Patterns of differential gene expression in a cellular model of human islet development, and relationship to type 2 diabetes predisposition

Human induced pluripotent stem cell (iPSC) lines from three independent individuals without diabetes were obtained from the StemBANCC consortium (www.​stembancc.​org) (see ESM Methods). The generation of lines SB Ad2 and SB Ad3 has previously been described [10]. A third line, SB Neo1, was generated from commercial fibroblasts obtained from a neonatal donor of European descent with no reported diagnosis of diabetes (CC-2509, tissue acquisition number 15819; Lonza, Walkersville, MD, USA). Characterisation of all three lines has been reported elsewhere [10, 11]. All lines were free of mycoplasma.

All tissue samples for reprogramming were collected with full informed consent. Ethical approval for the StemBANCC study (UK) was received from the National Research Ethics Service South Central Hampshire A research ethics committee (REC 13/SC/0179).

The iPSC lines were cultured in mTeSR1 medium (StemCell Technologies, Vancouver, BC, Canada) at 37°C under 5% CO₂, and passaged as single cells every 3–4 days or when confluent. In vitro differentiation involved the timely addition of recombinant growth factors and small molecules to sequentially generate cells representing key developmental stages of the endocrine pancreas: definitive endoderm, primitive gut tube, posterior foregut, pancreatic endoderm, endocrine progenitors, endocrine-like cells and beta-like cells. The differentiation protocol was carried out as described by Rezania and colleagues [9] with some modifications (ESM Tables 1, 2). All three iPSC lines were differentiated once, in parallel, using the same culture and differentiation media (ESM Methods).

The efficiency of in vitro differentiation was evaluated by measuring the expression of stage-specific markers indicative of the development of the endocrine pancreas. For each specific stage, these were: definitive endoderm (SRY-box 17 [SOX17] and octamer-binding transcription factor 4 [OCT4, also known as POU5F1]); pancreatic endoderm (NK6 homeobox 1 [NKX6-1] and pancreas/duodenum homeobox protein 1 [PDX1]); and endocrine-like cells (NKX6-1, insulin [INS] and glucagon [GCG]) (ESM Fig. 1). Methods for flow cytometry were as previously described [10], and details of antibodies are listed in ESM Table 3.

Cells were harvested and RNA extracted using TRIzol Reagent (ThermoFisher Scientific, Paisley, UK) as per the manufacturer’s guidelines. Library preparation and sequencing was performed at the Oxford Genomics Centre (Wellcome Centre for Human Genetics, Oxford, UK) as previously described [10]. RNA sequencing libraries were sequenced to a mean read depth of 148 (±12) million reads per sample. Reads were mapped to human genome build hg19, with GENCODE v19 (https://​www.​gencodegenes.​org/​releases/​19.​html) as the transcriptome reference, using STAR v.2.5 [12], followed by gene-level quantification with featureCounts from the Subread package v.1.5 (http://​subread.​sourceforge.​net/) [13] (ESM Methods).

Principal component analysis was used to cluster samples with those from previously published studies [10, 14]. Correlation of gene expression patterns across all stages was calculated using the weighted gene co-expression network analysis (WGCNA) package (v.1.51) in R (v.3.3.2) (ESM Methods) [15, 16].

Analysis was performed on 15,221 autosomal protein-coding and long intergenic non-coding RNA (lincRNA) genes present in Ensembl Genes v88 (http://​mar2017.​archive.​ensembl.​org/​index.​html) with more than one count per million in all donors of at least one differentiation stage (ESM Table 4). Genes were normalised using the voom function within the limma package (v.3.32.5) in R [17]. The eBayes function in limma was used for differential expression analysis, comparing all the differentiation stages with iPSC as the baseline, and adjusting for donor effects. We adjusted p values for multiple testing (q values) using the Benjamini–Hochberg method [18].

To define stage-specific marker genes, differentially expressed genes (q < 0.01) with an absolute log₂ fold change (FC) > 1 were assigned to the stage in which they were most upregulated compared with the baseline iPSC profile. When the log₂FC was negative for all contrasted stages, the gene was assigned to iPSCs (ESM Table 5). For comparison with the previously reported protocol [10], published data were reprocessed in an analogous manner for the stages shared between the protocols (ESM Methods; ESM Tables 6, 7).

Differentially expressed genes in each stage were tested for enrichment in gene ontology terms for biological processes using the GOstats package (v. 2.40.0) in R [19]. All genes tested for differential expression were used as background. Significant gene ontology terms (q < 0.05) were retained (ESM Table 8).

For transcription factor enrichment, upstream regulators for the differentially expressed genes were predicted using the iRegulon (v. 1.3) Cytoscape plugin (ESM Methods) [20]. Motifs and chromatin immunoprecipitation (ChIP) sequencing tracks were ranked based on the normalised enrichment score (NES), with only those with an NES > 3 (corresponding to a false discovery rate (FDR) of 3–9%) being considered. Enriched motifs were then matched to transcription factors known to bind them (ESM Table 9).

Enrichment analysis was implemented in two ways: as a hypergeometric test in R (using all genes tested for differential expression as background) or using the gene-scoring function in MAGENTA [21] followed by a gene set enrichment analysis (GSEA) [22, 23] (ESM Methods).

For the hypergeometric test, we analysed the differentially expressed genes from each differentiation stage for enrichment in genes mapping to type 2 diabetes or fasting glucose GWAS signals, which were defined as protein-coding and lincRNA genes located within specified distance bins (0, 50, 100, 200 or 500 kb) surrounding the credible intervals for trait-associated loci. Credible intervals were defined by the boundaries of the 99% credible sets of variants [24] from DIAGRAM (96 loci) [25] and ENGAGE (16 loci) [26] consortium data, respectively (ESM Table 10). A subset of 15 loci was considered to influence type 2 diabetes via beta cell dysfunction; these loci included ones causing hyperglycaemia, reduced insulin processing and secretion, and reduced fasting proinsulin levels [27, 28] (ESM Table 11, ESM Methods).

For the analysis with MAGENTA and GSEA, we mapped SNPs from the type 2 diabetes GWAS meta-analysis from DIAGRAM (96 loci) [25], and the ranked list of p values for each gene was tested in GSEA (ESM Methods).

To determine whether the differentiated cells followed normal islet development, we profiled gene expression patterns across iPSC and seven subsequent developmental stages in lines from three independent donors (SB Ad2, SB Ad3 and SB Neo1) differentiated in parallel. Each iPSC line successfully generated cells recapitulating key developmental stages of the endocrine pancreas as confirmed by the expression of known marker genes from developing and adult beta cells (ESM Fig. 2) [10].

Principal component analysis of the transcriptome showed that the beta-like cells generated in the current study clustered more closely with in vivo-matured islet-like cells [14] than cells from earlier differentiations [10] (Fig. 1, ESM Fig. 3). Differential expression analysis comparing transcriptomic profiles obtained from differentiations under current and previous protocols (see Methods) showed increasing divergence with differentiation stage (from 17 genes showing differential expression in iPSCs to 2095 at the endocrine-like cell stage) (ESM Table 7). Gene ontology analysis indicated that genes displaying increased expression at the endocrine-like cell stage (in comparisons of the current vs previous protocols) were enriched for terms including ‘regulation of insulin secretion’ (q = 2.3 × 10⁻⁴) and ‘hormone transport’ (q = 2.0 × 10⁻⁵).

Overall, cells generated in this study, compared with those previously reported [10], are more aligned to cells that have been further matured in vivo [14] (the current benchmark for most functionally mature endocrine pancreas-like cells). This reveals how advances in differentiation protocols are reflected in the transcriptome, particularly in the later stages of differentiation where there is a clear increase in the expression of genes essential for beta cell function and identity. This is the case for MAFA, which was completely absent in our previous differentiation protocol, and INS, whose high expression indicates the correct differentiation towards the last stage of beta cell development.

To characterise the transcriptomic landscape of each developmental stage in the in vitro-differentiated cells produced in this study, we assigned significantly differentially expressed genes to the stage at which they were most upregulated: if expression peaked in iPSCs, the gene was assigned to that stage (see Methods). We detected 9409 significantly differentially expressed genes (q < 0.01, absolute log₂FC > 1) across all stages, ranging in number from 623 in the primitive gut tube stage to 2773 in iPSCs (ESM Table 5). Known developmental marker genes, such as NEUROG3 in endocrine progenitors and INS in beta-like cells, were correctly assigned to their canonical stages. Gene ontology analysis of the sets of differentially expressed genes (ESM Table 8) showed enrichment in biological terms such as ‘hormone transport’ in endocrine-like cells (q = 0.047) and ‘regulation of insulin secretion’ in beta-like cells (q = 2.0 × 10⁻⁴).

The expression patterns of monogenic diabetes genes can point towards stages at which disruption of islet development has long-term consequences for glucose homeostasis. Of 28 genes implicated in monogenic or syndromic diabetes [1], 24 were differentially expressed in at least one stage of the in vitro-differentiated model. Nine mapped to the latest beta-like cell stage, but the other 15 showed significant upregulation earlier in differentiation (ESM Table 12). GATA6, for example, was differentially expressed at the definitive endoderm stage (log₂FC = 9.5, q = 7.6 × 10⁻¹¹), whereas GATA4 was differentially expressed in posterior foregut cells (log₂FC = 8.2, q = 1.9 × 10⁻¹¹); the later expression of GATA4 could contribute to the less severe phenotype of individuals carrying GATA4 vs GATA6 mutations [29, 30].

The differentiation model used in this study also sheds light on the developmental role of monogenic diabetes genes with lesser described roles. LMNA, for example, encodes a nuclear membrane protein involved in chromatin structure and nuclear stability; it has been implicated in the function and development of many tissues [31]. The diabetes in carriers of the LMNA mutation is mostly driven by altered adipose tissue deposition and insulin resistance [32]. However, the profile of LMNA expression during in vitro islet differentiation (peaking in pancreatic endoderm; log₂FC = 1.1, q = 3.1 × 10⁻³) may indicate an additional impact on islet development [33].

The developmental competence of differentiating cells is in part driven by a subset of transcription factors that initiate and regulate changes in response to external stimuli, as highlighted by the many monogenic diabetes genes that are also transcription factors. To identify potential upstream transcriptional regulators active at each stage of islet development, we performed a WGCNA and determined the enrichment of transcription factor binding motifs and ChIP sequencing signals near differentially expressed genes using iRegulon (see Methods; ESM Table 9). This analysis confirmed the impact of well-established developmental transcriptional regulators such as the monogenic diabetes gene HNF1B, which showed iRegulon enrichment of its targets at the primitive gut tube stage (NES 3.0–5.7 [see Methods]). Some of these HNF1B targets also have known effects on pancreas development (SMAD7 [34], ID2 [35]), on mature islet function and on the development of other tissues that also arise from the gut tube (GGCX) [36].

Analysis of the sets of stage-specific differentially expressed genes also highlighted the targets of transcription factors with less-well studied roles in human islet development. For example, expression of the transcriptional repressor REST peaks in the intermediate steps of in vitro differentiation and declines at the endocrine-like cell and beta-like cell stages, with reciprocal expression patterns seen among its predicted targets. These targets include genes encoding neurexins (NRXN1, NRXN2) and subunits of the glutamate receptor channels (GRIA1, GRIA2, GRID1, GRIK2) implicated in insulin exocytosis [37, 38]. Correlation of gene expression with WGCNA assigns REST to the same cluster as TCF7L2 and other genes from the Wnt signalling pathway, such as TCF7, TCF3 and TCF12 [39]. This pathway is important for islet development and is targeted in many in vitro differentiation protocols [8, 9]. These data therefore indicate that REST is likely to be an important transcriptional regulator of human islet development, both in intermediate (pancreatic endoderm, endocrine progenitor) and later (endocrine-like cell, beta-like cell) [40] stages of differentiation, as has also been recently suggested by studies in mice and humans [41, 42].

TCF7L2 maps to the type 2 diabetes-associated locus with the largest common effect on disease risk [1]. Analysis of TCF7L2 targets (as assessed by ChIP sequencing with iRegulon) shows marked enrichment at the posterior foregut stage (NES = 3.4) that mirrors that of TCF7L2 expression (log₂FC = 1.2; q = 8.5 × 10⁻¹⁰). The expression of several other Wnt family members also peaks at the posterior foregut stage; these include the coactivator CREBBP, the binding sites of which are significantly enriched in type 2 diabetes-associated loci [43], and HHEX, which maps to a prominent type 2 diabetes-risk locus and is implicated in foregut development [44]. In the developing embryo, cells of the posterior foregut can differentiate into liver as well as endocrine pancreas [45]. Alleles associated with risk of type 2 diabetes within the TCF7L2 and HHEX loci may influence early expression of these genes, which could affect development in multiple metabolic tissues. This view is supported by cellular and murine studies indicating that TCF7L2 regulates beta cell development and function [46], including via indirect effects in supporting tissues [47], as well as affecting hepatic function [48]. Similarly, Hhex is essential for the differentiation of the posterior foregut into the liver in mice [44], yet is also thought to regulate delta cell identity and function in islets [49].

Thus, several key functional candidates mapping within type 2 diabetes GWAS signals, in addition to those which overlap known monogenic diabetes genes, appear to be active during this early critical window of pancreatic development. Studying these and other diabetes-relevant genes in stem cell-derived models can help to decipher the role of multiorgan developmental effects on pathogenesis of diabetes. By integrating the differential expression data with genomic annotations on transcription factor binding and clustering of longitudinal expression, we identified novel potential regulators orchestrating gene expression patterns within the different developmental stages. Such transcriptomic analysis can also illuminate the mechanisms of action for monogenic diabetes genes and inform the search for novel MODY genes that influence the same pathways.

Most of the more than 100 type 2 diabetes susceptibility loci identified to date [1] map to non-coding regions of the genome and are likely to exert their effects through altered regulation of nearby genes. We examined the transcriptomic data for evidence of developmental stage-specific enrichment of genes near these loci.

We first concentrated on genes whose coding sequence was at least partly contained within 99% credible intervals from type 2 diabetes GWAS fine-mapping efforts on the basis that these represented a set of genes likely to be substantially enriched for type 2 diabetes effector transcripts (see Methods). Of the 117 genes so defined, most (86; 73%) showed differential expression that peaked before the final beta-like cell stage (ESM Table 13); the stages of maximal differential expression were widely distributed. GSEA, which considers the strength of association at type 2 diabetes GWAS signals (see Methods), demonstrated enrichment of the type 2 diabetes GWAS loci with largest effect for differentially expressed genes at the posterior foregut stage (q = 0.002, Fig. 2a). This enrichment remained significant (q = 0.001) if the GWAS genes also implicated in monogenic diabetes (ESM Table 12) were excluded. Using a complementary GSEA approach that ranked the strength of differential expression of each gene (in q value) per stage, we compared the most differentially expressed genes at each stage for enrichment among type 2 diabetes GWAS loci; this analysis highlighted the beta-like cell stage (q = 0.033, Fig. 2a). This enrichment was no longer significant (q = 0.151) after monogenic diabetes genes had been excluded.

As an additional analytical approach, we performed a hypergeometric test for enrichment in the same set of 117 type 2 diabetes credible interval genes (see Methods). As opposed to the GSEA method above, this analysis does not consider the strength of differential expression (or of association with type 2 diabetes) above the significance threshold. This test again demonstrated that genes showing differential expression at the beta-like cell stage were enriched (compared with background) for location within type 2 diabetes credible intervals (permuted p value =0.049; Fig. 2b). Excluding the monogenic diabetes genes, and those that fell in the same credible interval, from the differentially expressed genes at each stage removed the significance of the beta-like cells (permuted p value =0.302). We repeated the enrichment test using a subset of 15 type 2 diabetes GWAS loci for which the evidence from physiological studies points most emphatically to risk of type 2 diabetes mediated via reduced insulin secretion (ESM Table 11) [27, 28]. In this analysis, enrichment for genes differentially expressed at the beta-like cell stage became more significant (permuted p value =0.007; Fig. 2b). This enrichment was reduced (but not eliminated; permuted p value =0.03) after excluding the monogenic diabetes genes and those within the same credible interval. Using the same approach of sampling from the hypergeometric distribution, we also detected enrichment for genes mapping to credible intervals for 16 loci significantly associated with fasting glucose (permuted p value =0.0002; Fig. 2b). Earlier stages of differentiation did not show significant enrichment for genes within type 2 diabetes or fasting glucose credible intervals. Nevertheless, the assignment of differentially expressed genes to a specific stage may lead to a wide distribution of signal that dilutes the power to detect significant enrichment at stages before the beta-like cell stage.

Type 2 diabetes-associated signals falling in non-coding regions have a presumed regulatory function: some may map to tissue-specific enhancers acting some distance away from their effector transcripts [50]. However, consistent with observations that most regulatory GWAS effects operate at relatively short distances [3], we found attenuation of these enrichment signals as we extended the analyses to include genes mapping at increasing distance from the credible intervals (see Methods), both for genes in all type 2 diabetes credible intervals and for the subset implicated in beta cell function (ESM Fig. 4).

The notable overlap between monogenic diabetes genes and those mapping within type 2 diabetes-associated loci supports the hypothesis that some component of type 2 diabetes susceptibility arises through impairment of islet development [1], concretely in the posterior foregut stage. The final stage in the islet development model (featuring cells expressing genes encoding the machinery to support glucose-stimulated insulin secretion) is also enriched for genes mapping to GWAS signals for both type 2 diabetes and fasting glucose. These data are consistent with the concept that type 2 diabetes-associated loci act both on the adult islet and during earlier developmental stages.

In summary, this study demonstrates how characterisation of gene expression during human islet differentiation can identify potential novel transcriptional regulators of the differentiation process, and provide insights into developmental aspects underlying inherited predisposition to type 2 diabetes. Further refinement of in vitro models of endocrine pancreas development will allow more detailed interrogation of the genes and pathways influencing islet development and function in humans. Mechanistic analyses of the contribution of candidate regulators of islet development to long-term islet function is enhanced by recent advances in clustered regularly interspaced short palindromic repeats- (CRISPR-) based approaches that allow their experimental manipulation in in vitro systems [51]. Stem cell-derived islets may also serve as a cost-effective platform for drug screening in research into treatment of diabetes, and could provide material for transplant into individuals with diabetes [8, 9].