Epigenetic landscapes in human pancreatic islets reveal distinct drivers for adaptation to age and type 2 diabetes Authors: Lucas Maurin, Lorella Marselli, Mathilde Boissel, Vincent Pascat, Mara Suleiman, Emma Henriques, Carmela De Luca, Lijiao Ning, Marie Fourcot, Marta Tesi, Bénédicte Toussaint, Souhila Amanzougarene, Mehdi Derhourhi, Frédérik Oger, Frederic Burdet, Mickael Canouil, Miriam Cnop, Mark Ibberson, Amélie Bonnefond, Piero Marchetti, Philippe Froguel, Amna Khamis Published: 2026 ABSTRACT Age is the strongest risk factor for type 2 diabetes, yet their independent contribution to pancreatic islet dysfunction remains unclear. We integrate DNA methylation, transcriptomic, and genotyping data from 144 islet donors. We identify 996 age- and 902 T2D-associated CpGs with minimal overlap, and 251 age- and 310 diabetes CpG target genes, usually distant from the CpG. Age-linked CpGs are enriched in promoters, form co-regulated gene modules, link to beta-cell function, including insulin secretion. Diabetes-associated CpGs are enriched in enhancer/non-regulatory regions, and modules suggest stress-induced epigenetic drift. CpG-gene associations are independent of genetic variation. Mendelian randomisation supports a causal role for age-associated CpGs regulating KLHL42, a T2D GWAS locus. A blood-based methylation risk score based on age-linked CpGs correlates with insulin secretion and improves diabetes classification when combined with genetic risk (AUC = 0.91). Altogether, age is associated with a coordinated epigenetic programme, whereas diabetes links to a heterogeneous, stress-related epigenetic signature. Introduction Age is the strongest known risk factor for type 2 diabetes (T2D)1. T2D is characterized by progressive dysfunction of insulin-producing pancreatic islet cells2,3, in a context of whole-body insulin resistance. Both genetic predisposition (nature) and non-genetic influences, including environmental exposures, aging, and T2D itself, contribute to progressive islet cell dysfunction, presumably in part through epigenetic mechanisms (nurture), driving both the onset and progression of the disease and its associated complications1,4. Among epigenetic mechanisms, DNA methylation is the most widely studied and has been shown to play a crucial role in gene regulation5. Epigenetic clocks, based on selected CpG methylation measurement, are used to accurately predict chronological age6,7. Deviation of epigenetic clocks from chronological age has been linked to mortality and age-related disease incidence, including metabolic disorders8–11. Epigenome-wide association studies (EWAS) have consistently identified age-associated CpGs in or near the FHL2, ELOVL2, and KLF14 genes, across multiple tissues, including human pancreatic islets12–14. However, whether these age-linked methylation changes are adaptive, deleterious, or context-dependent remains unclear, particularly with regard to the progression to T2D. Multiple challenges currently hinder the translation of epigenetic findings into mechanistic insight. First, most differentially methylated CpGs lie in non-coding regions, complicating causal gene identification. Second, DNA methylation is highly tissue-specific15,16. Third, the contribution of epigenetic regulation is often masked by underlying genetic variation. Several studies have explored islet methylomes in the context of age or T2D independently4,12, yet none have integrated these data with gene expression and genetic variation to dissect the functional consequences of methylation changes. In particular, the extent to which age- and T2D-associated CpGs converge (or not) on shared molecular targets and contribute to β-cell adaptation to age or to dysfunction in T2D is largely unknown. In theory, modifying DNA methylation is a promising target for prevention or treatment of T2D, given its dynamic nature and responsiveness to environmental and metabolic cues. In this study, we integrated DNA methylation, RNA sequencing (RNA-seq), and genotyping data from islets obtained from 144 organ donors, of which 21 had T2D. By adjusting for genetic effects, we aimed to isolate methylation-driven regulatory changes associated with age, T2D, and modulating β-cell function. Results Epigenome-wide association studies reveal distinct age- and T2D-associated methylation profiles Pancreatic islet DNA methylation was profiled using the Illumina MethylationEPIC array in all individuals, and islet RNA-seq and genotyping data were generated from a subset of individuals. We used the insulin stimulation index (ISI), an indicator of β-cell insulin secretory response to glucose, as a functional proxy of islet performance (Table 1).Table 1Clinical characteristics of cohortCharacteristicN = 144aAge (years)73 (61, 79)SexFemale60 (42%)Male84 (58%)Body mass index (kg/m2)24.5 (22.9, 26.9)Diabetes statusNon-diabetic123 (85%)Type 2 diabetes21 (15%)Insulin stimulation index (µU/mL)2.50 (1.70, 3.50)Islet purity70 (60, 80)an (%); Median (IQR). Table 1Clinical characteristics of cohortCharacteristicN = 144aAge (years)73 (61, 79)SexFemale60 (42%)Male84 (58%)Body mass index (kg/m2)24.5 (22.9, 26.9)Diabetes statusNon-diabetic123 (85%)Type 2 diabetes21 (15%)Insulin stimulation index (µU/mL)2.50 (1.70, 3.50)Islet purity70 (60, 80)an (%); Median (IQR). Clinical characteristics of cohort an (%); Median (IQR). We first performed two independent EWAS to identify CpG sites significantly associated with age and T2D status. For age, we adjusted for sex, BMI, islet purity, T2D status, and cellular composition. We identified 996 CpG sites significantly associated with age (Bacon-corrected FDR < 0.05), of which the majority (85%) were hypermethylated (Fig. 1A; Supplementary Data 1). Importantly, the average methylation of these CpGs increased with advancing age, suggesting a global shift towards hypermethylation with age (P < 2 × 10−16; Fig. 1C). Notably, we replicated 60% of CpG sites from a previously published pancreatic islet EWAS for age12, all of which were consistent in direction of effect, demonstrating the reproducibility of methylation patterns in age (Supplementary Table 1). In contrast, replication rates relative to published CpG sets from blood17, liver13, and skeletal muscle18 EWAS were substantially lower (7.9%, 1.9%, and 0.3%, respectively; Supplementary Data 1–5).Fig. 1Overview of DNA methylation patterns associated with age and T2D in human pancreatic islets.Volcano plots showing epigenome-wide association study (EWAS) results for (A) age (n = 144, 60 females and 84 males) and (B) T2D (n = 120, 52 females and 68 males). The x-axis displays the effect estimates from the linear model, and the y-axis shows the corresponding two-sided −log10(P values). P-values were adjusted by false discovery rate (FDR) correction. Top CpGs are annotated, based on their annotated gene. When no gene is annotated the CpG is labelled. In A, orange points indicate FDR-significant CpGs and purple points indicate non-significant CpGs. In B, turquoise points indicate FDR-significant CpGs and blue points indicate non-significant CpGs. C Association between chronological age and the average methylation level (beta value) across all age-associated CpGs assessed using a linear model (n = 144, P-value < 2 × 10−16). D Boxplot comparing mean methylation levels across T2D-associated CpGs in non-diabetic (ND, n = 99) versus T2D (n = 21) individuals assessed using a linear model (P-value = 1.51 × 10−11). Centre line, median; box limits, 25th and 75th percentiles; whiskers, 1.5×interquartile range, all points are shown. Density plots showing the distribution of average methylation percentage across all (E) age-associated and (F) T2D-associated CpGs. Colour bars represent methylation bins (2% per bin), with the colour gradient indicating methylation levels (from low to high %). The red line in E, F shows a density smoothing curve. Source data are provided as a Source Data file. Overview of DNA methylation patterns associated with age and T2D in human pancreatic islets. Volcano plots showing epigenome-wide association study (EWAS) results for (A) age (n = 144, 60 females and 84 males) and (B) T2D (n = 120, 52 females and 68 males). The x-axis displays the effect estimates from the linear model, and the y-axis shows the corresponding two-sided −log10(P values). P-values were adjusted by false discovery rate (FDR) correction. Top CpGs are annotated, based on their annotated gene. When no gene is annotated the CpG is labelled. In A, orange points indicate FDR-significant CpGs and purple points indicate non-significant CpGs. In B, turquoise points indicate FDR-significant CpGs and blue points indicate non-significant CpGs. C Association between chronological age and the average methylation level (beta value) across all age-associated CpGs assessed using a linear model (n = 144, P-value < 2 × 10−16). D Boxplot comparing mean methylation levels across T2D-associated CpGs in non-diabetic (ND, n = 99) versus T2D (n = 21) individuals assessed using a linear model (P-value = 1.51 × 10−11). Centre line, median; box limits, 25th and 75th percentiles; whiskers, 1.5×interquartile range, all points are shown. Density plots showing the distribution of average methylation percentage across all (E) age-associated and (F) T2D-associated CpGs. Colour bars represent methylation bins (2% per bin), with the colour gradient indicating methylation levels (from low to high %). The red line in E, F shows a density smoothing curve. Source data are provided as a Source Data file. For T2D, we studied a subset of 120 individuals (21 with T2D), matched for age (P = 0.40), sex (P = 0.13), and BMI (P = 0.11; Supplementary Table 2). Expectedly, ISI was significantly decreased in T2D donor islets as compared to non-diabetic donor islets (P < 0.001) (Supplementary Table 2). We performed an EWAS with T2D status, and adjusted for age, sex, BMI, islet purity, and cellular composition (see Methods). We found 902 significantly associated CpG sites (Bacon-corrected FDR < 0.05; Supplementary Data 6). In contrast to age, T2D-associated CpGs were mostly hypomethylated (95%) (Fig. 1B; Supplementary Data 6), in keeping with an earlier study4. We replicated 395 CpG sites from a previously published EWAS in pancreatic islets with T2D4, of which 99.4% were consistent in direction of effect (Supplementary Data 7). Although these overlapping CpGs constitute a small fraction of the much larger set reported by Rönn et al. (~14,000 CpGs, i.e., 0.2%). We found that average methylation across T2D-associated CpGs was lower in individuals with T2D compared to controls (P < 1.51 × 1011; Fig. 1D). Only five CpGs overlapped between T2D and age. Given that age and T2D were included as covariates in the respective models, such limited overlap is expected. To assess this, we repeated each EWAS without adjusting for the other trait. This resulted in only a modest increase in overlap (nine shared CpGs), indicating that the limited intersection is not driven by covariate adjustment but reflects largely independent age- and T2D-associated epigenetic signatures in pancreatic islets. Overall, age-associated CpGs were predominantly located in lowly methylated regions (74.7% of CpGs at 0–25% methylation) (Fig. 1E), whereas T2D-associated CpGs were mostly located in highly methylated regions (63.7% of CpGs at 75–100% methylation) (Fig. 1F). Age-associated CpGs were predominantly located within CpG islands (63%), whereas T2D-associated CpGs were largely enriched in open sea regions (79%) (Supplementary Fig. 1A–B). With respect to gene context, age-associated CpGs were more frequently located in promoter-proximal regions (TSS200 and TSS1500), while T2D-associated CpGs were predominantly mapping to gene bodies (Supplementary Fig. 1C–D). Integration with the pancreatic islet regulome19 further revealed that age-associated CpGs were enriched in active promoters (40%), whereas T2D-associated CpGs were more frequently located in unannotated chromatin states (54%) (Supplementary Fig. 1E–F). Together, these findings reveal that age and T2D imprint distinct, non-overlapping epigenetic signatures onto the human islet methylome. Multi-omic data integration reveals CpGs linked to nearby gene expression To identify the transcriptional consequences of age- and T2D-associated DNA methylation changes, we performed cis-expression quantitative trait methylation (eQTM) analysis in pancreatic islets using matched DNA methylation and RNA-seq data from the same samples (n = 120). We tested associations between each CpG and the expression of nearby genes within ±1 Mb, while adjusting for sex, BMI, islet purity, cellular composition, and including genotype (for each genetic variant within 1 Mb) to account for the putative influence of single nucleotide polymorphisms (SNPs) on methylation-mediated regulation20 (Fig. 2A). All age or T2D CpGs overlapping with methylation QTLs (mQTLs) were excluded to pinpoint CpG effects independent of genetic variation.Fig. 2Multi-omic integration identifies CpG–gene regulatory relationships in pancreatic islets.A Schematic overview of the analytical workflow. DNA methylation levels were measured in pancreatic islets from 144 human donors using Illumina EPIC arrays (EWAS). Transcriptomic profiles were generated by RNA-seq, and genetic variants were profiled using Omni2.5 M arrays. Integration of CpG-gene pairs was performed through expression quantitative trait methylation (eQTM) mapping within ±1 Mb of each CpG site (n = 120). Created in BioRender. Gsour, A. (2026) https://BioRender.com/ldrfs0u. B Number of significant CpGs identified through EWAS (age or T2D) and their corresponding eQTM hits. C Genomic proximity of CpGs to the nearest gene for age- and T2D-associated hits, classified as either “nearest” (within gene body/promoter) or “distant” (>10 kb). D, E Distribution of eQTM associations by chromatin state and effect direction (same vs. inverse correlation). Source data are provided as a Source Data file. Multi-omic integration identifies CpG–gene regulatory relationships in pancreatic islets. A Schematic overview of the analytical workflow. DNA methylation levels were measured in pancreatic islets from 144 human donors using Illumina EPIC arrays (EWAS). Transcriptomic profiles were generated by RNA-seq, and genetic variants were profiled using Omni2.5 M arrays. Integration of CpG-gene pairs was performed through expression quantitative trait methylation (eQTM) mapping within ±1 Mb of each CpG site (n = 120). Created in BioRender. Gsour, A. (2026) https://BioRender.com/ldrfs0u. B Number of significant CpGs identified through EWAS (age or T2D) and their corresponding eQTM hits. C Genomic proximity of CpGs to the nearest gene for age- and T2D-associated hits, classified as either “nearest” (within gene body/promoter) or “distant” (>10 kb). D, E Distribution of eQTM associations by chromatin state and effect direction (same vs. inverse correlation). Source data are provided as a Source Data file. This analysis identified 251 age- (25.2% of age CpGs) and 310 T2D-associated (34.3% of T2D CpGs) unique CpGs linked with gene expression in cis (Fig. 2B; Supplementary Data 8–9; Supplementary Figs. 2 and 3). Notably, only a small fraction of these CpGs (10.1% for age, 25.8% for T2D) targeted their nearest gene, underscoring the need for multi-omic analyses to uncover long-range regulatory effects (Fig. 2C; Supplementary Figs. 4 and 5). Chromatin context distinguishes age- and T2D-linked CpGs To gain insight into the regulatory context of CpG–gene interactions, we examined the chromatin states of eQTM-associated CpGs, based on previously published epigenomic maps in human islets19. Age- and T2D-associated CpGs displayed distinct regulatory profiles: age-linked CpGs were predominantly located in active promoters (Fig. 2D), whereas T2D-linked CpGs were largely located in active enhancer regions or unassigned (Fig. 2E). To confirm these trends, we examined the enrichment of eQTM-identified CpGs compared to all CpGs in the Illumina array, across chromatin states. This analysis supported the functional specificity of age-associated CpGs, which showed significant enrichment in promoters (Chi2 P = 5.80 × 10-4) and depletion from unassigned regions (P = 4.61 × 10−11; Supplementary Fig. 6). In contrast, T2D-associated CpGs were significantly enriched in active enhancers (P = 4.36 × 10−4) and depleted from promoters (P = 1.37 × 10−3; Supplementary Fig. 7). Overall, we show distinct chromatin states between age- and T2D-related eQTMs. For each CpG–gene pair, we classified the relationship as either inverse (i.e., increased methylation associated with decreased expression) or same direction (i.e., increased methylation associated with increased expression) (Fig. 2D–E). Notably, we observed a roughly balanced proportion of inverse and positive correlations across all regulatory regions, suggesting that DNA methylation can act as either a gene activator or repressor, highlighting the complexity of methylation–transcription regulation in human pancreatic islets. Co-expression network analysis identifies age- and T2D-specific epigenetic gene modules Pathway enrichment analysis of age- and T2D-associated eQTM target genes identified broad biological processes, including RNA processing and intracellular signalling for age, and metal ion binding and axon guidance–related pathways for T2D (Supplementary Fig. 8). To understand the biological relevance of age- and T2D-associated target genes, we applied weighted gene co-expression network analysis (WGCNA)21 to define modules of co-regulated genes linked via eQTMs. This unsupervised approach groups genes into co-expression modules based on expression similarity across samples and enables the identification of coordinated transcriptional programmes and molecular pathways. Separate networks were constructed for age- and T2D-associated eQTM genes to pinpoint trait-specific programmes. This resulted in seven age-related gene modules (A1-A7) and six T2D-related modules (T1-T6) (Fig. 3A–B; Supplementary Figs. 9–12; Supplementary Data 10–11). We annotated each module based on enriched pathways and published literature. Age-associated modules were predominantly enriched for genes involved in insulin secretion (A1), protein synthesis (A4), and cellular metabolic regulation (A3), whereas T2D-associated modules spanned a broader range of functions including ribosomal and translational machinery (T3), endocrine cell identity (T4), nuclear regulatory pathways (T6) and ion channel activity (T2). These findings suggest that age-linked epigenetic regulation preferentially targets beta-cell intrinsic (physiological) processes, while T2D-associated changes extend stress-related, homoeostatic and differentiation-related deleterious pathways.Fig. 3Functional characterization of epigenetically regulated gene modules in pancreatic islets.Pie charts representing the number of genes per cluster identified through WGCNA analysis of age-associated (A) and T2D-associated (B) CpGs linked to gene expression (eQTM). Clusters were annotated based on pathway enrichment. Age-related modules (A1–A7) are shown in pink–purple; T2D-related modules (T1–T6) are shown in green–blue. C, D Heatmaps showing associations between pooled cluster gene expression and insulin stimulation index (ISI) in non-diabetic donors (n = 121, 53 females and 68 males). Values in each cell correspond to two-sided P-values from linear models; associations are colour-coded by effect estimate (scale bar, right). No adjustment for multiple comparisons was applied. Source data are provided as a Source Data file. Functional characterization of epigenetically regulated gene modules in pancreatic islets. Pie charts representing the number of genes per cluster identified through WGCNA analysis of age-associated (A) and T2D-associated (B) CpGs linked to gene expression (eQTM). Clusters were annotated based on pathway enrichment. Age-related modules (A1–A7) are shown in pink–purple; T2D-related modules (T1–T6) are shown in green–blue. C, D Heatmaps showing associations between pooled cluster gene expression and insulin stimulation index (ISI) in non-diabetic donors (n = 121, 53 females and 68 males). Values in each cell correspond to two-sided P-values from linear models; associations are colour-coded by effect estimate (scale bar, right). No adjustment for multiple comparisons was applied. Source data are provided as a Source Data file. To assess the metabolic relevance of these gene modules, we correlated each one with ISI (only in ND individuals). Among age modules, A1 (insulin secretion) showed a positive correlation with ISI (P = 0.04) (Fig. 3C), consistent with its enrichment in genes central to insulin secretion. For T2D modules, none of the modules showed a correlation with ISI (Fig. 3D). Together, these results identify key age- and T2D-specific gene modules under epigenetic regulation in pancreatic islets, with distinct physiological associations pointing to divergent mechanisms of metabolic dysfunction. Distinct functional and islet-specific gene programmes in age and T2D modules To investigate the functional relevance of the epigenetically regulated gene modules, we first assessed their tissue specificity by computing average expression enrichment in pancreatic islets compared to all other GTEx tissues. Among age-associated modules, both “A1: insulin secretion” and “A3: cellular and metabolic regulation” showed islet enrichment when compared to other tissues (Fig. 4A).Fig. 4Functional characterization of epigenetically regulated gene modules in relation to pancreatic islet specificity and insulin secretion.A, B Cluster-level enrichment of gene expression in human pancreatic islets compared to all other tissues. GTEx-based expression enrichment was calculated as: Mean log₁₀ fold-change (Islets / Other tissues), and averaged per module. The colours do not represent any biological parameter or statistical value. A Age-associated clusters (A1–A7) and (B) T2D-associated clusters (T1–T6) are ranked by enrichment. Positive values indicate islet-specific expression. Association between the expression of individual genes in selected modules and insulin stimulation index (ISI) for (C) A1 cluster (age-associated) and (D) T4 cluster (T2D-associated) assessed using a linear model. Each dot represents a gene, with x-axis showing module membership (eigengene connectivity) and y-axis showing correlation with ISI. Source data are provided as a Source Data file. Functional characterization of epigenetically regulated gene modules in relation to pancreatic islet specificity and insulin secretion. A, B Cluster-level enrichment of gene expression in human pancreatic islets compared to all other tissues. GTEx-based expression enrichment was calculated as: Mean log₁₀ fold-change (Islets / Other tissues), and averaged per module. The colours do not represent any biological parameter or statistical value. A Age-associated clusters (A1–A7) and (B) T2D-associated clusters (T1–T6) are ranked by enrichment. Positive values indicate islet-specific expression. Association between the expression of individual genes in selected modules and insulin stimulation index (ISI) for (C) A1 cluster (age-associated) and (D) T4 cluster (T2D-associated) assessed using a linear model. Each dot represents a gene, with x-axis showing module membership (eigengene connectivity) and y-axis showing correlation with ISI. Source data are provided as a Source Data file. We next correlated the expression of each gene within each age-associated module to insulin secretion, quantified by ISI, to prioritise genes most functionally relevant to β-cell function. ISI-linked genes were significantly overrepresented in “A1: insulin-secretion” module (P = 2 × 10−4), with tightly co-regulated genes (module membership > =0.7) that were associated with ISI (unadjusted p < 0.05; Fig. 4C; Supplementary Table 12). This includes already known genes OPRD1 (delta opioid receptor, whose rare mutations associate with T2D risk22), CYP27A1 (a mitochondrial cytochrome P450 enzyme involved in cholesterol metabolism), ICA1 (Islet Cell Autoantigen 1, involved in secretory vesicle trafficking)23, and DIRAS2 (DIRAS family, GTP-binding RAS-like 2, involved in beta cell survival)24. There are arguments that some of these genes represent core drivers of epigenetically adaptive insulin secretion with age to maintain euglycemia. We also show here that the expression of eQTM OPRD1 (that inhibits insulin secretion when activated) is decreased with age. Among the T2D-associated eQTM modules, “T4: endocrine cell identity” displayed the highest islet enrichment (P = 2.2 × 10−7; Fig. 4B). This module included a range of ISI-correlated genes, including positive associations with genes such as SLC8A1 (a calcium/sodium exchanger, and a key regulator of intracellular Ca²⁺), MTUS2 and EXOC2 (involved in brain development) and ACAT1 (an enzyme linked to lipid metabolism) (Fig. 4D, Supplementary Data 13). Genetic influence and causal effects at age- and T2D-associated CpGs on gene expression Next, we investigated the extent to which age- and T2D-associated CpGs are under genetic influence. We performed eQTM analyses and identified CpGs which simultaneously target a gene and were associated with nearby genetic variants. Whilst the majority of age- and T2D-associated CpG–gene pairs (>70%) showed no evidence of genetic influence, an important subset was influenced by nearby genetic variants. Specifically, 43 age-associated (Supplementary Data 14) and 64 T2D-associated CpGs (Supplementary Data 15) were genetically modulated (Fig. 5A), highlighting a group of hybrid CpGs shaped by both genetic and environmental influences.Fig. 5Genetic support and causal inference for age- and T2D-associated CpG–gene relationships in human pancreatic islets.A Proportions of epigenetically regulated CpG–gene pairs identified by eQTM analysis that show no evidence of genetic influence (“eQTM only”) versus those also associated with nearby genetic variants (“eQTM + mQTL”), shown separately for age- and T2D-associated CpGs. B Directional concordance of epigenetic (eQTM) and genetic (colocalisation) effects on gene expression for selected age- and T2D-associated target genes. Bars represent the proportion of CpG–gene associations showing positive (up; green) or negative (down; red) effects on gene expression. Genes marked with an asterisk (*) indicate loci with evidence of both epigenetic and genetic regulation. C Mendelian randomization (MR) analysis testing the causal effect of DNA methylation on gene expression for two age-associated CpGs (cg23222095 and cg19333839) targeting KLHL42 in human pancreatic islets (n = 124, 52 females and 72 males). MR estimates were obtained using the inverse-variance weighted (IVW), robust IVW, simple and weighted median, and MR Egger methods. For each method, the point indicates the causal effect estimate and error bars indicate the standard error. Two-sided P-values are shown on the plot. No adjustment for multiple comparisons was applied. D Genomic context of the KLHL42 locus showing the position of age-associated CpGs, nearby enhancer elements, and cis-mQTL instruments. Source data are provided as a Source Data file. Genetic support and causal inference for age- and T2D-associated CpG–gene relationships in human pancreatic islets. A Proportions of epigenetically regulated CpG–gene pairs identified by eQTM analysis that show no evidence of genetic influence (“eQTM only”) versus those also associated with nearby genetic variants (“eQTM + mQTL”), shown separately for age- and T2D-associated CpGs. B Directional concordance of epigenetic (eQTM) and genetic (colocalisation) effects on gene expression for selected age- and T2D-associated target genes. Bars represent the proportion of CpG–gene associations showing positive (up; green) or negative (down; red) effects on gene expression. Genes marked with an asterisk (*) indicate loci with evidence of both epigenetic and genetic regulation. C Mendelian randomization (MR) analysis testing the causal effect of DNA methylation on gene expression for two age-associated CpGs (cg23222095 and cg19333839) targeting KLHL42 in human pancreatic islets (n = 124, 52 females and 72 males). MR estimates were obtained using the inverse-variance weighted (IVW), robust IVW, simple and weighted median, and MR Egger methods. For each method, the point indicates the causal effect estimate and error bars indicate the standard error. Two-sided P-values are shown on the plot. No adjustment for multiple comparisons was applied. D Genomic context of the KLHL42 locus showing the position of age-associated CpGs, nearby enhancer elements, and cis-mQTL instruments. Source data are provided as a Source Data file. To assess whether any of these genetically influenced target genes were also implicated by genome-wide association studies (GWAS) of T2D and related metabolic traits, we compared our results with a recently published colocalization analysis performed in more than 400 human pancreatic islet samples25. This comparison identified eight overlapping target genes for age and eight for T2D, with no genes shared between the two conditions (Fig. 5B). Among age-associated targets, KLHL42 showed concordant directionality between colocalization and eQTM effects and was also influenced by an mQTL. In the context of T2D, we identified SLC8A1, that we linked to endocrine cell identity module, and the established T2D effector gene STARD10. Notably, most T2D-associated target genes showed discordant directions between colocalization and eQTM effects, whereas age-associated targets exhibited more consistent directional effects. To investigate whether age- and T2D-associated CpG–gene pairs show evidence of causal regulation of gene expression, we performed Mendelian randomization (MR) using independent cis-mQTLs as instruments. Among age-associated CpGs, two CpGs showed strong and consistent evidence of a causal effect on gene expression across multiple MR approaches. For both cg23222095 and cg19333839 targeting the same gene (KLHL42), IVW and robust IVW estimates were significant after multiple-testing correction (FDR < 0.05), with consistent direction and magnitude across median-based methods. Instrument strength was high (median F-statistics > 20), no evidence of heterogeneity was detected (Cochran’s Q p > 0.1), and leave-one-out (LOO) analyses demonstrated stability of the causal estimates, indicating that the associations were not driven by single SNPs (Fig. 5C, Supplementary Data 16). These two CpGs lie in close proximity within the first intron of the gene. The most significant cis-mQTLs map to a super-enhancer (Fig. 5D). Another targeting GSTT2B also showed strong MR evidence, although it was instrumented by only three SNPs and was therefore interpreted more cautiously (Supplementary Data 16). In contrast to age, MR provided limited evidence for causal effects among T2D-associated loci, as none remained significant after correction for multiple testing (Supplementary Data 17). Age methylation risk scores discriminates T2D status and associate with islet function To evaluate the value of methylation changes for T2D discrimination we constructed a methylation risk score (MRS) to classify T2D cases in our cohort. This score was based on CpGs that were significantly associated with age in our and a previous islet study12, and a blood-based study,17. We only selected CpGs that were located in CpG islands and in islet regulatory regions19, with strong age effect sizes (>0.2), ensuring biological relevance. This resulted in 6 CpGs (Supplementary Data 18). As expected, in the donor islets, the MRS showed a strong positive correlation with age (P = 2.2 × 10−16; Fig. 6A), confirming that the score captures age-related methylation change. Additionally, we found a significant inverse correlation of the MRS with ISI (Spearman’s P = -0.36, P = 0.00025; Fig. 6B), demonstrating that higher MRS scores are linked to reduced insulin secretion capacity, and thus that islet cellular age associates with insulin secretion in response to glucose.Fig. 6Contribution of genetic variants to DNA methylation.A Scatter plot of age versus methylation risk score (MRS) in human pancreatic islets (n = 124, 52 females and 72 males, r = 0.76, P < 2.2×10−16). B Scatter plot of insulin stimulation index (ISI) versus MRS (n = 101, r = −0.36, P-value < 0.00025). C Receiver operating characteristic (ROC) curve assessing the ability of MRS to classify type 2 diabetes (T2D) status in human pancreatic islets (123 non-diabetic controls, 21 T2D donors). The reported metric is the area under the curve (AUC). D Scatter plot of ISI versus polygenic risk score (PGS) for T2D (n = 101, 41 females and 60 males; r = −0.29, P-value = 0.0037). E ROC curve assessing the performance of MRS, PGS, and the combination of MRS + PGS for discrimination of T2D. The three lines represent models based on MRS alone, PGS alone, or the integration of both scores. The associations in A, B, and D were assessed using a correlation test, and the line represents the linear regression fit. Source data are provided as a Source Data file. Contribution of genetic variants to DNA methylation. A Scatter plot of age versus methylation risk score (MRS) in human pancreatic islets (n = 124, 52 females and 72 males, r = 0.76, P < 2.2×10−16). B Scatter plot of insulin stimulation index (ISI) versus MRS (n = 101, r = −0.36, P-value < 0.00025). C Receiver operating characteristic (ROC) curve assessing the ability of MRS to classify type 2 diabetes (T2D) status in human pancreatic islets (123 non-diabetic controls, 21 T2D donors). The reported metric is the area under the curve (AUC). D Scatter plot of ISI versus polygenic risk score (PGS) for T2D (n = 101, 41 females and 60 males; r = −0.29, P-value = 0.0037). E ROC curve assessing the performance of MRS, PGS, and the combination of MRS + PGS for discrimination of T2D. The three lines represent models based on MRS alone, PGS alone, or the integration of both scores. The associations in A, B, and D were assessed using a correlation test, and the line represents the linear regression fit. Source data are provided as a Source Data file. Next, we evaluated its ability to discriminate T2D status using receiver operating characteristic (ROC) analysis. The MRS discriminates T2D status in our cohort, achieving an AUC of 0.88 (95% confidence intervals [CI]: 0.80–0.96) (Fig. 6C), compared to age alone (AUC = 0.66) (Supplementary Fig. 13). In addition, we calculated a polygenic risk score (PGS) in the samples using all available SNP estimates from a recently published genome wide association study (GWAS) in T2D, composed of 2,535,601 controls and 428,452 cases26 (Supplementary Data 19). As with the age MRS, our PGS was negatively correlated with ISI (Spearman’s p = −0.29, P = 0.0037; Fig. 6D). We found that PGS alone yielded a similar AUC of 0.76 (CI: 0.669–0.863), which is comparable to previous published data (Fig. 6E)27. Finally, we evaluated the additive value of combining PGS and MRS scores for T2D, which yielded improved and strong predictive performance (AUC = 0.91; CI: 0.84–0.98) (Fig. 6E; Supplementary Data 20). These results highlight the complementary value of combining genetic susceptibility with epigenetic state to refine individual-level risk assessment for T2D. Discussion Insufficient insulin secretion is a key driver for the development of T2D, particularly in the context of obesity and aging-related insulin resistance28. While the overall “performance” of pancreatic islets is largely determined by genetic factors from conception, as clearly shown by rare and frequent DNA variants29, non-genetic factors, including age and lifestyle also modulate β-cell insulin secretion30,31. These factors initially trigger adaptive responses to maintain euglycemia (e.g., in presence of overweight), but may ultimately contribute to β-cell exhaustion and dysfunction. Epigenetic mechanisms such as DNA methylation provide a dynamic interface between genetic predisposition and environmental exposures4,12. In this context, DNA methylation may contribute both to adaptive responses required for maintaining euglycemia and to the amplification of cellular defects under conditions of metabolic stress, such as T2D. To elucidate how age and T2D influence the islet methylome and function, we integrated matched methylation, expression, and genotyping data from organ donors. Importantly, our models were adjusted for genetic effects, allowing us to isolate methylation-driven transcriptional changes linked to age, T2D, and islet function. Our findings reveal two divergent epigenetic programmes: a stable, coordinated remodelling with age vs. a heterogeneous, stress-linked reprogramming in T2D. Age-associated CpGs were highly reproducible across studies4,12, enriched in promoters, and formed tightly co-expressed modules regulating insulin secretion. These changes may represent a conserved transcriptional adaptation to aging. In contrast, T2D-associated CpGs were predominantly hypomethylated, enriched in enhancers or unassigned regions, suggesting a profile likely driven by metabolic stress. Hypomethylation in highly methylated regions, as observed in our T2D EWAS, is known to de-repress repetitive elements and amplify inflammatory signalling32,33, a hallmark of β-cell dysfunction in T2D. These divergent patterns were mirrored at the eQTM gene network level. Coherently, age effector genes from cluster A1: insulin secretion, including OPRD1, DIRAS2 and ICA1. Indeed, OPRD1, which encode the δ-opioid receptor, control glucose-stimulated insulin secretion22,34,35. In contrast, effector genes within the T2D-associated T4: endocrine cell identity module lack known links to diabetes, suggesting that this module may reflect a maladaptive or compensatory transcriptional response rather than a conserved disease mechanism. Together, these data suggest that aging remodels the islet epigenome in a targeted, programmatic fashion, while T2D progression involves a more context-specific epigenetic drift. This model aligns with the clinical course of T2D, which generally emerges after age 45 and may reflect the genetically determined limitations in islet functional capabilities in presence of insulin resistance, despite some adaptive functional epigenetically determined changes, in a general context of β-cell decline due to increased cellular age. In this context, the eventual failure of pharmacologic interventions after years of efficacy may reflect the overwhelming accumulation of epigenetic modifications in islets that induce inflammation and stress. This study also highlights the medical relevance of combining PGS for T2D and methylation-based cellular age score both measurable from blood DNA to predict both insulin secretory capacity (as both MRS and PGS correlate with ISI) and to predict accurately T2D occurrence. Unlike genetics, which is largely fixed at birth, the epigenome is modifiable. DNA methylation inhibitors, such as Azacytedine, are already used to treat acute myeloid leukaemia, and dietary and lifestyle interventions have been shown to influence DNA methylation patterns36,37. Given that DNA methylation changes contribute to both physiological adaptation and disease progression, understanding the precise molecular mechanisms that drive these modifications will be crucial for uncovering novel strategies to preserve β-cell function, slow T2D progression, and optimize therapeutic interventions. Methods Pancreatic islet preparation Isolated islets were prepared by collagenase digestion and density gradient purification38 from the pancreas of 144 brain-dead organ donors, of which 60 females and 84 males, with an age range of 22 to 96 years. The characteristics of the donors are detailed in Table 1. These samples were collected before November 2021 within the IMIDIA, RHAPSODY and T2DSystems consortia39–41, with written consent obtained by next-of-kin’s and with approval of the local ethics committee in Pisa. T2D status diagnosis was based on medical history and the American Diabetes Association criteria (2019). ISI of the pancreatic islets in this cohort was assessed in vitro by measuring islet insulin secretion capacity after high and low glucose stimulation, as previously reported42. Differences in clinical characteristics between ND and T2D individuals were assessed using the Wilcoxon rank sum test. DNA extraction, methylation arrays and statistical analysis DNA was extracted from pancreatic islets using the DNeasy Blood & Tissue kit (Qiagen, Germany). Bisulphite conversion was performed in a total of 500 ng of DNA from our samples using the EZ DNA Methylation kit (5001; Zymo Research) and subjected to Illumina’s MethylationEPIC array, which covers 850,000 CpG sites throughout the genome. Methylation array data was imported using the minfi R package (version 1.44.0)43. Quality control (QC) steps removed CpG probes if they were: located on sex chromosomes cross-hybridising, non-cg or had a detection threshold p-value greater than 0.01. Samples with less than 95% probes with a detection p-value lower than 0.01 were excluded. Probe-design biases and batch effects were normalised using R packages Enmix (version 1.26.9)44 and SVA (version 3.46.0)45 respectively. Following QC, 809,700 probes remained. Sample call rate threshold was set at 95%. Following QC, 144 samples remained for further analysis. Principal component analysis (PCA) PCA was calculated to reduce the dimensionality of the DNA methylation dataset and to assess the influence of clinical variables on methylation variation and sample-to-sample variability. This was performed using the flashpcaR R package (version 2.1)46. The 20 first principal components (PCs) were computed, and an ANOVA was used to test the association of each variable with each PC (Supplementary Fig. 14). Outliers were identified based on Euclidean distance from the centre. The squared sum of standardized scores across the first 10 PCs was calculated for each sample, and those exceeding the 75th percentile plus three times the interquartile range (IQR) were flagged as outliers. Epigenome-wide association study (EWAS) We applied linear regression models to associate age and T2D, independently, with CpG methylation level in probes present on the MethylationEPIC array. Potential confounding variables were identified from our PCA analysis, i.e., any variable significantly associated with PC1 was flagged as a potential confounder. T2D status, age, sex and islet purity were associated with PC1 (Supplementary Fig. 15). In addition, as DNA methylation is highly variable between cell types, we estimated and adjusted for cellular composition using the R package RefFreeEWAS (version 2.2)47. We adjusted for J-1 cell types (where J is the number of cell types defined), and to avoid collinearity, we removed the cell type with the lower estimated proportion. The EWAS models used are highlighted below.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\rm{CpG}}}}\left({{{\rm{M\; value}}}}\right) \sim \, {{{\rm{Age}}}}+{{{\rm{Sex}}}}+{{{\rm{T}}}}2{{{\rm{D}}}\; {{\rm{status}}}}+{{{\rm{BMI}}}}+{{{\rm{Islet\; purity}}}} \\ +{{{\rm{Cellular}}}\; {{\rm{composition}}}}$$\end{document}CpGM value~Age+Sex+T2Dstatus+BMI+Islet purity+Cellularcomposition1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\rm{CpG}}}}\left({{{\rm{M}}}}\,{{{\rm{value}}}}\right) \sim \,{{{\rm{T}}}}2{{{\rm{D}}}}\,+{{{\rm{Sex}}}}+{{{\rm{Age}}}}\,+{{{\rm{BMI}}}}+{{{\rm{Islet}}}}\,{{{\rm{purity}}}} \\ +{{{\rm{Cellular}}}}\,{{{\rm{composition}}}}$$\end{document}CpGMvalue~T2D+Sex+Age+BMI+Isletpurity+Cellularcomposition β-values denote methylation levels, where 0 indicates 0 % methylation and 1 indicates 100 % methylation. For the analyses, β-values were transformed to the more statistically robust M-values48. EWAS was conducted using the limma R package49. The bias- and inflation-corrected P-values were then adjusted for multiple testing using the false discovery rate (FDR) method from Benjamini-Hochberg. In order to control for false positives due to genomic inflation, we measured and handled the potential inflation and bias in our results using the Bacon method (R package Bacon), a Bayesian method based on estimation of the empirical null distribution50. For all EWAS, a Bacon-FDR p-value of <0.05 was used to determine statistical significance. RNA-sequencing RNA-sequencing was performed in three centres: (1) University of Lund, (2) Centre for Molecular Bioengineering (CMCB) in Dresden, and (3) at LIGAN genomics centre in Lille. For 142 pancreatic islet samples of the same cohort using 200 ng of RNA using the KAPA mRNA HyperPrep kit (in Lille; Roche Sequencing) or the Illumina SmartSeq protocol (CMDB and University of Lund). The libraries were sequenced in 2×75 bp paired-end reads (University of Lund, EGID) or single-ended 75 bp paired-end reads (CMCB) using the Illumina HiSeq 2500 or Illumina NextSeq500 systems. The average read per sample was of 72,000,000. These were mapped to the human genome on hg37 using STAR (version 2.7.10a). Sequence demultiplexing was performed using the bcl2fastq Conversion Software (Illumina; version v2.20.0.422). The QC was performed using the FastQC software (version v0.11.9). The removal of adaptor sequences and low-quality bases was performed with Trimmomatic (version v0.39)51. Genes were quantified using htseq-counts (version 0.9.1), read with tximport R package (version 1.26.1) and normalised using the method “vst” from the R package DESeq252. To account for technical variation across sequencing centres (University of Lund, CMCB, and LIGAN), batch effects were corrected using the ComBat algorithm from the sva R package (version 3.46.0), with sample origin specified as the batch variable. Genotyping Genotyping data were generated from DNA extracted from whole pancreas samples from the same study participants. DNA extraction was performed with the DNeasy Blood & Tissue kit (Qiagen). Genotyping was done with the GWAS Illumina HumanOmni2.5 arrays on the Illumina iScan. Genotypes were called with GenomeStudio software. SNPs were excluded based on the following criteria: call rate<0.9%, minor allele frequency (MAF) < 0.01%, and a Hardy-Weinberg equilibrium p < 1×10−4. Samples were excluded if they featured: discordant sex, a heterozygosity rate greater than four times the standard deviation from the mean heterozygosity rate, or a relatedness (determined by identity by descent)>0.2. QC resulted in the exclusion of 20 samples and 116,241 SNPs. Imputation was performed with the Haplotype Reference Consortium Panel53. Therefore, the following integrative analyses with genotyping data were performed in n = 120 individuals (52 females and 68 males). Expression quantitative trait methylation (eQTM) Loci To address the contribution of DNA methylation changes to gene expression whilst adjusting for the influence of the genotype on CpGs, we regressed all age- or T2D-associated CpGs identified in the EWAS against the expression of all genes within a ± 1 Mb window (n = 120). eQTM analyses were performed by modelling CpG methylation levels as the outcome and the expression of nearby genes (±1 Mb) as the explanatory variable. This approach was used to identify CpG sites whose methylation is tightly associated with local transcriptional variation, independently of genetic effects. The following formulae were applied:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\rm{CpG}}}}({{{\rm{M}}}}\; {{{\rm{value}}}})\, \sim \,{{{\rm{Gene}}}}\,{{{\rm{expression}}}}+{{{\rm{Ag}}}}{{{\rm{e}}}}+{{{\rm{Sex}}}}+{{{\rm{BMI}}}}+{{{\rm{T}}}}2{{{\rm{D}}}}\,{{{\rm{status}}}} \\+{{{\rm{Islet}}}}\,{{{\rm{purity}}}}+{{{\rm{Cellular}}}}\,{{{\rm{c}}}}{{{\rm{omposition}}}}+{{{\rm{Genotype}}}}\,$$\end{document}CpG(Mvalue)~Geneexpression+Age+Sex+BMI+T2Dstatus+Isletpurity+Cellularcomposition+Genotype2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\rm{CpG}}}}({{{\rm{M}}}}\; {{{\rm{value}}}})\, \sim {{{\rm{Gene}}}}\; {{{\rm{expression}}}}+{{{\rm{T}}}}2{{{\rm{D}}}}\; {{{\rm{status}}}}+{{{\rm{BMI}}}}+{{{\rm{Sex}}}}+{{{\rm{Age}}}} \\+{{{\rm{Islet}}}}\; {{{\rm{purity}}}}+{{{\rm{Cellular}}}}\; {{{\rm{c}}}}{{{\rm{omposition}}}}+{{{\rm{Genotype}}}}$$\end{document}CpG(Mvalue)~Geneexpression+T2Dstatus+BMI+Sex+Age+Isletpurity+Cellularcomposition+Genotype This resulted in trio associations (SNP-CpG-gene expression). Multiple testing was addressed using a hierarchical strategy, where FDR < 0.05 was considered significant at each step. CpG–trait associations were corrected for multiple testing, and eQTM associations were evaluated using gene-level FDR correction that accounts for the number of CpG–SNP tests per gene. To determine DNA methylation sites that were not associated with nearby gene expression, we focused on 1/ significant DNA methylation and gene expression associations, and 2/ removed any associations of significant CpGs with nearby genetic variants (methylation quantitative trait loci, or mQTL). To determine the relationship between the trait and gene expression we computed the “effect direction”. The effect direction of the association was determined by extracting the trait and gene estimate and multiplying the estimate signs, yielded positively and negatively regulated genes with age and T2D (Supplementary Table 3). CpGs identified in the age or T2D eQTM analyses were overlapped with islet-specific regulatory regions generated by Miguel-Escalada et al.19. The enrichment of eQTM-identified CpGs for promoters, inactive and active enhancers, and unassigned regions, relative to all CpGs present in the Infinium MethylationEPIC array was assessed using a Chi2 test. Cis-methylation quantitative trait loci (mQTL) analysis To identify CpG sites under genetic control, we performed cis-mQTL analysis using the same set of individuals with both genotype and methylation data (n = 124, 52 females and 72 males). Linear regression models were applied using a linear model, testing associations between each CpG and all SNPs located within a ± 1 Mb window around the CpG. The model used was:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\rm{CpG}}}}({{{\rm{M}}}}\; {{{\rm{value}}}})\, \sim \, {{{\rm{SNP}}}}+{{{\rm{T}}}}2{{{\rm{D}}}}\; {{{\rm{status}}}}+{{{\rm{BMI}}}}+{{{\rm{Sex}}}}+{{{\rm{Age}}}} \\ +{{{\rm{Islet}}}}\; {{{\rm{purity}}}}+{{{\rm{Cellular}}}}\; {{{\rm{c}}}}{{{\rm{omposition}}}}$$\end{document}CpG(Mvalue)~SNP+T2Dstatus+BMI+Sex+Age+Isletpurity+Cellularcomposition To derive a set of independent mQTL signals, we applied LD-based clumping to significant variants on a chromosome-by-chromosome basis. Pairwise LD estimates (r2) were obtained using the LDlink LDmatrix function with the CEU reference panel and GRCh37 genomic build54. Within each chromosome, SNPs were iteratively evaluated and variants showing r2 > 0.1 with a retained SNP were excluded. P-values were adjusted for multiple testing using the Benjamini–Hochberg FDR method, with significant mQTLs defined as those with FDR < 0.05. CpG sites identified as significant mQTLs were excluded from the eQTM analysis to focus on associations between methylation and gene expression independent of genetic variation. Relationship with insulin stimulation index (ISI) To assess the relationship between the expression of eQTM target genes to ISI, we applied linear models to each gene, correcting for islet purity and sample origin (n = 101, 41 females and 60 males). The model is described below:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\rm{Target\; gene}}}} \sim {{{\rm{ISI}}}}+{{{\rm{Age}}}}+{{{\rm{Sex}}}}+\,{{{\rm{BMI}}}}+{{{\rm{Islet\; purity}}}}$$\end{document}Target gene~ISI+Age+Sex+BMI+Islet purity Significance threshold was set at P < 0.05. Weighted gene co-expression network analysis (WGCNA) To uncover biologically meaningful structure among epigenetically regulated genes, we performed weighted gene co-expression network analysis (WGCNA; R package version 1.73) separately for age- and T2D-associated genes identified via eQTM analysis. An unsigned adjacency matrix was constructed using a soft-thresholding power of 10, selected based on scale-free topology criteria. Network modules were detected using hierarchical clustering with dynamic tree cutting (hybrid method), setting the minimum module size to 10 genes. Modules with highly correlated genes (dissimilarity <0.15) were subsequently merged. Identified modules were assigned biological themes based on literature search and the GeneCards database55. To identify biologically relevant modules, we assessed the association between module gene expression and clinical traits. For each module, the expression of constituent genes was pooled and associated with ISI using linear models, described below:5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\rm{Module}}}}\, \sim {{{\rm{ISI}}}}+{{{\rm{Age}}}}+{{{\rm{Sex}}}}+\,{{{\rm{BMI}}}}+\,{{{\rm{Islet}}}}\,{{{\rm{purity}}}}$$\end{document}Module~ISI+Age+Sex+BMI+Isletpurity To evaluate the relative enrichment of each WGCNA module in pancreatic islets, we used gene expression data from the GTEx Portal. For each gene, expression in pancreatic islets was compared with expression across all other GTEx tissues, and a log₁₀ fold-change (islets / other tissues) was calculated. Gene-level fold-change values were then averaged across all genes within each module to derive a module-level measure of islet enrichment. Mendelian randomisation To assess whether DNA methylation at age- or T2D-associated CpG sites exert a causal effect on gene expression in human pancreatic islets, we performed Mendelian randomisation (MR) analyses using cis-acting methylation quantitative trait loci (cis-mQTLs) as exposure instrumental variables. For the outcome instrumental variables, we performed cis-acting Expression trait quantitive loci (eQTL) analyses (n = 124, 52 females and 72 males). The following model was applied:6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\rm{Gene}}}} \sim {{{\rm{SNP}}}}+{{{\rm{T}}}}2{{{\rm{D}}}\; {{\rm{status}}}}+{{{\rm{BMI}}}}+{{{\rm{Sex}}}}+{{{\rm{Age}}}}+{{{\rm{Islet}}}\; {{\rm{purity}}}}$$\end{document}Gene~SNP+T2Dstatus+BMI+Sex+Age+Isletpurity P-values were adjusted for multiple testing using the Benjamini–Hochberg FDR method. eQTLs with an adjusted P < 0.05 were excluded from the MR analyses. MR analyses were conducted in R using the MendelianRandomization package (v 0.10.0). For each CpG–gene pair showing evidence of genetic regulation in the eQTM analysis, cis-mQTLs located within ±1 Mb of the CpG site were selected as candidate instruments. Genetic variants were harmonised between mQTL (exposure) and eQTL (outcome) datasets based on allele alignment, and duplicated instruments were removed. Only CpG–gene pairs with at least three independent instruments were retained for MR analysis. Instrument strength was evaluated using F-statistics. For each CpG–gene pair, the median, mean, and minimum F-statistics were reported, and the number of instruments with F < 10 was recorded to assess potential weak-instrument bias. MR analyses were performed using multiple complementary methods to ensure robustness to violations of MR assumptions. These included: Inverse-variance weighted (IVW) MR with a random-effects model, Robust IVW and penalized IVW approaches, Weighted median, penalized weighted median, and simple median methods, MR-Egger regression, including robust and penalized variants. Causal estimates were considered more reliable when direction and magnitude were consistent across multiple MR methods. Heterogeneity among instruments was assessed for IVW-based methods using Cochran’s Q statistic, with corresponding degrees of freedom and p-values reported. To evaluate the influence of individual instruments, leave-one-out (LOO) sensitivity analyses were performed for the random-effects IVW model, whereby MR was re-estimated after sequential removal of each instrument. Multiple-testing correction was applied to MR p-values using the Benjamini–Hochberg false discovery rate (FDR) procedure, performed separately within each MR method across all tested CpG–gene pairs. An FDR-adjusted p < 0.05 was considered statistically significant. Methylation risk score (MRS) To assess the ability of DNA methylation changes to predict T2D status in our cohorts, we generated an MRS. For this, we selected CpGs associated (Bacon FDR p < 0.05) with age in our cohort that were (1) located in CpG islands, (2) located in regulatory islet regions, and (3) an age estimate > 0.2 to ensure biological relevance. The selected CpGs were intersected with CpGs previously associated with age in (1) a previously published article for pancreatic islets and age12 and (2) associated with a recently published blood-based EWAS17. This criterion resulted in 6 CpGs associated in all studies (Supplementary Data 18). To generate MRS for each individual, we applied a simple additive model, whereby the nominal β-value of each CpG was added. A logistic regression model was fitted to evaluate the association between a CpG methylation score and T2D status. The model was constructed with T2D as the binary outcome and CpG score as the predictor using a binomial family. Model performance was assessed using a Receiver Operating Characteristic (ROC) curve from the pROC R package (version 1.18.5)56, and the Area Under the Curve (AUC) was computed to quantify predictive accuracy. The resulting MRS was associated to age and ISI using Spearman’s correlation. Polygenic risk score (PGS) We generated PGS using the PRSice-2 tool57 using genotyping data from our organ donors. We utilised summary statistics from a recent T2D GWAS26 To identify lead variants for each genome-wide significant locus, we applied a pruning and thresholding approach with the following parameters: a clumping window of 250 kb, a linkage disequilibrium threshold of r² <0.1, and SNP P-value thresholds ranging from 5 × 10-8 to 0.5. This filtering process resulted in the inclusion of 1475 independent SNPs. The PGS was then computed using an additive model, incorporating the SNP effect estimates from the GWAS summary statistics. Model performance was assessed as described above. Risk score integration We integrated our MRS to our PGS in order to assess the additive nature of these scores. We performed k-fold cross validation to limit overfitting. In this process, the dataset was partitioned into ten approximately equal subsets. For each of the ten iterations, one subset was held out as the validation set, and the remaining nine subsets were used to train the PGS model. Model performance was assessed as described above. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Description of Additional Supplementary Files Supplementary Data 1 Supplementary Data 2 Supplementary Data 3 Supplementary Data 4 Supplementary Data 5 Supplementary Data 6 Supplementary Data 7 Supplementary Data 8 Supplementary Data 9 Supplementary Data 10 Supplementary Data 11 Supplementary Data 12 Supplementary Data 13 Supplementary Data 14 Supplementary Data 15 Supplementary Data 16 Supplementary Data 17 Supplementary Data 18 Supplementary Data 19 Supplementary Data 20 Reporting Summary Transparent Peer Review file Supplementary Information Description of Additional Supplementary Files Supplementary Data 1 Supplementary Data 2 Supplementary Data 3 Supplementary Data 4 Supplementary Data 5 Supplementary Data 6 Supplementary Data 7 Supplementary Data 8 Supplementary Data 9 Supplementary Data 10 Supplementary Data 11 Supplementary Data 12 Supplementary Data 13 Supplementary Data 14 Supplementary Data 15 Supplementary Data 16 Supplementary Data 17 Supplementary Data 18 Supplementary Data 19 Supplementary Data 20 Reporting Summary Transparent Peer Review file Source data Source Data for Main Figures Source Data for Supplementary Figs. Source Data for Main Figures Source Data for Supplementary Figs.