GWAS of type 1 diabetes vs control

The study design is illustrated in Fig. 1. The baseline characteristics of the discovery, replication and validation cohorts are shown in ESM Table 1. In the discovery cohort, compared with controls, individuals with type 1 diabetes were younger on average, and 53.6% were male. In the discovery cohort, GWAS of 1303 type 1 diabetes participants and 2236 controls identified significant type 1 diabetes genetic associations both within and outside of the HLA region (Fig. 2). A total of 5817 SNPs reached the threshold of p<1.00×10−5, including 369 located outside the HLA region. We identified 17 SNPs that showed consistency with previous reports in populations of European ancestry, reaching a significance level of p<1×10−5 (ESM Table 2). These included the chromosomal regions 2q33.2 (CTLA4), 6q21.2 (HLA), 11p15.5 (INS-IGF2), 12q13.2 (RAB5B-SUOX-RPS26-ERBB3), 14q32.2 (VRK1), 18p11.21 (PTPN2) and 18q22.2 (CD226). Using our GWAS data and publicly available GWAS summary statistics from European populations, we found a moderate genetic correlation (rg=0.41) between European and Chinese ancestries, indicating a certain degree of population heterogeneity.

Subsequently, we performed GWAS validation using an independent replication cohort consisting of 501 participants with type 1 diabetes and 853 control individuals from Hong Kong. Meta-analysis of the top SNPs in the discovery cohort and replication cohort revealed 202 SNPs that reached genome-wide significance (p<5.00×10−8). In addition to the SNPs within the HLA region, there were eight SNPs in non-HLA regions (Table 1), including chromosomal regions 2q33.2 (CTLA4), 7p14.3 (BMPER), 10p15.1 (IL2RA), 11p15.5 (INS-IGF2), and 12q13.2 (SUOX-RPS26-SH2B3-ATXN2). We identified a SNP rs10232170 in BMPER as a possible novel risk locus. Although it was not replicated in the replication cohort, it reached genome-wide significance in the meta-analysis (p=9.897×10⁻9; Table 1 and ESM Fig. 2). Additionally, we identified pathways of the validated signals in type 1 diabetes. We found that the pathways of ‘MHC class II receptor activity’ and ‘peptide antigen binding’ were the most enriched biological processes in participants with type 1 diabetes when compared with control participants (ESM Fig. 3). Furthermore, when we compared our results with previously published data from the Japanese population, the most significant susceptibility loci remained in the HLA region (ESM Table 3).

Table 1 Genomic regions achieving genome-wide significance validated in replication cohort (p<5.00×10−8) HLA type 1 diabetes associations

We identified tag SNPs for 13 HLA DQA1-DQB1 haplotypes that were associated with type 1 diabetes in the discovery cohort (ESM Table 4), following the approach used in a previous study [13]. Interestingly, DQA1*0301-DQB1*0302 (DQ8) was not found to be significantly associated with type 1 diabetes in our data, despite its significant association in studies involving populations of European ancestry. After identifying DQ haplotypes, we detected and used their interactions to refine the estimation of type 1 diabetes risk. We identified 15 DQ haplotypes with significant interaction terms and calculated the OR specific to each interaction (ESM Table 5).

We identified 12 HLA region SNPs that were not representative of DR-DQ alleles but each was independently associated with type 1 diabetes (ESM Table 6). Some known type 1 diabetes-associated alleles in the HLA class I or II (except for DQ) gene regions, as well as some intergenic loci, were also present. By combining the (13 + 12) total HLA region SNPs, the C-GRS achieved good discrimination in the discovery cohort (ROC AUC=0.864).

Non-HLA type 1 diabetes associations

We identified eight non-HLA loci, each of which, again, was independently associated with type 1 diabetes (ESM Table 7). Together, these non-HLA SNPs discriminated type 1 diabetes in the discovery cohort (ROC AUC=0.641).

The C-GRS using 33 SNPs is highly discriminative of type 1 diabetes vs control in Chinese individuals

The final combined type 1 diabetes C-GRS used 33 SNPs, including 13 DR-DQ haplotypes, 12 other HLA SNPs and eight non-HLA SNPs. This C-GRS showed high discrimination of type 1 diabetes, with ROC AUC=0.876 (Fig. 3) and PRAUC=0.820 (ESM Fig. 4) in the discovery cohort. We found no genome-wide significant SNP associations when including C-GRS as a covariate, indicating good capture of type 1 diabetes-associated information (Fig. 4). The efficiency (or cost) of the C-GRS depends on the SNP count. We sequentially added SNPs outside the DR-DQ region, ranking them by p value to compare AUC discriminative ability, and found no significant AUC increase from 29 to 30 SNPs (AUC=0.874 vs AUC=0.875, p>0.05, Fig. 3b). Subsequently, we stratified the analysis by age of onset and calculated the AUC. We found that C-GRS shows a higher AUC in the youth-onset group (AUC=0.911) compared with the adult-onset group (AUC=0.849), with a significant difference (p=3.446×10−7, ESM Fig. 5).

Fig. 3

The diagnostic efficacy of C-GRS and its association with clinical indicators. (a) AUC results from ROC analysis for C-GRS in the discovery cohort. (b) The gradual improvement in the power of the C-GRS as additional SNPs were included. The blue shaded area represents the 95% CI. (c–g) A higher C-GRS was associated with an earlier age (c), lower BMI (d), lower fasting (e) and postprandial C-peptide levels (f) and higher proportion of multiple positive autoantibodies (g) at diagnosis. Box plots show median ± quartiles, and the whiskers extend from the hinge to the largest or smallest value no further than 1.5-fold of the IQR. *p<0.05, **p<0.01, ***p<0.001. Ab, autoantibody; FCP, fasting C-peptide; PCP, postprandial C-peptide

Fig. 4

GWAS After adjustment for C-GRS. (a, b) Manhattan plot showing the genome-wide association results for type 1 diabetes in the HLA region (a) and the non-HLA region (b). (c, d) Quantile–quantile plots were used to assess the association between type 1 diabetes risk and the HLA region (c) and the non-HLA region (d). The observed p values are plotted as a function of the expected p values. The blue areas represent the 95% CI derived from a null distribution of p values

Additionally, we validated the C-GRS in the validation cohort. We observed a significant difference in the C-GRS when comparing participants with type 1 diabetes and the control participants (p=4.483×10−81, Fig. 5a). Notably, using only the SNPs representing the HLA region, we achieved an AUC of 0.859 in discriminating between participants with type 1 diabetes and control participants (Fig. 5b). When using non-HLA region loci alone to distinguish between individuals with type 1 diabetes and control individuals, an AUC of 0.598 was obtained. When using all SNPs for discrimination, the AUC was 0.871 (Fig. 5b).

Fig. 5

The validation of the discrimination performance of C-GRS in an independent cohort. (a) The C-GRS of type 1 diabetes was significantly higher than that of control individuals and those with type 2 diabetes in the validation cohort. (b) The type 1 diabetes C-GRS has high discriminatory power in distinguishing individuals with type 1 diabetes from both control individuals. (c) C-GRS demonstrated strong discriminatory ability between individuals with type 1 diabetes and those with type 2 diabetes. ***p<0.001. T1D, type 1 diabetes; T2D, type 2 diabetes

Next, we evaluated the performance of the previously established GRS2, derived mainly from populations of European ancestry [13], in our cohort (Fig. 5 and ESM Fig. 6). We found that when using the HLA region loci and all loci of GRS2 to distinguish between type 1 diabetes and control individuals separately, the AUCs were 0.758 and 0.773 (Fig. 5b), respectively. Subsequently, we compared the performance of GRS2 with the newly constructed C-GRS to differentiate control individuals from those with type 1 diabetes in the validation cohort. We found that our newly established C-GRS improved the discriminative power for type 1 diabetes vs control compared with GRS2 (p=2.657×10−6).

Associations of type 1 diabetes C-GRS with clinical features

We stratified the C-GRS by tertiles in the discovery cohort. Compared with the low C-GRS group, the high C-GRS group showed an earlier age of type 1 diabetes diagnosis (median [IQR]: 26.0 [14.0–37.0] vs 15.0 [8.0–23.0] years, p=1.520×10−11, Fig. 3c), lower BMI (median [IQR]: 19.2 [17.3–21.2] vs 17.7 [15.8–20.0] kg/m2, p=4.875×10−7), lower levels of fasting C-peptide (median [IQR]: 100.8 [38.4–184.0] vs 79.0 [31.1–157.0] pmol/l, p=7.300×10−5) and lower levels of 2 h postprandial C-peptide (median [IQR]: 197.5 [98.7–396.2] vs 157.3 [55.9–298.4] pmol/l, p=7.640×10−7, Fig. 3d–f). Additionally, the proportion of individuals with multiple autoantibody positivity was higher in the high C-GRS group (34.1% vs 42.9%, p=4.030×10⁻2, Fig. 3g).

The type 1 diabetes C-GRS discriminates between clinically defined type 1 diabetes and type 2 diabetes

Next, we calculated C-GRS in a validation cohort, which included 262 individuals with type 1 diabetes (median [IQR]: 4.374 [2.358–5.466]) and 1080 individuals with type 2 diabetes (median [IQR]: 0.049 [−1.545 to 1.672]), with strict clinical criteria employed to define each type of diabetes. The C-GRS was highly discriminative of type 1 diabetes from type 2 diabetes (p=7.099×10⁻72; Fig. 5a) with an AUC of 0.869 (Fig. 5c). When only using HLA region loci alone or non-HLA GRS, the AUC was 0.857 or 0.588. The DeLong test demonstrated that, compared with GRS2, C-GRS exhibited superior performance in distinguishing between type 1 diabetes and type 2 diabetes in the Chinese population (0.869 vs 0.793, p=4.003×10−5).

We then determined cutoff values for the clinical classification of individuals with type 1 diabetes or type 2 diabetes. ESM Table 8 provides a set of examples of C-GRS cutoffs, sensitivities and specificities. Scores with 95% specificity for type 1 diabetes or type 2 diabetes were identified as being useful for further classification. A C-GRS of >1.211 was indicative of type 1 diabetes, with 95% specificity and 55% sensitivity (i.e. a C-GRS of 1.211 was at the 95th centile of a control population and 45th centile of the type 1 diabetes cohort). A C-GRS of < −0.407 (38th centile of C-GRS in validation cohort) was indicative of type 2 diabetes, with 95% specificity and 45% sensitivity. The 33-SNP C-GRS also provided excellent discrimination between adult-onset type 1 diabetes and type 2 diabetes (AUC=0.818).