Standardized Measurement of Type 1 Diabetes Polygenic Risk Across Multiancestry Population Cohorts Free

Type 1 diabetes (T1D) polygenic risk scores (PRS) effectively discriminate T1D from both type 2 and monogenic diabetes (1) and can aid identification of T1D in the population, serving as a valuable tool for screening and intervention trials (2). The Genetic Risk Score 2 (GRS2) (3) combining 67 single nucleotide variants is widely adopted in translational studies, yet absence of a consistent method for generation is a barrier to broader clinical and research utility. Study-specific differences limit the availability of variants, particularly those in the HLA region (6p21.3) with strong disease-modifying effects. Differences in genotyping coverage and accuracy arise from the choice of technology (arrays or whole-genome sequencing [WGS]) and access to reference panels used to infer missing genotypes (imputation), often constrained by global data access restrictions. This variability limits standardized risk assessment across research and clinical studies. We present GRS2x, an accessible approach for standardized measurement of T1D polygenic risk.

We used three widely available reference panels (1000 Genomes, Haplotype Reference Consortium [HRC], and Trans-Omics for Precision Medicine [TOPMed] [versions TOPMed-r2 and TOPMed-r3]) to replace variants from GRS2. Insertions and deletions (more than two base pairs), multinucleotide variants, and variants not in all reference panels were replaced (19 of 67), with linkage disequilibrium (as previously described [3]). Groups of variants partitioned by genetic locus (HLA class II, HLA class I, and non-HLA) were used to define partitioned polygenic scores (pPS), with total GRS2x calculated as their sum (Fig. 1A). We implemented algorithms to address two critical limitations: 1) incomplete sets of variants prohibit standardized risk assessment and 2) variants marking HLA haplotypes with imperfect linkage disequilibrium can lead to uncertain predictions (more than two HLA haplotypes) in individuals, often resulting in exclusion. Mean effect contribution for missing variants was estimated on the basis of Hardy-Weinberg equilibrium, based on reference frequencies (TOPMed). After assigning HLA predictions, we iteratively removed least likely genotypes based on population frequency (NMPD [National Marrow Donor Program]). We then applied normalization to rescale GRS2x based on theoretical minimum and maximum relative risk values. Performance in identifying cases was assessed using area under the receiver operating characteristic curve (AUROC).

We measured case-control discriminative performance of GRS2x using array genotyping from the Type 1 Diabetes Genetics Consortium (T1D 6,483, control 9,247; age of onset <20 years) after TOPMed imputation. Discrimination of T1D improved marginally (AUROC: GRS2x 0.924 vs. GRS2 0.922; P = 2.96 × 10⁻¹⁰) and remained consistent across reference panels, matching results previously reported for GRS2 after 1000 Genomes plus HRC imputation (3) (Fig. 1C).

We further validated GRS2x, with stratification by genotyping technology of the UK Biobank (UKB) (4). We replicated published case-control definitions (T1D 408, control 365,622) (3) from electronic health records (EHR) for European-ancestry individuals with array (n = 367,318) and/or WGS (n = 366,030) genotyping available. Performance of GRS2x in identifying T1D was stable (AUROC: array 0.931 vs. WGS 0.932), despite the absence of three variants in the array data set (Fig. 1C). In individuals with genotypes from both sources, GRS2x was highly concordant (R = 0.96) and rank order of individual-level risk predictions was preserved (ρ = 0.96). GRS2x further resolved conflicting HLA class II genotypes, allowing score assignment for additional individuals (n = 5,550 [1.14%]). Most T1D discrimination was driven by pPS encoding HLA class II (AUROC 0.90), over HLA class I and non-HLA variants (AUROC 0.55 and 0.75, respectively) (Fig. 1B).

For assessment of GRS2x multiancestry performance, case-control definitions (T1D 147, control 317,817) were generated in the All of Us Research Program (AoU) (5) (EHR-recorded data: T1D, age at EHR incidence <30 years, prescribed insulin, no noninsulin glucose-lowering agents). Sensitivity and specificity were limited by incomplete EHR data and the use of age at recorded incidence as a proxy for juvenile onset. Identification of T1D in the multiancestry population (MA) was strong but reduced in comparison with European-only ancestry population (EUR) (n_T1D = 103) (AUROC: MA 0.860, EUR 0.895) (Fig. 1C), though sample sizes were insufficient for further ancestry stratification. Rank-based concordance between GRS2x and GRS2 was marginally lower for non-Europeans (ρ_EUR = 0.906, ρ_non-EUR = 0.900; P < 2.22 × 10⁻¹⁶), suggesting subtle ancestry-related shifts in individual-level risk precision.

We have demonstrated that GRS2x enables standardized measurement of T1D polygenic risk across diverse studies. While complementary efforts aim to enhance discriminative performance, addressing practical limitations provides immediate utility in reducing study-specific variability. Standardization is crucial for clinical translation; however, validated reference ranges must be derived before thresholds for clinical action can be applied. Stabilizing individual-level precision is also critical for clinical consistency. Both individual- and population-level performance may improve with incorporation of greater ancestry-related allelic diversity, particularly from HLA, where attributable risk is variable by ancestry and age of onset. Since our study was limited to early-onset cases, performance for late-onset cases is less clear; however, immediate utility for early-life screening studies is evident. GRS2x offers a robust, unified method for T1D polygenic risk assessment for ongoing and future studies.