Implicating type 2 diabetes effector genes in relevant metabolic cellular models using promoter-focused Capture-C

We used a modified protocol from the Human Cell Design (France) direct protocol for culturing EndoC-BH1 cells. Cells were expanded in DMEM low glucose (1 g/ml) (Gibco, USA) with Glutamax, 2% BSA, 50 μmol/l 2(β)-mercaptoethanol, 10 mmol/l nicotinamide, 5.5 μmol/l transferrin, 6.7 ng/ml sodium selenite, 100 U penicillin/100 μg/ml streptomycin/0.25 μg/ml amphotericin B (1×, Gibco). Plates were coated with either Matrigel or MaxGel ECM (Sigma, USA) and 2 μg/ml fibronectin according to the manufacturer’s instructions: Cells were passaged when confluence was observed or otherwise passaged by two-thirds every 2 weeks. Cells were negative for mycoplasma and validated by expression of insulin. The other cell lines were maintained under standard conditions. See the electronic supplementary material (ESM) Methods ‘Cell culture’ for details for each cell type and for the SGBS cell differentiation.

The generation of promoter-focused Capture-C, assay for transposase-accessible chromatin (ATAC-seq) and RNA-seq libraries for EndoC-BH1 was in line with methods described previously [13]. The libraries for HepG2 and a preadipocyte cell line (SGBS_undiff) that was differentiated in vitro (SGBS_diff) were generated in line with our previous reports [14]. For full details see ESM Methods ‘ATAC-seq’. We considered the set of open chromatin regions (OCRs) as the ATAC-seq peaks identified in at least two replicates in each cell line (ESM Fig. 1). We used standard methods for generation of 3C libraries [14‐20]. See ESM Methods ‘Promoter-focused Capture-C’ and ‘Hi-C library preparation and analysis’ for details. The capture library was re-annotated under Gencode V30 (https://​www.​gencodegenes.​org/​) at both 1-fragment and 4-fragment resolution. We defined promoter OCRs as the set of OCRs overlapping the region within −1500/+500 bp of the gene transcription start site (TSS). This range may include both proximal and distal promoter elements but is generally not in the range where a chromatin contact at our resolution is able to distinguish these [15].

RNA was isolated from ~1 million cells of each cell type using TRIzol reagent (Invitrogen, USA), purified using the Directzol RNA Miniprep Kit (Zymo Research, USA) and depleted of contaminating genomic DNA using DNAse I. Purified RNA was checked for quality on a Bioanalyzer 2100 (Agilent, USA) using the Nano RNA Chip (catalogue no. 5067-1511, Agilent) and samples with RNA integrity number (RIN) >7 were used for RNA-seq library preparation. RNA samples were depleted of ribosomal RNA using the QIAseq Fastselect RNA removal kit (Qiagen, Germany). Samples were then processed for the preparation of libraries using the SMARTer Stranded Total RNA Sample Prep Kit (Takara Bio, USA) according to the manufacturer’s instructions. See ESM Methods ‘RNA-seq’ for details.

See ESM Methods ‘Partitioned LD score regression’, ‘Epigenome roadmap enrichment’, ‘GWAS data integration’ and ‘Transcription factor analysis’ for details. The resulting graphs were generated using R [21] (v4.0.2, https://​www.​r-project.​org/​) and ggplot2 (v3.3.0, Bioconductor; https://​www.​bioconductor.​org/​). Tracks were visualised using pyGenomeTracks [22] (v3.5, https://​github.​com/​deeptools/​pyGenomeTracks). Network diagrams were constructed using Cytoscape (v3.8.2, https://​cytoscape.​org/​).

EndoC-βH1 cells and reagents, including the maintenance medium (OPTIβ1), starvation medium (OPTIβ2) and coating matrix (βCoat), were purchased from Human Cell Design (France). First, culture dishes were prepared following the manufacturer’s instructions and coated with βCoat. EndoC-BH1 cells were thawed in a 37°C water-bath for 1–2 min, resuspended and plated. For resuspension, 1 ml of OPTIβ1 medium was slowly added to the cells before transferring the cell suspension to a separate tube of 8 ml of OPTIβ1 medium. Following this, the cell suspension was then centrifuged for 5 min at 500 g at room temperature. The supernatant was discarded, and the cell pellet was resuspended with 1 ml of OPTIβ1 medium by carefully detaching the cell pellet from the bottom of the tube by gently pipetting up and down. Then, 1 ml of OPTIβ1 was added to the cell suspension and gently homogenised. Next, live cells were counted using Countess 3 (Thermo Fisher Scientific, USA). Then, 1.6 × 10⁶ cells or 4.2 × 10⁶ cells were plated in 6 cm dishes or 10 cm dishes, respectively. EndoC-βH1 cells were routinely split every 7 days to a density of 70,000 cells per cm². The growth rate of EndoC-βH1 cells stabilised after the second passage of cells; therefore, all experiments were at a minimum of the third passage. The replicates were each within ±2 passages of each other.

Prior to cell seeding, four 12-well plates were coated with 500 µl of βCoat coating matrix (Human Cell Design), following the manufacturer’s instructions. These plates were then prepared in a 37°C, 5% CO₂ incubator for 3 h prior to cell seeding. Approximately 125,000 cells were uniformly added to each well and allowed to grow at 37°C, 5% CO₂ for 4 days. On the fifth day the appropriate siRNA treatments (non-targeting 50 nmol/l, fragile X syndrome-related protein 2 (FXR2; 50 nmol/l) or single-pass membrane and coiled-coil domain-containing protein 4 (SMCO4; 37.5 nmol/l) were performed. For details see ESM Methods ‘EndoC-BH1 siRNA transfection’. Then, 48 h later, the cells were transferred from the siRNA/OPTIβ1 medium to OPTIβ2 starvation medium (Human Cell Design) and placed in the incubator for 24 h. Following this, the cells were preincubated for the glucose stimulation by removing the OPTIβ2 and adding 1 ml/well of βKrebs (Human Cell Design)/BSA for 60 min. Next, the stimulation step began by removing the βKrebs/BSA and adding the corresponding treatment (0 mmol/l glucose, 20 mmol/l glucose, 0 mmol/l glucose + 45 µmol/l 3-isobutyl-1-methylxanthine [IBMX] or 20 mmol/l glucose + 45 µmol/l IBMX) following manufacturer’s instructions for 40 min. For collecting the secreted insulin, 800 µl of supernatant was collected from each well and these samples were stored at 4°C until the centrifugation step. For cell insulin content, the remaining medium was removed and 1 ml of cold lysis buffer (5 mol/l NaCl, 0.2 mol/l EGTA, 1 mol/l Tris pH 8.0, glycerol, Triton X-100 and H₂O) was added to each well for 5 min, then cells were manually sheared by pipetting up and down to complete the lysis. Next, these samples were kept at 4°C until centrifugation at 700 g for 5 min at 4°C. Finally, 300 µl from each sample was transferred into a new microtube and stored at −20°C until the Human Insulin ELISA was performed. For quantifying insulin content, we used the Mercodia Human Insulin ELISA kit (catalogue no. 10-1113-01, Mercodia, Sweden) to evaluate insulin secretion in EndoC-βH1 cells. Full details are available in ESM Methods ‘Human insulin ELISA’. This experiment was performed four times, with the technical replicates with the mean of each experiment treated as biological replicates.

Total RNA was extracted from EndoC-βH1 cells 72 h post transfection using Invitrogen TRIzol (Thermo Fisher Scientific, catalogue no. 15596018) and the Zymo Research Direct-zol RNA MiniPrep Plus kit (catalogue no. R2070), following the manufacturer’s instructions.

The quantity and purity of RNA extracted were assessed using a NanoDrop 2000 spectrophotometer (Thermo Fisher Scientific, USA).

First-strand cDNA synthesis was performed using the SuperScript VILO cDNA Synthesis Kit (Thermo Fisher Scientific, catalogue no. 11754050), following the manufacturer’s instructions. The quantity and purity of cDNA synthesised were assessed using a NanoDrop 2000 spectrophotometer.

Reverse transcriptase quantitative PCR (RT-qPCR) was performed using TaqMan Fast Advanced Master Mix (catalogue no. 4444556, Applied BioSystems, USA) and analysed with the Agilent Technologies AriaMX Real-Time System. Knockdown efficiency was assessed by RT-qPCR of the target genes using the following primers: FXR2 (Taqman FXR2, catalogue no. 4331182, assay ID Hs00191579_m1_FAM, Thermo Fisher), SMCO4 (Taqman SMCO4, catalogue no. 4351372, assay ID HS04980617_g1_FAM, Thermo Fisher). The endogenous control gene was beta-actin (Taqman ACTB, catalogue no. 4331182, assay ID HS01060665_g1_VIC, Thermo Fisher). For details see ESM Methods ‘Quantitative RT-PCR’. The endogenous control gene was used to normalise gene expression following the ΔΔC_t method [23]. The non-template control did not show amplification.

Research was conducted in accordance with the Children’s Hospital of Philadelphia’s ethics and compliance policies. The researchers had access to deidentified biospecimens and datasets.

To gain insights into key aspects of the epigenetic landscape of type 2 diabetes-relevant cell models, we analysed high-resolution promoter-focused Capture-C and ATAC-seq data in triplicate for the EndoC-BH1 and HepG2 cell lines, along with SGBS cells both at the pre-adipocyte stage and the in vitro differentiated adipocyte-like state (SGBS_undiff, SGBS_diff). We also leveraged the gene expression in each cell line using bulk RNA-seq [14].

For analysing the promoter-centric chromatin architecture, we identified Capture-C-defined contacts with gene promoters both at the level of individual restriction fragments (1frag) and by in silico binning four consecutive fragments (4frag) in order to increase the power to call distant interactions, as previously performed [14, 20, 24], yielding 53,689 1frag and 92,152 4frag promoter contacts for EndoC-BH1, respectively, and ranging from 80,449 to 229,880 1frag and 155,569 to 293,618 4frag promoter contacts for the other cell lines (Fig. 1a, ESM Table 1).

We leveraged ATAC-seq to identify OCRs, which potentially act as cREs to influence gene expression. We called OCRs as the set of reproducible peaks in at least two of three replicates. We identified 68,596 OCRs in EndoC-BH1 cells, 73,296 in HepG2, 74,296 in SGBS_undiff and 52,609 in SGBS_diff.

We then annotated OCRs to genes if they either overlapped with a promoter (−1500/+500 bp) or intersected the other end called by promoter-focused Capture-C, which we termed promoter interacting regions (PIRs). We found that 28% of OCRs contacted at least one gene promoter (Fig. 1b, ESM Table 2), which is in line with previous reports [14, 20]. We next compared the distribution of distances between the putative enhancer and promoter-connected region fragments and found that they were comparable (mean: EndoC-BH1, 133,320.9 bp; HepG2, 109,703.7 bp; SGBS_undiff, 155,686.6 bp; SGBS_diff, 207,733.6 bp) (Fig. 1c).

Next, we determined the degree of similarity across promoter–OCR connections between cell types by comparing the magnitude of overlap between the PIR-OCR-connected regions using the Jaccard index, i.e. the ratio of nucleotides located in both annotations to the union present in either set of annotations. We found that the four cell types displayed less than 30% overlap (Fig. 1d), suggesting that the promoter landscapes have distinct cell type-specific epigenetic and chromatin conformational features that contribute to transcriptional differences across these cell lines. Also consistent with previous reports, we observed bait-to-bait contacts ranging from 10% to 29% of called loops (ESM Table 3).

To further validate our OCR–gene contacts, we compared the expression of genes with a promoter contacting at least one OCR vs that of genes not in contact with any OCR. Consistent with our prior promoter-focused Capture-C datasets [16, 17, 19], genes with a promoter in contact with OCRs were generally expressed at significantly higher levels (unpaired Mann–Whitney U test, two-sided p<2.2 × 10⁻¹⁶) than those without OCR contacts across cell types (Fig. 2a). Similarly, we assessed the degree of similarity to GTEx expression profiles of equivalent tissue types and found that the genes contacted in our cell lines were expressed at higher levels in the corresponding tissue in that public domain dataset [25] (Fig. 2b). These results suggest that promoter contacts with OCRs are associated with increased gene expression.

We then compared the chromatin states identified by the epigenome roadmap with the open connected regions defined by ATAC-seq and promoter-focused Capture-C. As expected, we found at least fourfold enrichment for active and bivalent TSSs and enhancers, given that OCRs can represent either active or ‘poised’ regulatory elements (Fig. 2c). The shared TSS/promoter signature represents the number of bait-to-bait interactions present, suggesting either that some genes are co-regulated or that promoters of genes not expressed in a particular cell type can act as enhancers, as previously reported by others [26].

We subsequently examined the VISTA database for curated validated enhancers in different embryological tissues [27]. However, there were only six experimentally validated pancreatic enhancers within the database. Despite this limitation, we observed in our data that a known human enhancer, hs1977, was in contact with the ABCC8 promoter (Fig. 2d, e). ABCC8 encodes a component of an ATP-sensitive potassium channel expressed in beta cells and modulates glucose-dependent insulin secretion [28]. This enhancer region was accessible only in EndoC-BH1 cells, while expression was observed in the pancreas in the VISTA database, although several of the in situs display staining in neural tissues in addition to the pancreas. We also identified a liver enhancer, hs1752, expressed in the liver, heart and other abdominal tissues with contacts to PCYOX1L, which encodes prenylcysteine oxidase 1 like (Fig. 2f, g). Prenylcysteine oxidases are enzymes that scavenge free cysteines from a metabolic pathway involved in the degradation of prenylated proteins [29]. This cRE was also connected to the promoter of several non-coding RNAs: miR-143, miR-145 and AC131025.2. Taken together, these results support the use of chromatin features in these cell lines as a valid model to investigate cREs for type 2 diabetes-related traits.

Next, we sought to investigate the degree that cellular models are enriched for heritability associated with metabolic traits. To this end, we performed partitioned LD score regression [30] to test for enrichment of putative cREs (promoter OCRs + PIR-OCRs) (Fig. 3, ESM Table 4). We observed significant enrichment of EndoC-BH1 cREs for type 2 diabetes, fasting glucose, BMI and fetal body weight. HepG2 cREs were enriched for coronary artery disease and fasting glucose levels. SGBS_diff cREs were enriched for WHR, fasting insulin, plasma triglyceride and HDL-cholesterol levels, while SGBS_undiff cREs were enriched for WHR. cREs across all cell types were enriched for height. In addition, we included recent GWAS studies from Alzheimer’s disease and systemic lupus erythematosus as negative controls, and as expected we did not observe enrichment for these neural and immune GWAS traits in these principally metabolic cell types [30‐32]. Therefore, these results supported the utility of these cell models in investigating the cis-regulatory architecture for the type 2 diabetes-related traits.

To implicate type 2 diabetes causal variants impacting cREs, we curated the set of lead sentinel SNPs from the two most recent European and trans-ancestral GWAS reports for the disease [33, 34]. We identified proxies in high LD with the reported lead variants for each signal (r²>0.8) and intersected their genomic coordinates with OCRs and PIRs (Table 1). The subsequent 810 implicated genes from this variant-to-gene (V2G) mapping strategy, corresponding to 370 type 2 diabetes sentinels, revealed that the majority of putative causal variants did not contact the gene nearest to the sentinel, and in most cases where the nearest gene was implicated other gene promoters were also additionally contacted (Fig. 4a, ESM Table 5). Of these 810 genes implicated by this V2G mapping approach, 79.9% were specific to one cell line (130 EndoC-BH1, 311 HepG2, 123 SGBS_diff, 83 SGBS_undiff) while 163 (20.1%) were shared in at least two cell types and 53 (6.5%) across all cell types (Fig. 4b). We compared the list of genes with phenotypes identified in mice, and found significant enrichment for nervous system, haematopoietic system, growth/size body region and mortality/ageing (ESM Fig. 2, ESM Table 6).

One obvious mechanism by which non-coding variants can influence gene expression is through disruption of TF binding sites. To investigate this possibility, we predicted whether any of the V2G-implicated open proxies that contacted genes contained known TF binding motifs. This approach led to the identification of 288 proxies predicted to alter binding affinity (Fig. 4c, ESM Table 7). Most notably, we observed that several binding sites for zinc finger TF KLF and GLI families were disrupted by variants identified in EndoC-BH1. Seventeen of the TFs with motifs impacted by type 2 diabetes variants are predicted effector transcripts in at least one cell type (HES2, ZNF384, STAT6, NR2C1, ONECUT1, TCF12, SOX15, TP53, HNF1B, TCF4, TCF3, OSR1, PPARG, NFKB1, CREB3, ZBTB6, ZBTB26). We compared the motifs with publicly available chromatin immunoprecipitation (ChIP) data from ENCODE, to determine if there is evidence of TF binding at the genomic location in various cellular contexts (ESM Fig. 3, ESM Table 8). We also checked the expression of implicated TFs and observed 280 of 447 (62.6%) of the predicted TFs with disrupted motifs had transcripts per million mapped reads (TPM) >1 and 187 had TPM >10 in EndoC-BH1 cells, suggesting that a majority of predicted TFs are expressed (ESM Fig. 4).

We then examined the genes contacting proxies regardless of the status of predicted disruption of TF binding sites. While some signals were only predicted to have one proxy interacting with one gene, others yielded multiple proxies with contacts to multiple genes (Fig. 4d). While this could suggest a subset of variants acting through multiple genes, further work is necessary to functionally validate these predictions. We note that we did not observe a trend towards those multi-gene signals corresponding to more statistically significant type 2 diabetes GWAS loci.

Next, we compared our analysis with previous type 2 diabetes functional prioritisation approaches, yielding moderate agreement but highlighting a potentially higher confidence gene set for further investigation [3, 8, 35‐37] (Fig. 4e, ESM Table 9). Additionally, we compared our findings with a recent CRISPR interference (CRISPRi) screen identifying potential type 2 diabetes effectors, noting eight overlapping genes, seven associated with decreased insulin content and one with increased content (decreased: CHD4, PRPF18, GMEB1, CREB3, PITPNM2, SIN3A and ATP6 V1C1; increased: FADS1). However, this overlap did not reach strict statistical significance (Fisher’s test, one-sided, p=0.09). Moreover, we compared our list with one generated using chromatin conformation in pancreatic cells, finding 100 genes annotated as type 2 diabetes effectors in beta cell Hi-C also present in our dataset (ESM Table 10) [9].

Subsequently, we compared our results with eQTL associations of nearby genes [5, 25]. Among 221 eQTL-linked genes associated with type 2 diabetes risk loci in pancreatic/islet tissues, only 12 had open proxies in chromatin contact with their promoters in EndoC-BH1 cells: SMCO4, DOC2A, OPRL1, TUFM, YBEY, MTMR11, UBE2D3, PLEKHA1, GPSM1, RNF6, ABCD9 and AGFG2. Six genes were implicated by both liver eQTLs and our promoter-focused Capture-C in HepG2 cells, while 25 genes were linked to adipose tissue eQTLs and our Capture-C in at least one adipose model (SGBS_undiff, SGBS_diff) (ESM Table 11). Colocalisation analyses between GTEx v7 and DIAMANTE type 2 diabetes GWAS signals [33] revealed 22 genes colocalised (posterior inclusion probability (PIP) >0.85) in the pancreas, with only DOC2A also implicated by our Capture-C approach. We detected 11 genes with eQTL colocalisation in the liver, of which MAN2C1, AP3S2 and CEP68 were also implicated by promoter contacts in HepG2. Additionally, 54 genes were identified in subcutaneous or visceral adipose tissue, with AP3S2, CALR and NDUFAF6 also implicated by our V2G approach in SGBS_diff cells, while PABPC4, PLEKHA1, AP3S2 and DCAF16 were implicated in the SGBS_undiff dataset (ESM Table 12). Although there is limited overlap, the intersection of these methods can better prioritise genes for functional validation.

In our analysis pipeline, two genes, SMCO4 and FXR2, were among those identified as potential type 2 diabetes candidate genes (Fig. 5a, b) in both EndoC-BH1 and prior primary beta cell chromatin maps [9]. Additionally, an eQTL associated with SMCO4 colocalises with a type 2 diabetes signal marked by rs57235767 in islets [5]. To investigate the impact of these two genes on insulin secretion in EndoC-BH1 cells (Fig. 5c), we performed targeted knockdown experiments of SMCO4 and FXR2. Confirmation of SMCO4 and FXR2 knockdown was achieved using RT-qPCR (Fig. 5d). Our analyses revealed a significant increase in insulin secretion in SMCO4 knockdown cells relative to non-targeting siRNA controls (~18.0% increase, paired two-tailed t test p=0.0243) (Fig. 5e). Notably, there was no significant difference when IBMX was added to the stimulation medium (ESM Fig. 5). IBMX raises cellular levels of cAMP to stimulate secretion of insulin. These results suggest that that SMCO4, but not FXR2, inhibits glucose-stimulated insulin secretion in EndoC-BH1 cells, while having no significant impact on IBMX-stimulated secretion.