Gervis: None. K. Lorenz: None. H.J. Taylor: None. R. Mandla: None. B.F. Voight: Other Relationship; Eli Lilly and Company. J.I. Rotter: None. C. Yap: None. C.N. Spracklen: None. E. Zeggini: None. J.M. Mercader: None. A. Morris: None. M. Udler: Research Support; Novo Nordisk. Doris Duke (241537)

Introduction and Objective: Polygenic scores (PS) can predict risk for type 1 diabetes (T1D) by integrating genetic information across the genome. However, most PS are derived from a single population and perform best when applied to genetically similar populations. Selecting the optimal PS is challenging in admixed populations. Here, we investigated whether a T1D PS derived from multiple ancestries could accurately predict risk across diverse populations. Methods: We collected summary statistics from a recent T1D genome-wide association study, which included individuals with African (AFR), Admixed American (AMR), or European (EUR) genetic ancestry. As variation in HLA genes accounts for ~50% of T1D genetic risk, we modeled T1D risk separately for HLA and non-HLA variants. We used PRS-CS to construct a T1D PS with over 1 million variants. We applied the score in Mass General Brigham Biobank and evaluated performance with the area under the receiver operating characteristic curve (AUC), controlling for age, sex, and 10 principal components of ancestry. We compared this T1D PS to T1D GRS2, which was previously developed using data from EUR ancestry populations. Results: The novel multi-ancestry PS was strongly associated with T1D risk, with an odds ratio of 16.9 for those in the top 5% of the distribution compared to the remaining 95% (P < 2×10-16). Among 53,885 participants with EUR ancestry (370 with T1D), the AUC of the T1D PS was 0.87, which was not significantly different from T1D GRS2 (AUC = 0.87, P = 0.84). However, among 6,872 participants with AFR or AMR ancestry (38 with T1D), the AUC of the T1D PS was 0.90, which was significantly higher than T1D GRS2 (AUC = 0.86, P = 0.04). Conclusion: Our multi-ancestry T1D PS improves T1D risk prediction in non-EUR ancestries, while maintaining high predictive power in EUR populations. These results underscore the need to generate genetic data from non-EUR ancestries to improve T1D prediction. A.J. Deutsch: None. A.S. Bell: None. D.A. Michalek: None. S. Onengut-Gumuscu: None. S.S. Rich: None. J.C. Florez: Research Support; Novo Nordisk. A. Manning: None. J.M. Mercader: None. M. Udler: Research Support; Novo Nordisk. NIH/NIDDK (K23 DK140643; NIH/NHGRI (U01HG011723)

Deep learning models leveraging electronic health records (EHR) for opportunistic screening of type 2 diabetes (T2D) can improve current practices by identifying individuals who may need further glycemic testing. Accurate onset prediction and subtyping are crucial for targeted interventions, but existing methods treat the tasks separately, thus limiting clinical utility. In this paper, we introduce a novel deep metric learning (DML) model that unifies both tasks by learning a latent space based on sample similarity. In onset prediction, the DML model predicts the onset of T2D 7 years later with an AUC of 0.754, outperforming logistic regression (AUC 0.706), clinical risk factors (AUC 0.693), and glycemic measures (AUC 0.632). For subtyping, we identify three subtypes with varying prevalences of obesity-related, cardiovascular, and mental health conditions. Additionally, the subtype with fewer comorbidities shows earlier metformin initiation and a greater reduction in HbA1c. We validated these findings using data from 300 U.S. hospitals in the All of Us program (T2D, n = 7567) and the Massachusetts General Brigham Biobank (T2D, n = 3298), demonstrating the transferability of our model and subtypes across cohorts.

BACKGROUND:Islet autoantibodies form the foundation for type 1 diabetes (T1D) diagnosis and staging, but heterogeneity exists in T1D development and presentation. We hypothesized that autoantibodies can identify heterogeneity before, at, and after T1D diagnosis, and in response to disease-modifying therapies.METHODS:We systematically reviewed PubMed and EMBASE databases (6/14/2022) assessing 10 years of original research examining relationships between autoantibodies and heterogeneity before, at, after diagnosis, and in response to disease-modifying therapies in individuals at-risk or within 1 year of T1D diagnosis. A critical appraisal checklist tool for cohort studies was modified and used for risk of bias assessment.RESULTS:Here we show that 152 studies that met extraction criteria most commonly characterized heterogeneity before diagnosis (91/152). Autoantibody type/target was most frequently examined, followed by autoantibody number. Recurring themes included correlations of autoantibody number, type, and titers with progression, differing phenotypes based on order of autoantibody seroconversion, and interactions with age and genetics. Only 44% specifically described autoantibody assay standardization program participation.CONCLUSIONS:Current evidence most strongly supports the application of autoantibody features to more precisely define T1D before diagnosis. Our findings support continued use of pre-clinical staging paradigms based on autoantibody number and suggest that additional autoantibody features, particularly in relation to age and genetic risk, could offer more precise stratification. To improve reproducibility and applicability of autoantibody-based precision medicine in T1D, we propose a methods checklist for islet autoantibody-based manuscripts which includes use of precision medicine MeSH terms and participation in autoantibody standardization workshops.

Importance Immune checkpoint inhibitors (ICIs) have revolutionized cancer care; however, accompanying immune-related adverse events (irAEs) confer substantial morbidity and occasional mortality. Life-threatening irAEs may require permanent cessation of ICI, even in patients with positive tumor response. Therefore, it is imperative to comprehensively define the spectrum of irAEs to aid individualized decision-making around the initiation of ICI therapy. Objective To define incidence, risk factors, and clinical spectrum of an irreversible and life-threatening irAE: ICI-induced diabetes. Design, Setting, and Participants This cohort study, conducted at an academic integrated health care system examined 14 328 adult patients treated with ICIs, including 64 patients who developed ICI-induced diabetes, from July 2010 to January 2022. The data were analyzed from 2022 to 2023. Cases of ICI-induced diabetes were manually confirmed; detailed clinical phenotyping was performed at diagnosis and 1-year follow-up. For 862 patients, genotyping data were available, and polygenic risk for type 1 diabetes was determined. Main Outcomes and Measures For ICI-induced diabetes cases and controls, demographic characteristics, comorbidities, tumor category, and ICI category were compared. Among ICI-induced diabetes cases, markers of glycemic physiology were examined at diagnosis and 1-year follow-up. For patients with available genotyping, a published type 1 diabetes polygenic score (T1D GRS2) was calculated. Results Of 14 328 participants, 6571 (45.9%) were women, and the median (range) age was 66 (8-106) years. The prevalence of ICI-induced diabetes among ICI-treated patients was 0.45% (64 of 14 328), with an incidence of 124.8 per 100 000 person-years. Preexisting type 2 diabetes (odds ratio [OR], 5.91; 95% CI, 3.34-10.45) and treatment with combination ICI (OR, 2.57; 95% CI, 1.44-4.59) were significant clinical risk factors of ICI-induced diabetes. T1D GRS2 was associated with ICI-induced diabetes risk, with an OR of 4.4 (95% CI, 1.8-10.5) for patients in the top decile of T1D GRS2, demonstrating a genetic association between spontaneous autoimmunity and irAEs. Patients with ICI-induced diabetes were in 3 distinct phenotypic categories based on autoantibodies and residual pancreatic function, with varying severity of initial presentation. Conclusions and Relevance The results of this analysis of 14 328 ICI-treated patients followed up from ICI initiation determined the incidence, risk factors and clinical spectrum of ICI-induced diabetes. Widespread implementation of this approach across organ-specific irAEs may enhance diagnosis and management of these conditions, and this becomes especially pertinent as ICI treatment rapidly expands to treat a wide spectrum of cancers and is used at earlier stages of treatment.

Type 2 diabetes (T2D) is a multifactorial disease with substantial genetic risk, for which the underlying biological mechanisms are not fully understood. In this study, we identified multi-ancestry T2D genetic clusters by analyzing genetic data from diverse populations in 37 published T2D genome-wide association studies representing more than 1.4 million individuals. We implemented soft clustering with 650 T2D-associated genetic variants and 110 T2D-related traits, capturing known and novel T2D clusters with distinct cardiometabolic trait associations across two independent biobanks representing diverse genetic ancestral populations (African, n = 21,906; Admixed American, n = 14,410; East Asian, n =2,422; European, n = 90,093; and South Asian, n = 1,262). The 12 genetic clusters were enriched for specific single-cell regulatory regions. Several of the polygenic scores derived from the clusters differed in distribution among ancestry groups, including a significantly higher proportion of lipodystrophy-related polygenic risk in East Asian ancestry. T2D risk was equivalent at a body mass index (BMI) of 30 kg m-2 in the European subpopulation and 24.2 (22.9-25.5) kg m-2 in the East Asian subpopulation; after adjusting for cluster-specific genetic risk, the equivalent BMI threshold increased to 28.5 (27.1-30.0) kg m-2 in the East Asian group. Thus, these multi-ancestry T2D genetic clusters encompass a broader range of biological mechanisms and provide preliminary insights to explain ancestry-associated differences in T2D risk profiles.