Animals

Female NOD/ShiLtJ (hereafter, NOD; RRID: IMSR_JAX:001976) mice were purchased from The Jackson Laboratory (Bar Harbour, ME, USA) for in vivo intervention studies and ex vivo mechanistic studies. Female NOD/ShiLt-Tg(Foxp3-EGFP/cre)1cJbs/J (hereafter, NOD-Foxp3EGFP; RRID: IMSR_JAX:008694) transgenic mice were used for Treg suppression studies [24]. Chromium-release assays were accomplished using cells from NOD/ShiLtJ-Tcra-AI4.Tcrb-AI4.Thy1a.Rag1−/− mice (hereafter, NOD-AI4; [25]). Animals treated with mAbs were housed in animal biosafety level 2 (aBSL-2) containment, whereas all others were housed in animal biosafety level 1 (aBSL-1) containment. All animals were housed in specific pathogen-free facilities at the University of Florida, with food and water available ad libitum. All studies were conducted in accordance with protocols approved by the University of Florida Institutional Animal Care and Use Committee (UF IACUC) and in accordance with the Guide for the Care and Use of Laboratory Animals [26].

Preparation of single-cell suspensions

Isolation of cells from relevant tissues, including the spleen, thymus, pancreatic-draining lymph nodes (pLN) and pancreas, was necessary for in vitro and ex vivo assessment of the mAb mechanism of action. For all tissues except the pancreas, homogenous single-cell suspensions were generated by mechanically dissociating organs and filtering through a 70 µm membrane filter. Cell suspensions were washed with complete RPMI (cRPMI; RPMI 1640 media Phenol Red with l-Glutamine 139 [Lonza, Basel, CH-BS, Switzerland], 5 mmol/l HEPES [Gibco, Waltham, MA, USA], 5 mmol/l Minimum essential medium 140 Non-Essential Amino Acids [MEM NEAA; Gibco], 2 mmol/l Glutamax [Gibco], 50 µg/ml penicillin 141 [Gibco], 50 µg/ml streptomycin [Gibco], 20 mmol/l sodium pyruvate [Gibco], 50 mmol/l 2-mercaptoethanol [Sigma-Aldrich, St Louis, MO, USA], 20 mmol/l sodium hydroxide [Sigma-Aldrich] and 10% FBS [Genesee Scientific, El Cajon, CA, USA]) and pelleted by centrifugation (350 × g for 7 min). Pancreas tissues were minced into 1 mm pieces and incubated in cRPMI media with 1 mg/ml collagenase IV (Gibco) for 18 min at 37°C. After digestion, pancreas suspensions were washed with cRPMI and strained through a 40 μm filter. Lysis of erythrocytes was accomplished by resuspension of cell pellets in ACK Lysing Buffer (Gibco) for 5 min at 4°C before quenching in 1× PBS (Gibco). Cell viability was quantified by staining with acridine orange/propidium iodide at a 1:1 dilution in a Cellometer slide before reading on an Auto2000 Cellometer (Nexcelom Biosciences, Lawrence, MA, USA).

Administration of anti-CD226 and isotype control mAbs

Blocking mAbs against CD226 (BioLegend, San Diego, CA, USA; clone 480.1, RRID: AB_2876467) and its corresponding rat IgG2a isotype control (BioLegend; RRID: AB_11147167) were obtained in a low-endotoxin, azide-free formulation. mAbs were diluted in PBS (Gibco) to 1.33 µg/µl for i.p. administration. Mice received three 200 µg doses of mAb at 49, 53 and 56 days of age and were monitored daily for potential reactions.

Anti-CD226 mAb blockade validation

For in vitro validation, NOD-Foxp3EGFP splenocytes were resuspended at 0.5 × 106 cells/ml and incubated with 20 µg/ml IgG2a isotype control (BioLegend) or anti-CD226 mAb (BioLegend) for 30 min in cRPMI at 37°C. Cells were stained with Live/Dead Near-IR viability dye (Invitrogen, Waltham, MA, USA) per the manufacturer's protocol. Then, Fc receptors were blocked with ⍺-CD16/CD32 (BD Pharmingen, Franklin Lakes, NJ, USA; RRID: AB_394656) for 5 min at 4°C to prevent non-specific binding before extracellular staining with an antibody cocktail consisting of anti-mouse CD4-PerCP-Cy5.5 and CD8⍺-BV711, and a fluorophore-conjugated anti-CD226 mAb (clone TX42.1), for 30 min at 23°C (clone, RRID, concentration and manufacturer information are provided in electronic supplementary material [ESM] Table 1). Data were collected on an Aurora 5L (16UV-16V-14B-10YG-8R) spectral flow cytometer (Cytek, Freemont, CA, USA) and analysed using FlowJo software (TreeStar, Ashland, OR, USA; version 10.6.1).

T cell proliferation assays

Whole splenocytes isolated from 8- to 12-week-old female NOD mice were labelled with 5 µmol/l CellTrace Violet (CTV; Thermo Fisher, Waltham, MA, USA) as recommended by the manufacturer's protocol. Following proliferation dye staining, cells were resuspended at 0.5 × 106 cells/ml and incubated with 20 µg/ml IgG2a isotype control (BioLegend) or anti-CD226 mAb (BioLegend) for 30 min in cRPMI at 37°C. Then, 0.25 × 106 cells were stimulated with either plate-bound ⍺-CD3 (BioLegend; RRID: AB_11149115; 2 µg/ml) and plate-bound ⍺-CD28 (BioLegend; RRID: AB_11147170; 1 µg/ml) or plate-bound ⍺-CD3 (BioLegend; 2 µg/ml) and plate-bound CD155-Fc (BioLegend; 1 µg/ml). Following 96 h of culture, supernatants were stored at −20°C for ELISA as described below, and cells underwent viability staining and ⍺-CD16/CD32 blocking, as described above, followed by surface staining with anti-mouse CD4-PerCP-Cy5.5 and CD8⍺-BV711 for 30 min at 23°C (ESM Table 1). Data were collected on a Cytek Aurora 5L spectral flow cytometer. The detailed gating strategy is shown in ESM Fig. 1. The proliferation of CD4+ and CD8+ T cells was established by the division index (DI) method using proliferation modelling on FlowJo software (TreeStar; version 10.6.1).

ELISA

To determine whether anti-CD226 mAb blockade modulates secretion of the proinflammatory IFN-γ or anti-inflammatory IL-10 cytokines, ELISAs were performed on culture supernatants from the in vitro T cell proliferation assay described above. Briefly, culture supernatants were diluted 1:2 for IL-10 and 1:100 for IFN-γ, and ELISAs were performed using the Mouse IL-10 and Mouse IFN-γ OptEIA Kits (BD Biosciences, Franklin Lakes, NJ, USA) according to the manufacturer's protocol. Colorimetric analyses were performed in duplicate at 450 nm with a λ correction at 570 nm on a SpectraMax M5 microplate reader (Molecular Devices, San Jose, CA, USA).

Flow cytometry

We used 1–2 × 106 cells from each tissue for flow cytometry and these cells were stained with Live/Dead Near-IR viability dye. Cells were blocked with ⍺-CD16/CD32 (BD Pharmingen) and Brilliant Stain Buffer (BD Biosciences) for 5 min at 4°C before extracellular staining with an antibody cocktail consisting of anti-mouse CD4-PerCP-Cy5.5, CD8-BV711, CD25-Alexa Fluor 700, CD44-PE, CD62L-APC and CD226-BV650 for 30 min at 23°C (ESM Table 1). Next, cells were fixed and permeabilised using the eBioScience Foxp3 Transcription Factor Staining Buffer Set (Invitrogen) according to the manufacturer's instructions. Permeabilised cells were stained with anti-mouse forkhead box protein P3 (FOXP3)-Alexa Fluor 488, Helios-Pacific Blue and Ki-67-PE-Cy-7 antibodies overnight at 4°C (ESM Table 1). Data were collected on a Cytek Aurora 5L spectral flow cytometer and analysed using FlowJo software (TreeStar; version 10.6.1). Gating strategies were determined using fluorescence-minus one (FMO) and unstained controls. Detailed gating strategies for Tregs and T cell memory subsets are shown in ESM Fig. 2.

Anti-CD226 mAb blockade persistence

To assess the in vivo persistence of anti-CD226 mAb, 1–2 × 106 splenocytes underwent viability staining and ⍺-CD16/CD32 blocking, as previously described, before surface staining with anti-mouse CD4-PerCP-Cy5.5, CD8-BV711 and NKp46-Alexa Fluor 647 and anti-rat IgG2a-FITC for 30 min at 23°C (ESM Table 1). Flow cytometry data were collected and analysed as described above. The gating strategy to assess anti-CD226 mAb persistence is shown in ESM Fig. 3.

Histology

Pancreases were collected from 12-week-old mice at necropsy and fixed overnight in a buffered 10% formalin solution. Samples underwent paraffin-embedding, and three sections (250 µm steps) were obtained for H&E staining. Digital whole-slide scans of pancreas sections were obtained using an Aperio CS Scanner (Leica Biosystems, Wetzlar, Germany). Two blinded observers completed insulitis scoring for at least 45 islets per mouse (with one islet defined as >10 endocrine cells) according to previously published guidelines [27].

Intervention study

Beginning at 7 weeks of age, concurrent with the initiation of mAb administration, blood glucose levels of female NOD mice were monitored weekly until 30 weeks of age, using an AlphaTrak glucometer (Zoetis, Parsippany, NJ, USA) to measure samples obtained from a tail vein bleed. Mice with blood glucose ≥13.9 mmol/l were retested the following day, and those with two consecutive blood glucose levels ≥13.9 mmol/l were diagnosed with autoimmune diabetes. Body mass measurements were recorded weekly to monitor for possible impacts of treatment on the growth and overall health of the animals. Mice were humanely euthanised by CO2 asphyxiation and cervical dislocation at diabetes onset or study conclusion.

p-STAT5 Phosflow assay

To determine whether mAb blockade alters the janus kinase 2 (JAK2)–signal transducer and activator of transcription 5 (STAT5) pathway in Tregs, we quantified phosphorylated STAT5 (p-STAT5) levels following stimulation with recombinant human IL-2 (rhIL-2; Roche, Basel, CH-BS, Switzerland; [28]). Single-cell suspensions of splenocytes were obtained from 12-week-old NOD mice 5 weeks after in vivo treatment with IgG2a isotype control or anti-CD226 mAb, as described above. Cells were plated at 1 × 106 cells/ml cRPMI. They were stimulated with 10 IU/ml of rhIL-2 for 0, 15 or 60 min before immediately fixing cells with warmed CytoFix Buffer (BD Biosciences) for 10 min at 37°C. Following fixation, cells underwent viability staining, as described above, before permeabilising cells for 30 min at 4°C with chilled Perm Buffer III (BD Biosciences). Following Fc blocking with anti-CD16/32, cells were stained with anti-mouse CD4-PE-Cy7, CD8-BV711, Foxp3-AF488 and p-STAT5-AF647 overnight at 4°C (ESM Table 1). Flow cytometry data were collected and analysed as described above.

Suppression assays

To assess the potential impact of anti-CD226 mAb blockade on the suppressive capacity of mouse Tregs, we conducted in vitro suppression assays [29] using serial dilutions of fresh splenic Tregs with irradiated autologous whole splenocytes and autologous CD4+ Tconvs that respectively served as antigen-presenting cells (APCs) and responder T cells (Tresps). Briefly, whole splenocytes were isolated from 8- to 12-week-old female NOD-Foxp3EGFP mice. The splenocyte suspension was split so that half of the cells received 137Cs gamma irradiation at a dose of 3000 centigrays (cGy), after which cells were concentrated at 1.0 × 106 cells/ml cRPMI, and 50,000 cells were added to each well of a 96-well U-bottom plate to provide APC-mediated stimulation alongside soluble ⍺-CD3 (BioLegend; 0.5 µg/ml) and ⍺-CD28 (BioLegend; 0.5 µg/ml).

Concurrently, Treg and Tconv cells were isolated from the remaining splenocytes by enriching for CD4+ T cells using the EasySep Mouse CD4+ T Cell Isolation Kit (StemCell, Vancouver, BC, Canada) according to the manufacturer's instructions. CD4+ T cell-enriched splenocytes were stained with anti-CD4-PE-Cy7 (ESM Table 1) to allow for both Foxp3+ Treg and Foxp3− Tconv isolation by sorting for CD4+GFP+ and CD4+GFP− cells, respectively, using a FACSMelody Cell Sorter (Becton Dickinson, Franklin Lakes, NJ, USA; ESM Fig. 4). Following isolation, Tconvs were labelled with CTV, as described above, and concentrated at 1 × 106 cells/ml cRPMI, and 50,000 cells were added to each well. Before co-culture, Tregs were split and incubated at a concentration of 1 × 106 cells/ml cRPMI for 30 min at 37°C in cRPMI with 10 µg/ml anti-CD226 mAb (BioLegend) or IgG2a isotype control (BioLegend). After incubation, cells were washed with cRPMI to remove any unbound mAb, then co-cultured with the autologous Tconvs at the following Treg:Tresp ratios: 1:1, 1:2, 1:4, 1:8, 0:1. Following 72 h of co-culture, cells underwent viability dye staining, Fc blocking with anti-CD16/32, surface staining with anti-mouse CD4-PerCP-Cy5.5 and flow cytometry data collection as previously described (ESM Table 1). The detailed gating strategy is shown in ESM Fig. 5. Percentage suppression of CD4+GFP− Tresps was established by the DI method using proliferation modelling on FlowJo (TreeStar; version 10.6.1).

IGRP206–214 tetramer staining

Single-cell suspensions obtained from collagenase-digested pancreas from female NOD mice were assessed for frequency of islet-specific glucose-6-phosphatase catalytic subunit-related protein (IGRP)206–214-reactive CD8+ T cells, 5 weeks following treatment with IgG2a isotype control or anti-CD226 mAb in vivo. Cells underwent viability dye staining and Fc blocking before concurrent surface antibody (anti-mouse CD4-PerCP-Cy5.5 and CD8-BV711; ESM Table 1) and tetramer staining (10 nM IGRP206–214-BUV395 tetramer) for 1 h at 37°C in the presence of 75 nM dasatinib [30]. Flow cytometry data were collected and analysed as described above with a detailed gating strategy shown in ESM Fig. 6.

Chromium-release assay

To evaluate whether anti-CD226 mAb blockade could sufficiently reduce the cytotoxicity of autoreactive CD8+ T cells and subsequently decrease pancreatic beta cell killing, chromium-release assays were performed in the format previously described by Chen et al [31]. Briefly, whole splenocytes isolated from 3- to 4-week-old NOD-AI4 mice, reported to possess diabetogenic insulin-reactive CD8+ T cells [32], were concentrated at 5 × 106 cells/ml in cRPMI and incubated with either IgG2a isotype control or anti-CD226 mAb (BioLegend; 20 µg/ml) for 30 min at 37°C. After incubation, splenocytes were diluted to a concentration of 2 × 106 cells/ml cRPMI and activated over 3 days in the presence of 0.1 µmol/l AI4 mimotope (YFIENYLEL; GenScript, Piscataway, NJ, USA) and Teceleukin rhIL-2 at a concentration of 25 IU/ml cRPMI to selectively expand CD8+ T cells. Following activation, autoreactive CD8+ T cells were transferred into complete DMEM (cDMEM; DMEM [Lonza], 5 mmol/l HEPES [Gibco], 5 mmol/l MEM NEAA [Gibco], 50 µg/ml penicillin 141 [Gibco], 50 µg/ml streptomycin [Gibco], 0.02% BSA [Sigma-Aldrich] and 10% FBS [Genesee Scientific]) and were added to flat-bottom 96-well plates containing pre-seeded 51CrNa2O4-labelled (1.8 × 105 Bq/well; Revvity, Waltham, MA, USA) murine NIT-1 pancreatic beta cells (RRID: CVCL_3561; [33]) at the following effector:target cell ratios: 0:1, 1:1, 2:1, 5:1, 10:1, 25:1. Following 16 h of co-culture, supernatants were removed and transferred to 6 × 50 mm lime glass tubes. Lysates of adherent cells were collected using a 2% SDS wash and transferred into separate tubes. 51Cr activity, measured in counts per minute (CPM), was assessed for both fractions on a Wizard 1470 automatic gamma counter (Revvity). The specific lysis of NIT-1 cells was calculated as follows: \(\%\;\mathrm\;\mathrm\;=\;\mathrm\;\left[\frac\;\mathrm\;\mathrm\right)}\;\mathrm\;\mathrm\right)\;+\;\left(\mathrm\;\mathrm\;\mathrm\right)}\right]\;-\;\mathrm\;\left[\frac\;\mathrm\;\mathrm\right)}\;\mathrm\;\mathrm\right)\;+\;\left(\mathrm\;\mathrm\;\mathrm\right)}\right]\)

Single-cell sequencing

Single-cell suspensions from the pancreas and pLN were obtained from 12-week-old female NOD mice following treatment with IgG2a isotype control (pLN: n=3; pancreas: n=3) or anti-CD226 mAb (pLN: n=5; pancreas: n=2) to prepare sequencing libraries on the 10x Genomics platform. CD3+ T cells were enriched within each sample using the Mouse T cell EasySep kit (StemCell) per the manufacturer’s instructions. Each sample underwent Fc blocking with anti-CD16/32 for 10 min at 4°C before staining with oligonucleotide-tagged antibodies for CD4, CD8⍺, CD44, CD62L and T cell immunoreceptor with Ig and ITIM domains (TIGIT; clone and DNA barcode information is provided in ESM Table 2) for 30 min at 4°C. Following staining, cells were washed three times with PBS + 1.0% BSA before loading 5000 CD3+ enriched cells onto a Chromium Next GEM Chip K (10x Genomics, Pleasanton, CA, USA) to generate Gel Beads-in-Emulsions (GEMs) using a 10x Chromium Controller (10x Genomics). Gene expression (GEX), V(D)J and feature barcode (FB) libraries were generated using the Chromium Next GEM Single Cell 5′ v2 and Chromium Single Cell Mouse TCR Amplification Kits (10x Genomics). All libraries were sequenced using the NovaSeq XPlus Illumina platform with a minimum sequencing depth of >19,214 reads/cell for GEX libraries, >5271 reads/cell for V(D)J libraries and >6053 reads/cell for FB libraries.

Pre-processing of 10x Genomics sequencing data

The Cell Ranger (version 3.0.0; 10x Genomics) multi-pipeline was used to generate raw feature barcode matrices by processing raw sequencing reads from GEX and FB libraries as well as annotated full-length transcripts (contigs) by processing the raw sequencing reads of V(D)J libraries. Briefly, GEX sequencing reads were aligned to a reference genome (mm10-2020-A) using STAR (version v_2.5.1b; [34]). Confidently mapped reads sharing the same unique molecular identifier (UMI), 10x barcode and feature were collapsed, with the number of reads per feature saved in the raw FB matrix. Contigs with V(D)J segment labels were aligned to a reference genome (GRCm38-alts-ensembl-7.0.0) to identify complementarity-determining region 3 (CDR3) sequences. Productive contigs sharing the same UMI and barcode were saved as filtered contig annotations.

Quality control of scRNA-seq and CITE-seq data

Raw FB matrices of cells sequenced from each tissue were imported into R (https://www.r-project.org/; version 4.3.1) using the Read10X function in the Seurat package (version 5.0.1). Live cell-containing droplets were distinguished using gene (80–2000) and protein library size (1.0–3.0) as well as mitochondrial content (<15%). Cellular indexing of transcriptomes and epitopes by sequencing (CITE-seq) data were normalised using the denoised and scaled by background (dsb) method with the DSBNormalizeProtein function, without isotype controls, in the dsb package (version 1.0.3). To remove variation in our single-cell RNA sequencing (scRNA-seq) depth between cells across each sample, we performed normalisation and variance stabilisation using residuals from negative binomial regression with the proportion of mitochondrial genes as a covariate, as described by Choudhary and Satija [35], using the SCTransform (SCT) version 2 function in Seurat with the glmGamPoi method [36]. A principal component analysis (PCA) dimensionality reduction was run on each sample, and the number of statistically significant principal components (PCs) was identified. Using generated artificial doublets integrated into the dataset at a proportion of 0.25, the proportion of artificial nearest neighbours (pANN) was determined for each PC neighbourhood size (pK). Cells with the highest pANN values were identified as predicted doublets and were removed from each sample using the doubletFinder function in the DoubletFinder package (version 2.0.4).

Sample integration and clustering and differential gene expression analysis

To perform an integrative analysis of shared T cell phenotypes across each organ, samples from each mouse were combined into lists (Pancreas: 3× Isotype, 2× anti-CD226 mAb; pLN: 3× Isotype, 5× anti-CD226 mAb) to select 3150 integration features not including T cell receptor (TCR)-α/β variable genes and ribosomal genes before using the SelectIntegrationFeatures function in Seurat. Using the Seurat v4 integration workflow, the SCTransform residuals for each integration feature were used to create integrated Seurat objects for each organ. Cells with related transcriptomic profiles from each organ were clustered using the first 30 PCs in the integrated assay. Briefly, the FindNeighbors function in Seurat was used to generate a nearest neighbour graph before identifying clusters by the Louvain method of community detection with a resolution of 1.0, using the FindClusters function in Seurat. Differential gene expression (DGE) between clusters was evaluated on the RNA assay using a Wilcoxon rank sum test. Faster implementation was achieved using the Presto package (v. 1.0.0), executed with the FindAllMarkers function in Seurat. Clusters with gene expression profiles suggestive of other immune subsets besides T cells (e.g. B lymphocytes, eosinophils), apoptotic cells or epithelial cells were excluded from the dataset using the subset function in Seurat. Following the exclusion of cellular debris and contaminating subsets, cells were re-clustered, as described above, based on the first 13 PCs with a resolution of 0.77 for the pancreas and the first ten PCs with a resolution of 0.7 for the pLN. This yielded 15 T cell clusters from the pancreas with 17,617 cells and 12 from the pLN with 54,229 cells. After excluding contaminating pancreas-specific transcripts [37], differentially expressed genes (DEGs) between treatment conditions within clusters were assessed using the Wilcoxon rank sum test with Bonferroni correction with an adjusted p value ≤0.05 (pLN: ESM Table 3; pancreas: ESM Table 4). DEGs displayed in cluster heatmaps represent aggregated expression per treatment condition of a curated selection of T cell genes. Criteria for selection included presence within the top 50 DEGs per cluster and significant differential expression between treatment groups (false discovery rate [FDR]-adjusted p<0.05) in at least two clusters. Cell cluster annotations were assigned using a priori knowledge and relevant literature based on DGE between clusters (ESM Table 5).

Differential abundance analysis

To identify differences in T cell cluster abundance between treatment conditions, we performed a differential abundance (DA) analysis for both the pancreas and pLN [38, 39]. Briefly, the edgeR package (version 3.42.4) was used to fit negative binomial generalised linear models (GLMs). Quasi-likelihood (QL) dispersions were calculated from GLM deviances before p values were determined using the glmQLFTest function in the edgeR package. The log fold change (FC) was determined based on the DA of anti-CD226 mAb-treated mice normalised to isotype-treated mice.

TCR repertoire analysis and mapping of IGRP-reactive clonotypes

To evaluate differences in TCR repertoires between treatment conditions, the scRepertoire package (version 2.0.0) was used to analyse V(D)J libraries. Briefly, clones were assembled by associating contigs with single-cell barcodes using the combineTCR function before merging the clonal information with the processed scRNA-seq data using the combineExpression function. To identify IGRP-reactive T cells, the TCR-⍺ and TCR-β CDR3 sequences were compared with the CDR3 sequences of NOD IGRP206–124-specific CD8+ T cells previously described by Kasmani et al [40] using the Biostrings package (version 2.68.1). CDR3 sequences within one amino acid mismatch of either the TCR-⍺ or TCR-β chain of the previously published sequences were mapped as IGRP-reactive. Relative clonal expansion was established by determining the proportion each clone represented of the entirety of the TCR repertoire, where X represents the proportion of the repertoire occupied by a single clonotype: medium: 0.001 < X <0.01; small: 1 × 10−4 < X <0.001; rare: 0 < X <1 × 10−4; not applicable (NA): no CDR3 sequence for cell barcode.

Data visualisation and statistical analysis

Single-cell data were processed and visualised using the following R packages: Seurat [41], dsb [42], DoubletFinder [43], Presto [44], edgeR [45], SingleCellExperiment (version 1.22.0; [46]), scRepertoire [47], Biostrings [48], DittoSeq (version 1.12.2; [49]) and EnhancedVolcano (version 1.18.0; [50]). Statistical analyses were performed using Prism software (version 10.3.1; GraphPad, San Diego, CA, USA) for all other data. Unless otherwise stated, flow cytometric data were analysed by two-way ANOVA, and ELISA data were analysed by one-way ANOVA, with Bonferroni’s post hoc test for multiple testing correction. In vitro suppression and chromium-release assay curves were also analysed by two-way ANOVA with Bonferroni’s post hoc test for multiple testing correction, with AUC values compared using paired t tests [51]. A χ2 test was used to compare insulitis severity, and a Fisher’s exact test was used to identify differences in IGRP-reactive pancreatic CD8+ T cells by T cell receptor sequencing (TCR-seq). The logrank (Mantel–Cox) test was used to identify significant differences in disease incidence between treatment groups and to calculate an HR. We considered p values ≤0.05 significant. For experiments where intra-sample pairing could not be accomplished (i.e. in vivo and ex vivo analyses), cages were randomised using simple randomisation to assign treatment groups.