The T1DGRS derived from WGS is highly correlated to T1DGRS from array genotypes overall but lower in non-European genetic ancestries.
The correlation between T1DGRS_WGS and T1DGRS_1000G was 0.9815 (95% CI 0.9813, 0.9817) (Fig. 1A). In contrast, T1DGRS_WGS showed a higher, near-perfect correlation to T1DGRS_TOPMed (0.99555; 95% CI 0.99551, 0.99560) (Fig. 1B). The correlation was also higher across all genetic ancestries compared to T1DGRS_1000G, particularly for the African and South Asian genetic ancestries (0.9892; 95% CI 0.9891, 0.9893 and 0.9861; 95% CI 0.9860, 0.9862 respectively) (Fig. 1C). The results were consistent for both HLA and non-HLA components of the GRS (Suppl. Figure 1).
The differences in minor allele frequencies contribute to the discrepancies between WGS-based and array-based T1DGRS.
The T1DGRS_WGS values were on average 0.043 SD lower than their corresponding T1DGRS_1000G values (95% CI -0.044, -0.042; p < 10− 300), and 0.0028 SD lower than corresponding T1DGRS_TOPMed values (95% CI -0.0033, -0.0023; p < 10− 31) (Suppl. Figure 2A). Bland-Altman analysis on all individuals revealed no proportional bias between T1DGRS_WGS and T1DGRS_array (Suppl. Figure 3). T1DGRS_1000G was much lower across all ancestries whereas T1DGRS_TOPMed much closer to T1DGRS_WGS (mean difference range 0.003–0.019 SD) (Suppl. Figure 2A). In line with these results, the minor allele frequencies for 9 variants were statistically different between the WGS and the 1000 Genomes-imputed array genotypes after Bonferroni correction, whereas only 2 variants were different for the TOPMed-imputed array genotypes (Fig. 2, Suppl. Table 2). Further investigation revealed that the differences were predominantly from the non-European genetic ancestries and the lower frequency variants in 1000 Genomes-imputed data (Suppl. Figure 4, Suppl. Table 2). These data together suggest that differences in minor allele frequencies contribute to the discrepancies between WGS-based and array-based T1DGRS.
Array-based T1DGRS categorised up to 5% of cases differently at clinically relevant T1DGRS risk thresholds.
We next assess the impact of different method on clinically utilised thresholds using T1DGRS_WGS as a reference. Of the 74,633 individuals with T1DGRS_WGS < 50th centile (threshold for excluding T1D in clinical practice) 2,6, 4.4% (95% CI 4.3%, 4.6%) would categorise to have T1DGRS ≥ 50th centile using T1DGRS_1000G and 1.1% (95% CI 1.0%, 1.2%) using T1DGRS_TOPMed (Table 1). Similarly, of the 14,924 individuals with T1DGRS_WGS ≥ 90th centile (threshold for T1D screening in the population) 2, 4.2% (95% CI 3.8%, 4.5%) were found to have score below this threshold using T1DGRS_1000G and 2.3% (95% CI 2.0%, 2.5%) using T1DGRS_TOPMed. Overall, the accuracy of T1DGRS_1000G at the 50th and 90th centile thresholds were 97.2% and 98.3%, and higher for T1DGRS_TOPMed (99% and 99.5% respectively). In line with this higher accuracy, the discriminative ability of the T1DGRS for European T1D cases (n = 112) and European controls was similar across all the methods (ROC AUC 0.93; 95% CI 0.9, 0.95 for all)(Suppl. Figure 5). We were unable to perform this analysis in non-European ancestries due to lack of T1D cases in our study cohort.
Table 1
Contingency tables showing the accuracy of array-derived T1DGRS at clinically useful levels of disease risk, using WGS-derived T1DGRS as a reference.
| | T1DGRS_1000G centiles % (N) | T1DGRS_TOPMed centiles % (N) |
| | < 50th | ≥ 50th | < 50th | ≥ 50th |
T1DGRS_WGS centiles (N) | < 50th (74,633) | 95.6 (71,316) | 4.4 (3,317) | 98.9 (73,817) | 1.1 (816) |
≥ 50th (74,632) | 1.07 (802) | 98.93 (73,830) | 0.86 (640) | 99.14 (73,992) |
| | < 90th | ≥ 90th | < 90th | ≥ 90th |
T1DGRS_WGS centiles (N) | < 90th (134,341) | 98.6 (132,410) | 1.4 (1,931) | 99.7 (133,953) | 0.3 (388) |
≥ 90th (14,924) | 4.2 (616) | 95.8 (14,308) | 2.3 (337) | 97.7 (14,587) |
WGS-derived T1DGRS were lower for individuals of African and South Asian ancestries compared to European ancestry.
Next generation sequencing methods like WGS are well-suited to overcome biases introduced by the imputation reference panels in array-genotyping allowing better comparison of GRS across the ancestries. We therefore compare the T1DGRS across genetic ancestries using T1DGRS_WGS. We found that the T1DGRS_WGS was substantially lower for individuals of African ancestry (-0.89 SD, 95% CI -0.92, -0.85; p < 10− 300) and South Asian ancestry (-0.28 SD, 95% CI -0.31, -0.24; p < 10− 58) compared to individuals of European ancestry (Fig. 3). This was found to correlate with the difference in allele frequency in non-Europeans compared to Europeans (Suppl. Figure 6, Suppl. Table 2). Use of the European ancestry-based population threshold of < 50th centile (threshold for excluding T1D in clinical practice) will identify additional 35.3% of African as having as having low genetic risk of T1D (95% CI 33.8%, 36.7%; p < 10− 290) and additional 14.1% South Asians (95% CI 12.4%, 15.7%; p < 10− 60). Similarly, use of the European ancestry-based population risk threshold of ≥ 90th centile (threshold for T1D screening in the population) will only identify 0.71% (95% CI 0.41%, 1.13%; p < 10− 300) of individuals of African ancestry and 6.4% (95% CI 5.6%, 7.2%; p < 10− 13) of South Asians individuals rather than expected 10%. Results were consistent in both the HLA and non-HLA components of the score, with greater differences observed in non-HLA score across genetic ancestries (Suppl. Figure 7).