https://orcid.org/0000-0002-0703-0742
Abstract
This study aims to identify BMI-associated genes by integrating aggregated summary information from different omics data.
We conducted a meta-analysis to leverage information from a genome-wide association study (n = 339 224), a transcriptome-wide association study (n = 5619), and an epigenome-wide association study (n = 3743). We prioritized the significant genes with a machine learning-based method, netWAS, which borrows information from adipose tissue-specific interaction networks. We also used the brain-specific network in netWAS to investigate genes potentially involved in brain-adipose interaction.
We identified 195 genes that were significantly associated with BMI through meta-analysis. The netWAS analysis narrowed down the list to 21 genes in adipose tissue. Among these 21 genes, six genes, including FUS, STX4, CCNT2, FUBP1, NDUFS3, and RAPSN, were not reported to be BMI-associated in PubMed or GWAS Catalog. We also identified 11 genes that were significantly associated with BMI in both adipose and whole brain tissues.
This study integrated three types of omics data and identified a group of genes that have not previously been reported to be associated with BMI. This strategy could provide new insights for future studies to identify molecular mechanisms contributing to BMI regulation.
Introduction
The prevalence of obesity has increased globally in the past several decades, especially in developed countries. According to the World Health Organization (https://www.who.int/news-room/facts-in-pictures/detail/6-facts-on-obesity), the prevalence of obesity has almost tripled since 1975, and an estimated 2.8 million people die yearly from obesity or overweight-related diseases. Excess weight impacts both physical and mental health and may lead to specific diseases like type 2 diabetes, hypertension, and several forms of cancers [1]. Identifying obesity-associated factors will enhance our understanding of the mechanism underlying obesity and may provide potential intervention targets for disease prevention. Many large genome-wide association studies (GWAS) have identified genomic loci associated with body mass index (BMI), a commonly used measurement of adiposity. However, there is still considerable unexplained heritability, and more studies of the molecular contributions to BMI are warranted.
Due to recent advances in high-throughput technology, different kinds of omics data have become increasingly popular and aided in identifying BMI-associated genes. Epigenome-wide Association Studies (EWAS) associate epigenetic markers with phenotypes of interest [2]. Transcriptome-wide association analysis (TWAS) aims to identify genes associated with complex human traits due to genetically regulated transcriptional activities using gene expression data [3]. Analyzing different types of omics data provides alternative approaches to gain insight into the underlying biological mechanism of various traits, specifically BMI in our study.
An integrative omics approach offers an opportunity to improve our understanding of obesity by applying a multidimensional strategy. In this study, we borrowed the idea from a previously published article and implemented the concept of omics data integration in the context of BMI [4]. We hypothesized that leveraging information from different omics data types would reveal further insights beyond the analysis of single omics platforms. Our goal is to identify BMI-associated genes using an integrative approach with aggregated summary data from genetic, transcriptomic, and epigenetic data.
Results
In our study, we first collapsed the summary statistics from three omics studies to the gene level. Then, we performed a meta-analysis using three omics data to identify genes significantly associated with BMI. Finally, we performed tissue-specific gene prioritization using the Network-wide Association Study (netWAS) [5]. We presented the workflow in Fig. 1. The details of the analysis are presented in Materials and Methods.
Flowchart of the multi-omics data integration. Results from the GWAS and EWAS were first collapsed into genes. Meta-analysis was then performed to identify genes that are significantly associated with BMI. netWAS was conducted to prioritize BMI-associated genes. Alt text: Flowchart illustrating the integration process of GWAS (genome-wide association study), EWAS (epigenome-wide association study), and TWAS (transcriptome-wide association study) for the identification of tissue-specific genes associated with BMI (body mass index).
Omics data and study characteristics
We first compared aggregated data with previously reported results separately obtained from GWAS and EWAS. The GWAS aggregation analysis yielded 23 935 genes, and 149 of them were significant (P-value < 2.09 × 10−6) (Supplementary Table 1), while the previous GWAS research on BMI discovered 97 BMI-associated loci. 19 822 genes had available p-values after methylation association results were collapsed at the gene level, and 82 of them were significant (P-value < |$2.5\times{10}^{-6}$|) (Supplementary Table 2), while the previous EWAS study reported 85 BMI-associated genes based on differentially methylated CpG sites. By performing TWAS with 17 318 transcripts, we found 2962 genes significant (P-value < |$0.05/\mathrm{17,318}=2.9\times{10}^{-6}$|) (Supplementary Table 3).
Meta-analysis of GWAS, EWAS, and TWAS and enrichment analysis
We retained 14 994 genes with available information in all three omics data, and 195 of them were significantly associated with BMI in the meta-analysis (P < |$0.05/\mathrm{14,994}=3.3\times{10}^{-6}$|) (Supplementary Table 4), and 125 of them were documented to be BMI-associated according to the GWAS Catalog and PubMed (Supplementary Table 5). Among the top ten GO terms based on the identified 195 genes, six (GO:0051402; GO:0007274; GO:0007212; GO:0070997; GO:0048483; GO:0043525) were associated with brain functions (Supplementary Table 6).
Adipose-specific and brain-specific BMI-associated genes
We implemented netWAS to narrow the gene list of 195 from the meta-analysis to 21 adipose-specific BMI-associated genes (Fig. 2 and Supplementary Table 7). Among these 21 genes, 15 have been associated with BMI in multiple association studies (SULT1A1 [6], CALCR [6], INO80E [7], SHISA4 [8], GTF3A [9], NCAM1 [10], DANSE1 [11], MADD [12], ZNF12, DENND1A [13], MAP2K3 [6], ADCY9 [14], GNB1 [10], SULT1A2 [15], and TAOK2 [11]). The remaining six have not been associated with BMI based on our search of the GWAS Catalog and PubMed (FUS, STX4, CCNT2, FUBP1, NDUFS3, and RAPSN). We also leveraged brain tissue-specific network information through netWAS and identified 53 BMI-associated genes specified in the whole brain (Supplementary Table 8). These two gene lists based on adiposity and brain tissue-specific networks shared 11 genes (MYL4NCAM1, FUS, SULT1A1, SULT1A2, MADD, GTF3A, STX4, GNB1, CALCR, NDUFS3, FUBP1).
BMI-related genes identified with netWAS. In adipose tissue, 21 genes were identified as BMI-associated, and six have not been reported to be BMI-associated in the GWAS Catalog and PubMed. Fifty-three genes were identified as BMI-associated in the whole brain tissue. Besides, 11 genes are significantly associated with BMI in both adipose tissue and whole brain tissue. Alt text: Vien-diagram illustrating the genes associated with BMI identified in this study, indicating their presence in the previous publications.
Discussion
We performed a meta-analysis of summary statistics from three BMI-related omics studies and found 195 genes that were significantly associated with BMI. We further applied netWAS and narrowed our meta-analysis results to 21 BMI-associated genes specified in adipose tissue, including six genes whose associations with BMI were not reported in PubMed or the GWAS Catalog. Over-representative analysis of the 195 genes suggested that the top GO terms tend to be brain-related. By incorporating the brain-specific network with netWAS, 53 genes were significantly associated with BMI in the whole brain tissue, and 11 were also identified as BMI-associated in adipose tissue.
As a confirmation of our results, we compared our gene list with the results from the GWAS we used for analysis [10] and another expanded GWAS on BMI [16], which additionally included ~450 000 individuals. We mapped significant SNPs (97 in the GIANT 2015 and 941 in the GIANT 2018) to genes where they are located for comparison. Among the 195 genes in our results, 34 are among the list obtained from GIANT 2015 summary data [10]. Twenty of the remaining 63 genes are a subset of the gene list from the second paper [16]. This result further indicates that integrating information from different omics data could increase our power for gene identification. We also used some known monogenic obesity genes as true positives [17], and five among our 195 identified BMI-associated genes overlapped with the reported 14 monogenic obesity genes, which are ADCY3, BDNF, MC4R, POMC, and SH2B1. SH2BI is also one of the prioritized genes in brain-tissue-specific analysis.
With adipose tissue-specific networks, our prioritized list of 21 genes confirmed 15 previously reported BMI-associated genes, and some of them were also demonstrated in different biological experiments. For example, SULT1A1 and SULT1A2 are sulfotransferase genes from the SULT family responsible for the sulfation of many exogenous compounds. 17 beta-estradiol (E2) is a medium-size substrate of SULT1A1 [18] and SULT1A2, and the increase in E2 level is associated with obesity [19]. SULT1A1 is also a key enzyme in adipocytes responsible for the biotransformation of resveratrol (an obesity treatment). Researchers found that SULT1A1 was upregulated upon adipocyte differentiation, and knocking down SULT1A1 had an antilipolytic effect in human adipocytes by treatment of resveratrol [20]. For CALCR, hyperphagic obesity in mice was caused by the disruption of direct leptin action on CALAR-expressing leptin receptor neurons [21]. Another animal experiment showed that CALCR-containing neurons in nucleus tractus solitarius (NTS) mediate a non-aversive suppression of food intake and silencing CALCR-containing neurons in NTS blunted food intake suppression by gut peptides and nutrients, leading to increased food intake and obesity [22].
While the remaining six genes of the 21, including FUS, STX4, CCNT2, FUBP1, NDUFS3, and RAPSN, have not been reported to be BMI-associated in GWAS Catalog and PubMed, some of them were associated with other metabolic traits. STX4 was associated with multiple metabolic traits like triglyceride levels [23], systolic blood pressure [13], and body fat percentage [24]. FUBP1 [25], NDUFS3 [26], and RAPSN [6] have been associated with waist-hip ratios. One study suggested that silencing CCNT2 in human adipocytes may decrease leptin secretion and mRNA expression of several genes involved in adipogenesis [27].
The result of the over-representative analysis implied the potential connection between BMI-associated genes and brain function, which was not merely a coincidence. Brain-adipose or brain-obesity crosstalk has long attracted attention. A previous study suggested that the brain affects adipose tissue metabolism through the regulation of leptin, while the adipose tissues, in turn, inform the brain of energy storage through hormones [28, 29]. Multiple studies suggested the relationship between the overlapped genes and brain functions or diseases. For example, MADD played an important role in Ca2 +− dependent neurotransmitter release and exocytosis [30], and was also associated with Alzheimer’s disease [31]. NCAM1 was identified as an extracellular vesicle for the excitatory neuron and might involve neurodegenerative disease progression [32]. Some other genes were identified to associate with brain-related diseases, such as the progression rate of Parkinson’s (MYL4) [33] and primary brain tumors (SULT1A1) [34]. Future studies could focus on the role of these genes in brain-adipose interaction.
Our study used the genome-wide association results for each omics data, rather than pre-filtering each summary statistic to get significant units first and then to focus on only those significant units during the integration process. Our approach helped preserve the integrity of all potential signals that may be identified by integration analysis. Moreover, our study incorporated adipose and brain tissue-specific information in the analysis and gained insight into possible brain-adipose interactions. It could be fundamental in interpreting the underlying biological pathways that affect BMI with its most relevant tissues or the effect of BMI on other issues through the pathways. Our study has several limitations. First, this study was restricted to European ancestry due to the availability of EWAS and TWAS data. Thus, the genes we found may need further validation in the sample of other ancestries. Second, the multi-omics data in our analysis are from whole blood due to limited access to data from other tissues. Future studies could consider using omics data from more adiposity-relevant tissues such as adipose, brain, etc. Third, as a proof of concepts demonstrating the utilization of the integrative approach of multi-omics data, we did not consider the potential correlation between summary statistics across different omics data in a meta-analysis, which may lead to inflated type-I errors. Last, the way we mapped the summary statistics to gene level might not be accurate. For example, we used the minimal p-value approach to aggregate EWAS summary statistics from the GpG site level to the gene level, but we didn’t consider the case in which multiple CpG sites affect gene expression cooperatively. While encouraged by the support of many previously published works for our results, there are still some genes whose biological associations with BMI have not yet been revealed. Further experimental validation is warranted for elucidating their roles in BMI regulation.
In conclusion, we leveraged summary statistics from different omics data and identified multiple BMI-associated genes, including genes not associated with BMI in the GWAS Catalog or PubMed. We further refined our list of genes potentially involved in adipose-brain interaction. These findings could be the potential focus of future obesity and obesity-brain interaction studies.
Materials and methods
Omics studies of BMI
We used aggregated summary results from three previously published association studies of BMI: GWAS [10], EWAS [35], and TWAS [36]. The workflow was presented in Fig. 1, which was referenced from an integrative omics approach published in 2020 [4].
For the GWAS data, we used the summary statistics from a large-scale meta-analysis based on the Genetic Investigation of ANthropometric Traits (GIANT) consortium [10]. It had a sample size of 339 224 and was European ancestry-specific.
The EWAS summary statistics were obtained from the preliminary work of a Mendelian randomization study on the causal relationships between methylation probes and BMI [35]. The analyses were based on 3743 participants from two cohorts: the Framingham Heart Study (FHS) and the Lothian Birth Cohorts (LBC). Each cohort performed the EWAS separately, and then the results were meta-analyzed using methods that weighted the p-value by sample size [37].
The TWAS results were based on an analysis of array-based gene expression data from FHS. Specifically, it included 2440 participants from the FHS Offspring cohort [38] and 3179 participants from the FHS Generation 3 cohort [39]. The previously published work provided the details of the analysis [36]. In brief, whole blood gene expression was measured by the Affymetrix Human Exon 1.0st Array. We conducted linear mixed-effects models with a random effect specified by the kinship matrix to account for familial relatedness. We used gene expression level as a dependent variable and BMI as the primary predictor, adjusting for age at blood draw, sex, cohort, and technical covariates.
Aggregating to gene-level using omics data
We converted all summary statistics to the gene level before integration. We aggregated GWAS summary statistics from single nucleotide polymorphisms (SNPs) to gene level using GCTA fastBAT software [40]. Under the null hypothesis of no association, fastBAT models the Z-statistics of the SNPs within one gene region as a multivariable normal distribution with a mean of zero and linkage disequilibrium (LD) matrix as the correlation matrix. The LD information came from the whole genome sequencing data of 4165 FHS subjects participating in Freeze 4 of the Trans-Omics for Precision Medicine (TOPMed) program. For aggregating EWAS summary statistics from the CpG site level to the gene level, we used the adjusted minimal p-value approach [41], where |${P}_{adjust}=1-{\left(1-{P}_{min}\right)}^L$|. |$L$| stands for the number of CpG sites mapped to the same gene, and |${P}_{min}$| stands for the minimal P-value among the CpG sites mapped to the gene. For the case that one CpG site may map to multiple genes based on Illumina 450 k annotation, it would be included in |${P}_{adjust}$| calculations of all the genes it mapped to. In our TWAS result, we assigned the minimum p-value to a gene when more than one transcript was mapped to it.
Meta-analysis
After generating gene-based aggregation statistics for GWAS, EWAS, and TWAS, we conducted a meta-analysis using a fixed-effects model based on the gene-based association p-value, ignoring the direction of associations, and all the genes with complete P-values from three omics studies were included in the meta-analysis. As shown in the formula below, |${Z}_i$| and |${N}_i$| are the z-scores and the number of individuals from different omics studies, and the z-scores are meta-analyzed with weights of the square root of each omics study’s sample size. We reported significant genes from a meta-analysis using a Bonferroni-corrected threshold.
Over-representative analysis
Functional enrichment analysis plays an important role in aiding the interpretation of high-throughput omics data analysis. We used a web tool, WebGestalt (WEB-based Gene Set Analysis Toolkit), for functional enrichment analysis [42]. It covers functional categories from various databases and allows us to perform Over Representation Analysis with the Gene Ontology (GO) biological process database to gain an overall understanding of the genes significantly associated with BMI in our meta-analysis. We used “affy huex 10 st vs” as the reference set and focused on the top ten GO terms.
Tissue-specific gene prioritization using netWAS
We further performed the Network-wide Association Study (netWAS) [5] to prioritize the BMI-associated genes. The resulting list of genes from NetWAS would be a subset of the significant BMI-associated genes in our meta-analysis. Given that the precise actions of genes frequently depend on their tissue context, comprehending the gene activities within distinct tissues is of great importance. NetWAS combines gene-level summary statistics from meta-analysis and tissue-specific networks established based on 14 000 distinct publications to identify outcome-gene association more accurately. Specifically, genes with P-values less than 0.05/the number of genes (n = 14 994) were considered “pre-positive.” Those genes with a positive value in the netWAS score calculated by the support vector machine (SVM) were considered “post-positive.” The value of netWAS score indicates the distance from the SVM separating the hyperplane in the positive direction. Genes with both pre-positive and post-positive were considered BMI-associated genes. With netWAS, we generated a list of BMI-associated genes specified in adipose tissue.
We also generated a list of BMI-associated genes specified in brain tissue using netWAS. Some previous studies reported the connection between BMI and cognitive functioning in brain tissue [43] or suggested a brain-adipose interaction [28, 44]. We also estimated the genetic correlation between BMI and Alzheimer’s disease using LD score regression (LDSC) analysis [45, 46] with previously published data [10, 47]. The result showed a significant genetic correlation of −0.2644 with a P-value of 0.0066 between BMI and Alzheimer’s disease. Therefore, we further utilized the brain-specific network in netWAS to investigate the genes potentially associated with brain-adipose interaction.
We also performed NetWAS with various tissues (blood tissues, skeleton muscle tissues, liver and pancreas tissues). The related results are presented in Supplementary Tables 9–13. To streamline the discussion, we only focus on the results from adipose and brain tissues.
Conflict of interest statement
No conflict of interest is claimed for all authors.
Funding
This research was partly supported by grants NIDDK R01DK122503 and by NHLBI contracts N01-HC-25195 and HHSN268201500001I.
References
Author notes
Jingxian Tang and Hanfei Xu joint first author
