Systematic functional characterization of non-coding regulatory SNPs associated with central obesity

Summary

Central obesity is associated with higher risk of developing a wide range of diseases independent of overall obesity. Genome-wide association studies (GWASs) have identified more than 300 susceptibility loci associated with central obesity. However, the functional understanding of these loci is limited by the fact that most loci are in non-coding regions. To address this issue, our study first prioritized 2,034 single-nucleotide polymorphisms (SNPs) based on fine-mapping and epigenomic annotation analysis. Subsequently, we employed self-transcribing active regulatory region sequencing (STARR-seq) to systematically evaluate the enhancer activity of these prioritized SNPs. The resulting data analysis identified 141 SNPs with allelic enhancer activity. Further analysis of allelic transcription factor (TF) binding prioritized 20 key TFs mediating the central-obesity-relevant genetic regulatory network. Finally, as an example, we illustrate the molecular mechanisms of how rs8079062 acts as an allele-specific enhancer to regulate the expression of its targeted RNF157. We also evaluated the role of RNF157 in the adipogenic differentiation process. In conclusion, our results provide an important resource for understanding the genetic regulatory mechanisms underlying central obesity.

Graphical abstract

We used STARR-seq to evaluate 2,034 prioritized SNPs associated with central obesity and identified 141 SNPs with allelic enhancer activity. Further analysis prioritized 20 key transcription factors mediating the central obesity genetic regulatory network. We also illustrate how rs8079062 affects RNF157 expression and evaluate the role of RNF157 in adipogenesis.

Introduction

Obesity is a medical condition where excessive body fat accumulation results in adverse health effects. While overall obesity confers a threat to individual health, depot-specific accumulation of fat is also of great importance in determining this threat. Importantly, central obesity, which refers to a state of excessive accumulation of visceral fat primarily in the abdominal area,¹ has been recognized as an independent risk factor for many diseases (e.g., cardiometabolic diseases), regardless of overall obesity.^2,3 It was also reported to be significantly and positively associated with elevated all-cause mortality risk,⁴ emphasizing its important role in determining health outcomes. Central obesity is commonly measured by waist-to-hip ratio adjusted for BMI (WHRadjBMI), which is a highly heritable trait with estimated heritability of 30%–60%, underscoring the significant genetic component in its determination.⁵ Therefore, elucidating the genetic mechanisms of central obesity holds the potential to unravel novel avenues for prevention and therapeutic intervention.

Genome-wide association studies (GWASs) have identified more than 300 susceptibility loci associated with central obesity.^5,6 However, due to linkage disequilibrium (LD), these loci correspond to thousands of genetic variants. Furthermore, most of the trait-associated variants are in non-coding regions with unknown function, which complicates the functional interpretation. Accumulating evidence suggests that these non-coding variants are enriched in enhancers in relevant cell types,^7,8 which implies their regulatory roles in controlling target gene expression. Enhancers often act by looping with their target promoter. Integration of epigenomic annotation, expression quantitative trait locus (eQTL) analysis, and chromatin interaction has been widely used to decipher the regulatory mechanisms of non-coding variants.^9,10 For example, a BMI-associated signal that resides within an intronic region of FTO was found to regulate the expression of IRX3 and IRX5 by modulating enhancer activity.¹¹ Moreover, recently developed massively parallel reporter assays (MPRAs) or self-transcribing active regulatory region sequencing (STARR-seq) have increased the throughput at which regulatory sequences can be tested for functional potential.^12,13 For example, Joslin et al.¹⁴ have utilized MPRA and promoter capture high-throughput chromosome conformation capture (Hi-C) data to systematically prioritize BMI-related genes from GWASs. Hansen et al.¹⁵ also utilized MPRA to characterize allele-sensitive enhancers for female central obesity GWAS data. While there may be female-specific genetic risk factors of central obesity, genetic determinants of central obesity are probably shared between the sexes. However, systematic function dissection of non-coding regulatory variants associated with overall central obesity have not been performed.

In this study, we used STARR-seq to systematically assess the activity of the non-coding regulatory SNPs associated with central obesity. An overview of the current study is shown in Figure 1. First, based on fine-mapping and epigenomic annotation analysis, we selected 2,034 prioritized single-nucleotide polymorphisms (SNPs) for subsequent STARR-seq experiments. Data analysis identified 141 SNPs with allelic enhancer activity. For these SNPs, we identified their target genes through Hi-C, eQTL, and colocalization analyses. For the target genes, we analyzed the correlations between their gene expression levels and central-obesity-related traits. We also explored their potential to serve as therapeutic targets for central obesity by performing Mendelian randomization (MR) analysis with protein quantitative trait locus (pQTL) data from the Iceland cohort (sample size n = 35,559). Transcription factor (TF) enrichment analysis of these SNPs with allelic enhancer activity further identified key TFs mediating central-obesity-relevant genetic regulatory networks. Finally, using a series of functional experiments, we validated the molecular mechanism of rs8079062 in regulating the expression of RNF157. Collectively, these results provide important insights for understanding central obesity risk mechanisms.

Figure 1.

The pipeline of the current study

After fine-mapping, we investigated 9,127 non-coding SNPs. For 2,034 potential regulatory SNPs located in ATAC/H3K27ac/H3K4me1 peaks, we investigated the allele-specific activity of the SNPs using STARR-seq. After data analysis, 141 SNPs with allelic enhancer activity were identified. Target genes for these SNPs were identified using Hi-C, eQTL, and colocalization analyses. For the target genes, we analyzed their correlation with central-obesity-related traits. We also explored their potential to be served as therapeutic targets for central obesity by performing Mendelian randomization analysis. Transcription factor (TF) enrichment analysis of these SNPs with allelic enhancer activity further prioritized key TFs mediating central-obesity-relevant genetic regulatory networks. Finally, we validated the molecular mechanism of one SNP in regulating the expression of its target, RNF157.

Material and methods

GWAS summary data

The GWAS summary data for central obesity we used was the largest European GWAS summary dataset (n up to 694,649) until May of 2021.⁶ We obtained the dataset from the website https://zenodo.org/record/1251813#.XVYC3OgzZPZ. This dataset was derived from meta-analyses of the UK Biobank dataset (n = 484,563) and the publicly available GWAS data generated by the Genetic Investigation of ANthropometric Traits (GIANT) consortium. Only SNPs with minor allele frequency (MAF) > 0.01% and imputation quality (info) > 0.3 were retained in our analyses.

GWAS locus definition

We first selected lead SNPs attaining genome-wide significant evidence of association (p < 5 × 10⁻⁸) in the summary data. Loci were defined as 500 kb upstream and downstream of the lead SNPs. If two loci overlapped, they were merged as a single locus.

Fine-mapping

For each locus, distinct association signals were then selected by using the conditional and joint multiple-SNP (COJO) analysis implemented in the Genome-wide Complex Traits Analysis (GCTA) software¹⁶ with the parameter --cojo-slct --cojo-p 5e−8. Briefly, the most significant SNP within the locus was initially chosen to condition the effects of other SNPs. Subsequently, the SNPs with minimum conditional p value lower than 5 × 10⁻⁸ were added to the selected sets. All selected SNPs were then jointly fit in a model and the SNP with the largest p value that was greater than the cutoff was dropped. This process was repeated until no SNPs could be added or removed from the model. In reference to Morris et al.,¹⁷ the LD information used for conditional analysis was obtained from a reference population including genotype data from 50,000 randomly selected White British UK Biobank participants (UK Biobank data application no. 46387). All UK Biobank participants provided informed consent.¹⁸ Signals in the major histocompatibility complex (MHC) region were excluded from fine-mapping analyses. For each distinct signal, we first defined a genomic region of 500 kb on either side of the variant. FINEMAP¹⁹ was then used to identify the most likely number of causal SNPs per signal in a Bayesian framework. FINEMAP¹⁹ was run with default parameters except that the number of maximum causal configurations was set to 10. Each SNP was assigned a log10 Bayes factor (log10BF) as a measure of its posterior probability for being causal, and the cutoff was set as log10BF > 2 in this study. SNPs achieving conditional independence were also included in the final fine-mapped SNP set. The fine-mapped SNPs were annotated with ANNOVAR²⁰ to get their genomic region information.

Chromatin state enrichment analysis

Enrichment analysis was performed for 15 chromatin states of adipose (download link: https://egg2.wustl.edu/roadmap/data/byFileType/chromhmmSegmentations/ChmmModels/coreMarks/jointModel/final/E063_15_coreMarks_dense.bed.gz) and adipocyte (download link: https://egg2.wustl.edu/roadmap/data/byFileType/chromhmmSegmentations/ChmmModels/coreMarks/jointModel/final/E023_15_coreMarks_dense.bed.gz) between fine-mapped non-coding SNPs and random background autosomal SNPs. The random SNPs were generated using SNPsnap.²¹ SNPs were matched for MAF, number of SNPs in LD, gene density, and distance to the nearest gene, allowing for a maximum deviation of 10% for all four criteria. Similarly, when testing the enrichment of potential regulatory SNPs in Hi-C-interacting anchors, we also selected randomly matched SNPs using SNPsnap.²¹ The statistical significance was calculated using Fisher’s exact test.

Selecting prioritized SNPs falling within potential adipocyte/adipose enhancer or open chromatin regions

To select prioritized SNPs falling within potential adipocyte enhancer or open chromatin regions for further STARR-seq experiments, we annotated the fine-mapped SNPs with epigenetic peaks of adipose/adipocytes from public datasets. These datasets included our previously published H3K27ac/H3K4me1/ATAC-seq peaks in adipocytes (GEO: GSE151324),²² chromatin state enhancer regions (15-state model state 7) of adipocyte (E023) or adipose (E063) from the Roadmap²³ project, and ATAC-seq peaks in adipocytes (GEO: GSE129574²⁴) or adipose (GEO: GSE123038,²⁵ GSE110734,²⁶ and GSE172647²⁷). SNPs that overlapped with at least one H3K27ac/H3K4me1/ATAC-seq peak or enhancer region were considered as prioritized SNPs.

STARR-seq

Generally, there are two strategies for STARR-seq library construction: short fragments (≤230 bp)^28,29 obtained from oligonucleotide synthesis technology or long fragments (≥500 bp)^13,30 sourced from sheared whole-genome DNA. As summarized by Das et al.,³¹ shorter fragments are more suited to fine-mapping the enhancer effects of individual TF binding sites, while longer fragments are optimal for genome-wide screens or enhancer discovery. Since a major aim of our study is to assess the allelic enhancer activity of SNPs, we chose the short fragment from oligonucleotide synthesis, as recommended by the company (GenScript, USA), including a 120-bp DNA sequence and 30-bp adaptor. This strategy can directly obtain the fragments containing both reference and alternative alleles for all prioritized SNPs.

Using prioritized SNPs, the 120-bp DNA sequence centered on each SNP was extracted from the human genomic context, which included reference and alternative alleles. The 15-bp identical adapter sequence was further added on the 5′ and 3′ regions for library construction (Table S1). We synthesized 150-bp-long DNA fragments using an oligonucleotide CustomArray (Table S1) (GenScript, USA). First, oligonucleotides were amplified using adapter primers (Table S1) (NEB, USA) and cloned into the linearized hSTARR-seq_ORI vector (Addgene, USA) as previously described,³² and then we extracted the plasmid library for cell transfection. Second, a total of 15 μg STARR-seq plasmid library was transfected into 1.0 × 10⁷ preadipocytes according to the manufacturer’s protocol. Preadipocytes derived from human subcutaneous adipose tissue of healthy subjects were purchased from BLUEFBIO (Shanghai, China). Three biological replicates were performed. Third, STARR-seq libraries were prepared for Illumina sequencing. The plasmid library was used to construct an input sequencing library before cell transfection. In brief, 100 ng plasmid library was treated with reverse-transcript unique molecular identifier (RT-UMI) (Table S1) as follows: 98°C for 30 s, 63°C for 30 s, and 72°C for 2 min. Subsequently, two rounds of amplification were applied using Vector-F and Read1-R or barcode primers, respectively (Table S1). On the other hand, total RNA was extracted and treated with DNase I for output sequencing construction 24 h post-transfection. The poly(A) RNA was further isolated for cDNA synthesis (NEB, USA). Next, specific reverse transcription was performed using RT-UMI (Table S1) (NEB, USA) and treated with RNase A (Takara, Japan). Similarly, two rounds of amplification were performed using a junction primer³³ and Read1-R or barcode primers, respectively (Table S1). Sequencing libraries were generated following the manufacturer’s protocol (NEB, USA) and sequenced with paired-end reading on an Illumina NovaSeq 6000 platform.

STARR-seq data analysis and identification of SNPs with allelic enhancer activity

The FASTQ files for each replicate were separated according to the 8-bp barcode sequence information. The 120-bp SNP element sequences were extracted and aligned to our selected SNP sequences using Bowtie 2.4.1.³⁴ Only 100% matched reads were used for counting unique molecular identifiers (UMIs) for each SNP reference/alternative allele. SNP alleles with no expression in any one replicate or with pooled input/output counts less than 10 were removed for subsequent analysis.

As in Tippens et al.,³⁵ enhancer activity analysis for each fragment was performed using the limma package (v.3.34.9).³⁶ Three replicates of the DNA STARR-seq library were sequenced and included in the differential analysis. Briefly, read counts were first normalized with the voom function. A linear model was then fitted to each fragment, and empirical Bayes moderated t statistics were used to assess the difference. Raw p values were adjusted using the Benjamini-Hochberg method. Fragments with enhancer activity were defined as those showing significantly higher expression in output than in input samples (fold change [FC] > 1.5, false discovery rate [FDR] < 0.05). To identify SNPs with allelic enhancer activity, we used Fisher’s exact test to compare allelic reads ratio between input and output libraries for the enhancer-region SNPs. The threshold for significant allelic expression change was set as FDR < 0.05.

Dual-luciferase reporter assay

We randomly selected eight SNPs of STARR-seq for validation in human preadipocytes purchased from BLUEFBIO (Shanghai, China). To be consistent with the STARR-seq experiment, we also chose 120-bp fragments for our dual-luciferase reporter assays. Briefly, these 120-bp fragments containing reference or alternate alleles of SNPs were amplified from a plasmid library or genomic DNA (Table S1). Each fragment was inserted into the laboratory-modified pGL3-SCP1-promoter vector. For each SNP, three biological replicates were performed, and ∼3.0 × 10⁵ cells were used for each replicate. Luciferase activity was measured via target SNP-pGL3-promoter plasmid cotransfected with Renilla plasmid using the ViaFect transfection reagent (Promega, USA) in a 24-well chamber. After 48 h, firefly luciferase fluorescence and Renilla were tested with Promega GloMax Navigator (Promega, USA). Firefly luciferase was normalized to Renilla measurements and then the fold change over the pGL3-promoter vector was calculated to determine enhancer activity.

Target gene identification

To identify target genes for the SNPs with allelic enhancer activity, we first used Hi-C and eQTL analysis to obtain potential target genes. Then GWAS-eQTL colocalization analysis was performed, and only genes with PP4 (the probability of a shared causal variant for GWAS and eQTL) over 0.5 were finally retained as the target genes. In our analysis, we retained only protein-coding and long non-coding genes annotated in GENCODE v.19. Using the precomputed median expression across individuals for the Genotype-Tissue Expression (GTEx) dataset,³⁷ we kept only the genes with median transcripts per million (TPM) > 5 in subcutaneous or visceral omentum adipose tissue in our analyses. The GTEx genotype data were applied from the dbGaP database (phs000424.v8.p2).

Potential target genes derived from Hi-C or capture Hi-C data

We used previously published Hi-C or capture Hi-C data to identify potential target genes of the SNPs with allelic enhancer activity. We assessed whether the SNP location (2 kb surrounding the SNP position) and gene promoter had significant loop or chromatin interaction. The Hi-C data we used included our in-house Hi-C data from human mesenchymal stem cell induced adipocytes (GEO: GSE151324)²² and capture Hi-C interactions from four other sources: human white adipocytes differentiated from primary preadipocytes (GEO: GSE110619),³⁸ human fat tissue (GEO: GSE86189),³⁹ human mesenchymal stem cell induced adipocytes (GEO: GSE140782),⁴⁰ and adipocytes differentiated from Simpson-Golabi-Behmel syndrome preadipocytes.¹⁴ Detailed information on these data is shown in Table S2.

Potential target genes derived from eQTL analysis

eQTL analyses were performed using data from subcutaneous adipose and visceral omentum adipose tissues. For subcutaneous adipose, we performed a meta-analysis on the data from the GTEx consortium v.8 (n = 581)³⁷ and the METSIM study (n = 434)⁴¹ by using the conventional inverse-variance-weighted meta-analysis in the SMR software package⁴² (v.1.03). All METSIM participants provided informed consent.⁴¹ The visceral omentum adipose eQTL data were obtained from the GTEx consortium v.8 (n = 469).³⁷

GWAS-eQTL colocalization analysis

We used coloc⁴³ to assess whether the detected GWAS signal and eQTL association shared the same causal variant. This method used a Bayesian statistical test to estimate five posterior probabilities (PP0, PP1, PP2, PP3, and PP4). A larger value of PP4 indicates a higher probability of a shared causal variant for GWAS and eQTL. Genes with PP4 over 0.5 were considered to support the colocalization in this study.

Gene association with cardiometabolic traits in 770 METSIM individuals

Using data from GEO: GSE70353,⁴⁴ we performed Spearman correlation analysis to investigate the association between the target genes and cardiometabolic traits, including WHR, seven glycemic traits, six lipid-related traits, two inflammatory traits, two blood pressure traits, and one kidney function trait. For the target genes whose expression values were significantly associated with WHR, we checked the International Mouse Phenotyping Consortium database (IMPC; https://www.mousephenotype.org/) to confirm whether these genes have been reported to be related to obesity-related traits in mouse experiments. The DrugBank (https://go.drugbank.com/) and Therapeutic Target Database (http://db.idrblab.net/ttd/) were also used to check whether the protein products of these genes are targeted by active ingredients for central-obesity-related diseases.

MR analysis to identify potential therapeutic targets

pQTL data

Summary data of pQTL data were obtained from the resource released by Ferkingstad et al. (https://www.decode.com/summarydata/).⁴⁵ The data were derived from 35,559 Icelanders.

Instrument variable selection

For each exposure, we used the clumping algorithm in PLINK (v.1.9)⁴⁶ to select independent SNPs (r² threshold = 0.001, window size = 1 Mb, and p < 1 × 10⁻⁵) as instrument variables (IVs). The 1000G European data (phase 3) were used as the reference for LD estimation. SNPs with MAF < 0.01 or SNPs within long LD regions⁴⁷ in the genome (https://genome.sph.umich.edu/wiki/Regions_of_high_linkage_disequilibrium_(LD)) were removed. We used the RadialMR⁴⁸ package, which identifies outlying genetic instruments via modified Q statistics, to further exclude outlying pleiotropic SNPs.

Two-sample MR analyses

Two-sample MR analyses were performed to explore the causal relationships between the plasma protein levels and central obesity. The inverse variance weighted (IVW) method,⁴⁹ which performs meta-analysis on the SNP-specific Wald estimates using multiplicative random effects, was considered as the primary MR method. If the balanced pleiotropy assumption is violated, the causal estimates of IVW will be biased. Therefore, four other complementary MR methods were conducted to enhance the reliability of the results. MR-RAPS accounts for systematic and idiosyncratic pleiotropy and can provide a robust causal effect inference with many weak instruments.⁵⁰ The MR-Egger method⁴⁹ is based on the INSIDE (instrument strength independent of the direct effects) assumption and estimates the causal effect through the slope coefficient of the Egger regression. The weighted median method assumes that at least 50% of the total weight of the instrument comes from valid instruments⁵¹ and estimates the causal effect from the median of the weighted empirical density function of individual IV effect estimates. The weighted mode method assumes that the most common causal effect is consistent with the true causal effect and provides a consistent effect estimate when the largest number of similar IV estimates come from valid instruments.⁵² We also performed directionality check analysis of causal relationships by the MR-Steiger test.⁵³ The analyses of all the above MR methods were carried out using the TwoSampleMR package (v.0.4.26) in R.⁵⁴ For the significant MR results, the intercept term of the MR-Egger method was used to estimate the directional pleiotropic effect⁵⁵ (p < 0.05).

TF enrichment analysis

To identify TFs that preferentially bind one allele over another, we extracted genome sequences encompassing a different allele of each SNP and used FIMO from the MEME Suite toolkit⁵⁶ for binding affinity analysis using default parameters. The motifs were collected from five public databases: JASPAR,⁵⁷ HOCOMOCO,⁵⁸ SwissRegulon,⁵⁹ TRANSFAC,⁶⁰ and the TF profiles released by Jolma et al.⁶¹ TFs with median TPM > 5 in subcutaneous or visceral omentum adipose tissue in the GTEx dataset³⁷ were retained. To identify key TFs involved in the regulation roles of SNPs with allelic enhancer activity, we performed enrichment analysis for all predicted allelic binding motifs by comparing SNPs with allelic enhancer activity against other enhancer-region SNPs using Fisher’s exact test. TFs binding on at least five SNPs with allelic enhancer activity were subjected to the enrichment analysis. We kept all TFs with p < 0.05 and fold change >2 as significant enriched TFs.

Cell culture

Human embryonic kidney 293T cells (HEK293T), 293A, and mouse embryonic fibroblasts (3T3-L1) were purchased from National Collection of Authenticated Cell Cultures (Shanghai, China). Human preadipocytes derived from the subcutaneous white adipose tissue of healthy subjects were purchased from BLUEFBIO (Shanghai, China). Cells were cultured in DMEM (HyClone, USA) with 10% fetal bovine serum (FBS; Biological Industries, Israel), 100 U/mL penicillin, and 0.1 mg/mL streptomycin (Solarbio, USA) at 37°C in a 5% CO₂ incubator.

Genotyping of SNPs

To confirm the genotype of rs7119797 and rs8079062, genomic DNA was isolated from 293A cells or human preadipocytes purchased from BLUEFBIO (Shanghai, China). The fragment harboring these two SNPs was amplified and sequenced (Table S1).

Dual-luciferase reporter assay of rs8079062

The 120-bp putative enhancer fragment centered on the reference or alternate allele of rs8079062 and the 2,148-bp RNF157 promoter (1,986 bp upstream to 161 bp downstream of the transcription start site [TSS]) were amplified from a plasmid library or healthy human genomic DNA (Table S1). The putative enhancer and promoter fragments were cloned into the pGL3-basic vector (Promega, USA). Three biological replicates were performed, and ∼3.0 × 10⁵ cells were used for each biological replicate. The cotransfection and measurement of luciferase activity were the same as described under “dual-luciferase reporter assay.”

CRISPRi of SNP-harboring regions

We applied CRISPR inhibition (CRISPRi; lenti-CRISPR-KRAB-PuroR vector) to inhibit the regulatory activity of fragments harboring SNPs of interest. Two single-guided RNAs (sgRNAs) were designed for each targeted region using CHOPCHOP (http://chopchop.cbu.uib.no) and CRISPOR (http://crispor.tefor.net) (Table S1). Annealed sgRNA was cloned into the pUC19-hU6-sgRNA vector and then was digested by KpnI and SpeI (Takara, Japan). Next, the target sgRNA was inserted into the lenti-CRISPR-KRAB-PuroR vector by KpnI and NheI (Takara, Japan). We cotransfected 3 μg plasmids, including the CRISPRi plasmid with pCMV-VSV-G and psPAX2 (Addgene, USA) (1:1:1 molar ratio), into HEK293T in a six-well chamber. After 48 and 72 h, lentiviral supernatant was collected and further infected into 0.8–1.0 × 10⁶ target cells (human preadipocytes purchased from BLUEFBIO, Shanghai, China). To facilitate infection, we added Polybrene (Solarbio, China) to the viral medium at a concentration of 2 μg/mL. Then 1.5 μg/mL puromycin was used to select positive human preadipocytes for 72 h. Finally, total RNA was harvested for real-time qPCR analysis using ∼1.0 × 10⁶ positive cells. Three biological replicates were performed.

ChIP-qPCR assay

First, ∼1 × 10⁷ cells were cross-linked with 37% formaldehyde, and the reaction was terminated by adding glycine. The cells were then washed with ice-cold 1× PBS and harvested using a cell scraper. Second, 0.5 μL micrococcal nuclease was added to each immunoprecipitation (IP) and incubated for 25 min at 37°C with frequent mixing. The digestion was halted by adding 0.5 M EDTA, and the tubes were chilled on ice for 1–2 min. The nuclear pellet was resuspended with 1× chromatin IP (ChIP) buffer and sonicated using a JY92-IIN ultrasonic homogenizer (SCIENTZ, China) to fragment the chromatin, applying three sets of 10-s pulses in 1.5-mL tubes containing up to 500 μL lysate. Between pulses, samples were incubated on wet ice for 30 s. The supernatant was collected after centrifugation. Third, 10 μL digested chromatin was used as the 2% input sample. For each IP, 500 μL digested chromatin was transferred to a 1.5-mL tube and the corresponding IP antibody was added. For the negative control, 2 μL (2 μg) immunoglobulin G (IgG) was added to the IP sample. For the TFs (anti-ELF-1, sc-133096, and anti-LBP1, sc-81310, Santa Cruz Biotechnology, USA), 2 μg of each antibody was added to the respective IP samples. The reactions were incubated overnight at 4°C with rotation. Protein G magnetic beads were added to each IP reaction and incubated for 4 h at 4°C with rotation. The beads/antibody/chromatin complexes were washed with low- and high-salt wash buffer. Chromatin was then eluted from the antibody/protein G magnetic beads for 30 min at 65°C with gentle vortexing. To all tubes, including the 2% input sample, reverse cross-linking was achieved by adding 5 M NaCl and proteinase K, followed by overnight incubation at 65°C. Finally, the purified DNA samples were used to verify the binding affinity by using real-time qPCR (primers shown in Table S1).

Lentivirus gene overexpression

Full-length cDNAs of ELF1 (GenBank: NM_172373.4), UBP1 (GenBank: NM_001128160.2), and Rnf157 (GenBank: NM_027258.2) were amplified from human preadipocytes purchased from BLUEFBIO (Shanghai, China) or 3T3-L1 cells and further cloned into laboratory-modified lenti-CMV- EF1a-PuroR vector (Table S1). A total of 3 μg plasmids, including the plasmid overexpressing the gene of interest and two helper plasmids (pCMV-VSV-G and psPAX2), was cotransfected into HEK293T cells in a six-well chamber. After 48 and 72 h, the lentiviral supernatant was collected and used to infect 0.8–1.0 × 10⁶ target cells. Empty lenti-CMV-EF1a-PuroR vector was used as negative control. Subsequently, puromycin was used to select positive cells (1.5 μg/mL in human preadipocytes and 2 μg/mL in 3T3-L1). Finally, total RNA and protein were extracted for real-time qPCR and western blot after puromycin selection (Table S1). Three biological replicates were performed.

Short hairpin RNA knockdown

Independent oligonucleotides targeting ELF1, UBP1, and Rnf157 were designed and cloned into the laboratory-modified lenti-shRNA-miR30 backbone (Table S1). Similarly, the short hairpin RNA (shRNA) plasmid and two helper plasmids (pCMV-VSV-G and psPAX2) were cotransfected into HEK293T cells in a six-well chamber. After 48 and 72 h, we collected the lentiviral supernatant and infected human preadipocytes purchased from BLUEFBIO (Shanghai, China) or the 3T3-L1 cell line. The shRNA-NC vector was used as negative control (Table S1). Total RNA and protein were extracted for detecting the mRNA and protein expression.

DNA/RNA isolation and quantitative real-time PCR

We extracted genomic DNA from different cells according to the manufacturer’s protocol (TIANGEN Biotech, China). Total RNA was isolated (Fastagen, China) and further reversed transcribed using the PrimeScript RT reagent kit (Takara, Japan). real-time qPCR was applied according to the manufacturer’s instruction (Biomake, China) by the CFX Connect Real-Time PCR Detection System (Bio-Rad, USA). β-actin (human/mouse) was used as an endogenous control to normalize the differences in samples (primers shown in Table S1).

Western blotting

Cells were lysed with cold RIPA lysis solution containing SDS-PAGE loading buffer (Epizyme Biomedical Technology, China). The protein was denatured by boiling for 15 min and separated by 10% SDS-PAGE. Then the sample was electrotransferred onto polyvinylidene fluoride (PVDF) membranes. Blocked with 5% non-fat milk for 2 h at room temperature, the membranes were incubated at 4°C with corresponding primary antibodies overnight. After three washes with Tris Buffered Saline with Tween-20 (TBST), the membranes were incubated with horseradish peroxidase (HRP)-labeled goat anti-mouse/rabbit IgG (H + L) (Epizyme Biomedical Technology, China) for 1 h at room temperature. The results were visualized by using an Omni-ECL Femto Light Chemiluminescence Kit (Epizyme Biomedical Technology, cat. no. SQ201, China) in MiniChemi610 (Beijing Sage, China). Antibodies used in western blotting included anti-UBP1 (Santa Cruz Biotechnology, sc-81310, USA), anti-CEBPa (HuaBio, ET1612-46, China), anti-PPARg (Cell Signaling Technology, cat. no. 24435, USA), anti-RNF157 (Immunoway, cat. no. YT6295, USA), and anti-β-actin (Cell Signaling Technology, cat. no. 4970, USA).

3T3-L1 cell differentiation

Briefly, 3T3-L1 cells were cultured in DMEM with 10% FBS and 1% penicillin/streptomycin for 2–3 days, and confluent cells were induced in DMEM containing 10% FBS and 1% penicillin/streptomycin, dexamethasone (1.0 μM, Beyotime, China), 3-isobutyl-1-methylxanthine (IBMX, 0.5 mM, Beyotime, China), rosiglitazone (2.5 μM, Solarbio, USA), as well as insulin (10 μg/mL, Solarbio, USA). Two days after the induction, the inducing medium was changed, and the cells were maintained in DMEM containing 10% FBS and insulin (10 μg/mL) for 2 additional days. Total RNA and protein were collected for real-time qPCR and western blot at different time points. In addition, lipid accumulation in the cells was detected by oil red O staining following the manufacturer’s protocol (ZHHC, China).

Results

SNP prioritization based on fine-mapping and epigenomic annotation

For central obesity, we used the largest European GWAS summary dataset (n up to 694,649)⁶ until May of 2021 for subsequent analysis. LD score regression (LDSC)⁶² was used to evaluate tissue-level enrichment of SNP heritability for central obesity and BMI (summary data derived from up to 806,834 individuals⁶) using expression data from the GTEx database³⁷ and the Franke lab dataset.^63,64 Different from BMI, which is mainly enriched in tissues of the central nervous system, central obesity is most significantly associated with the adipose tissue (Figure S1; Table S3). Therefore, the Hi-C, eQTL, or epigenomic data we used for the current study are mainly derived from adipose tissue or adipocytes.

Loci were defined as 500 kb upstream and downstream of SNPs attaining genome-wide significant evidence of association (p < 5 × 10⁻⁸) in the summary data. If two loci overlapped, they were merged into a single locus. Finally, a total of 362 loci for central obesity were identified. For each locus, we performed a stepwise model selection procedure to identify independent signals using the GCTA software¹⁶ with the parameter --cojo-slct --cojo-p 5e−8. As shown in Figure 2A, we detected a total of 614 independent signals. After the fine-mapping analysis, a total of 9,277 SNPs (Table S4) achieving either conditional independence or a high posterior probability for causality (log10BF > 2) were retained and are referred to as “fine-mapped SNPs” for subsequent analysis. 150 SNPs were located within exons of protein-coding genes, and only 31 SNPs (26 loci, Table S5) were further predicted to produce altered proteins. As expected, the remaining 9,127 non-coding SNPs of central obesity were found to be most significantly enriched in the state of enhancer (Figure S2).

Figure 2.

Prioritized SNPs for the STARR-seq experiment and selective validation of the SNPs with allelic enhancer activity

(A) Selection process of the prioritized SNPs.

(B) Compared with other fine-mapped SNPs or matched random background SNPs selected using SNPsnap, the 2,034 prioritized SNPs have higher PhyloP conservation scores (^∗∗∗p < 0.001, Wilcoxon test).

(C) Using five previously published Hi-C or capture Hi-C datasets, we checked the proportion of prioritized SNPs falling within Hi-C interacting anchors. A significantly higher proportion of prioritized SNPs overlapped with interacting anchors compared with other fine-mapped SNPs or matched random background SNPs. The capture Hi-C data for Simpson-Golabi-Behmel syndrome (SGBS) human preadipocytes-derived adipocytes were obtained from the study performed by Joslin et al.¹⁴

(D) We randomly selected five SNPs with allelic enhancer activity to examine their regulatory activity by dual-luciferase reporter assay. Four of the five SNPs showed significant differences between alleles (p < 0.05; the directions of effect are also the same). Error bars represent standard deviation. ^∗p < 0.05, ^∗∗p < 0.01, ^∗∗∗p < 0.001; paired Student’s t test.

(E) We randomly selected three SNPs without allelic enhancer activity to examine their regulatory activity by dual-luciferase reporter assay. No significant allelic difference was detected for these three SNPs (ns, not significant; paired Student’s t test). Error bars represent standard deviation.

SNPs falling within adipocyte/adipose enhancer or open chromatin regions were considered as prioritized SNPs and used for subsequent STARR-seq experiments. The data sources for enhancers included our previously published H3K27ac and H3K4me1 peaks in adipocytes (GEO: GSE151324)²² and the chromatin state enhancer regions (15-state model state 7) of adipocytes (E023) or adipose (E063) from the Roadmap²³ project. The data sources for open chromatin regions included ATAC-seq peaks in adipocytes (GEO: GSE151324²² and GSE129574²⁴) or adipose (GEO: GSE123038,²⁵ GSE110734,²⁶ and GSE172647²⁷).

As shown in Figure 2A, we obtained 9,277 fine-mapped SNPs from the fine-mapping analysis. After excluding 150 exonic SNPs, the remaining 9,127 non-coding fine-mapped SNPs were subjected to enhancer or open chromatin region annotation. 1,591 SNPs from 303 loci were identified within enhancer regions, whereas 919 SNPs from 288 loci resided in open chromatin regions. Among them, 476 SNPs were annotated with both enhancer and open chromatin signals. Consequently, a total of 2,034 SNPs from 326 loci were selected. Detailed information on the prioritized SNPs is shown in Table S6 and Figure S3.

We examined conservation scores to check whether the prioritized SNPs are conserved in primate evolution. Compared with other fine-mapped SNPs or matched random background SNPs selected using SNPsnap,²¹ the prioritized SNPs have higher PhyloP conservation scores (Figure 2B, Wilcoxon test, p = 1.98 × 10⁻¹⁴ and 8.57 × 10⁻¹⁶, respectively). We also checked the proportion of prioritized SNPs falling within Hi-C interacting anchors. Detailed information on the Hi-C or promoter capture Hi-C data we used are described under material and methods. In each dataset, there are significantly higher proportions of prioritized SNPs overlapped with interacting anchors (Figure 2C). These results suggested that the prioritized SNPs we selected are likely to play important regulatory roles in the development of central obesity.

Identifying SNPs with allelic enhancer activity through STARR-seq

Identification of SNPs with allelic enhancer activity

A common mechanism by which non-coding SNPs lead to disease risk is target gene expression regulation mediated by alterations in enhancer function. To identify SNPs with allelic enhancer activity, we employed STARR-seq to test the 2,034 prioritized SNPs. The experiment pipeline is shown in Figure 1. Detailed information on the sequences is shown in Table S7. Correlation analysis validated the concordance among the three biological replicates (r² > 0.8, Figure S4A). We evaluated the correlation between the input and the output libraries for all prioritized fragments in the three replicates, yielding moderate correlations (r² ranged from 0.55 to 0.59, Figure S4B). However, a higher correlation between enhancer fragments (r² ranged from 0.82 to 0.88, Figure S4B) was observed. This could be explained by the enhancer fragments being more functionally active compared with all fragments, leading to log2FC distributions with fewer variations and higher correlations between input and output libraries.

Enhancer activity analysis was performed for the STARR-seq data using the limma package (v.3.34.9)³⁶ (details under material and methods). A total of 791 SNPs were identified as enhancer-region SNPs, based on the observation that at least one of their allele’s corresponding fragments exhibited enhancer activity (FC > 1.5, FDR < 0.05, Table S8). For these 791 enhancer-region SNPs, we further used Fisher’s exact test to compare the allelic reads ratio between the input and the output libraries. The analysis revealed that 141 SNPs exhibited significant differences in enhancer activity between the two alleles, thereby qualifying them as SNPs with allelic enhancer activity (FDR < 0.05, Table S9). We checked the located regions for these SNPs and found that these 141 variants represent causal variant candidates for 97 GWAS risk loci (29.75% of all 326 tested loci) (Table S9). In the results of fine-mapping analysis, a total of 3,107 fine-mapped non-coding SNPs were identified for these 97 loci. After the epigenomic annotation and STARR-seq process, we reduced the number of causal variants from an average of 32 fine-mapped non-coding SNPs to an average of 2 SNPs per locus (Figure S5).

Validation of enhancer activity of random selected SNPs with dual-luciferase reporter assay

We randomly selected five SNPs with allelic enhancer activity and three enhancer-region SNPs without allelic enhancer activity to examine their regulatory activity by dual-luciferase reporter assay in the preadipocytes derived from the subcutaneous white adipose tissue of healthy subjects (purchased from BLUEFBIO, Shanghai, China). Four of the five SNPs indeed showed significant differences between alleles (p < 0.05, Figure 2D, the directions of effect are also the same). Meanwhile, no significant allelic difference was detected for the three enhancer-region SNPs without allelic enhancer activity (Figure 2E). These results supported the credibility of our STARR-seq analysis results. We noticed that the FC values measured by dual-luciferase reporter assay were different from those obtained from the STARR-seq experiment. The differences might be because the principles of the two methods are different. In STARR-seq, enhancer activity is assessed by the transcriptional fold change, quantified through the output RNA level relative to input DNA. In contrast, the dual-luciferase reporter assay creates a plasmid that enables the enhancer sequence to govern luciferase reporter gene transcription. This approach evaluates enhancer activity at the protein level, specifically through the quantification of luciferase enzyme activity.

Target genes of the SNPs with allelic enhancer activity

Target gene identification

We next tried to pinpoint potential target genes for these SNPs with allelic enhancer activity. Considering that enhancer variants might regulate target genes through long-range chromatin interaction, we first used Hi-C or capture Hi-C datasets to obtain potential target genes. Five Hi-C or capture Hi-C datasets from adipocyte/adipose tissue were included in the analysis (details on formation in Table S2), including our in-house Hi-C data from human mesenchymal stem cell induced adipocytes (GEO: GSE151324),²² capture Hi-C interactions obtained from human white adipocytes (GEO: GSE110619),³⁸ human fat tissue (GEO: GSE86189),³⁹ human mesenchymal stem cell induced adipocytes (GEO: GSE140782),⁴⁰ and human adipocytes differentiated from Simpson-Golabi-Behmel syndrome preadipocytes.¹⁴ Genes whose promoter has a significant loop or chromatin interaction with the location of the SNPs with allelic enhancer activity (2 kb surrounding the SNP position) in at least one dataset were considered as potential target genes. Finally, from Hi-C or capture Hi-C data analysis, we obtained 241 potential target genes (Table S10).

To complement the identification of target genes, we also performed cis-eQTL analysis using data from subcutaneous adipose and visceral omentum adipose tissues. A total of 280 potential eQTL target genes were identified (Table S11). Collectively, a total of 453 potential target genes were identified. These genes were found to be enriched for functional annotations such as lipid storage, fat cell differentiation, and lipid homeostasis (Figure 3A), supporting the importance of the identified genes in central-obesity-related disease etiology. For these potential target genes, we further performed colocalization analysis to prioritize target genes that share the same causal variant between cis-eQTL and GWAS association. Finally, a total of 41 target genes with the probability of shared causal variant for GWAS and eQTL (PP4) over 0.5 remained (Figure 3B; Table S12). The regulatory relationship between 15 SNP-gene pairs was supported by both Hi-C and eQTL analyses (Figure 3B; Table S12). For example, a known obesity gene, AHNAK, was potentially regulated by the SNP rs7119797, which was located in an enhancer about 150 kb away from the promoter (Figure 3C). Previous studies^65,66 have shown that Ahnak^−/− mice have a strong resistance to high-fat diet-induced obesity.

Figure 3.

Target genes for the SNPs with allelic enhancer activity

(A) Gene ontology biological process enrichment analysis results for the potential target genes.

(B) The link between SNPs with allelic enhancer activity and target genes with colocalization. Purple lines indicate that the regulation relationship between SNP and target gene is supported by both Hi-C and eQTL analysis. Those genes are also marked in purple background. The number in the outer ring indicates the chromosome number.

(C) Hi-C interactions between the AHNAK promoter and a distal enhancer (∼150 kb). Epigenetic annotation in adipose/adipocytes was visualized using the WashU EpiGenome Browser (https://epigenomegateway.wustl.edu/browser/).

(D) Correlation analysis results between the 41 target genes and central-obesity-related traits in the METSIM cohort. Asterisks (^∗) in the heatmap indicate the correlation results with p < 0.05. 37 genes were detected in the expression array, and 20 of them were associated with central obesity. Among these 20 genes, the numbers of genes associated with lipid-related traits, glycemic traits, blood pressure, kidney function, and inflammatory traits were 19, 19, 17, 2, and 19, respectively. In addition, 9 of these 20 genes have been confirmed to be related to obesity-related traits in mouse experiments (in red) and 4 genes are the target genes for approved or investigational drugs (marked with pink box).

(E) Summary Mendelian randomization (MR) estimates derived from the inverse-variance weighted method. Plasma protein levels of MFAP2, TMEM132C, and NID2 were used as exposure, and significant associations were detected for them with central obesity. The x axis represents the betas and the bars around the points are 95% confidence intervals.

Understanding phenotype association through gene expression and function

To understand the association between the above target genes and central obesity, we first studied whether the expression of the gene itself in subcutaneous adipose was correlated with the central-obesity-related traits in the METSIM cohort. Among the 41 identified target genes, 37 were detected in the expression array, and 20 of them were associated with central obesity (Figure 3D; Table S13). Among these 20 genes, the number of genes associated with lipid-related traits, glycemic traits, blood pressure, kidney function, and inflammatory traits were 19, 19, 17, 2, and 19, respectively (Figure 3D; Table S13). In addition, 9 of these 20 genes have been confirmed to be related to obesity-related traits in mouse experiments (Figure 3D; Table S14), and 4 genes are the target genes for approved or investigational drugs (Figure 3D; Table S15). For example, GPBAR1 was supported by both mouse experiments and a drug database. Mutation of Gpbar1 in mice resulted in abnormal adipose tissue amount, abnormal cholesterol homeostasis, and impaired glucose tolerance (MGI: 2653863). GPBAR1 is the target gene for sodium taurocholate (Table S15), which is a drug for treating type 2 diabetes (T2D) in phase 1/2 trials.

Potential therapeutic target discovery

To test whether the above 20 genes associated with central obesity could be served as potential therapeutic targets, we performed MR analysis using pQTL data from the Iceland cohort (sample size n = 35,559). pQTL data for only three genes (MFAP2, TMEM132C, and NID2) were available, and we found evidence for causal associations of all three of these plasma proteins with central obesity (Figure 3E; Table S16). For example, higher levels of plasma MFAP2 were associated with a higher level of central obesity (β = 0.0315, confidence interval [CI] from 0.0178 to 0.0451, p = 6.31 × 10⁻⁶). The MR-Steiger test showed that there was no evidence for reverse causation (Table S16). No significant evidence of pleiotropy was detected by the MR-Egger’s intercept test (Table S16).

Key TFs mediating the central-obesity-relevant genetic regulatory network

Enrichment analysis identified 20 key TFs

To explore candidate TFs mediating the regulatory roles of the 141 SNPs with allelic enhancer activity, the MEME Suite toolkit⁵⁶ was used to predict allelic binding motifs. All TFs with predicted genotype-dependent motif disruption for the SNPs with allelic enhancer activity are listed in Table S17. Further enrichment analysis detected 20 key TFs; thus, a significantly higher proportion of SNPs with allelic enhancer activity are within the binding sites of these TFs compared to other enhancer-region SNPs (Figure 4A). Most of these TFs have been reported to be involved in obesity-related biological processes (Table 1), with five exceptions (ELF1, ZFX, THAP1, TCF12, and ETV6). Using the SNPs with allelic enhancer activity, the 20 key TFs, and potential target genes of the SNPs, we constructed a network to illustrate the regulatory roles of these TFs (Figure 4B).

Figure 4.

TFs mediating the central-obesity-relevant genetic regulatory network

(A) Transcription factor (TF) enrichment analysis identified 20 key TFs (p < 0.05 and FC > 2). The threshold line for significance of p = 0.05 is shown in blue, whereas the cutoff line for FC = 2 is shown in yellow. Seven TFs with adjusted p < 0.1 determined by the Benjamini-Hochberg method are shown in red.

(B) The regulatory network was constructed with the Cytoscape software using SNPs with allelic enhancer activity, key TFs, and all potential target genes. The protein-protein interactions in adipocytes from the TissueNet database (v.2)⁶⁷ were used to connect the target genes. For genes with GWAS-eQTL colocalization, node size is proportional to its number of connections with other nodes.

(C) ELF1 motif binding the sequences surrounding rs7119797.

(D) The color-highlighted peak shows the homozygous CC genotype of rs7119797 in human preadipocytes purchased from BLUEFBIO (Shanghai, China). ChIP-qPCR assay confirmed that ELF1 was recruited by rs7119797-C (^∗∗∗p < 0.001, paired Student’s t test). Error bars represent standard deviation.

(E) The expression of AHNAK was significantly depressed in ELF1-knockdown preadipocytes and significantly enhanced in preadipocytes overexpressing ELF1. Error bars represent standard deviation (^∗p < 0.05, ^∗∗p < 0.01, and ^∗∗∗p < 0.001, paired Student’s t test).

Table 1.

Support from published studies for the involvement of 20 key transcription factors in obesity-related biological processes

TF name	Related function	Supporting study
ZBTB7C	adipocyte differentiation	Choi et al.68
KLF2	adipogenesis	Banerjee et al.69
CTCF	adipogenesis	Chen et al.70
KLF9	thermogenesis of brown and beige fat	Fan et al.71
ELF1	–	–
KLF4	adipose tissue inflammation	Bulut et al.72
ZFX	–	–
WT1	browning of epididymal fat and metabolic dysfunction	Kirschner et al.73
SP1	lipid and glucose homeostasis	Chen et al.74
THAP1	–	–
PATZ1	adipogenesis	Patel et al.75
BCL6B	energy dissipation	Hiraike et al.76
EGR3	insulin resistance	Zapata-Bustos et al.77
ELK4	proliferation of adipose-derived mesenchymal stem cells	Pepin et al.78
TCF12	–	–
ETV6	–	–
VEZF1	diabetes-associated endothelial dysfunction	Kuschnerus et al.79
ELF2	lipoapoptosis	Akazawa et al.80
EBF1	adipogenesis	Jimenez et al.81
ZBTB7B	adipose inflammation and fibrosis	Zhao et al.82

Validating the regulatory effect of ELF1 on AHNAK expression

Among the five TFs without well-studied functional roles in obesity-related biological processes, ELF1 showed the most significant enrichment binding on SNPs with allelic enhancer activity (Figure 4A; Table 1). In addition, one of its target genes is our above-mentioned known obesity gene AHNAK (Figure 4B). Therefore, we chose ELF1 for further functional validation.

Motif analysis showed that ELF1 might specifically bind to rs7119797-C (Figure 4C) to regulate the expression of AHNAK. To validate the binding of ELF1 to rs7119797-C, we performed SNP genotyping, and the result showed that our preadipocyte is homozygous CC genotype for rs7119797 (Figure 4D). Further ChIP-qPCR assay confirmed that ELF1 bound at this locus (Figure 4D). We also performed genotyping for the 293A cell, and it is heterozygous CT genotype for rs7119797 (Figure S6A). ChIP-qPCR confirmed that ELF1 was preferentially recruited by the C allele compared with the T allele (Figure S6A, p < 0.001). We used CRISPR-dCas9 to suppress the activity of this enhancer region. Significantly decreased AHNAK expression (p < 0.001, Figure S6B) was detected in enhancer-suppressed cells compared with the wild-type cells, confirming that AHNAK was the target gene. In addition, the expression of AHNAK was significantly depressed in ELF1-knockdown preadipocytes purchased from BLUEFBIO (Shanghai, China) and significantly enhanced in preadipocytes overexpressing ELF1 (Figure 4E). Taken together, our results supported the idea that the TF ELF1 can bind to the rs7119797-C allele, resulting in increased AHNAK expression.

Functional validation of the regulation effect of rs8079062 on RNF157

rs8079062 acts as a SNP with allelic enhancer activity regulating RNF157 expression

We also chose the rs8079062-RNF157 pair for validation. The regulation relationship between rs8079062 and RNF157 was supported by both Hi-C and eQTL analysis. How the expression of RNF157 was regulated and whether RNF157 could affect the adipogenic differentiation process was still unknown.

The SNP rs8079062 located in the q25.1 region of chromosome 17 was detected as a SNP with allelic enhancer activity. Its A allele showed higher enhancer activity than the G allele (Table S9). Among the Hi-C datasets, rs8079062 and its targeted RNF157 are located within the same topologically associated domain of adipocytes (GEO: GSE151324).²² The interaction of rs8079062 and RNF157 is also observed in the capture Hi-C dataset from human fat tissue (GEO: GSE86189)³⁹ (Figure 5A). The eQTL analysis also supported the idea that there is a significant correlation between rs8079062 and RNF157 in both subcutaneous and omentum adipose (Table S11, p = 1.63 × 10⁻²¹ and 6.77 × 10⁻¹⁷, respectively). Using genotype and adipose expression data from the GTEx project (V8),³⁷ we observed that subjects with the “AA” genotype have higher RNF157 expression than other subjects (Figure S7).

Figure 5.

rs8079062 acts as a SNP with allelic enhancer activity regulating RNF157 expression

(A) Hi-C interactions between the RNF157 promoter and the enhancer where rs8079062 is located. Epigenetic annotation in adipose/adipocytes was visualized using the WashU EpiGenome Browser.

(B) Dual-luciferase reporter assay for the RNF157 promoter (RNF157-P) containing the region surrounding either rs8079062-G or rs8079062-A or for RNF157-P (baseline control) in preadipocytes. Error bars represent standard deviation.

(C) Effect of suppressing the region containing rs8079062 on RNF157 expression by CRISPR-dCas9 in preadipocytes purchased from BLUEFBIO (Shanghai, China). Error bars represent standard deviation.

(D) UBP1 motif binding the sequences surrounding rs8079062. ChIP-qPCR assay for the UBP1 binding with the rs8079062-A allele in preadipocytes purchased from BLUEFBIO (Shanghai, China). Error bars represent standard deviation.

(E) Effect of UBP1 overexpression on RNF157 expression (real-time qPCR and western blot). Error bars represent standard deviation.

(F) Effect of UBP1 knockdown on RNF157 expression (real-time qPCR and western blot). Three independent shRNAs (shRNA1, shRNA2, and shRNA3) and shRNA-NC (negative control) were used. Error bars represent standard deviation.

^∗p < 0.05, ^∗∗p < 0.01, and ^∗∗∗p < 0.001, paired Student’s t test.

Using the H3K27ac ChIP-seq data during the adipocyte differentiation process released by Rauch et al.,⁸³ we annotated rs8079062 and found that it is consistently in the peak of H3K27ac (Figure S8). We next compared the regulatory activity of genomic fragments containing different genotypes of rs8079062 using dual-luciferase reporter assays in preadipocytes purchased from BLUEFBIO (Shanghai, China). We found that both alleles of rs8079062 could drive the expression of RNF157 (p < 0.001, Figure 5B). In particular, the rs8079062-A allele had a significantly enhanced effect compared with the rs8079062-G allele (p < 0.01), which is consistent with the STARR-seq results. Genotyping assay showed that the preadipocyte is homozygous AA genotype for rs8079062 (Figure 5C). We used CRISPR-dCas9 to suppress the activity of this enhancer region, and significantly decreased RNF157 (p < 0.01, Figure 5C) was detected, indicating that RNF157 was the target gene of this enhancer. We also assessed whether the CRISPR-dCas9 experiment would affect the expression of genes neighboring RNF157. UBALD2 and FOXJ1 neighbor RNF157 upstream and downstream, respectively. FOXJ1 is barely expressed in adipose tissue.⁸⁴ As for UBALD2, no significant difference in its expression was observed after doing the CRISPR-dCas9 experiment (Figure S9).

UBP1 binds to rs8079062-A to regulate RNF157 expression

We next explored the functional mechanism underlying rs8079062 as a SNP with allelic enhancer activity regulating RNF157. Among the TFs predicted to be specifically binding to the A allele (the allele with higher enhancer activity) of rs8079062, UBP1 showed the highest expression level in adipose tissue. Therefore, we chose UBP1 for further validation. The motif analysis result showing that UBP1 specifically binds to the rs8079062-A allele is illustrated in Figure 5D. ChIP-qPCR confirmed that UBP1 was recruited by the rs8079062-A allele (Figure 5D). In addition, the expression of RNF157 was significantly enhanced in preadipocytes overexpressing UBP1 (Figure 5E). Next, we suppressed the expression of UBP1 by three independent shRNAs in preadipocytes, and a significant decline in RNF157 expression was detected at both RNA and protein levels (Figure 5F). The above results support the idea that the TF UBP1 can bind the A allele of rs8079062, which elevates the enhancer activity and increases RNF157 expression.

Abnormal expression of Rnf157 affects the adipogenic differentiation process

To investigate the role of Rnf157 in the adipogenic differentiation process, we first induced the adipogenic differentiation of 3T3-L1 cells for 8 days. We detected decreased expression of Rnf157 in the adipogenic differentiation process (Figure 6A). Therefore, we hypothesized that abnormal expression of Rnf157 might affect adipogenic differentiation. To test our hypothesis, we first overexpressed this gene in 3T3-L1 cells using the lentivirus overexpression assay. As shown in Figures 6B and 6C, during the adipogenic differentiation process, the adipogenic differentiation marker genes (Cebpa and Pparg) were significantly downregulated in cells with overexpressed Rnf157. In contrast, when the expression of Rnf157 was knocked down using shRNA plasmids, the expression of Cebpa and Pparg was promoted (Figures 6D and 6E).

Figure 6.

Abnormal expression of Rnf157 affects the adipogenic differentiation process

(A) The mRNA and protein levels of Rnf157 and adipogenic differentiation marker genes (Cebpa and Pparg) in the normal adipogenic differentiation process of 3T3-L1 cells (days 0, 2, 4, and 8). β-actin was used as loading control. Error bars represent standard deviation.

(B) Effect of Rnf157 overexpression on mRNA levels of Cebpa and Pparg during 3T3-L1 adipogenic differentiation. The lentiviral empty vector was used as the negative control. Error bars represent standard deviation.

(C) Effect of Rnf157 overexpression on protein levels of Cebpa and Pparg and oil red O staining on day 8 of adipogenic differentiation. β-actin was used as loading control for western blotting.

(D) Knockdown of Rnf157 increased the mRNA levels of Cebpa and Pparg during 3T3-L1 adipogenic differentiation (days 0, 2, 4, and 8). Error bars represent standard deviation.

(E) Knockdown of Rnf157 increased the protein levels of Cebpa and Pparg and oil red O staining on day 8 of adipogenic differentiation. β-actin was used as loading control for western blotting.

(F) A schematic representation elucidating how rs8079062 affects the expression of RNF157 and how RNF157 might be involved in the development of central obesity.

^∗p < 0.05, ^∗∗p < 0.01, and ^∗∗∗p < 0.001; n,: not significant, paired Student’s t test.

The expression of RNF157 in adipose tissue is associated with central obesity traits

As shown in Figure 3D, in the dataset of 770 METSIM individuals, the expression of RNF157 is also negatively associated with central obesity and triglyceride, HOMA-β, insulin, and proinsulin levels. Meanwhile, the expression of RNF157 is positively associated with HDL-C, adiponectin, and Matsuda index (Table S13). Consistently, we found that the expression of RNF157 in visceral adipose tissue from diabetic subjects was significantly lower than in heathy controls (GEO: GSE16415, Figure S10). In addition, the expression level of Rnf157 in mesenteric adipose tissue from mice fed a high-fat diet over 8 weeks was significantly lower than in mice fed a normal diet (GEO: GSE39549, Figure S10). For diet-induced obese rats, the expression level of Rnf157 was significantly increased in subcutaneous adipose compared with the rats losing weight successfully after Roux-en-Y gastric bypass surgery (GEO: GSE8314, Figure S10). From these results taken together, a potential regulatory model between rs8079062 and RNF157 is shown in Figure 6F.

Discussion

GWASs have identified hundreds of susceptibility loci associated with central obesity. However, most of these loci are located in non-coding regions, and each locus contains many plausibly causal variants due to LD. Here, after fine-mapping and epigenomic annotation, we selected 2,034 prioritized SNPs to evaluate their enhancer activity through STARR-seq. The results showed that 38.89% of the SNPs examined in this study have enhancer activity in human preadipocytes. Furthermore, 141 SNPs at 94 loci have allelic enhancer activity. These findings supported the theory that genetic risk variants might be involved in the development of central obesity though transcriptional perturbation of critical target genes.

Among the target genes of SNPs with allelic enhancer activity, the adipose tissue expression levels of over 50% of the genes were associated with central obesity. Further MR analysis using plasma pQTL data supported MFAP2, TMEM132C, and NID2 as potential drug targets for central obesity. Protein encoded by MFAP2 is a component of extracellular matrix microfibrils. Consistent with our findings, a previous study⁸⁵ has observed elevated MFAP2 expression levels in the white adipose tissue of obese human subjects. Similarly, they also observed that the Mfap2 transcript was significantly increased in the adipose tissue of obese mice fed a high-fat diet. However, global knockout of Mfap2 resulted in adipocyte hypertrophy and predisposition to metabolic dysfunction.⁸⁵ Therefore, keeping the expression of MFAP2 at an appropriate level is important to reduce the risk of obesity. As for TMEM132C and NID2, studies about their involvement in the development of central obesity are still limited. Further research is warranted to decipher the mechanism through which they protect against central obesity.

Another finding of this study is that SNPs with allelic enhancer activity are likely to alter the binding of many TFs. Enrichment analysis identified 20 key TFs in the regulatory network. 80% (16/20) of these TFs have been proved to be involved in obesity-related biological processes, with five exceptions (ELF1, ZFX, THAP1, TCF12, and ETV6). We selectively validated the allelic-specific binding of ELF1 to rs7119797 and further showed that it regulates the expression of a known obesity gene, AHNAK.^65,66 ELF1 encodes an E26 transformation-specific related TF. Previous studies^86,87 for this gene mainly focus on its roles in cancers and the maintenance of the hematopoietic system. Our results showed that in preadipocytes, overexpression of ELF1 could enhance the expression of AHNAK, while ELF1 knockdown resulted in significant decline of AHNAK expression, supporting its involvement in the development of obesity.

We also demonstrated that rs8079062 acted as a SNP with allelic enhancer activity to regulate the expression of RNF157. The importance of RNF157 in central obesity is highlighted by the fact that we observed that its expression levels in adipose tissue correlated with central obesity. In addition, we also observed that Rnf157 was downregulated in the adipogenic differentiation process of 3T3-L1 cells. According to information from the IMPC mouse database, global Rnf157-knockout (Rnf157^−/−) mice showed increased circulating amylase levels (MGI: 2442484). The main function of amylase is to cleave starch into smaller polysaccharides in the process of digestion.⁸⁸ Elevated serum amylase level is typically a sign of acute pancreatitis. As a key enzyme for starch metabolism, mammalian α-amylase has been considered as a therapeutic target for obesity-related complications. The α-amylase inhibitor acarbose has been used as a prescription drug to tackle hyperglycemia in the treatment of T2D.⁸⁹ Several clinical studies^89,90 have shown that acarbose could improve insulin sensitivity in individuals with T2D and reduce cardiovascular risks in subjects with metabolic syndrome. Therefore, targeting RNF157 may be a potential therapeutic strategy for treating obesity. Although we observed lower RNF157 expression in obese mice and people with diabetes compared to healthy individuals, Kosaka et al.⁹¹ reported higher RNF157 expression in both visceral and subcutaneous adipose tissue of Wistar Ottawa Karlsburg W (RT1u) (WOKW) rats. This discrepancy in direction might be because the WOKW rat model is different from the high-fat-induced mouse model. With the MHC RT1u haplotype, WOKW rats only develop polygenetically inherited metabolic syndrome, and they do not develop any age-related metabolic syndrome complications such as diabetes or neuropathy. According to the results of Kosaka et al.,⁹¹ another potential pathway in which RNF157 may play a role is regulating autophagy to prevent the development of obesity-related metabolic complications (e.g., neuropathy or diabetes).⁹¹

Limitations of this study should be addressed. First, in the STARR-seq experiment, we validated the activity of only the 2,034 prioritized SNPs located in enhancer/open chromatin regions. That is, we did not estimate the full scope of possible functional variants. However, the prioritized SNPs we selected have higher conservation scores and higher proportions overlapped with Hi-C interaction anchors, suggesting that the prioritized SNPs are more likely to play important regulatory roles than other SNPs. Second, since short fragments are more suited to fine-mapping the enhancer effects of individual TF binding sites,³¹ we chose 120-bp fragments for STARR-seq and subsequent dual-luciferase reporter assay to fit our study’s objective. Given that the enhancer activity of identical enhancers can vary depending on the length of their flanking sequences,⁹² we might not be able to detect enhancers that function in longer fragments. Further studies with longer fragments might provide new insights for the function of the prioritized SNPs. Third, the regulatory model we proposed for RNF157 could not exclude the contribution of other SNPs to its expression, but instead illustrates the regulatory mechanism of rs8079062. Last, the GWAS summary dataset we used was derived from European populations. Therefore, the prioritized SNPs in our study might not be related to central obesity in other populations.

In summary, using STARR-seq, we systematically uncovered the regulatory activity of 2,034 prioritized SNPs associated with central obesity in adipocytes. Our integrative analyses reveal specific molecular mechanisms underlying allelic transcriptional regulation of important target genes. We hope that our findings can help accelerate the translation of discoveries from GWASs into insights of biological and clinical value, thereby enhancing our understanding of the mechanisms underlying the development of central obesity.

Data and code availability

STARR-seq data have been submitted to GEO and are publicly available as of the date of publication (GEO: GSE271195).

Acknowledgments

This study is supported by the National Natural Science Foundation of China (32070588, 32370653, and 32100416), Innovation Capability Support Program of Shaanxi Province (2022TD-44), Key Research and Development Project of Shaanxi Province (2022GXLH-01-22), Fundamental Research Funds for the Central Universities, and Opening Research Fund from the Key Laboratory of Shaanxi Province for Craniofacial Precision Medicine Research (College of Stomatology, Xi'an Jiaotong University, 2024LHM-KFKT006). No funding bodies had any role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. This study is also supported by the High-Performance Computing Platform and Instrument Analysis Center of Xi’an Jiaotong University.

Author contributions

T.-L.Y. and S.-S.D. designed the study. S.-S.D., Y.-Y.D., and Y.G. wrote and edited the manuscript. S.-S.D., R.-J.Z., S.-H.T., K.Y., W.S., X.-F.C., F.J., and R.-H.H. collected and analyzed the data. Y.-Y.D., Y.-Y.J., J.-X.C., and X.-T.H. performed the functional experiments. T.-L.Y., Y.L., and Z.L. reviewed the manuscript. T.-L.Y. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.

Declaration of interests

The authors declare no competing interests.

Published: January 2, 2025

Web resources

• Central obesity GWAS summary data, https://zenodo.org/record/1251813#.XVYC3OgzZPZ

• GTEx eQTL summary and gene expression data, https://www.gtexportal.org/home/

• METSIM subcutaneous adipose eQTL summary data, https://mohlke.web.unc.edu/data/metsim-adipose-tissue-cis-eqtl-summary-statistics/