Page 1 of 51 Diabetes

CDKN2A/B T2D GWAS risk-SNPs impact locus expression and proliferation in human islets

Yahui Kong1, Rohit B. Sharma1, Socheata Ly1, Rachel E. Stamateris1, William M. Jesdale2 and Laura C. Alonso1

Diabetes Center of Excellence in the Department of Medicine1, and the Department of Quantitative Health Sciences2, University of Massachusetts Medical School, Worcester MA

Running title CDKN2A/B T2D SNPs impact human islet biology

Corresponding author Laura C. Alonso 7744553640 (phone) 5088563803 (fax) AS72047, Division of Diabetes 368 Plantation Street, Worcester, MA 01605 [email protected]

Keywords Aging, ANRIL, beta cell mass, Cdkn2A, Cdkn2B, CDKN2B-AS, insulin secretion, oncogene, p14, p15, , , p15INK4B, p16INK4A, pancreatic beta cell, proliferation

Abbreviations ACTB, betaactin gene ANRIL, antisense noncoding RNA in the INK4 locus ARF, alternate reading frame CCND2, D2 CDK, cyclin dependent kinase CDKN2A, cyclin dependent kinase inhibitor 2, encodes p14ARF and p16INK4A CDKN2B, cyclin dependent kinase inhibitor 2, encodes p15INK4B CDKN2B-AS, cyclin dependent kinase inhibitor 2B antisense eQTL, expression quantitative trait loci GAPDH, Glyceraldehyde3Phosphate Dehydrogenase GWAS, genome wide association studies lncRNA, long noncoding RNA MTAP, 5methylthioadenosine phosphorylase PCNA, proliferating cell nuclear antigen SNP, single nucleotide polymorphism

Diabetes Publish Ahead of Print, published online February 6, 2018 Diabetes Page 2 of 51

ABSTRACT

Genomewide association studies link the CDKN2A/B locus with T2D risk, but mechanisms increasing risk remain unknown. The CDKN2A/B locus encodes inhibitors p14, p15, and p16, MTAP, and ANRIL, a lncRNA. The goal of this study was to determine whether CDKN2A/B T2D riskSNPs impact locus , insulin secretion, or beta cell proliferation, in human islets. Islets from nondiabetic donors (n=95) were tested for SNP genotype (rs10811661, rs2383208, rs564398, rs10757283), gene expression (p14, p15, p16, MTAP, ANRIL, PCNA, KI67, CCND2), insulin secretion (n=61) and beta cell proliferation (n=47). Intriguingly, locus were coregulated in islets in two physically overlapping cassettes: p14p16-ANRIL, which increased with age, and MTAPp15, which did not. Riskalleles at rs10811661 and rs2383208 were differentially associated with expression of ANRIL, but not p14, p15, p16 or MTAP, in agedependent fashion, such that younger homozygousrisk donors had higher ANRIL expression, equivalent to older donor levels. We identified several riskSNP haplotype combinations that may impact locus gene expression, suggesting possible mechanisms by which SNPs impact locus biology. Riskallele carriers at ANRIL coding SNP rs564398 had reduced beta cell proliferation index. In conclusion, CDKN2A/B locus SNPs may impact T2D risk by modulating islet gene expression and beta cell proliferation. Page 3 of 51 Diabetes

T2D risk has a strong genetic component. Significant research investment has identified >100 genomic regions that influence T2D risk in human populations (1–3). Most T2D risk SNPs are noncoding, and the mechanism by which they impact local genome biology remains unclear for most loci (3). Risk alleles may act in multiple ways, interacting with other genes and polymorphisms in tissuespecific manner. Genomewide eQTL studies seek to identify how polymorphisms impact biology at any given locus (1,4–7); however, depth of information at individual loci is limited in genomewide studies. Most T2D SNPs influence risk by impacting islet biology (8), but the cost and inaccessibility of human islets, and poor utility of nonhuman models to study the , have slowed progress in clarifying mechanisms.

SNPs at the CDKN2A/B genomic locus impact risk of T2D, and related diseases such as gestational diabetes, cystic fibrosis related diabetes and posttransplant diabetes, across ethnicities and cultures, suggesting a central diabetogenic mechanism (9). Multiple SNPs in different linkage blocks at the CDKN2A/B locus confer T2D risk (9); mechanisms impacting risk remain unknown. The CDKN2A/B locus encodes four genes (Figure 1): MTAP, CDKN2A, CDKN2B and a longnoncoding RNA named ANRIL. CDKN2A and CDKN2B are well studied, encoding cell cycle inhibitors (p14 and p16 are splice variants of CDKN2A, p15 is encoded at CDKN2B) that impact aging, senescence, and tumorigenesis via regulation of Rb (retinoblastoma) and (10,11). Three T2D SNPs at this locus, rs10811661, rs2383208 and rs10757283, are noncoding, located downstream of known genes; rs2383208 and rs10811661 are in one linkage block and rs10757283 is in a separate linkage block immediately downstream. A fourth SNP, rs564398, about 100,000 bp upstream of these, falls within exon2 of ANRIL. These SNPs were identified in large population studies seeking to identify genomic regions associated with T2D risk (12–14); for more details please see (9). The three downstream SNPs are mostly associated with T2D risk and not other diseases; the rs564398 SNP is also associated with coronary heart disease and glaucoma (15). The absolute magnitude of T2D risk is low with all identified SNPs (at this and other loci); for example, reported odds ratio for the linkage region containing rs10811661 and rs2383208 ranges from 1.181.46 (4,12,14,16,17). Weaker odds ratios were seen for rs564398 1.121.26 e.g. (13); intriguingly, multiple studies show that rs564398 is associated with T2D risk in Diabetes Page 4 of 51

Caucasian but not Asian populations (18). A riskrisk haplotype of rs10811661/rs2383208 and rs10757283 conferred an odds ratio of 1.24, with stronger association than individual risk alleles (13). Although each T2D SNP is in linkage disequilibrium with multiple other SNPs, fine mapping has not identified linked SNPs with greater disease association than these GWASidentified SNPs (1,4). The causal SNP in any of these linkage blocks is not yet known.

In human populations, rs10811661 risk allele is associated with reduced insulin secretory capacity after oral or IV glucose challenge (16,19–22). Insulin secretory capacity is a composite endpoint influenced by beta cell mass, insulin production, glucose sensing and stimulussecretion coupling (23), factors that cannot currently be effectively separated in living human subjects. Intriguingly given the aging and senescence roles played by CDKN2A/B genes, the impact of rs10811661 on T2D risk was influenced by subject age (18). SNPs at this locus also influence insulin sensitivity and biology of other metabolic tissues, demonstrating the complexity of even a single genomic locus on T2D biology (9).

Since human studies suggest that CDKN2A/Blocus SNPs impact T2D risk, at least in part, by reducing insulin secretory capacity, we hypothesized that locus SNPs influence pancreatic islet biology. Here we present a detailed analysis of CDKN2A/B biology in nondiabetic human islets. We identified two overlapping coregulated gene sets: p14p16ANRIL, and p15MTAP. p14-p16-ANRIL expression, but not p15-MTAP expression, increased with donor age. Of the four T2D riskSNPs tested, rs2383208 and rs10811661 risk alleles were associated with inappropriate high expression of the ANRIL lncRNA in samples from younger donors. No other SNPgene interaction was identified, but our data suggest certain SNP haplotypes that may impact locus gene expression in combinatorial fashion. Finally, riskalleles at rs564398 were associated with reduced beta cell proliferation index, suggesting a functional implication for this SNP, and perhaps the ANRIL lncRNA, in accrual or maintenance of human beta cell mass.

RESEARCH DESIGN AND METHODS

Human islets Page 5 of 51 Diabetes

Human islets were obtained from the NIHNIDDKsupported Integrated Islet Distribution Program at the City of Hope, or from a collaborative group headed at Vanderbilt (24). Human islet studies were determined by the University of Massachusetts Institutional Review Board to not qualify for IRB review or exemption because they do not involve the use of human subjects. Deidentified islet samples from 95 nondiabetic subjects were live shipped in Prodo islet transport media transport. Donors (Supplemental Table 1) included 42 females, 48 males and 5 without sex reported, age 40 +/ 16 years, ethnicity 1 Asian, 8 Black or African American, 14 Hispanic/Latino, 66 White, and 6 unknown. Upon receipt, islets were plated in islet culture medium (RPMI, 10% FBS, 5 mmol/L glucose, penicillin/streptomycin) and incubated at 37 degrees, 5% CO2 overnight to recover from isolation and shipment. Following recovery, 800 IEQ were handpicked, washed in PBS

containing 100 nmol/L Na3VO4, and flash frozen at 80°C in 200 islet equivalent (IEQ) aliquots for future DNA and RNA analysis. Additional islets from a subset of donors were cultured as described below for glucosestimulated proliferation.

Genotyping DNA and RNA were extracted from flashfrozen 200 IEQ aliquots using the Norgen RNA/DNA/ purification kit (Norgen Biotek Corp., Ontario, Canada) following the manufacturer's protocol. Genotyping for four CDKN2A/B SNPs: rs564398 (C/T), rs10811661 (C/T), rs2383208 (G/A) and rs10757283 (C/T) was performed in duplicate using commercial (C_2618017_10, C_31288917_10, C_15789011_10, C_31288916_10) TaqMan® SNP genotyping assays (Thermo Fisher Scientific, Waltham, MA, USA) on Biorad (C1000 Touch Thermal cycler) or Eppendorf (Realplex cycler) realtime PCR platforms, using 20ng of DNA in a 10µl reaction volume under conditions recommended by the manufacturer. SNP determination was confirmed by both allelic discrimination and by manual CT value assessment for all samples and all SNPs. Minor allele frequencies (MAF) in our cohort were in agreement with expected MAF based on the 1000 genomes project (25) (Supplemental Table 2), and the observed haplotype frequency of SNP combinations predicted similar linkage disequilibrium to 1000 genomesreported values for these SNPs (Supplemental Table 3)(26).

Diabetes Page 6 of 51

Gene expression assays Total RNA was reverse transcribed using SuperScript IV VILO Master Mix kit (Thermo Fisher Scientific, Waltham, MA, USA). The expression levels of target genes in human islets were quantitatively assessed in duplicate using Taqman® validated human gene expression assays (Thermo Fisher Scientific). Primers/probes used were: ANRIL, Hs04259476_m1; p15, Hs00793225_m1; p14, Hs99999189_m1; p16, Hs02902543_mH; MTAP, Hs00559618_m1; KI67, Hs01032443_m1; PCNA, Hs00696862_m1; CCND2, Hs00153380_m1; ACTB, Hs01060665_g1; GAPDH, Hs02758991_g1. ACTB and GAPDH were used as endogenous reference to normalize gene expression. Reproducibility of duplicate measurements was high, as assessed by the R2 of the correlation between duplicates and by the absolute value of the Relative Percentage Difference between the duplicates (Supplemental Figure 1). Transcript expression levels were presented as log2transformed expression (CT).

Human islet culture experiments Human islets cultured overnight in islet culture medium were dispersed to single cells using singleuse–apportioned 0.05% trypsin and plated on uncoated glass coverslips (Fisherbrand) as described (27–29). Dispersed cells were cultured in islet culture medium containing either 5mM or 15mM glucose for 96 hours, with 20 g/ml bromodeoxyuridine (BrdU) included for the entire time. After culture, the islet cells were fixed for 10 min in 4% paraformaldehyde (Sigma). Immunofluorescence staining was performed after unmasking in 1N HCl for 25 minutes at 37°C for insulin, (Abcam, ab7842 or Dako, A0564012), BrdU (Abcam, ab6326) and DAPI as described (27–29). βCell proliferation, defined as the percent of insulin staining cells that were also BrdUlabeled, was quantified on blinded images (30). Data were expressed as the proliferation index, calculated as the ratio of %BrdU+ βcells in 15mM glucose divided by the %BrdU+ βcells in 5mM glucose.

Statistics Univariate analyses were performed using GraphPad Prism and expressed as mean ± SD. P values were determined by twotailed Student t test when comparing 2 conditions, with F test to compare variances, one way ANOVA with Tukey posttest for correction for multiple comparisons when comparing >2 conditions, or by linear regression when assessing the Page 7 of 51 Diabetes

relationship between two continuous variables. Multivariable linear models were performed to examine gene expression (p14, p15, p16, ANRIL, MTAP) simultaneously adjusted for donor sex, race/ethnicity, age (continuous) and body mass index (continuous); additional models further adjusted for expression of the other gene products. Missing values were modeled with a missing indicator, replacing unknown values with sample means for linear variables. Infrequent or unknown race/ethnicity were grouped in a residual category. RNA expression associated with SNPs was estimated in linear multivariable models adjusting for demographics (as above), in two fashions: first, setting the population with no risk alleles as the common reference group, and second by estimating the linear effect on a perallele basis (treating the number of risk alleles as additive). Insulin secretion index was estimated as a function of demographic variables as described above, and by including each SNP as a predictor of insulin secretion index. Interpretation of these models can be found in the Supplemental Materials. P<0.05 was considered significant, although this may be too generous for the exploratory analyses with multiple comparisons. For the haplotype hypothesisgenerating analyses, the false discovery rate (FDR), calculated by the original method of Benjamini and Hochberg, was set at 10%, based on our estimation that a hypothesis with 90% likelihood of being correct warranted experimental followup.

RESULTS CDKN2A/B locus gene expression is coordinately regulated in human islets

To understand the context of how T2D riskSNPs might impact biology at this locus in human islets, we first quantified expression of all locus genes (Figure 1A). Validated Taqman probes were chosen that could independently quantify transcripts including MTAP, p14 (CDKN2A, ARF), p15 (CDKN2B, INK4B), p16 (CDKN2A, INK4A) and the ANRIL (CDKN2BAS1) lncRNA. p14 and p16 are splice variants of CDKN2A, sharing exons 23 but with different first exons; exons 23 are in different reading frames and thus p14 and p16 encode entirely different with different functions (31). The ANRIL probe spans exons 56, thus detecting all known isoforms. In this cohort of islet samples from 95 unique nondiabetic donors (Supplemental Table 1), RNA abundance of p14, p16 and ANRIL were highly correlated with each other (Figure 1, normalized to ACTB, and Supplemental Figure 2, Diabetes Page 8 of 51

normalized to GAPDH). In contrast, abundance of p15, despite being physically located within the first intron of ANRIL, was poorly (ACTB normalization) or not (GAPDH normalization) correlated with p14, p16 or ANRIL. On the other hand, p15, but not p14, p16 or ANRIL, was highly correlated with MTAP expression. When the data were examined in multivariable linear models, integrating donor characteristics such as age into the model, again p14-ANRIL and p14-p16 were highly correlated, as were p15-MTAP (Supplemental Table 4). These results suggest two independent but overlapping coregulatory cassettes at the CDKN2A/B locus in human islets, with p14p16ANRIL in one and p15MTAP in the other.

Age-dependent gene expression increase of p14, p16 and ANRIL but not p15 or MTAP In many tissues, including islets, some CDKN2A/B locus genes increase with advancing age (9,10,32). In this cohort of human islets, expression of p14, p16 and ANRIL showed a modest positive correlation with donor age, whereas p15 and MTAP did not (Figure 2AE). Donor body mass index (BMI) could potentially confound the impact of age on gene expression; however, BMI was similar across donor ages (Figure 2F). Furthermore, we observed no correlation between donor BMI, sex or ethnicity and expression of any CDKN2A/B locus gene in univariate analysis (Supplemental Figures 35). Multivariable linear models integrating age, sex, race and BMI confirmed a positive correlation between p14, p16 and ANRIL with donor age, and confirmed a lack of impact of sex, race or BMI on locus gene expression (Supplemental Table 4). Scatter plots of gene expression versus age showed that some genes were expressed in very low abundance in islets from juvenile (age < 10 years) donors, with points falling well below the linear regression curve (Figure 2A, C, D). Focused analysis of juvenile (<10 years) versus adolescent/adult (>10 years) islets (Figure 2GK) revealed that expression of p14, p16 and ANRIL were markedly lower in juvenile islets, but MTAP and p15 were not, again suggesting altered regulatory characteristics of these two genes relative to other locus genes. Interestingly, an F test showed that the variances were reduced in juvenile islets (see standard deviation bars in Figure 2 GK) for p14 (p<0.0001), p16 (p<0.001) and ANRIL (p<0.0001) but not for p15 or MTAP (p=ns for both), despite the much smaller sample size, again suggesting fundamentally different biology of the juvenile samples. In sum, older age increased expression of p14, p16 and ANRIL, but not p15 or Page 9 of 51 Diabetes

MTAP, and the youthassociated suppression was exaggerated in islets from very young donors.

T2D risk-SNPs at rs10811661 and rs2383208 increased ANRIL expression in an age- defined subset of islet samples We next tested whether T2Drelated SNPs at CDKN2A/B impact locus gene expression. Validated Taqman genotyping procedures ascertained and confirmed the genotype of all n=95 preps for four T2D SNPs: rs564398 (hg38 chr9:22029548), rs2383208 (hg38 chr9:22132077), rs10811661 (hg38 chr9:22134095) and rs10757283 (hg38 chr9:22134173). Measured minor allele frequencies (MAF; Supplemental Table 2) were similar to reported MAFs for ethnicitymatched populations, supporting genotyping accuracy. SNPs rs2383208 and rs10811661 were tightly linked, with only 2 of 95 samples differing in our cohort, consistent with the linkage disequilibrium reported in LDpair and HaploReg (26,33,34) (Supplemental Tables 3 and 5). In raw analysis across the entire cohort, no SNP genotype correlated with abundance of any locus transcript by univariate (Figure 3) or multivariable (Supplemental Table 6) analysis. Since donor age impacted expression of p14, p16 and ANRIL, we assessed whether age interfered with the assessment of SNP effect on gene expression. Mean age was not significantly different between genotypes for any SNP (data not shown). However, expressing transcript abundance as a function of age revealed that for ANRIL, but not for p14 or p16, the agedependent increase was genotypedependent, evident only in protectiveallelecarrying samples at rs2383208 (Figure 4) and rs10811661 (Supplemental Figure 6). Homozygousrisk samples had high levels of ANRIL across all ages >10 years (Figure 4B, flat slope of AA regression line even despite the influence of the juvenile samples). Agegenotype interaction was not observed for any locus gene for rs10757283 or rs564398 (not shown). When the samples were reanalyzed using a different methodology, binning by quartiles, it was again evident that samples with protectiveallele at rs2383208 or rs10811661 showed an agedependent increase in ANRIL, but homozygous risk samples did not. In contrast, for p16 the slope of the agedependentgene expression regression lines (Figure 4A) and binning analysis (Figure 4C) were similar across genotypes. To test whether ANRIL abundance was inappropriately increased by homozygousrisk at rs2383208 or rs10811661 in samples from young donors, samples between the ages of 10 (to Diabetes Page 10 of 51

exclude juveniles, which were all suppressed independent of genotype) and 50 (defined by the intersection of the linear regression curves in 4B), was stratified by genotype (Figure 4E F). ANRIL, but not p16, abundance was significantly increased in younger homozygousrisk samples compared with protectiveallele carriers. Taken together, T2D homozygousrisk genotype at rs2383208 or rs10811661 prematurely increased ANRIL expression in islets of younger donors to olderdonor levels.

SNPs may interact with each other to combinatorially influence gene expression This cohort was not powered to perform subgroup analyses to definitively detect gene expression impact of SNP haplotype combinations. For example, a sample size analysis using our ANRIL expression mean and standard deviation for rs2383208 genotypes reveals that we would require n=54 per subgroup to achieve a power of 80% and type 1 error of 0.05. Given the diminishing number of samples as we partition by haplotype, we do not approach this sample size for subanalyses. Instead, we analyzed our dataset using a false discovery rate (FDR) approach to prioritize hypotheses to test in future studies such as ex vivo promoterenhancer experiments. We estimated that a risk tolerance of 90% likelihood that a hypothesis was correct would support future experimental investment. We then stratified our gene expression data by all SNP haplotype combinations and analyzed each comparison for likelihood of difference, defined by a FDR of 10% (Figure 5 and Supplemental Figure 7). By these criteria, we determined that genotype at rs564398 and rs10757283 may influence the impact of rs2383208 and/or rs10811661 on gene expression. For rs564398: in homozygous protective rs564398, but not riskallelecontaining samples, protective alleles at rs2383208 and rs10811661 may decrease abundance of p16 compared with homozygousrisk carriers. For rs10757283: in homozygousrisk rs10757283 samples, but not protectiveallele samples, protectiveallele at rs2383208 or rs10811661 may decrease abundance of p15. These observations suggest that individual SNPs may contribute risk by impacting locus biology only in the presence of other locus SNP genotypes, support investment in future experiments to test specific combinations, and help narrow which combinations to target.

CDKN2A/B T2D SNPs did not impact glucose-stimulated insulin secretion Page 11 of 51 Diabetes

A subset of islet samples (n=61) were tested for insulin secretion stimulation index at their respective islet isolation centers. Insulin secretion index was positively correlated with BMI but showed no relationship with donor age, sex, or islet isolation center by univariate (Supplemental Figure 8) or multivariable (Supplemental Table 7) analysis. When insulin secretion index was stratified by SNP genotype, contrary to the hypotheses that CDKN2A/B T2D riskSNPs impair glucose sensing, insulin production or stimulussecretion coupling, samples with T2D riskalleles did not show evidence for impairment in ex vivo insulin secretion in this cohort (Figure 6 and Supplemental Table 8).

Risk alleles at rs564398 reduced glucose-induced beta cell proliferation Since CDKN2A/B locus genes are best known for inhibiting the cell cycle, we assessed transcript markers of proliferation (KI67, PCNA and CCND2) in this cohort, as well as the actual rate of Sphase entry in growthstimulatory culture conditions, by BrdU labeling, in a subset of samples. Surprisingly, although PCNA and CCND2 showed a high degree of correlation with each other, KI67 did not correlate with either PCNA or CCND2 (Figure 7A C). No SNP genotype was correlated with abundance of PCNA, CCND2 or KI67 (not shown). Transcript level is only a surrogate for proliferation, and lacks sensitivity in a tissue with a very low frequency of proliferation events and a mixture of cell types. To measure actual cell cycle entry in beta cells we cultured a subset (n=47) of islet preparations in low (5mM) or high (15mM) glucose and quantified nuclear BrdU incorporation in insulin positive cells (Supplemental Figure 9). BrdU incorporation rate in 5mM glucose was nominally correlated with basal PCNA abundance, but not with KI67 or CCND2 (Supplemental Figure 10). As previously observed (27,28,35), glucose increased human beta cell proliferation (p<0.0001, not shown). To test whether any T2D SNP genotype impacted beta cell proliferation, the proliferation index (ratio of BrdU+ beta cells in 15mM compared to 5mM glucose) was stratified by SNP identity. Genotype at rs2383208, rs10811661 and rs10757283 did not influence the proliferation index (Figure 7DE). However, genotype at rs564398 was strongly associated with the human beta cell proliferation index, with homozygousprotective alleles showing approximately doubled stimulation of proliferation by 15mM glucose compared with islet samples harboring risk alleles at this SNP (Figure 7F).

Diabetes Page 12 of 51

DISCUSSION We have performed a comprehensive analysis of how one T2D GWAS locus associated with insulin secretory capacity in human populations influences human islet biology. In n=95 islet samples we quantified locus gene expression, SNP genotype, donor characteristics, beta cell function (insulin secretion) and beta cell proliferation. We have made several important observations. First, the locus contains two distinct gene cassettes that are physically overlapping but have different regulatory characteristics, one agedependent and the other not. Juvenile islets have markedly suppressed expression of p14p16ANRIL but not p15 MTAP. Second, individual T2D SNPs at the locus do not substantially alter expression of locus genes, but subtle agedependent influences are detected, which, contrary to expectations, impact the ANRIL lncRNA rather than the most prominent locus product, p16. SNPs may interact with each other to influence gene expression, increasing the complexity of genotype interpretation and raising intriguing mechanistic hypotheses. Third, genotype did not impact insulin secretion index. Finally, riskallele at rs564398, which is located within a transcribed exon of ANRIL, decreased the beta cell proliferation index. This work improves understanding of how CDKN2A/B T2D SNPs impact human islet biology, and suggests the influence of the locus on human insulin secretory capacity may be effected via beta cell mass rather than function.

How SNPs influence T2D risk is an important question in genetics today (3). Although the CDKN2A/B locus is associated with a number of different disease syndromes (36), the SNPs we selected to study are most strongly related to T2D risk, with the exception of rs564398, which is also associated with coronary disease and glaucoma (15). CDKN2A/B SNPs are associated with a range of diabetesrelated syndromes beyond T2D, such as gestational diabetes and transplantrelated diabetes (9). Since this locus is active in many different cell types, and the coronary disease risk region is mostly nonoverlapping with the T2D risk region (36), it is likely that CDKN2A/B impact on diseases other than T2D is mediated by effects outside the islet. Given the association of rs10811661/rs2383208 with impaired insulin production capacity, to identify possible T2D risk mechanisms we performed our study in islets. Our data highlight the complexity of genetic inputs to human metabolism. Even starting with a genomic locus repeatedly associated with disease risk across ethnicities Page 13 of 51 Diabetes

and T2Drelated syndromes (9), with in vivo evidence that the islet is the riskmediating tissue (16,19–21), abundant preclinical locus knowledge in model systems (9,32), and a fairly large sample size of the relevant tissue, the impact of riskSNPs is subtle. As with many studies in human islets, our data illustrate the marked variability from one donor to the next, which reflects the variability of outbred human populations. We incorporated donor parameters such as age and BMI in our analyses, but could not measure many potential other confounding premortem influences such as insulin sensitivity (liver, muscle, brain, fat), coexisting diseases and medications, environmental effects (diet, stresses, toxins), exercise history, prenatal events, and others. Islet stress related to donor demise, isolation, shipping and culture may also introduce variability (37). In this context, subtle effects observed in this challenging system may reflect large effects in certain subpopulations, or small effects present uniformly across variable conditions.

The presence of two gene expression cassettes at the locus, only one of which is age dependent, suggests interesting biology. p16 is well known to increase with age in many tissues and organisms (10). Our observation that p15 abundance did not increase with age in human islets conflicts with published results in other tissues (38,39). The relationship between MTAP and CDKN2A/B locus genes has not been extensively studied, but these are contained in the same HiC defined topologically associating domain in the human genome (Supplemental Figure 11) (40,41). Whether coregulation of p15 and MTAP has functional importance in the islet remains unknown.

rs2383208/rs10811661 impacted ANRIL abundance in agedependent fashion. Islets containing homozygousrisk alleles showed a “premature aging” phenotype, with young risk allele islets having ANRIL levels similar to older protectiveallele islets. The low MAF for these SNPs precluded comparison between homozygousprotective and homozygousrisk, which might have revealed a larger effect size. We do not currently know whether one of these, or another SNP in linkage with them, is causative. Fine mapping of this region has not revealed SNPs with greater impact on T2D risk (1,4), but islet ANRIL abundance was not the endpoint in those studies. These SNPs fall near a known regulatory region downstream of the 3’ end of ANRIL which may regulate ANRIL transcription. Mechanistically, how higher Diabetes Page 14 of 51

ANRIL abundance in islets might increase T2D risk is unknown. In other cell types, ANRIL is proproliferative (42), an effect mediated by ANRILdependent suppression of locus cell cycle inhibitors p14, p15 and p16 (43). Our RNA analyses did not detect any hint of negative correlation between ANRIL and p14/p16 in islets; in fact, the strong positive correlation between these transcripts calls into question whether ANRIL negatively regulates other locus genes in human islets.

Beyond SNP regulation of ANRIL abundance, a second observation also points to a role for ANRIL in human islets: rs564398, which influences the beta cell proliferation index, is a transcribed polymorphism in this longnoncoding RNA. Although the causative SNP remains unknown, it is possible that rs564983 itself impacts lncRNA activity. ANRIL is a complicated gene, with 20 exons and at least 14 reported isoforms (42). Some ANRIL variants are circular (44). Exon2 is not contained in all isoforms, but is generally associated with linear variants (44). Whether rs564398 identity impacts ANRIL isoform production, splicing, stability, or interaction with other cellular DNA, RNA or protein, to regulate human beta cell proliferation, remains unknown. Genotype at rs564398 did not correlate with expression of cell cycle genes, or BrdU incorporation, under basal conditions; rs564398 may be a marker for beta cell responsiveness to proliferationinducing conditions rather than increased proliferation in unstimulated conditions. Importantly, there are many SNPs in linkage with rs564398, and any of these, or a combination of these, could be influencing biology instead of rs564398 itself. Also important, while rs564398 has been repeatedly confirmed to be associated with T2D in Caucasian populations, it has little to no relationship to T2D risk in Asian populations (45–47). Taken together with the ‘premature aging’ influence of rs10811661/rs2383208 on ANRIL expression, CDKN2A/B SNPs may impact T2D risk by adversely impacting accrual of beta cell mass during early adulthood.

Combinatorial SNP regulation of gene expression increases the complexity of how genomic variation may impact cellular function. Wholegenome studies restricting locus analyses to single ‘lead’ SNPs cannot detect biology related to two or more local SNPs interacting with each other. The mechanism by which SNPs interact to regulate CDKN2A/B locus gene expression in human islets is unknown. Interaction between rs10757283 and rs2383208 or Page 15 of 51 Diabetes

rs10811661 linked SNPs to regulate p15 abundance could be via modulation of transcription factor occupancy or epigenetic regulation of the enhancer region in which they are located (48). Intriguingly, our observed interaction between these SNPs is supported by a complementary LDblock analysis which revealed that a haplotype consisting of rs2383208/rs10811661 and rs10757283 was associated with T2D risk (13). There are multiple SNPs in linkage disequilibrium with rs10811661 and also with rs10757283; actual causal polymorphisms are unknown. A physical or functional interaction between the ANRIL lncRNA and this enhancer may mediate cooperation between rs564398 and rs2383208, rs10811661 or linked SNPs to regulate p16 expression. Our study cannot distinguish between incis and intrans interaction. Focused, celltypespecific studies are needed to determine how SNP combinations influence locus gene expression.

This study adds to the body of knowledge debating the relative influence of beta cell mass versus function on T2D risk. Our study found that CDKN2A/B SNPs did not influence glucosestimulated insulin secretion, both in univariate analysis and in multivariable models incorporating donor age, sex, race and BMI, and islet isolation center. This is perhaps surprising given the in vivo data linking risk allele at rs10811661 with impaired insulin secretion (16,22). Since in vivo insulin secretion is a composite outcome of both mass and function, CDKN2A/B SNPs may impact beta cell mass but not beta cell function. This concept is in agreement with the widespread assumption that CDKN2A/B SNPs influence beta cell proliferation because of the known cell cycle inhibitory effects of locus genes p14, p15 and p16 (49). Our rs564398 data are the first demonstration, to our knowledge, of a CDKN2A/B locus SNP impacting beta cell proliferation.

Our study has caveats. Multiple cell types are found in human islets. Other than the proliferation measurements, which were assessed in insulinpositive cells, all other studies were performed on whole islets. We did not assess the relative proportion of islet cells that were beta cells, and to the extent that gene expression may be cell type specific, variable cellular makeup may have influenced results. In addition, there is considerable heterogeneity even among beta cells (50,51). Repeating our current analyses on sorted beta cells, or on single cells, exceeds our current resources. The insulin secretion data have caveats; using the Diabetes Page 16 of 51

IIDPreported insulin secretion index introduces technical variability, although benefits from freshly isolated, preshipment tissue. Our confirmation that insulin secretion correlated with BMI, but not with isolation center, are reassuring in this regard.

In sum, this work provides new information about how CDKN2A/B T2D SNPs impact islet biology, suggests the ANRIL lncRNA may play a role in human islets, and uncovers a link between a T2D SNP and beta cell proliferation. Further studies into the CDKN2A/B locus to develop a mechanistic understanding of how these SNPs impact islet biology to influence T2D risk could one day open the door for using personalized genomic information to inform T2D subtype definitions and therapeutic choice.

ACKNOWLEDGMENTS LCA devised and planned the experiments. YK performed the majority of the experiments; SL, RBS and RES also performed experiments. YK and LCA analyzed the data. WMJ performed the multivariable linear modeling. LCA, the guarantor of this work, wrote the manuscript; all authors viewed and had the opportunity to edit and approve of the manuscript. The authors have no conflicts of interest with the work contained in this manuscript. Human pancreatic islets were provided by the NIDDKfunded Integrated Islet Distribution Program (IIDP) at City of Hope, as well as Sambra Redick and David Harlan from the UMass Medical School, and Alvin C. Powers from Vanderbilt University. We are grateful to Ahmet Rasim Barutcu, from the Broad Institute and Harvard University, for helpful guidance with the TAD analysis. We would like to thank the Beta Cell Biology Group at the University of Massachusetts Medical School for many helpful discussions. This work was supported by NIH/NIDDK: R01DK095140 (LCA), DK104211 (ACP), DK106755 (ACP), 2UC4DK098085 (IIDP) and by the American Diabetes Association grant #115BS003 (LCA) in collaboration with the Order of the Amaranth. Page 17 of 51 Diabetes

REFERENCES

1. Morris AP. Fine Mapping of Type 2 Diabetes Susceptibility Loci. Curr Diab Rep [Internet]. 2014 [cited 2015 Dec 30];14(11). Available from: http://www.ncbi.nlm.nih.gov/pmc/articles/PMC4169615/

2. Morris AP, Voight BF, Teslovich TM, Ferreira T, Segrè AV, Steinthorsdottir V, et al. Largescale association analysis provides insights into the genetic architecture and pathophysiology of type 2 diabetes. Nat Genet. 2012 Sep;44(9):981–90.

3. Fuchsberger C, Flannick J, Teslovich TM, Mahajan A, Agarwala V, Gaulton KJ, et al. The genetic architecture of type 2 diabetes. Nature. 2016 Aug 4;536(7614):41–7.

4. Gaulton KJ, Ferreira T, Lee Y, Raimondo A, Mägi R, Reschen ME, et al. Genetic fine mapping and genomic annotation defines causal mechanisms at type 2 diabetes susceptibility loci. Nat Genet. 2015 Dec;47(12):1415–25.

5. van de Bunt M, Manning Fox JE, Dai X, Barrett A, Grey C, Li L, et al. Transcript Expression Data from Human Islets Links Regulatory Signals from GenomeWide Association Studies for Type 2 Diabetes and Glycemic Traits to Their Downstream Effectors. PLoS Genet. 2015 Dec;11(12):e1005694.

6. Fadista J, Vikman P, Laakso EO, Mollet IG, Esguerra JL, Taneera J, et al. Global genomic and transcriptomic analysis of human pancreatic islets reveals novel genes influencing glucose metabolism. Proc Natl Acad Sci. 2014 Sep 23;111(38):13924–9.

7. Varshney A, Scott LJ, Welch RP, Erdos MR, Chines PS, Narisu N, et al. Genetic regulatory signatures underlying islet gene expression and type 2 diabetes. Proc Natl Acad Sci U S A. 2017 Feb 28;114(9):2301–6.

8. Gaulton KJ. Mechanisms of Type 2 Diabetes Risk Loci. Curr Diab Rep. 2017 Sep;17(9):72.

9. Kong Y, Sharma RB, Nwosu BU, Alonso LC. Islet biology, the CDKN2A/B locus and type 2 diabetes risk. Diabetologia. 2016 Aug;59(8):1579–93.

10. Kim WY, Sharpless NE. The regulation of INK4/ARF in cancer and aging. Cell. 2006 Oct 20;127(2):265–75.

11. Sharpless NE, Sherr CJ. Forging a signature of in vivo senescence. Nat Rev Cancer. 2015 Jul;15(7):397–408.

12. Scott LJ, Mohlke KL, Bonnycastle LL, Willer CJ, Li Y, Duren WL, et al. A genome wide association study of type 2 diabetes in Finns detects multiple susceptibility variants. Science. 2007 Jun 1;316(5829):1341–5. Diabetes Page 18 of 51

13. Zeggini E, Weedon MN, Lindgren CM, Frayling TM, Elliott KS, Lango H, et al. Replication of GenomeWide Association Signals in UK Samples Reveals Risk Loci for Type 2 Diabetes. Science. 2007 Jun 1;316(5829):1336–41.

14. Saxena R, Voight BF, Lyssenko V, Burtt NP, de Bakker PIW, Chen H, et al. Genome wide association analysis identifies loci for type 2 diabetes and triglyceride levels. Science. 2007 Jun 1;316(5829):1331–6.

15. Cunnington MS, Santibanez Koref M, Mayosi BM, Burn J, Keavney B. 9p21 SNPs Associated with Multiple Disease Phenotypes Correlate with ANRIL Expression. PLoS Genet. 2010 Apr;6(4):e1000899.

16. Grarup N, Rose CS, Andersson EA, Andersen G, Nielsen AL, Albrechtsen A, et al. Studies of association of variants near the HHEX, CDKN2A/B, and IGF2BP2 genes with type 2 diabetes and impaired insulin release in 10,705 Danish subjects: validation and extension of genomewide association studies. Diabetes. 2007 Dec;56(12):3105–11.

17. Xu M, Bi Y, Xu Y, Yu B, Huang Y, Gu L, et al. Combined effects of 19 common variations on type 2 diabetes in Chinese: results from two communitybased studies. PloS One. 2010;5(11):e14022.

18. Peng F, Hu D, Gu C, Li X, Li Y, Jia N, et al. The relationship between five widely evaluated variants in CDKN2A/B and CDKAL1 genes and the risk of type 2 diabetes: a metaanalysis. Gene. 2013 Dec 1;531(2):435–43.

19. Hribal ML, Presta I, Procopio T, Marini MA, Stančáková A, Kuusisto J, et al. Glucose tolerance, insulin sensitivity and insulin release in European nondiabetic carriers of a polymorphism upstream of CDKN2A and CDKN2B. Diabetologia. 2011 Apr;54(4):795– 802.

20. Hart LM ‘t, SimonisBik AM, Nijpels G, Haeften TW van, Schäfer SA, Houwing Duistermaat JJ, et al. Combined Risk Allele Score of Eight Type 2 Diabetes Genes Is Associated With Reduced FirstPhase GlucoseStimulated Insulin Secretion During Hyperglycemic Clamps. Diabetes. 2010 Jan 1;59(1):287–92.

21. Dimas AS, Lagou V, Barker A, Knowles JW, Mägi R, Hivert MF, et al. Impact of Type 2 Diabetes Susceptibility Variants on Quantitative Glycemic Traits Reveals Mechanistic Heterogeneity. Diabetes. 2014 Jun 1;63(6):2158–71.

22. Jonsson A, Ladenvall C, Ahluwalia TS, Kravic J, Krus U, Taneera J, et al. Effects of Common Genetic Variants Associated With Type 2 Diabetes and Glycemic Traits on α and βCell Function and Insulin Action in Humans. Diabetes. 2013 Aug 1;62(8):2978– 83.

23. Brereton MF, Rohm M, Ashcroft FM. βCell dysfunction in diabetes: a crisis of identity? Diabetes Obes Metab. 2016 Sep;18 Suppl 1:102–9. Page 19 of 51 Diabetes

24. Dai C, Hant Y, Shostak A, Poffenberger G, Hart N, Prasad N, et al. Agedependent human βcell proliferation induced by GLP1 and calcineurin signaling. J Clin Invest. in press.

25. 1000 Genomes Project Consortium, Auton A, Brooks LD, Durbin RM, Garrison EP, Kang HM, et al. A global reference for human genetic variation. Nature. 2015 Oct 1;526(7571):68–74.

26. Machiela MJ, Chanock SJ. LDlink: a webbased application for exploring population specific haplotype structure and linking correlated alleles of possible functional variants. Bioinformatics. 2015 Nov 1;31(21):3555–7.

27. Sharma RB, O’Donnell AC, Stamateris RE, Ha B, McCloskey KM, Reynolds PR, et al. Insulin demand regulates β cell number via the unfolded protein response. J Clin Invest. 2015 Oct 1;125(10):3831–46.

28. Stamateris RE, Sharma RB, Kong Y, Ebrahimpour P, Panday D, Ranganath P, et al. Glucose induces mouse beta cell proliferation via IRS2, mTOR and but not the insulin receptor. Diabetes. (in press, 2016).

29. Pascoe J, Hollern D, Stamateris R, Abbasi M, Romano LC, Zou B, et al. Free fatty acids block glucoseinduced βcell proliferation in mice by inducing cell cycle inhibitors p16 and p18. Diabetes. 2012 Mar;61(3):632–41.

30. Alonso LC, Yokoe T, Zhang P, Scott DK, Kim SK, O’Donnell CP, et al. Glucose infusion in mice: a new model to induce betacell replication. Diabetes. 2007 Jul;56(7):1792–801.

31. Robertson KD, Jones PA. Tissuespecific alternative splicing in the human INK4a/ARF cell cycle regulatory locus. Oncogene. 1999 Jul 1;18(26):3810–20.

32. Krishnamurthy J, Ramsey MR, Ligon KL, Torrice C, Koh A, BonnerWeir S, et al. p16INK4a induces an agedependent decline in islet regenerative potential. Nature. 2006 Sep 28;443(7110):453–7.

33. Ward LD, Kellis M. HaploReg: a resource for exploring chromatin states, conservation, and regulatory motif alterations within sets of genetically linked variants. Nucleic Acids Res. 2012 Jan;40(Database issue):D930–4.

34. Gaunt TR, Rodríguez S, Day IN. Cubic exact solutions for the estimation of pairwise haplotype frequencies: implications for linkage disequilibrium analyses and a web tool “CubeX.” BMC Bioinformatics. 2007 Nov 2;8:428.

35. Levitt HE, Cyphert TJ, Pascoe JL, Hollern DA, Abraham N, Lundell RJ, et al. Glucose stimulates human beta cell replication in vivo in islets transplanted into NODsevere combined immunodeficiency (SCID) mice. Diabetologia. 2011 Mar;54(3):572–82. Diabetes Page 20 of 51

36. Jeck WR, Siebold AP, Sharpless NE. Review: a metaanalysis of GWAS and age associated diseases. Aging Cell. 2012 Oct;11(5):727–31.

37. Stiegler P, Stadlbauer V, Hackl F, Schaffellner S, Iberer F, Greilberger J, et al. Prevention of oxidative stress in porcine islet isolation. J Artif Organs Off J Jpn Soc Artif Organs. 2010 Apr;13(1):38–47.

38. Menthena A, Koehler CI, Sandhu JS, Yovchev MI, Hurston E, Shafritz DA, et al. Activin A, p15INK4b Signaling, and Cell Competition Promote Stem/Progenitor Cell Repopulation of Livers in Aging Rats. Gastroenterology. 2011 Mar;140(3):1009–20.

39. Wang F, Zhang L. p15INK4b regulates cell cycle signaling in hippocampal astrocytes of aged rats. Aging Clin Exp Res. 2016 Oct 1;28(5):813–21.

40. Zhou X, Maricque B, Xie M, Li D, Sundaram V, Martin EA, et al. The Human Epigenome Browser at Washington University. Nat Methods. 2011 Nov 29;8(12):989– 90.

41. Wang Y, Zhang B, Zhang L, An L, Xu J, Li D, et al. The 3D Genome Browser: a web based browser for visualizing 3D genome organization and longrange chromatin interactions | bioRxiv. BioRxiv [Internet]. 2017 [cited 2017 Dec 15]; Available from: https://www.biorxiv.org/content/early/2017/02/27/112268

42. Congrains A, Kamide K, Ohishi M, Rakugi H. ANRIL: molecular mechanisms and implications in human health. Int J Mol Sci. 2013;14(1):1278–92.

43. Aguilo F, Zhou MM, Walsh MJ. Long noncoding RNA, polycomb, and the ghosts haunting INK4bARFINK4a expression. Cancer Res. 2011 Aug 15;71(16):5365–9.

44. Burd CE, Jeck WR, Liu Y, Sanoff HK, Wang Z, Sharpless NE. Expression of linear and novel circular forms of an INK4/ARFassociated noncoding RNA correlates with atherosclerosis risk. PLoS Genet. 2010;6(12):e1001233.

45. Wu Y, Li H, Loos RJF, Yu Z, Ye X, Chen L, et al. Common variants in CDKAL1, CDKN2A/B, IGF2BP2, SLC30A8, and HHEX/IDE genes are associated with type 2 diabetes and impaired fasting glucose in a Chinese Han population. Diabetes. 2008 Oct;57(10):2834–42.

46. Ng MCY, Park KS, Oh B, Tam CHT, Cho YM, Shin HD, et al. Implication of genetic variants near TCF7L2, SLC30A8, HHEX, CDKAL1, CDKN2A/B, IGF2BP2, and FTO in type 2 diabetes and obesity in 6,719 Asians. Diabetes. 2008 Aug;57(8):2226–33.

47. Horikawa Y, Miyake K, Yasuda K, Enya M, Hirota Y, Yamagata K, et al. Replication of genomewide association studies of type 2 diabetes susceptibility in Japan. J Clin Endocrinol Metab. 2008 Aug;93(8):3136–41. Page 21 of 51 Diabetes

48. Pasquali L, Gaulton KJ, RodríguezSeguí SA, Mularoni L, MiguelEscalada I, Akerman I, et al. Pancreatic islet enhancer clusters enriched in type 2 diabetes riskassociated variants. Nat Genet. 2014 Feb;46(2):136–43.

49. Hannou SA, Wouters K, Paumelle R, Staels B. Functional genomics of the CDKN2A/B locus in cardiovascular and metabolic disease: what have we learned from GWASs? Trends Endocrinol Metab. 2015 Apr;26(4):176–84.

50. Gutierrez GD, Gromada J, Sussel L. Heterogeneity of the Pancreatic Beta Cell. Front Genet. 2017;8:22.

51. Roscioni SS, Migliorini A, Gegg M, Lickert H. Impact of islet architecture on βcell heterogeneity, plasticity and function. Nat Rev Endocrinol. 2016 Dec;12(12):695–709.

Diabetes Page 22 of 51

FIGURE LEGENDS

Figure 1. CDKN2A/B locus genes were expressed coordinately in human islets. (A) Diagram of the CDKN2A/B locus at 9p21, adapted from the UCSC genome browser GRCh38/hg38 assembly. Vertical arrows show the locations of T2D SNPs tested in this study, by linkage block: green (rs564398; leftmost), blue (rs2383208 and rs10811661; middle two) and red (rs10757283; rightmost). (BD) Abundance of p14, p16 and ANRIL were highly correlated in human islet samples. (EG) p15 abundance did not correlate with p16, and only marginally correlated with p14 and ANRIL. MTAP expression was marginally correlated with p14, p16 and ANRIL (H and not shown), but (I) highly correlated with p15 expression. mRNA abundance is expressed as deltaCt, normalized to ACTB. Line, p values and Rsquared values were calculated by linear regression; n=95 for all panels. Red lines highlight correlations with higher Rsquared values.

Figure 2. Abundance of p14, p16 and ANRIL, but not p15 or MTAP, was correlated with donor age and strongly reduced in juvenile islets. (AE) Consistent with prior observations, p16 mRNA abundance was positively correlated with donor age (in years). p14 and ANRIL were also correlated with age, but p15 and MTAP were not. Age accounted for only a small proportion of the variance in gene expression, even for p16. (F) BMI partitioned equally across donor age in this cohort (dotted lines demarcate BMI 1825 (normal weight), 2530 (overweight) and >30 (obese). (GK) Islets from juvenile donors (age <10 years) contained markedly less p14, p16 and ANRIL, but not p15 or MTAP, than older islets. mRNA abundance is expressed as deltaCt, normalized to ACTB. Mean +/ SD; p values and Rsquared values were calculated by linear regression (AF) or by Student’s Ttest (GK). All panels n=92; missing values are due to lack of donor information for age and BMI (3 samples).

Figure 3. In crude analysis, individual SNP identity did not impact expression of locus genes in human islets. Risk allele for each SNP, the rightmost genotype in each case, is in red. All comparisons were nonsignificant by ANOVA with correction for multiple comparisons. ANRIL showed a trend towards higher abundance in islets with homozygousrisk for rs10811661 (p=0.08) and rs2383208 (p=0.07) compared with protectiveallelecarrying samples. Page 23 of 51 Diabetes

rs108=rs10811661; rs238=rs2383208; rs107=rs10757283; rs564=rs564398. mRNA abundance is expressed as deltaCt, normalized to ACTB. Mean +/ SD; n=95 for all subpanels.

Figure 4. Age interacted with genotype at rs2383208 to determine ANRIL abundance; young donors with protective alleles had lower ANRIL expression. (AB) Expressing p16 (A) or ANRIL (B) abundance as a function of donor age, stratified by genotype, showed that unlike p16, agedependence of ANRIL was driven by samples of rs2383208GG+GA genotype and was absent in samples with AA (homozygous risk) genotype. (CD) Binning analysis of the cohort (nonjuvenile samples separated by quartiles) illustrated the agedependent ANRIL increase in GG+GA samples but not in homozygousrisk AA samples. Juveniles <10 years of age showed markedly lower abundance, independent of genotype. (EF) In younger donors (ages 1050; lower threshold defined by juvenile cutoff and upper threshold defined by the intersection of the regression curves in (B), which is 50.8) homozygousrisk increased ANRIL abundance. mRNA abundance is expressed as deltaCt, normalized to ACTB. Statistics by linear regression (AB), ANOVA (DF) with overall ANOVA significance in upper left corner and significant pairwise comparisons after correction for multiple comparisons labeled. Sample size: (AD) n=92 (3 samples missing age) and (EF) n=57 samples between the ages of 1050.

Figure 5. SNP combinatorial haplotypes may influence locus gene expression. (A) Schematic showing approximate locations of the T2D SNPs analyzed in this study, relative to the ANRIL gene. SNP colors, as in Figure 1A, indicate linkage disequilibrium. (BC): Protective alleles of rs10811661 (shown) and rs2383208 (Supplemental Fig 4) may decrease expression of p16 in homozygousprotective rs564398-CC samples. The same comparison for p14 did not meet FDR<10% (qvalue 17%); for ANRIL, FDR>20% (not shown). (DE): Homozygous risk alleles at both neighboring SNPs rs10757283 and rs10811661 may collaboratively increase p15 expression. The same comparison for MTAP showed FDR>20%. * indicates FDR<10%, our pre determined risk tolerance for future experiments exploring haplotype hypotheses. mRNA abundance is expressed as deltaCt, normalized to ACTB. n=95 for all panels. All other inter SNP comparisons, both shown and not shown (aside from those in Supplemental Figure 4), resulted FDR>10% or had insufficient data points to evaluate (defined as n=2 or fewer). FDR, false discovery rate. Diabetes Page 24 of 51

Figure 6. Insulin secretion was similar across SNP genotypes. 61 of the islet preparations were tested for glucosestimulated insulin release “stimulation index” at islet isolation centers; the IIDPderived insulin secretion index is plotted against donor genotype. No relationship is evident between T2D SNP genotype and IIDPreported insulin secretory index. rs108=rs10811661; rs238=rs2383208; rs107=rs10757283; rs564= rs564398. n=61 for all SNPs.

Figure 7. Risk allele at rs564398 suppressed glucose induction of beta cell proliferation. (A C) RNA abundance of proliferationrelated genes PCNA, CCND2 and KI67 in flashfrozen islets showed a strong correlation between PCNA and CCND2 (A) but not with KI67 (BC). (DF) rs564398, but not rs2383208, rs10811661 or rs10757283 was associated with beta cell proliferation. Islets containing 1 or 2 T2D risk alleles at rs564398 had lower proliferation index than islets containing homozygousprotective alleles at rs564398. Dispersed islets were cultured on glass coverslips for 96 hours in either 5mM glucose (unstimulated) or 15mM glucose (stimulated) with BrdU present for the whole 96 hours. Cultures were fixed, immunostained, imaged, blinded and the percent insulin(+) cells that were also BrdU(+) quantified by manual counting. Plotted is the proliferation index, which is the ratio of 15mM to 5mM. n=95 (AC) and n=43 (DF; 47 preps tested, but 4 preps had 0% BrdU+ in 5mM glucose and thus could not calculate an index). mRNA abundance is expressed as deltaCt, normalized to ACTB. Mean +/ SD; p values are by linear regression (AC) and ANOVA with correction for multiple comparisons (DF).

Page 25 of 51 Diabetes

Figure 1. CDKN2A/B locus genes were expressed coordinately in human islets. (A) Diagram of the CDKN2A/B locus at 9p21, adapted from the UCSC genome browser GRCh38/hg38 assembly. Vertical arrows show the locations of T2D SNPs tested in this study, by linkage block: green (rs564398; left-most), blue (rs2383208 and rs10811661; middle two) and red (rs10757283; right-most). (B-D) Abundance of p14, p16 and ANRIL were highly correlated in human islet samples. (E-G) p15 abundance did not correlate with p16, and only marginally correlated with p14 and ANRIL. MTAP expression was marginally correlated with p14, p16 and ANRIL (H and not shown), but (I) highly correlated with p15 expression. mRNA abundance is expressed as delta-Ct, normalized to ACTB. Line, p values and R-squared values were calculated by linear regression; n=95 for all panels. Red lines highlight correlations with higher R-squared values.

172x163mm (300 x 300 DPI)

Diabetes Page 26 of 51

Figure 2. Abundance of p14, p16 and ANRIL, but not p15 or MTAP, was correlated with donor age and strongly reduced in juvenile islets. (A-E) Consistent with prior observations, p16 mRNA abundance was positively correlated with donor age (in years). p14 and ANRIL were also correlated with age, but p15 and MTAP were not. Age accounted for only a small proportion of the variance in gene expression, even for p16. (F) BMI partitioned equally across donor age in this cohort (dotted lines demarcate BMI 18-25 (normal weight), 25-30 (overweight) and >30 (obese). (G-K) Islets from juvenile donors (age <10 years) contained markedly less p14, p16 and ANRIL, but not p15 or MTAP, than older islets. mRNA abundance is expressed as delta-Ct, normalized to ACTB. Mean +/- SD; p values and R-squared values were calculated by linear regression (A-F) or by Student’s T-test (G-K). All panels n=92; missing values are due to lack of donor information for age and BMI (3 samples).

172x163mm (300 x 300 DPI)

Page 27 of 51 Diabetes

Figure 3. In crude analysis, individual SNP identity did not impact expression of locus genes in human islets. Risk allele for each SNP, the right-most genotype in each case, is in red. All comparisons were non- significant by ANOVA with correction for multiple comparisons. ANRIL showed a trend towards higher abundance in islets with homozygous-risk for rs10811661 (p=0.08) and rs2383208 (p=0.07) compared with protective-allele-carrying samples. rs108=rs10811661; rs238=rs2383208; rs107=rs10757283; rs564=rs564398. mRNA abundance is expressed as delta-Ct, normalized to ACTB. Mean +/- SD; n=95 for all sub-panels.

172x163mm (300 x 300 DPI)

Diabetes Page 28 of 51

Figure 4. Age interacted with genotype at rs2383208 to determine ANRIL abundance; young donors with protective alleles had lower ANRIL expression. (A-B) Expressing p16 (A) or ANRIL (B) abundance as a function of donor age, stratified by genotype, showed that unlike p16, age-dependence of ANRIL was driven by samples of rs2383208-GG+GA genotype and was absent in samples with AA (homozygous risk) genotype. (C-D) Binning analysis of the cohort (non-juvenile samples separated by quartiles) illustrated the age-dependent ANRIL increase in GG+GA samples but not in homozygous-risk AA samples. Juveniles <10 years of age showed markedly lower abundance, independent of genotype. (E-F) In younger donors (ages 10-50; lower threshold defined by juvenile cutoff and upper threshold defined by the intersection of the regression curves in (B), which is 50.8) homozygous-risk increased ANRIL abundance. mRNA abundance is expressed as delta-Ct, normalized to ACTB. Statistics by linear regression (A-B), ANOVA (D-F) with overall ANOVA significance in upper left corner and significant pairwise comparisons after correction for multiple comparisons labeled. Sample size: (A-D) n=92 (3 samples missing age) and (E-F) n=57 samples between the ages of 10-50.

172x163mm (300 x 300 DPI)

Page 29 of 51 Diabetes

Figure 5. SNP combinatorial haplotypes may influence locus gene expression. (A) Schematic showing approximate locations of the T2D SNPs analyzed in this study, relative to the ANRIL gene. SNP colors, as in Figure 1A, indicate linkage disequilibrium. (B-C): Protective alleles of rs10811661 (shown) and rs2383208 (Supplemental Fig 4) may decrease expression of p16 in homozygous-protective rs564398-CC samples. The same comparison for p14 did not meet FDR<10% (q-value 17%); for ANRIL, FDR>20% (not shown). (D-E): Homozygous risk alleles at both neighboring SNPs rs10757283 and rs10811661 may collaboratively increase p15 expression. The same comparison for MTAP showed FDR>20%. * indicates FDR<10%, our pre- determined risk tolerance for future experiments exploring haplotype hypotheses. mRNA abundance is expressed as delta-Ct, normalized to ACTB. n=95 for all panels. All other inter-SNP comparisons, both shown and not shown (aside from those in Supplemental Figure 4), resulted FDR>10% or had insufficient data points to evaluate (defined as n=2 or fewer). FDR, false discovery rate.

172x163mm (300 x 300 DPI)

Diabetes Page 30 of 51

Figure 6. Insulin secretion was similar across SNP genotypes. 61 of the islet preparations were tested for glucose-stimulated insulin release “stimulation index” at islet isolation centers; the IIDP-derived insulin secretion index is plotted against donor genotype. No relationship is evident between T2D SNP genotype and IIDP-reported insulin secretory index. rs108=rs10811661; rs238=rs2383208; rs107=rs10757283; rs564= rs564398. n=61 for all SNPs.

172x163mm (300 x 300 DPI)

Page 31 of 51 Diabetes

Figure 7. Risk allele at rs564398 suppressed glucose induction of beta cell proliferation. (A-C) RNA abundance of proliferation-related genes PCNA, CCND2 and KI67 in flash-frozen islets showed a strong correlation between PCNA and CCND2 (A) but not with KI67 (B-C). (D-F) rs564398, but not rs2383208, rs10811661 or rs10757283 was associated with beta cell proliferation. Islets containing 1 or 2 T2D risk alleles at rs564398 had lower proliferation index than islets containing homozygous-protective alleles at rs564398. Dispersed islets were cultured on glass coverslips for 96 hours in either 5mM glucose (unstimulated) or 15mM glucose (stimulated) with BrdU present for the whole 96 hours. Cultures were fixed, immunostained, imaged, blinded and the percent insulin(+) cells that were also BrdU(+) quantified by manual counting. Plotted is the proliferation index, which is the ratio of 15mM to 5mM. n=95 (A-C) and n=43 (D-F; 47 preps tested, but 4 preps had 0% BrdU+ in 5mM glucose and thus could not calculate an index). mRNA abundance is expressed as delta-Ct, normalized to ACTB. Mean +/- SD; p values are by linear regression (A-C) and ANOVA with correction for multiple comparisons (D-F).

172x163mm (300 x 300 DPI)

Diabetes Page 32 of 51

ACTN GAPDH p14 A B C D p15 p<0.0001 p<0.0001 p<0.0001 p<0.0001 R2=0.85 R2=0.94 R2=0.95 R2=0.92 Ct value 1 Ct value 1 Ct value 1 Ct value 1 RPD RPD RPD RPD 1.13% 0.81% 0.65% 1.05% 16 18 20 22 24 26 28 30 32 34 36 22 24 26 28 30 32 34 16 18 20 22 24 16 18 20 22 24 16 18 20 22 24 26 28 30 32 34 36 22 24 26 28 30 32 34 Ct value 2 Ct value 2 Ct value 2 Ct value 2

E F G H p15 p16 ANRIL MTAP p<0.0001 p<0.0001 p<0.0001 p<0.0001 R2=0.92 R2=0.97 R2=0.87 R2=0.94 Ct value 1 Ct value 1 Ct value 1 Ct value 1 RPD RPD RPD RPD 1.05% 0.56% 1.30% 0.60% 24 26 28 30 32 34 25 30 35 40 22 24 26 28 30 32 22 24 26 28 30 32 34 22 24 26 28 30 32 34 24 26 28 30 32 34 25 30 35 40 22 24 26 28 30 32 Ct value 2 Ct value 2 Ct value 2 Ct value 2

I J K PCNA KI67 CCND2 p<0.0001 p<0.0001 p<0.0001 R2=0.88 R2=0.99 R2=0.96 Ct value 1 Ct value 1 Ct value 1 RPD RPD RPD 0.64% 0.60% 0.67% 24 26 28 30 32 34 20 25 30 35 40 20 22 24 26 28 30 24 26 28 30 32 34 20 25 30 35 40 20 22 24 26 28 30 Ct value 2 Ct value 2 Ct value 2

Supplemental Figure 1. Taqman gene expression duplicates show high reproducibility. For each CDKN2A/B locus gene, Taqman Ct values of replicate 1 (Ct value 1) and replicate 2 (Ct value 2) are plotted. The RPD values, the absolute value of the Relative Percentage Di erence, were calculated from the equa- tion RPD=(D1-D2)/(D1+D2)/2X100. RPDs for all genes were well within established goals for biological replicate variability of 10%. Line, p values and R-squared values were calculated by linear regression; n=95 for all panels. Page 33 of 51 Diabetes

A p<0.0001 B p<0.0001 C p<0.0001 D R2=0.42 R2=0.83 R2=0.37 p=ns p1 4 p1 6 p1 5 ANRI L 0.0005 0.0010 0.01 0.02 0.02 0.04 0.002 0.004

0.0000 0.0005 0.0010 0.000 0.002 0.004 0.00 0.01 0.02 0.00 0.01 0.02 ANRIL p14 p16 p16

E p=ns F p=ns G p<0.01 H p<0.0001 R2=0.08 R2=0.52 0.02 0.04 0.02 0.04 p1 5 p1 5 MTAP MTAP 0.02 0.04 0.02 0.04

0.000 0.002 0.004 0.0000 0.0005 0.0010 0.0000 0.0005 0.0010 0.00 0.02 0.04 p14 ANRIL ANRIL p15

Supplemental Figure 2. CDKN2A/B locus gene expression in human islets normalized to GAPDH. As observed when normalized to ACTIN, CDKN2A/B locus genes when normalized to islet GAPDH abundance revealed a high degree of correlation between p14-p16-ANRIL (A-C). p15 and MTAP showed little or no correlation with p14, p16 or ANRIL (D-G). However, p15 and MTAP expression correlated with each other. mRNA abundance is expressed as delta-Ct, normalized to GAPDH. Line, p values and R-squared values were calculated by linear regression; n=95 for all panels. Red lines high- light correlations with higher R-squared values. Diabetes Page 34 of 51

A B C p1 6 p1 4 p1 5 0.01 0.02 0.02 0.04 0.002 0.004

0 20 40 60 0 20 40 60 0 20 40 60 BMI BMI BMI

D E 0.02 0.04 MTAP ANRI L 0.0005 0.0010

0 20 40 60 0 20 40 60 BMI BMI

Supplemental Figure 3. Donor BMI did not correlate with locus gene expression. (A-E) Reported donor BMI was compared with islet gene expression by univariate linear regression analysis. No locus gene mRNA abundance was correlated with BMI. mRNA abundance is expressed as delta-Ct, normalized to actin. Statistical analysis was calculated by linear regres- sion; n=92 for all panels. Missing values include the 3 islet samples for which donor BMI was not available. Page 35 of 51 Diabetes

A B C p1 4 p1 5 p1 6 0.01 0.02 0.002 0.004 0.01 0.02 0.03

FM FM FM

D E F p<0.05 Ag e 0.02 0.04 MTAP ANRI L 0.0005 0.0010 Donor 0 20 40 60 80 FM FM FM

Supplemental Figure 4. Locus gene expression did not vary by donor sex. (A-E) No relationship was observed between locus gene expression and donor sex. F: female; M: male. (F) The mean age of male donors was signicantly lower than that of females, owing to an unfortunately high number of teenage and young-adult male donors. After age-matching the male and female cohorts (by removing all samples from both sexes with age < 27 years), reanalysis conrmed that no locus gene expression correlated with donor sex (data not shown). mRNA abundance is expressed as delta-Ct, normalized to actin. Statistical analy- sis was calculated by Student’s t-test; n=42 females and 48 males for all panels. Missing values include the 5 islet samples for which donor sex was not available. Diabetes Page 36 of 51

A B C p1 4 p1 5 p1 6 0.01 0.02 0.03 0.00 0.01 0.02 0.000 0.002 0.004 Black Black Black Asian Asian Asian White White White Hispanic Hispanic Hispanic D E F

80

60 Ag e 40 MTAP ANRI L

Donor 20 0.01 0.02 0.03

0.0000 0.0005 0.0010 0 Black Black Black Asian Asian Asian White White White Hispanic Hispanic Hispanic

Supplemental Figure 5. Locus gene expression did not di er by donor ethnici- ty. (A-E) No relationship was observed between locus genes and donor ethnicity. (F) Mean donor age was not di erent in the ethnicity categories. mRNA abundance is expressed as delta-Ct, normalized to actin. Statistical analysis was calculated by ANOVA; p=ns for all comparisons. Sample sizes are n=1 (Asian), n=8 (black), n=14 (hispanic) and n=66 (white). Missing values include the 6 islet samples for which donor ethnicity was not available. Page 37 of 51 Diabetes

A C E p=0.05 p<0.05 p=ns 0.025 p=0.03, R2=0.20 p=0.001, R2=0.16 0.020

0.015 p1 6 p1 6 p1 6 0.010

0.005

0.000 0.00 0.01 0.02 0.00 0.01 0.02 0 20 40 60 80 T TT C CC <10 <10

Donor age 15-31 32-44 45-51 52-68 15-31 32-44 45-51 52-68

B D p<0.01 p=ns F p=0.0002, R2=0.46 p<0.05 0.0015 p=ns p<0.01

p<0.05 0.0010 p<0.05 ANRI L ANRI L ANRI L 0.0005

0.0000 0.0000 0.0005 0.0010 0 20 40 60 80 T 0.0000 0.0004 0.0008 TT C CC <10 <10

Donor age 15-31 32-44 45-51 52-68 15-31 32-44 45-51 52-68 All panels: rs10811661 CC+CT TT C=protective, T=risk

Supplemental Figure 6. Age interacts with genotype at rs10811661 to determine ANRIL abundance; young donors with protective alleles had lower ANRIL expression. Similar analysis to Figure 4 for rs2383208. (A-B) Expressing p16 (A) or ANRIL (B) abundance as a function of donor age, stratied by geno- type, showed that unlike p16, age-dependence of ANRIL was driven by samples of rs10811661-CC+CT geno- type and was absent in samples with TT genotype. (C-D) Binning analysis of the cohort (non-juvenile sam- ples separated by quartiles) illustrated an age-dependent ANRIL increase in CC+CT samples but not in homozygous-risk TT samples. Juveniles <10 years of age showed markedly di erent biology, independent of genotype. (E-F) In younger donors (ages 10-50; lower threshold dened by juvenile cuto and upper thresh- old dened by the intersection of the regression curves in (B), which is 50.8) homozygous-risk increased ANRIL abundance. mRNA abundance is expressed as delta-Ct, normalized to actin. Statistics by linear regres- sion (A-B), ANOVA (D-F) with overall ANOVA signicance in upper left corner and signicant pairwise com- parisons after correction for multiple comparisons labeled. Sample size: (A-D) n=92 (3 samples missing age) and (E-F) n=57 samples between the ages of 10-50. Diabetes Page 38 of 51

A B 0.005 0.03 * 0.004 0.02 0.003

0.002 0.01 p1 6 p1 4 0.001 0.00 0.000

rs564398 CC CT TT rs564398 CC CT TT

C D 0.05 * 0.04 0.04 0.03 0.03 0.02 p1 5 0.02 MTAP

0.01 0.01

0.00 0.00 rs10757283 CC CT TT rs10757283 CC CT TT

All panels: rs2383208 GG+GA rs2383208 AA

Supplemental Figure 7. Haplotype analysis showing rs2383208 comparisons. Related to the analysis in Figure 5 for rs10811661. (A-B): For p16, in homozygous-protective rs564398-CC samples, homozy- gous-risk rs2383208 increased expression. The same comparison for p14 did not meet FDR<10% (q-value 17%); for ANRIL, FDR>20% (not shown). (C-D): rs10757283 and rs2383208 may collaboratively regulate p15 expression; for MTAP for the same comparison, FDR>20%. * indicates FDR<10%, our pre-determined risk tolerance for future experiments exploring haplotype hypotheses. mRNA abundance is expressed as delta-Ct, normalized to actin. n=95 for all panels. All other inter-SNP comparisons, both shown and not shown, resulted FDR >10% or had insucient data points to evaluate (dened as n=2 or fewer). FDR, False Discovery Rate. Page 39 of 51 Diabetes

A B 15 p<0.01 15 p=ns R2=0.13

10 10

5 5

Insulin secretion inde x 0 10 20 30 40 50 60 Insulin secretion inde x 0 0 20 40 60 80 Donor BMI Donor age

C p=ns D p=ns 10 10 8 6 5 4 2 0 Insulin secretion inde x 0 Insulin secretio n index F M I II III IV V VI Islet isolation center

Supplemental Figure 8. Insulin secretion index correlated with donor BMI but was not related to age, sex or isolation center. (A) Insulin secretion index positively correlated with donor BMI. Vertical dotted lines represent demarkations between normal weight (18-25), overweight (25-30) and obese (>30). (B-D) Insulin secretion index was not related to donor age (B), donor sex (C), or isolation center (D). Numerals I-VI refer to the six di erent islet isolation centers where the IIDP islet samples originated; we do not have insulin secretion data from any non-IIDP samples used in this study. (A-B): n=61; line, p values and R-squared values were calculated by linear regression. (C): n=31 (F), n=30 (M); mean +/- SD, p value by Student’s t-test. (D): n=61 samples from all isolation centers combined; mean +/- SD, ANOVA with Bonfer- roni correction for multiple comparisons. Diabetes Page 40 of 51

5 mM glucose 15 mM glucose

F F

F

F F

Insulin BrdU Dapi

Supplemental Figure 9. Images of dispersed human islet cells cultured for BrdU analysis. Human islets were rested overnight, dispersed using trypsin, and cultured on coverslips for 96 hours in 5mM or 15mM glucose with BrdU included for the entire 96 hours. Coverslips were xed in paraformaldehyde, immunostained for insulin and BrdU, mounted in Dapi-containing media, and imaged using uorescent microscopy. Images were blinded and manually counted for total insulin(+) cells and % of insulin(+) BrdU(+) cells to generate the data shown in Figure 7. F, BrdU-staining broblast. Arrows denote BrdU(+) Insulin(+) cells. Page 41 of 51 Diabetes

p<0.05 R2=0.10 p=ns p=ns KI6 7 PCN A CCND 2 0.00 0.05 0.10 0.15 0.000 0.005 0.010 0.015 0.020 0.00.000 0.0020.5 0.004 0.006 1.0 1.5 2.0 0.0 0.5 1.0 1.5 2.0 0.0 0.5 1.0 1.5 2.0 BrdU 5mM BrdU 5mM BrdU 5mM

Supplemental Figure 10. Cell cycle gene expression (on whole islets under basal (5mM) glucose culture conditions, same samples as all previous gene expression data), shown in relation to dispersed islet cell BrdU incorporation in basal (5mM) glucose conditions. PCNA was marginally correlated with BrdU%; KI67 and CCND2 were not. mRNA abundance is expressed as delta-Ct, normalized to actin. BrdU 5mM: % of insulin(+) cells that were also BrdU(+) cells, in 5mM glucose cultures. Statistical analysis was calculated by linear regression; n=46 all panels. Diabetes Page 42 of 51

A. resolution 25 kb

B. resolution 10 kb

Supplemental Figure 11. Topologically Associated Domain analysis of the human CDKN2A/B locus by Hi-C of lymphoblastoid cell line GM12878. Assmebly: hg19. (A) 25 kb resolution, and (B) 10 kb reso- lution. Data are from Rao et al, Cell 159(7):1665-1680 (2014), and analysis is by the Yue lab website http://promoter.bx.psu.edu/hi-c/view.php and "The 3D Genome Browser: a web-based browser for visual- izing 3D genome organization and long-range chromatin interactions." http://biorxiv.org/content/ear- ly/2017/02/27/112268, Biorxiv, 2017. Page 43 of 51 Diabetes

Supplemental Table 1: Donor characteristics and data obtained for each islet preparation SNP Insulin Beta cell Sex Age BMI Ethinicity Cause of Death genotype secretion proliferation and RNA M 42 22.76 Black Anoxia x x x M 40 38.91 White Unknown x x x F 54 22.6 White Cerebrovascular/stroke x x F 51 35.6 White Anoxia x x x F 38 33.1 White Anoxia x x x M 22 32.1 Hispanic Head trauma x x F 47 34.5 Hispanic Cerebrovascular/stroke x x F 39 45.2 White Anoxia x x x M 40 35.4 White Unknown x x M 46 29.3 Unknown Cerebrovasular/stroke x x M 22 40.2 White Head trauma x x M 1.8 18.7 White anoxia x F 0.2 20.8 Hispanic Anoxia x M 48 31.2 White Cerebrovascular/stroke x x Unknown Unknown Unknown Unknown Unknown x M 57 29.8 White Head Trauma x x F 45 27.4 White Cerebrovascular/stroke x x x F 51 28.7 White Cerebrovascular/stroke x x x M 25 33.8 Hispanic Head Trauma x x x F 50 41.3 Hispanic Cerebrovascular/stroke x x F 49 36.9 White Cerebrovascular/stroke x x x F 52 35.2 White Cerebrovascular/stroke x x F 47 29.9 Black Cerebrovascular/stroke x x F 61 31 Black Cerebrovascular/stroke x x F 32 39.4 White Unknown x x x F 8 16.1 White Cerebrovascular/stroke x F 29 21 White Head trauma x x x M 7 26.6 White Anoxia x M 52 29 White Anoxia x x F 52 31.4 White Cerebrovascular/stroke x x x M 28 32.8 White Cerebrovascular/stroke x x x M 15 23 White Head trauma x x x M 63 38.6 White Anoxia x x x M 35 32 Hispanic Cerebrovascular/stroke x x x M 24 29.4 White Head trauma x x F 36 42.7 White Anoxia x x x M 18 27.9 White Cerebrovascular/stroke x x x M 55 33.4 White Cerebrovascular/stroke x x F 45 32.9 White Cerebrovascular/stroke x x x F 39 22.8 White Cerebrovascular/stroke x x F 56 21.43 White Chronic back pain/stroke x M 19 34.1 Hispanic Head trauma x x x M 25 26 White Cerebrovascular/stroke x x x M 32 27.8 Black Head trauma x x x F 49 31.6 White Cerebrovascular/stroke x x F 58 19.2 White Anoxia x x x M 21 24.8 White Head trauma x x x F 45 26.6 White Cerebrovascular/stroke x x x F 56 33.4 Black Cerebrovascular/stroke x x x M 35 28.5 Asian Head trauma x x x M 30 56.8 White Anoxia x x M 63 21.9 White Cerebrovascular/stroke x x x Diabetes Page 44 of 51

F 54 30.1 White Cerebrovascular/stroke x x x M 20 19.8 White Anoxia x x x F 52 26.8 Black Cerebrovascular/stroke x x x M 37 30.5 White Head trauma x x x F 53 31.9 White Cerebrovascular/stroke x x x F 59 28.2 Hispanic Cerebrovascular/stroke x x x F 33 34.2 Black Cerebrovascular/stroke x x x F 40 23 Hispanic Cerebrovascular/stroke x x x M 47 31 White Anoxia x x x M 1.3 22 White Head trauma/Blunt Injury x F 44 34.5 White Cerebrovascular/stroke x x x F 47 36.4 Black Anoxia x x x M 15 24.6 Hispanic Head trauma x x M 49 26.47 White Head trauma/Blunt Injury x x F 59 22 White Cerebrovascular/stroke x x x M 59 26.8 White Anoxia x x x M 68 29.7 White Cerebrovascular/stroke x x x M 60 31.3 White Head trauma x x x M 60 37.9 White Anoxia x x x F 37 25.07 Hispanic Anoxia x Unknown Unknown Unknown Unknown Unknown x Unknown 29 42.2 Unknown Unknown x F 40 27.8 White Cerebrovascular/stroke x M 28 32 Hispanic Cerebrovascular/stroke x F 27 26.9 White Cerebrovascular/stroke x M 51 30.2 White Cerebrovascular/stroke x M 52 33.7 White Cerebrovascular/stroke x F 51 23.9 White Unknown x M 30 26.5 Hispanic Unknown x M 30 26.9 White Head trauma x F 63 36.6 White Cerebrovascular/stroke x Unknown Unknown Unknown Unknown Unknown x M 45 25.49 White Head trauma x M 54 39.2 White Blunt Head Trauma MVA x M 17 32.5 White Unknown x M 51 28.1 Hispanic Unknown x Unknown 56 26.7 Unknown Unknown x F 35 37 White Cerebrovascular/stroke x M 36 51.9 White Cerebrovascular/stroke x F 40 28.4 White Unknown x F 33 22.3 White Cerebrovascular/stroke x M 35 46.1 White Unknown x M 48 29 White Cerebrovascular/stroke x Page 45 of 51 Diabetes

Supplemental Table 2: Minor Allele Frequencies rs2383208 rs10757283 rs10811661 rs564398 CDKN2A/B Locus G/G C/C C/C C/C T2D SNPs tested G/A C/T C/T C/T A/A T/T T/T T/T Minor allele G T C C Risk allele A T T T All 16.3% 46.8% 15.3% 36.8% (n=95) Black Observed minor allele 6.3% 37.5% 0.0% 6.3% (n=8) frequency in this Hispanic human islet cohort 10.7% 42.9% 10.7% 28.6% (n=13) Caucasian 18.9% 50.0% 18.2% 42.4% (n=62) Observed minor allele frequency in 1000 EUR 17.3% 43.9% 16.8% 41.4% Genomes Supplemental Table 2. For each SNP the minor allele, risk allele (in red) and observed minor allele frequency (MAF) in this human islet cohort are described. These MAFs are similar to those reported in the large populations tested in the 1000 genomes project (data are shown for EUR, since we have too few samples in non-white categories for accurate comparison). Diabetes Page 46 of 51

Supplemental Table 3: Linkage Disequilibrium

Expected LD (R2) (from EUR population) rs564398 rs2383208 rs10811661 rs10757283 rs564398 0.006 0.007 0.016 rs2383208 0.006 0.952 0.256 LDpair rs10811661 0.007 0.952 0.258 Expected (EUR population) by rs10757283 0.016 0.256 0.258

Observed LD (R2) in our samples rs564398 rs2383208 rs10811661 rs10757283 rs564398 0.040 0.046 0.001 rs2383208 0.040 0.924 0.221 rs10811661 0.046 0.924 0.204 samples (n=95) Observed in our rs10757283 0.001 0.221 0.204 Supplemental Table 3. Expected LD (R2) (top chart) were calculated using the LDpair function on the NIH-supported LDlink website. The EUR population was selected since the majority of our samples were Caucasian. The observed LD (R2) in our samples (bottom chart) were calculated using the cubeX web tool at http://www.oege.org/software/cubex/. Page 47 of 51 Diabetes

Supplemental Table 4: Multivariable linear models testing relationships between donor characteristics and gene expression p14 p15 p16 ANRIL MTAP Sex Point SE p Point SE p Point SE p Point SE p Point SE p Female v. Male 0.0869 0.1578 0.5819 0.9949 1.0280 0.3331 0.8965 0.8675 0.3014 0.0410 0.0426 0.3362 0.6714 1.0426 0.5196 Race/ethnicity Black v. White -0.2308 0.2706 0.3937 -2.7309 11.7634 0.1215 -0.6622 1.4880 0.6563 0.0145 0.0731 0.8428 -2.8995 1.7884 0.1050 Other v. White -0.0723 0.2016 0.7198 1.2057 1.3136 0.3587 0.0125 1.1084 0.9910 -0.0177 0.0545 0.7454 -0.0933 1.3322 0.4836 Age, per year 0.0186 0.0051 0.0003 -0.0641 0.0332 0.0535 0.1099 0.0280 <0.0001 0.0030 0.0014 0.0304 -0.0178 0.0337 0.5961

BMI, per kg/m2 -0.0085 0.0107 0.4281 0.0383 0.0697 0.5829 0.0212 0.0588 0.7181 -0.0003 0.0029 0.9063 -0.0108 0.0707 0.8788

Above model, integrating mRNA expression of other locus genes p14 --- 2.8584 1.1627 0.0140 4.8178 0.3939 <0.0001 0.1584 0.0424 0.0002 -0.0666 1.2224 0.9566 p15 0.0209 0.0085 0.0140 - - - -0.0892 0.0533 0.0944 0.0000 0.0039 0.9992 0.5201 0.0900 <0.0001 p16 0.1269 0.0104 <0.0001 -0.3209 0.1918 0.0944 --- 0.0033 0.0074 0.6577 0.1248 0.1980 0.5285 ANRIL 0.8076 0.2164 0.0002 0.0028 2.7079 0.9992 0.6317 1.4258 0.6577 --- 3.8463 2.7320 0.1592 MTAP -0.0005 0.0086 0.9566 0.5005 0.0866 <0.0001 0.0334 0.0529 0.5285 0.0053 0.0038 0.1592 ---

Supplemental Table 4. Exploratory multivariable model testing for impact of donor characteristics (sex, race, age, BMI) on gene expression confirmed a positive relationship between donor age and islet abundance of p14, p15 (marginal association; inverse relationship), p16 and ANRIL. This model did not uncover an impact of sex, race or BMI on expression of these genes. Integrating expression of other locus genes into the model (lower rows) confirmed a significant positive correlation between p14-ANRIL, p14-p16 and p15-MTAP, as well as p14-p15 (weaker correlation). Point estimate is the difference in gene expression conferred by comparator condition to control condition, or the incremental increase in gene expression conferred by higher amount for linear input variables such as age, BMI and gene expression. SE, standard error of the point estimate. Other: combined all samples of non-white race, since sample size was too small to analyze for those other than black or white. BMI, body mass index. Diabetes Page 48 of 51

Supplemental Table 5: List of SNPs in linkage disequilibrium with SNPs tested in this study rs2383208 chr pos (hg38) LD(r²) LD(D') variant Ref Alt AFR freq AMR freq ASN freq EUR freq 9 22133285 0.93 0.98 rs10965250 GA 0.05 0.14 0.44 0.16 9 22134069 0.93 0.98 rs10811660 GA 0.05 0.14 0.44 0.16 9 22134095 0.93 0.98 rs10811661 TC 0.05 0.14 0.44 0.16 9 22134254 0.93 0.97 rs10811662 GA 0.1 0.15 0.43 0.17 9 22132730 0.95 0.99 rs10965247 AG 0.05 0.14 0.44 0.16 9 22132879 0.96 0.98 rs10965248 TC 0.07 0.14 0.44 0.17 9 22132699 0.97 0.99 rs10965246 TC 0.07 0.14 0.44 0.17 9 22132077 1 1 rs2383208 AG 0.17 0.14 0.4 0.17 rs10811661 chr pos (hg38) LD(r²) LD(D') variant Ref Alt AFR freq AMR freq ASN freq EUR freq 9 22136490 0.82 0.99 rs1333051 AT 0.08 0.09 0.15 0.14 9 22132077 0.93 0.98 rs2383208 AG 0.17 0.14 0.4 0.17 9 22132730 0.94 0.97 rs10965247 AG 0.05 0.14 0.44 0.16 9 22132699 0.96 0.99 rs10965246 TC 0.07 0.14 0.44 0.17 9 22132879 0.97 1 rs10965248 TC 0.07 0.14 0.44 0.17 9 22134254 0.99 1 rs10811662 GA 0.1 0.15 0.43 0.17 9 22133285 1 1 rs10965250 GA 0.05 0.14 0.44 0.16 9 22134069 1 1 rs10811660 GA 0.05 0.14 0.44 0.16 9 22134095 1 1 rs10811661 TC 0.05 0.14 0.44 0.16 rs10757283 chr pos (hg38) LD(r²) LD(D') variant Ref Alt AFR freq AMR freq ASN freq EUR freq 9 22134303 0.91 0.99 rs7019437 CG 0.23 0.45 0.63 0.41 9 22134652 0.98 0.99 rs7019778 AC 0.23 0.46 0.64 0.43 9 22133646 0.99 1 rs10217762 TC 0.2 0.45 0.63 0.43 9 22133985 0.99 1 rs10757282 TC 0.23 0.45 0.64 0.43 9 22134173 1 1 rs10757283 CT 0.45 0.47 0.64 0.43 rs564398 chr pos (hg38) LD(r²) LD(D') variant Ref Alt AFR freq AMR freq ASN freq EUR freq 9 22015998 0.81 0.91 rs1101329 CT 0.01 0.21 0.1 0.41 9 22043613 0.81 0.99 rs1412830 CT 0.01 0.19 0.1 0.37 9 22021173 0.82 0.91 rs597816 TC 0.01 0.21 0.1 0.41 9 22007358 0.84 -0.99 rs3217977 CA C 0.96 0.75 0.9 0.55 9 22056360 0.84 -0.93 rs7866783 AG 0.98 0.79 0.9 0.59 9 22011643 0.85 0.98 rs573687 GA 0.01 0.19 0.1 0.38 9 21999329 0.86 0.95 rs2811713 GA 0.01 0.19 0.18 0.39 9 22009699 0.86 -1 rs2069418 GC 0.98 0.77 0.9 0.55 9 22049480 0.88 -0.94 rs200059580 A ACT 0.99 0.8 0.9 0.59 9 22015466 0.89 0.98 rs1101330 CA 0.02 0.19 0.1 0.39 9 22051671 0.89 -0.95 rs944801 GC 0.99 0.79 0.9 0.59 9 22052735 0.89 -0.95 rs6475604 TC 0.99 0.8 0.9 0.59 9 22054041 0.89 -0.95 rs7030641 CT 0.99 0.79 0.9 0.59 9 22003368 0.9 -0.99 rs1063192 GA 0.99 0.79 0.82 0.57 9 22019130 0.9 1 rs523096 AG 0.01 0.23 0.1 0.43 9 22019674 0.9 1 rs518394 GC 0.01 0.23 0.1 0.43 9 22022377 0.9 0.99 rs581876 CT 0.01 0.19 0.32 0.39 9 22026078 0.9 1 rs615552 TC 0.01 0.22 0.1 0.43 9 22045318 0.9 -0.96 rs1360589 CT 0.99 0.79 0.9 0.58 9 22040766 0.94 -0.98 rs1333037 CT 0.99 0.79 0.9 0.58 9 22036113 0.95 -0.99 rs1008878 GT 0.97 0.79 0.9 0.58 9 22036368 0.95 -0.99 rs1556515 CT 0.97 0.79 0.9 0.58 9 22031006 0.96 -1 rs7865618 GA 0.99 0.8 0.9 0.58 9 22033367 0.96 -1 rs2157719 CT 0.99 0.8 0.9 0.58 9 22028213 0.98 1 rs142048183 CAT C 0.01 0.2 0.1 0.41 9 22043927 0.98 0.99 rs1412829 AG 0.01 0.21 0.1 0.41 9 22026595 0.99 1 rs613312 GA 0.01 0.2 0.1 0.41 9 22026640 0.99 1 rs543830 AT 0.01 0.2 0.1 0.41 9 22027403 0.99 1 rs599452 GA 0.01 0.2 0.1 0.41 9 22029081 0.99 1 rs679038 GA 0.02 0.2 0.1 0.41 9 22032153 0.99 1 rs634537 TG 0.02 0.21 0.1 0.41 9 22029548 1 1 rs564398 TC 0.01 0.2 0.1 0.41 Supplemental Table 5. Multiple SNPs are in linkage disequilibrium with the T2D-associated CDKN2A/B SNPs genotyped in this study. Data in this table include all SNPs in HaploReg with LD(r2) > 0.80, from the EUR population data (chosen because the majority of our samples were Caucasian). T2D SNPs genotyped for this study are in red font. Allele frequencies in AFR, AMR, ASN, EUR are included for reference. Pos, position. Ref, reference allele. Alt, alternate allele. Page 49 of 51 Diabetes

Supplemental Table 6: Multivariable linear models testing for SNP impact on gene expression p14 p15 p16 ANRIL MTAP point SE p point SE p point SE p point SE p point SE p rs2383208 A/A v. G/A 0.1151 0.1679 0.4931 2.0529 1.7650 0.0565 0.4085 0.9206 0.6572 0.0635 0.0450 0.1585 -0.4870 1.1066 0.6599 A/A v. G/G 0.2909 0.5227 0.5779 2.3846 3.3504 0.4766 3.1359 2.8652 0.2738 0.0817 0.1401 0.5598 -3.7195 -3.4444 0.2802 number of risk alleles 0.1234 0.1467 0.4003 1.8169 0.9414 0.0536 0.7264 0.8064 0.3677 0.0573 0.0394 0.1455 -0.8634 0.9693 0.3731

rs10811661 T/T v. C/T 0.0473 0.1742 0.7862 2.1538 1.1138 0.0531 0.0192 0.9540 0.9839 0.0566 0.0467 0.2259 -0.5684 1.1453 0.6197 T/T v. C/C 0.2730 0.5242 0.6025 2.4398 3.3510 0.4666 3.0297 2.8701 0.2911 0.0806 0.1406 0.5667 -3.7490 3.4458 0.2766 number of risk alleles 0.0729 0.1513 0.6297 1.8850 0.9679 0.0515 0.4498 0.8316 0.5886 0.0519 0.0406 0.2008 -0.9444 0.9963 0.3432

rs10757283 T/T v. C/T -0.0318 0.2011 0.8742 1.9756 1.2644 0.1182 0.5558 1.1087 0.6162 -0.0274 0.0548 0.6178 2.5824 1.3227 0.0509 T/T v. C/C -0.3073 0.2179 0.1586 -1.5806 1.3704 0.2488 -0.9627 1.2017 0.4231 -0.0600 0.0594 0.3127 1.8335 1.4336 0.2009 number of risk alleles -0.1637 0.1085 0.1316 -1.0181 0.7081 0.1505 -0.5668 0.6011 0.3458 -0.0302 0.0295 0.3058 0.7796 0.7215 0.2799

rs564398 T/T v. C/T 0.0252 0.1692 0.8814 0.4790 1.0970 0.6624 0.1187 0.9339 0.8988 -0.0102 0.0459 0.8236 -0.1792 1.1199 0.8729 T/T v. C/C 0.2564 0.2379 0.2810 2.2636 1.5427 0.1423 0.8902 1.3133 0.4979 -0.0336 0.0646 0.6032 1.2535 1.5749 0.4261 number of risk alleles 0.1049 0.1133 0.3547 0.9840 0.7350 0.1806 0.3712 0.6248 0.5525 -0.0153 0.0307 0.6183 0.4443 0.7510 0.5541 Supplemental Table 6. Exploratory multivariable model testing for impact of CDKN2A/B locus T2D SNP genotype, incorporating donor characteristics (sex, race, age, BMI from Supplemental Table 2) on gene expression failed to reveal a significant impact of any genotype on any gene expression. Point estimate is the difference in gene expression relative to protective genotype (individual comparisons) or incremental gene expression for each additional risk allele at that SNP (number of risk alleles). SE, standard error of the point estimate. Risk allele is depicted in red for each SNP. Diabetes Page 50 of 51

Supplemental Table 7: Multivariable models testing determinants of insulin secretion index mean difference in 95% confidence insulin secretion index interval p-value Female v. Male -0.25 ( -1.03 - 0.53 ) 0.5318

age (per year) 0.00 ( -0.03 - 0.03 ) 0.8951

African-American v. White 0.02 ( -1.09 - 1.13 ) 0.974

included in model 2

Isolation center not BMI (per kg/m ) 0.08 ( 0.03 - 0.14 ) 0.0029

Female v. Male -0.32 ( -1.05 - 0.41 ) 0.3873

age (per year) 0.00 ( -0.03 - 0.03 ) 0.9665

African-American v. White -0.10 ( -1.14 - 0.95 ) 0.8588 Isolation center included in model BMI (per kg/m2) 0.10 ( 0.02 - 0.17 ) 0.0171 Supplemental Table 7. Multivariable model testing for impact of donor characteristics (sex, age, race, BMI) on insulin secretion index confirmed a positive relationship between donor BMI and insulin secretion, but did not uncover an impact of sex, race or age. Integrating islet isolation center into the model using a generalized estimating equation (GEE) approach to adjust for potential clustering of insulin secretion measurements within islet isolation centers (lower rows) confirmed a significant positive correlation between BMI and insulin secretion index (0.10 units of insulin secretion index per BMI unit, p=0.0171). In a multivariable linear model adjusted for sex, age, ethnicity and BMI, mean insulin secretion values from center IV (see Supplemental figure 8) were higher than the mean values from the other centers, but after Bonferroni adjustment none of the between-center mean comparisons were statistically significant (not shown). BMI, body mass index. Page 51 of 51 Diabetes

Supplemental Table 8: Multivariable models testing for impact of SNP genotype on insulin secretion index rs10757283 rs10811661 rs2383208 rs564398 mean mean mean mean difference 95% difference 95% difference 95% difference 95% in insulin confidence in insulin confidence in insulin confidence in insulin confidence secretion interval p-value secretion interval p-value secretion interval p-value secretion interval p-value Female v. Male -0.32 ( -1.16 - 0.52 ) 0.4583 -0.25 ( -1.03 - 0.53 ) 0.5316 -0.25 ( -1.03 - 0.53 ) 0.5264 -0.24 ( -1.04 - 0.55 ) 0.5503 Age (per year) 0.00 ( -0.03 - 0.03 ) 0.7976 0.00 ( -0.03 - 0.03 ) 0.8704 0.00 ( -0.03 - 0.03 ) 0.8717 0.00 ( -0.03 - 0.03 ) 0.9685 African-American v. White 0.05 ( -1.07 - 1.17 ) 0.9330 0.07 ( -1.07 - 1.21 ) 0.9031 0.04 ( -1.07 - 1.16 ) 0.9387 0.19 ( -1.04 - 1.43 ) 0.7573 BMI (per kg/m2) 0.08 ( 0.03 - 0.14 ) 0.0030 0.08 ( 0.03 - 0.14 ) 0.0041 0.08 ( 0.03 - 0.14 ) 0.0039 0.09 ( 0.03 - 0.14 ) 0.0021

rs10757283 no risk alleles v. two -0.03 ( -1.16 - 1.11 ) 0.9655 one risk allele v. two 0.17 ( -0.93 - 1.27 ) 0.7661 rs10811661 one risk allele v. two 0.20 ( -0.74 - 1.13 ) 0.6773 rs2383208 one risk allele v. two 0.18 ( -0.71 - 1.07 ) 0.6961 rs564398 no risk alleles v. two 0.02 ( -1.15 - 1.19 ) 0.9692 one risk allele v. two 0.45 ( -0.42 - 1.32 ) 0.3123 Supplemental Table 8. In multivariable linear models adjusted for sex, age, race and BMI, and for potential clustering of insulin secretion index produced by different islet isolation centers (using a GEE approach), none of the four CDKN2A/B SNP genotypes had any significant association with insulin secretion index. These models confirmed the positive relationship between donor BMI and insulin secretion. BMI, body mass index.