Immunohistochemistry was performed using a standard immunoperoxidase approach, as previously described [21]. To examine multiple antigens within the same FFPE section, samples were probed in a sequential manner with up to three different antibodies (ESM Tables 3, 4). The mean fluorescence intensity (MFI) of stained antigens was measured using ImageJ Version 1.50b Java 1.8.0_77; https://imagej.nih.gov/ij/download.html. Some slides were processed with isotype control antisera to confirm the specificity of labelling (ESM Fig. 1). Frozen sections were stained using a standard immunofluorescence approach [22].

Using the Affymetrix Human Gene 2.0 ST array, CEL files were generated from both control and type 1 diabetic donors, as previously described [24]. Raw signal-intensity values from Affymetrix spike-in controls demonstrated that array hybridisation had been successful (i.e. bioB<bioC<bioD<Cre). Data quality was verified by measuring the positive vs negative area under the curve. Raw signal-intensity values from all arrays were robust multichip average background corrected, quantile normalised, median polish summarised and log2 transformed [25–27]. NetAffx-determined probe-set annotations for HLA genes (Affymetrix) were re-mapped according to RefSeq, release 73 (15 November 2015; see ftp://ftp.ncbi.nlm.nih.gov/refseq/release/release-catalog/archive/). For each HLA gene, where multiple mappings were possible (i.e. HLA-A, -B, -C and -F), probe sets were annotated according to eight major haplotypes incorporated into the human genome assembly, as previously described [28]. Because probe sets shared mappings, it was not possible to identify HLA subtypes uniquely using this gene chip; rather, transcript clusters were used to examine changes in global gene expression. Processing was carried out using the Partek Genomics Suite, version 6.5 (Partek, St Louis, MO, USA). The resulting normalised expression data for specific genes of interest were then subjected to analysis as described below.