Through targeted resequencing of GLIS3 (NM_001042413.2) in 5471 participants from RaDiO [7], we detected 105 rare coding GLIS3 variants, including one nonsense variant (p.Y627*), among 395 carriers (ESM Table 1). At this stage of analysis, the burden of rare variants was not associated with type 2 diabetes risk (pπ=0.20 [poverall=0.055], where pπ is the p value for the impact of the burden; with an OR of 0.85 [95% CI, 0.66, 1.1]; Table 1).

To assess the pathogenicity of the 105 variants, we used the criteria from the American College of Medical Genetics and Genomics (ACMG) [9]; notably we developed in vitro assays to address the pathogenic strong criterion number 3 (PS3). Plasmids including each GLIS3 variant were overexpressed along with their gene reporter assay to assess the ability of each mutant to bind to the 5′ flanking region of INS containing GLIS3 binding sites. Furthermore, these variants were evaluated in conjunction with the overexpression of MafA to determine their capacity for recruiting this transcription factor to the INS 5′ flanking region, consequently enhancing luciferase signalling. When compared with wild-type GLIS3, 49 variants decreased luciferase activity with and/or without the addition of MAFA in the system, and were considered loss-of-function (ESM Fig. 1).

In comparison with our functional results, the in silico pathogenicity prediction by REVEL had a poor sensitivity as only 4% of loss-of-function variants were predicted to be deleterious (ESM Table 1), but had a high specificity (ESM Table 1).

Following ACMG criteria including the PS3 criterion, 18 out of 105 variants, carried by 18 unrelated individuals of European ancestry, were found to be pathogenic or likely pathogenic (P/LP; Fig. 1 and ESM Table 2). Furthermore, P/LP variants were strongly enriched in the C-terminal part of GLIS3, i.e. the last two coding exons after accounting for exon length (p<5×10−6 with an OR>3.5; Fig. 1 and ESM Table 3).

Location of the 18 P/LP variants in GLIS3, shown using arrows. Ser-rich, regions enriched in serine; Pro-rich, region enriched in proline

We then assessed the effect of the burden of the 18 rare P/LP GLIS3 variants on type 2 diabetes risk. This burden was four times higher among participants with type 2 diabetes compared with controls (pπ=8.0×10−3 [poverall=3.0×10−3]; with an OR of 3.9 [95% CI 1.4, 12]; Table 1). We found an equal number of female and male participants carrying a P/LP GLIS3 variant, either with type 2 diabetes or not (data not shown). Interestingly, all participants with type 2 diabetes carrying a P/LP GLIS3 variant were treated with sulfonylureas, a class of medication that stimulates the secretion of insulin by activating ATP-dependent potassium channels. In contrast, 50% of non-carriers suffering from type 2 diabetes were treated with sulfonylureas. The sibling of one of the participants with type 2 diabetes (carrying the P/LP variant encoding p.A168G) also presented with type 2 diabetes and was also treated with sulfonylureas. The age at diagnosis, BMI and cholesterol levels were similar among P/LP variant carriers vs non-carriers with type 2 diabetes (ESM Table 4).

We then analysed the association between P/LP variants in GLIS3 (NM_001042413 [ENST00000381971] transcript) and type 2 diabetes risk in the Type 2 Diabetes Knowledge Portal (using the genetic association interactive tool) [10]. In 43,125 individuals from 52K and in 44,083 individuals from TOPMed, we only found eight P/LP variants per study (i.e. loss-of-function transcript effect estimator [LofTee] variants with a very low MAF in GnomAD; ESM Table 5) that had no effect on disease risk due to a low statistical power (p=0.13 with an OR of 1.7 and p=0.44 with an OR of 3.3, respectively; variable threshold test). No P/LP missense variants identified in our study were observed in either TOPMed or the 52K study. However, through a meta-analysis of RaDiO, TOPMed and 52K studies, we found an enrichment of P/LP GLIS3 variants among individuals with type 2 diabetes (p=5.6×10−5 with an OR of 2.1 [1.4, 2.9]).