The IMIDIA cohorts
We collected pancreatic specimens from two cohorts (Fig. 1a and ESM Table 6). One cohort consisted of 243 OD, including 204 non-diabetic and 39 with type 2 diabetes. As expected, blood fructosamine, a biomarker for glucose levels in the days preceding organ donation, was greater in type 2 diabetic (222 ± 72 μmol/l, n = 11) than in non-diabetic OD (180 ± 45 μmol/l, n = 46, p = 0.018). The second cohort included 201 PPP who underwent pancreatectomy for pancreatic diseases. Among PPP, 70 were non-diabetic, 54 had type 2 diabetes, 30 had IGT, and 46 had diabetes that was likely due to the pancreatic disorder leading to surgery (type 3c diabetes)  (see ESM Methods). A diagnosis of type 3c diabetes was made if diabetes was first detected <1 year before the symptoms, which led to surgery. Histopathology of the resected tissue did not reveal insulitis in any PPP. Assessment of insulin secretion by OD islets showed reduced release from type 2 diabetic beta cells in response to glucose or glibenclamide (known as glyburide in the USA and Canada), but not to arginine (ESM Fig. 1), complementing previous findings .
Islets were isolated from 141 OD (115 non-diabetic and 26 type 2 diabetic), and 117 PPP (37 non-diabetic, 16 IGT, 41 type 2 diabetic, 23 type 3 diabetic) by enzymatic digestion or LCM. Following filter selection (see ESM Methods) the transcriptomes of islets from 103 OD (84 non-diabetic and 19 type 2 diabetic), and 103 PPP (32 non-diabetic, 36 type 2 diabetic, 15 IGT, 20 type 3c diabetic) were profiled (Table 1). Hence, in total we profiled the islet transcriptomes of 116 non-diabetic OD (84) and PPP (32) and of 55 type 2 diabetic OD (19) and PPP (36). Between the OD and PPP groups, the non-diabetic and type 2 diabetic patients were similar in terms of their age, BMI, and mean duration of diabetes (Table 1). Among non-diabetic and type 2 diabetic PPP, the prevalence of chronic pancreatitis and of benign/malignant tumours was also similar (Table 1).
Differentially expressed genes in type 2 diabetic islets
In total, 4438 out of 29,529 probe sets, corresponding to 2976 unique genes, were differentially expressed in type 2 diabetic vs non-diabetic OD islets (false discovery rate [FDR] ≤0.05). Overall, the expression of 608 probe sets, corresponding to 444 gene annotations (of which 421 were regulated only in OD and not in PPP; ESM Table 7) changed ≥1.5 fold (Fig. 2a and GEO, accession number: GSE76896). Among the top 20 regulated genes, 18 were downregulated, and two were upregulated (ESM Table 8). A total of 1439 of 29,612 probe sets were differentially expressed in type 2 diabetic compared with non-diabetic PPP islets (FDR ≤0.05), corresponding to 1039 unique genes. Overall, the expression of 208 probe sets, corresponding to 136 gene annotations (of which 113 were regulated only in PPP and not in OD; ESM Table 7), changed ≥1.5 fold (Fig. 2a and GEO, accession number: GSE76896). Among the top 20 expressed genes, 12 were downregulated and eight upregulated (ESM Table 8). In type 2 diabetic OD islets, 69% (307/444) of the differentially regulated genes were downregulated, whereas 62% (84/136) of the differentially regulated genes were upregulated in type 2 diabetic PPP islets (Fig. 2a). Exocrine and ductal pancreatic markers were comparably low in OD and PPP islets (see ESM Table 5). Furthermore, islets isolated enzymatically from OD and PPP clustered together by principal component analysis (PCA) and separately from the cluster of islets isolated by LCM from the same OD and PPP (Fig. 1b, c), suggesting the influence of the isolation procedure (enzymatic for OD and LCM for PPP) rather than differences between OD and PPP. Comparing differentially expressed genes with pancreatic cancer transcriptomic signatures (see ESM Results) we found no evidence for contamination of PPP samples with cancer cells .
Dysregulated genes common to type 2 diabetic OD and PPP islets
To identify the most reproducible transcriptomic changes in type 2 diabetic islets independent from covariates such as islet retrieval procedure, tissue source and/or collecting centre, we focused on genes significantly dysregulated in type 2 diabetic islets in both cohorts. This allowed us to identify 23 genes with an FDR of ≤0.05 and a fold change of ≥1.5, of which 19 were dysregulated in the same direction in both type 2 diabetic OD and PPP islets (Table 2 and Fig. 2a). The reason for the regulation in opposite directions of DAB1, GAP43, PDK4 and RGS16 in type 2 diabetic OD and PPP islets is unclear. Fifteen genes were downregulated, including SLC2A2, ARG2, CHL1, PPP1R1A, TMEM37 (Fig. 2b–f), G6PC2 and CAPN13, while four were upregulated (KCNH8, FAM102B, FBXO32 and CD44). These 19 genes were correlated with stimulated insulin secretion from OD islets (ESM Fig. 2). Notably, nine of these genes, namely ANKRD23/39, ASCL2, HHATL, NSG1, PCDH20, SCTR, CD44, FAM102B and FBXO32 have not been previously reported to be dysregulated in type 2 diabetes.
Meta-analysis of a published transcriptomic dataset  revealed dysregulation (FDR ≤0.05) of 114 probe sets, corresponding to 94 unique genes. We searched this dataset for the 19 genes dysregulated in both our islet cohorts and found that CHL1, FFAR4 and SLC2A2 were downregulated with an FDR of ≤0.05 and a fold change of ≥1.5, while ANKRD23/39, ARG2, HHATL, PPP1R1A and UNC5D were downregulated with an FDR of ≤0.1 in type 2 diabetic islets, thus confirming the significant downregulation of CHL1, FFAR4 and SLC2A2 in type 2 diabetic OD islets in a different cohort.
Dysregulated genes in IGT and type 3c diabetic islets
The availability of islets from PPP with IGT or type 3c diabetes (Table 1) allowed us to investigate the expression of the 19 genes commonly dysregulated in type 2 diabetic islets according to the extent of hyperglycaemia (Fig. 2g). The expression of the genes differed significantly only between type 2 diabetic and non-diabetic PPP islets (FDR ≤0.05). Nonetheless, the fold changes between type 3c diabetic and non-diabetic PPP islets were in the same direction, albeit with smaller differences than between type 2 diabetic and non-diabetic PPP islets. The fold changes in IGT vs non-diabetic PPP islets were low (≤1.6); only three of the 19 genes had absolute fold changes >1.2. These diversities between islets from type 2 diabetics (both OD and PPP), IGT (PPP) and type 3c diabetics (PPP) may be due to idiosyncrasies of these conditions and/or different duration and severity of the hyperglycaemia.
Validation of selected genes
Some (SLC2A2, CHL1, PPP1R1A, ARG2 and TMEM37), but not all, of the 19 differentially expressed genes in type 2 diabetic OD and PPP islets were previously shown to be enriched in beta cells and altered in type 2 diabetes . Among the 19 genes dysregulated in type 2 diabetic islets, ARG2 and PPP1R1A were also differentially expressed in non-diabetic OD islets exposed ex vivo to high glucose (22.2 mmol/l) for 48 h, while CHL1, FBX032 and SLC2A2 showed a trend towards dysregulation (ESM Table 9). These results, although obtained with relatively few preparations (n = 3), suggest that the expression of several of the 19 signature genes changes within a relatively short time span upon islet exposure to a ‘hyperglycaemic’ milieu. Confirmation with more samples will be required to better ascertain the precise regulation of these genes in islets upon glucose treatment.
Consistent with recent RNA sequencing data of sorted adult beta cells (n = 7, non-diabetic cells) , in situ PCR on human pancreas sections confirmed islet expression of the 19 type 2 diabetes signature genes (ESM Fig. 3 shows images for three representative genes). As proof of principle, we verified protein expression and localisation of ARG2, PPP1R1A and TMEM37 in human pancreas. ARG2 was detected in a subset of insulin-positive and glucagon-positive cells (Fig. 3a, ESM Fig. 4). Conversely, PPP1R1A was co-localised with insulin-positive cells, but its expression was weaker in glucagon-positive cells. TMEM37 was co-localised with insulin-positive cells and some glucagon-positive cells. Analysis of islet alpha and beta cell-enriched fractions from five non-diabetic and four type 2 diabetic OD by RT-qPCR showed that ARG2, PPP1R1A and TMEM37 were enriched in beta cells and downregulated in type 2 diabetes (Fig. 3b–d).
Silencing of either Arg2 (ESM Fig. 5a) or Ppp1r1a (ESM Fig. 5b) in insulinoma INS-1 832/13 cells reduced insulin secretion (Fig. 3e, f), while the opposite was observed by silencing Tmem37 (Fig. 3g, ESM Fig. 5c), consistent with the latter being an inhibitory subunit of voltage-gated calcium channels . Accordingly, Tmem37 downregulation increased the proportion of cells with elevated intracellular Ca2+ ([Ca2+]i) concentrations (ESM Fig. 5d, e), and the peak [Ca2+] amplitudes (Fig. 3h, ESM Fig. 5f) after exposure to high glucose. Overexpression of mTmem37-V5 (ESM Fig. 5g, h) reduced insulin release (Fig. 3i) and [Ca2+]i under basal conditions (ESM Fig. 5i) or in response to high glucose or potassium (Fig. 3j). These data suggest that insulin secretion is inhibited by downregulation of ARG2 and PPP1R1A and enhanced by downregulation of TMEM37.
Distinct sets of upstream and downstream pathways are predicted for the OD and PPP cohorts
Among the enriched pathways related to dysregulated genes identified in type 2 diabetic islets (Fig. 4), several were previously found to influence beta cell function. ‘maturity onset diabetes of young (MODY) signalling’ and ‘neuropathic pain signalling’ were the only two pathways in common among the top 20 identified in each cohort separately. Differentially expressed probe sets separated for upregulation and downregulation were analysed for enriched gene ontologies. Interestingly, similar biological processes were enriched for downregulated probe sets in OD and PPP islets (ESM Table 10) with a strong focus on hormone secretion (ESM Table 11). Analysis of downstream functions using a literature-based prediction method  revealed a decrease in processes controlling cAMP concentrations, neurotransmitter release and synaptic transmission in OD islets, while pathways related to numbers of beta cells, islet cells and neuroendocrine cells were mostly affected in PPP islets (ESM Table 12).
The same method was used to identify putatively activated or inhibited upstream regulators of the regulated genes. For OD islets, the highest inhibition scores were found for ADCYAP1, NEUROD1, BDNF and PAX6 (ESM Table 13), supporting altered differentiation of beta cells during development and/or function and viability. Significant activation scores were found for FOXO1 and, in particular, REST, two transcriptional regulators, which preclude the differentiation of beta cells and/or the retention of their identity [32, 33]. HNF1A was the only transcription factor predicted to be significantly inhibited in PPP islets.
Analysis of potential upstream regulators revealed key transcription factors involved in beta cell dysfunction
Gene co-expression modules were identified for non-diabetic OD and PPP islets. Modules that significantly overlapped between OD and PPP were then correlated with clinical or functional traits (see ESM Methods). This identified a set of ten modules in which 14 out of 19 differentially expressed signature genes were enriched (hypergeometric p = 3.34 × 10−5); only CAPN13, FFAR4, NSG1, FAM102B and KCNH8 were absent from the selected modules. To investigate whether genes within modules share the same transcriptional control, we identified putative upstream transcription factors using two complementary bioinformatics approaches (Fig. 5a). The first used a literature-based prediction method , while the second used enrichment of predicted transcription factor binding sites in the promoter sequences of the module genes . The results were then used to generate two networks, one for each analysis, describing predicted upstream transcription factor interactions with their predicted target genes. The networks were merged into a single network containing 17 upstream transcription factors and 29 transcription factor–target gene interactions predicted by both approaches (Fig. 5b, c). Several modules were correlated with insulin and blood glucose levels (Fig. 5c and ESM Figs 6–8).
Of note, three of the 19 differentially expressed signature genes (PPP1R1A, SLC2A2 and CD44) were present in this network and were potential targets of PDX1. We therefore hypothesised that PDX1 and other transcription factors in the network regulate the differentially expressed genes in type 2 diabetes. The volcano plots in Figs 5d and e show the 17 transcription factors in OD and PPP islets, respectively. REST, FOXA1 and SP1 were upregulated and HLF was downregulated in OD islets. Meanwhile, PDX1, HNF1A and HLF were downregulated in PPP islets. HNF1A tended to be downregulated in OD islets. Other than HLF, these transcription factors, are known regulators of beta cell differentiation and function.
Validation of upstream regulators
Among the transcription factors identified above, PDX1, HNF1A and HLF were chosen for further analyses. RT-qPCR assays of human islet alpha and beta cell-enriched fractions confirmed that expression of PDX1 and HNF1A was reduced in type 2 diabetic beta cells (Fig. 6a, b). Although HLF was enriched in non-diabetic beta cells, its expression was not altered in type 2 diabetes (Fig. 6c), and its role was not further investigated. Both PDX1 and HNF1A were detected in the nucleus of EndoC-βH1 cells  (Fig. 6d). Silencing of HNF1A (Fig. 6e, ESM Fig. 9a, b) reduced mRNA levels of SLC2A2, PPP1R1A and TMEM37 (Fig. 6f). While SLC2A2 is a known target of HNF1A [36, 37], PPP1R1A and TMEM37 are not predicted to include a binding site for HNF1A within 5 kb upstream or downstream of their transcription start site (ESM Methods and ESM Table 14). Silencing of PDX1 reduced SLC2A2 mRNA levels but increased those of ANKRD23 and ARG2 (Fig. 6f). All three genes include one or more putative binding sites for PDX1. Chromatin immunoprecipitation assays did not provide evidence for binding of HNF1A to the promoters of UNC5D, FAM102B and CD44, while the promoter regions of ARG2 and SLC2A2, the latter being a well-established PDX1 target, were recovered with PDX1 (Fig. 6g).