Genetic Regulation of Adipose Gene Expression and Cardio-Metabolic Traits

Abstract

Subcutaneous adipose tissue stores excess lipids and maintains energy balance. We performed expression quantitative trait locus (eQTL) analyses by using abdominal subcutaneous adipose tissue of 770 extensively phenotyped participants of the METSIM study. We identified cis-eQTLs for 12,400 genes at a 1% false-discovery rate. Among an approximately 680 known genome-wide association study (GWAS) loci for cardio-metabolic traits, we identified 140 coincident cis-eQTLs at 109 GWAS loci, including 93 eQTLs not previously described. At 49 of these 140 eQTLs, gene expression was nominally associated (p < 0.05) with levels of the GWAS trait. The size of our dataset enabled identification of five loci associated (p < 5 × 10⁻⁸) with at least five genes located >5 Mb away. These trans-eQTL signals confirmed and extended the previously reported KLF14-mediated network to 55 target genes, validated the CIITA regulation of class II MHC genes, and identified ZNF800 as a candidate master regulator. Finally, we observed similar expression-clinical trait correlations of genes associated with GWAS loci in both humans and a panel of genetically diverse mice. These results provide candidate genes for further investigation of their potential roles in adipose biology and in regulating cardio-metabolic traits.

Introduction

Genome-wide association studies (GWASs) have identified many loci for complex metabolic and cardiovascular traits, yet the underlying genes and mechanisms by which they affect disease remain poorly characterized.^{1, 2} The genetic analysis of gene expression by identification of expression quantitative trait loci (eQTLs) in relevant tissues has proven useful to predict candidate genes at GWAS loci and biological pathways that are perturbed in affected individuals.^{3, 4, 5, 6} Subcutaneous adipose tissue serves as a buffering system for lipid energy balance, particularly fatty acids,^{7, 8} and might play a protective role in metabolic and cardiovascular disease risk.⁹

Subcutaneous adipose eQTL studies have implicated genes involved in obesity and metabolic traits.^{10, 11, 12, 13} Recent GWASs for type 2 diabetes (T2D), cholesterol and triglyceride levels, body mass index, waist-hip ratio, and adiponectin have reported subcutaneous adipose eQTLs that are coincident with specific GWAS loci.^{14, 15, 16, 17} Similarly, a recent large GWAS for waist-hip ratio identified loci that were enriched for genes expressed in subcutaneous adipose tissue and for putative regulatory elements in adipocyte nuclei.¹⁶ Many GWAS loci for these traits do not yet have clear candidate genes, in part because of the limited statistical power of existing eQTL studies.

trans-eQTL associations between variants and transcripts located far from each other or on different chromosomes can identify downstream disease genes, including those not implicated by GWASs. Identifying these distant relationships is difficult because of the multiple testing burden in humans. Studies of natural variation in mice have identified a number of “hotspot” loci associated with trans regulation of genes and clinical traits.^{18, 19} One of the first reported human trans-acting eQTLs involved the KLF14 transcription factor in adipose tissue.²⁰ The locus associated with T2D and HDL-cholesterol levels showed a cis-acting association with expression of KLF14 and ten distal genes.²⁰ Studies of gene expression in circulating monocytes or whole-blood cells have also provided evidence of trans regulation of gene expression with linkage to traits relevant to lipid metabolism, type 1 diabetes, hypertension, celiac disease, and cancer.^{21, 22}

In light of the widespread use of mice to help validate and gain mechanistic understanding of genes in GWAS loci, commonalities and differences in the regulatory networks of mice and humans are of clear relevance for studies of common diseases.⁶ A previous comparison between mouse and human adipose transcriptional networks detected a shared core-network module enriched for genes involved in the inflammatory and immune response causally associated with obesity-related traits.¹⁰ Recent analysis of DNase I hypersensitive sites and occupancy profiles of transcription-factor binding in humans and mice suggests the preservation of similar regulatory mechanisms for adipose gene expression in both species, and this preservation could be leveraged to inform disease pathways.^{23, 24}

We describe here the analysis of gene expression in 770 subcutaneous adipose samples from Metabolic Syndrome in Men (METSIM), a study of 10,197 men, 45–73 years of age, living in the Kuopio area of Finland. Study data include dense genotypes and extensive metabolic and cardiovascular traits, such as plasma lipids, inflammatory markers, glycemic traits, and anthropometric traits.²⁵ We identified cis-eQTLs at GWAS loci for cardio-metabolic traits, as well as trans-eQTL hotspots with at least five distant target genes. We also identified associations of gene expression with human clinical traits and compared our results with those of analogous studies in a set of 120 inbred strains of mice.

Subjects and Methods

METSIM Study Participants and Sample Characteristics

We analyzed samples from 770 males who are part of the METSIM study.²⁵ The Ethics Committee of the Northern Savo Hospital District approved this study, and all participants gave written informed consent. The population-based METSIM study included 10,197 men, aged 45–73 years and randomly selected from the population register of Kuopio town in eastern Finland (population 95,000). Every participant had a 1 day outpatient visit to the Clinical Research Unit at the University of Kuopio, including an interview on the history of previous diseases and current drug treatment and an evaluation of glucose tolerance and cardiovascular risk factors. After 12 hr of fasting, a 2 hr oral 75 g glucose tolerance test was performed, and the blood samples were drawn at 0, 30, and 120 min. Plasma glucose was measured by enzymatic hexokinase photometric assay (Konelab Systems reagents; Thermo Fischer Scientific), and insulin and pro-insulin were determined by immunoassay (ADVIA Centaur Insulin IRI no. 02230141; Siemens Medical Solutions Diagnostics). Evaluation of insulin sensitivity (Matsuda index) and insulin secretion (calculated from area under the curve between 0 and 30 min of glucose tolerance test with the formula InsAUC_0–30/GlucAUC_0–30) have been previously described.^{25, 26} Plasma levels of lipids were determined via enzymatic colorimetric methods (Konelab System reagents, Thermo Fisher Scientific). Plasma adiponectin was measured with the Human Adiponectin Elisa Kit (Linco Research), C-reactive protein with high sensitive assay (Roche Diagnostics GmbH, Mannheim, Germany), and interleukin 1 receptor agonist with immunoassay (ELISA, Quantikine DRA00 Human IL-1RA, R&D Systems). Serum creatinine was measured by the Jaffe kinetic method (Konelab System reagents, Thermo Fisher Scientific) and was used for calculating the glomerular filtration rate. Height and weight were measured to the nearest 0.5 cm and 0.1 kg, respectively. Waist circumference (at the midpoint between the lateral iliac crest and lowest rib) and hip circumference (at the level of the trochanter major) were measured to the nearest 0.5 cm. Body composition was determined by bioelectrical impedance (RJL Systems) in participants in the supine position. The characteristics of the study participants are shown in Table S1. 770 participants were recruited for adipose-tissue needle biopsies. 61 participants were diagnosed with impaired glucose tolerance, and 27 participants had newly diagnosed type 2 diabetes at the time of the tissue collection.

Genotyping and Imputation

Genotyping of METSIM samples was performed with the Illumina HumanOmniExpress BeadChip array and the Illumina HumanCoreExome at the Center for Inherited Disease Research. Markers with poor mapping, no founder genotypes, call rate < 95%, deviation from Hardy-Weinberg equilibrium (p < 10⁻⁶), or more than two alleles were removed from subsequent imputation. We carried out genotype imputation of the 681,789 directly genotyped variants that passed quality control by applying the Markov Chain Haplotyping algorithm (MaCH) and the reference panel from the Haplotype Reference Consortium (see Web Resources). After imputation, variants were filtered on the basis of imputation quality (MaCH r² > 0.3) and minor-allele frequency (MAF ≥ 0.01). 7,677,146 variants were retained for subsequent analysis.

Gene Expression Profiling

Total RNA from METSIM participants was isolated from adipose tissue via the QIAGEN miRNeasy kit, according to the manufacturer’s instructions. RNA integrity numbers (RINs) were assessed with the Agilent Bioanalyzer 2100 instrument, and 770 samples with RIN > 7.0 were used for transcriptional profiling. Expression profiling with the Affymetrix U219 microarray was performed at the Department of Applied Genomics at Bristol-Myers Squibb according to the manufacturer’s protocols. The probe sequences were re-annotated so that probes that mapped to multiple locations, contained variants with MAF > 0.01 in the 1000 Genomes Project European samples, or did not map to known transcripts on the basis of the RefSeq (version 59) and Ensembl (version 72) databases were removed; 6,199 probe sets were removed in this filtering step. For subsequent analyses, we used 43,145 probe sets that represent 18,155 unique genes. The microarray image data were processed with the Affymetrix GCOS algorithm via the robust multiarray average (RMA) method for determination of the specific hybridizing signal for each gene.

PEER Factor Analysis

We applied the probabilistic estimation of expression residuals (PEER) method²⁷ to infer and account for complex non-genetic factors affecting gene expression levels. This method is designed to detect the maximum number of cis-eQTLs. To optimize the discovery of trans-eQTLs within the same analysis, we performed PEER analysis by examining 10–50 inferred factors (Nk) at increments of five factors. We then used Matrix eQTL²⁸ to assess the genetic association with inverse normal-transformed PEER-processed residuals from RMA-normalized expression data. The numbers of cis- and trans-eQTLs obtained at different Nk levels are shown in Table S2. We examined variants on chromosome 7, including rs4731702 at the known master regulator KLF14,²⁰ to determine the number of trans-eQTL target genes by using various numbers of PEER factors. We selected Nk = 35 as a single analysis to maximize the number of target genes at this locus; this threshold captured the 94.8% of cis-eQTLs identified with 50 PEER factors. For downstream eQTL mapping, we used the inverse normal-transformed PEER-processed residuals after accounting for 35 factors.

eQTL Mapping

We performed eQTL mapping in 770 MESTIM individuals by using both FaST-LMM²⁹ and EPACTS and implementing a linear mixed model to account for the population structure among the samples.³⁰ Genotype dosages from all autosomal chromosomes and expression data that had passed the aforementioned quality-control measures were used. For FaST-LMM implementation, to improve power³¹ when testing all the variants on chromosome N for association, we constructed the kinship matrix by using the variants from all other chromosomes besides N. This procedure allowed us to include the variant being tested for association in the regression equation only once. Results obtained with FaST-LMM and EPACTS were similar. Results obtained from EPACTS analysis were used in the identification of coincident eQTL and GWAS signals. Results from the FaST-LMM analysis are available on our website.

eQTLs were defined as cis (local) if the peak association was within 1 Mb on either side of the exon boundaries of the gene or as trans (distal) if the peak association was at least 5 Mb outside of the exon boundaries. We used all association p values in the cis region to estimate the false-discovery rate (FDR-qvalue) by using the qvalue package in R. Variants with association p values < 2.46 × 10⁻⁴ corresponding to 1% FDR were considered significant. Considering the large number of analyses we performed to calculate trans-eQTL associations, we used the conservative Bonferroni-corrected p < 1.51 × 10⁻¹³ (0.05/[7.67 million variants × 43,145 probe sets]) to identify trans-eQTLs. To detect possible trans-eQTL hotspots, we report variants that are associated with at least five genes at a more liberal threshold of p < 5 × 10⁻⁸. These hotspots were visualized with Circos-0.66.³² LocusZoom was used for the regional visualization of eQTL results on the basis of linkage disequilibrium (LD) ascertained from the 770 METSIM samples.³³

Heritability Calculations

Heritability was estimated via a linear mixed model with the GCTA software.^{34, 35} In this approach, gene expression phenotypes are assumed to be generated by genetic and environmental components. The assumption behind the linear-mixed-model approach is that the covariance of the genetic component of the phenotypic data is proportional to the kinship or genetic similarity matrix between the individuals. The analysis provides estimates of σ²_u and σ²_e, the variances corresponding to the genetic and environmental components, respectively. The heritability is then the fraction of the variance accounted for by the genetics

$$h^2=\frac{\text{var}\left(u\right)}{\text{var}\left(u\right)+\text{var}\left(e\right)},$$

where var(u) and var(e) are, respectively, the genetic and residual variance components estimated by the restricted-maximum-likelihood approach using related individuals, and is computed for each probe set. A standardized kinship matrix, which has a mean of 1 along the diagonal, is used for these estimates so that they are consistent with the classical definition of heritability.³⁶

Evaluation of eQTLs in Another Adipose eQTL Cohort

eQTLs identified in METSIM and MuTHER cohorts³⁷ were compared in each study reciprocally. Of the ∼5 million significant variant-probe pairs identified in the METSIM study for cis associations (Table S3), 453,484 had an exact match for the variant in a corresponding variant-probe pair in the MuTHER study. A majority of the genes are represented by multiple probes on the Affymetrix microarray used in the METSIM study and the Illumina microarray used in the MuTHER study. Therefore, we required at least 50% of the variant-probe pairs for a given gene to have the same direction of effect in both studies at various p value thresholds (p < 5 × 10⁻² to 5 × 10⁻⁶) in the MuTHER study. Of the 11,053 significant variant-probe pairs identified in the METSIM study for trans associations (Table S3), 2,312 had an exact match for the variant in a corresponding variant-probe pair in the MuTHER study. We report the METSIM-detected trans-eQTLs for which the exact variant-gene pair exists and shows the same direction of effect in MuTHER at several p value thresholds (p < 5 × 10⁻² to 5 × 10⁻⁶) for which MuTHER data were available. cis- and trans-eQTLs detected in the MuTHER study were compared to those in the METSIM study according to the same criteria.

For the three trans-eQTL hotspots for which the peak variant or its LD proxy was available in the MuTHER dataset, we were able to find corresponding variant-probe pairs for 83 of the 106 genes. We report the variant-gene pairs with the same direction of effect size at p < 0.05 in the MuTHER study.

cis-eQTLs at GWAS Loci

We focused on the GWAS index variants reported to be associated at genome-wide significance (p < 5 × 10⁻⁸) with any of 91 cardio-metabolic diseases and related quantitative traits (Table S4). We downloaded these variants from the National Human Genome Research Institute–European Bioinformatics Institute GWAS Catalog (September 22, 2014) and included additional variants.^{16, 17} We estimated the number of loci for each trait group on the basis of the names of “reported gene(s)” and/or “mapped genes” shown in the catalog. Within the variants available in the METSIM eQTL data, we initially examined 115 variants at 68 loci for T2D, 114 variants at 68 loci for glycemic traits, 325 variants at 205 loci for obesity and related traits, 410 variants at 213 loci for lipid traits, 79 variants at 60 loci for metabolic syndrome, and 249 variants at 187 loci for other broadly defined cardiovascular risk traits. After removing duplicate variants present in at least one disease/trait group, we further examined the remaining 1,221 GWAS variants at approximately 680 loci for association with local gene expression levels.

For significant variant-transcript associations (FDR < 1%), if multiple probe sets mapped to the same transcript, we retained the probe sets for which the lead eSNP (the variant that exhibited the strongest evidence of association with the expression level of the corresponding transcript) had the largest pairwise LD r² estimate with the GWAS variant; LD was calculated from the 770 METSIM samples. If two or more highly correlated variants (LD r² > 0.8) were associated with the same transcript, we reported the variant with the largest LD r² with the lead eSNP. For two variants (pairwise LD r² = 0.79) associated with different GALNT2 probe sets, we kept the variant exhibiting higher LD r² with the GWAS index variant.

We further examined whether each eQTL was coincident with the GWAS signal. We defined GWAS-coincident eQTLs as loci with pairwise LD r² > 0.8 between the GWAS index variant and the lead eSNP. To evaluate association between the GWAS variant and the lead eQTL variant at each locus, we performed reciprocal conditional analyses; we tested association between the GWAS variant and transcript level when the lead eSNP was included in the model, and vice versa.

For 61 cardio-metabolic traits with publically available GWAS summary results^{14, 16, 17, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48} (Table S5), we applied the summary-data-based Mendelian randomization (SMR) method to propose relevant genes at the GWAS loci.⁴⁹ We performed a transcriptome-wide association for 61 cardio-metabolic traits and 24,383 probe sets with cis-eQTLs (<1 Mb) and focused on GWAS loci (p < 5 × 10⁻⁸). Probe sets with p_SMR < 2.1 × 10⁻⁶ (0.05/24,383 probe sets) were then tested for pleiotropy via heterogeneity in dependent instruments (HEIDI). Probe sets with p_HEIDI ≥ 0.05 were identified as potential gene targets of GWAS loci.

We also quantified the uncertainty in the inference that the cis-eQTL genes are causally mediating the association between the GWAS loci with cardio-metabolic traits by using the causal inference test (CIT).⁵⁰ We used the GWAS index variants, expression level of the cis-eQTL genes, and the inverse normalized GWAS trait measured in METSIM to infer a causal relationship wherein for gene expression mediates the association between GWAS variant and trait when p_CIT < 0.05.⁵⁰

Association between Gene Expression Level and Phenotypic Traits

We conducted regression analyses to evaluate the association between gene expression and 23 cardio-metabolic-related traits in up to 770 METSIM individuals. The RMA-normalized expression levels were inverse normal transformed after age and BMI were accounted for. We used non-PEER-corrected expression levels for this analysis because correction for PEER factors can remove broadly acting phenotypic effects on gene expression. Eight of the 35 PEER factors had significant correlation with BMI (|r| = 0.10 to 0.50, p = 2.8 × 10⁻³ to 3.5 × 10⁻⁵⁰). For genes coincident with GWAS loci, in addition to non-PEER-corrected associations, we calculated the association between PEER-corrected expression levels and traits. The 23 phenotypic traits examined included five obesity-related traits (BMI, waist-hip ratio, waist circumference, hip circumference, and fat-free mass), seven glycemic traits (glucose, insulin, proinsulin, HOMA-β, Hb1Ac, Matsuda index, and adiponectin), six lipid-related traits (total cholesterol, low-density lipoprotein cholesterol [LDL-C], high-density lipoprotein cholesterol [HDL-C], triglycerides, total fatty acids, and free fatty acids), two inflammatory traits (C-reactive protein and interleukin 1 receptor antagonist), two blood-pressure traits (systolic and diastolic blood pressure, respectively), and one kidney-function trait (glomerular filtration rate). The phenotypic traits were adjusted for age and BMI before inverse normal transformation. When examining the association between gene expression and BMI, we applied no adjustment for BMI for either the phenotypic trait (BMI) or gene expression level. We used FDR < 1% (equivalent p < 6.5 × 10⁻⁴) to define a significant association between gene expression level and phenotypic-trait level. Genes were clustered via hierarchical clustering based on Euclidian distance as implemented in the heatmap2 function of the gplots package in R.

Mouse Expression and Phenotypic-Trait Data

The mouse studies using the Hybrid Mouse Diversity Panel (HMDP) have been described previously (GEO: GSE42890).^{51, 52} In brief, mice from 120 strains were obtained from The Jackson Laboratory and were bred at the University of California, Los Angeles. Mice were maintained on a chow diet (Ralston Purina Company) until 8 weeks of age, when they were either continued on a chow diet (n = 185 mice) or given a high-fat, high-sucrose diet (Research Diets-D12266B) for an additional 8 weeks (n = 227 mice). Body composition analysis was performed by nuclear magnetic resonance with a Bruker Minispec. Retro-orbital blood was collected under isoflurane anesthesia after mice fasted for 4 hr. Plasma lipids, insulin, glucose, and HOMA-IR were determined as previously described.^{52, 53} The animal protocol for the study was approved by the Institutional Care and Use Committee (IACUC) at the University of California, Los Angeles. Total RNA from flash-frozen epididymal adipose samples from 228 male mice was hybridized to Affymetrix HT_MG-430A arrays and scanned according to standard Affymetrix protocols. To reduce the chances of spurious association results, we performed RMA normalization after removing all individual probes with variants and all probe sets containing eight or more variant-containing probes, which resulted in 22,416 remaining probe sets. Correlations of non-PEER-corrected expression levels with phenotypic traits were calculated with the biweight midcorrelation, which is robust to outliers.⁵⁴

Results

Genetic Variants Associated with Gene Expression in Subcutaneous Adipose Tissue

To identify genetic loci associated with transcript abundance in abdominal subcutaneous adipose tissue, we studied 770 extensively phenotyped men from the METSIM study. We analyzed ∼7.67 million variants and abundance of 43,145 probe sets corresponding to 18,155 unique genes. The mean narrow-sense heritability, h², of the probe sets was 0.27, suggesting significant genetic effects on adipose gene expression (Figure S1).

eQTL mapping identified cis-eQTLs (<1 Mb) for 12,400 genes at a 1% FDR (p < 2.46 × 10⁻⁴) (Figure S2 and Table S3). The larger number of adipose eQTLs observed in comparison to that in previous adipose eQTL studies might be due to the larger sample size, denser imputation, and/or analysis of a larger number of transcripts; some differences might be due to different microarray platforms, statistical methodology, and p value thresholds (Table S6).^{10, 11, 37}

30% of the cis-eQTLs discovered in METSIM at p < 5 × 10⁻⁶ showed consistent allelic direction of effect in the MuTHER study, and 79.1% of the cis-eQTLs reported in MuTHER at this threshold showed a consistent direction of effect in METSIM (Table S7).

Coincidence of cis-eQTLs and GWAS Loci

Hundreds of GWAS loci have been reported for cardio-metabolic diseases and related traits.⁵⁵ Given the value of identifying candidate genes at GWAS loci and the observation that many eQTLs are shared across tissues,⁵⁶ we investigated loci for a broad set of 91 cardio-metabolic diseases or traits (Table S4). Among the 1,221 GWAS signals (p < 5 × 10⁻⁸), we detected 944 initial unique GWAS variant and eQTL gene pairs (FDR < 1%, equivalent p < 2.4 × 10⁻⁴) (Table S8).

On the basis of pairwise LD r² > 0.8 between the GWAS variant and the variant that exhibited the strongest association with the gene expression level (lead eSNP), 140 (15%) of the 944 cis-eQTLs appear to be coincident with the GWAS signals. The coincidence is supported by reciprocal conditional analysis: after conditioning was performed on each lead eSNP, no GWAS variant remained significant (FDR < 1%), and after conditioning on each GWAS variant, 124 eSNP signals were no longer significant, and 16 were strongly attenuated (Δlog₁₀(p) of 8.5 to 112 ) (Table S8). Of the 140 cis-eQTLs at GWAS loci, 93 (66.4%) were not previously reported by large-scale GWASs that interrogated available cis-eQTLs^{14, 15, 16, 17, 38, 45, 48, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68} and 50 showed consistent direction of allelic effect at p < 0.05 in the MuTHER study. Table 1 shows 29 eQTLs for glycemic, obesity, and lipid traits at the LD threshold of r² > 0.9 between the GWAS variant and lead eSNP; three of these eQTLs were also identified with the SMR method. The full set of 944 adipose eQTLs at GWAS loci, evidence of their coincidence with the GWAS signals, and the corresponding p values in the MuTHER study are provided in Table S8.

Table 1.

Selected Adipose eQTLs Coincident with GWAS Signals for Cardiometabolic Risk

GWAS Variant	GWAS Trait	GWAS Locus	eQTL Gene	GWAS Variant					Lead eSNP						LD r2
				A1/A2	βinitial	pinitial	βcond	pcond	Lead eSNP	A1/A2	βinitial	pinitial	βcond	pcond
rs2013208	HDL cholesterol	RBM5	RBM6	C/T	−0.987	4.4E−113	0.120	6.9E−01	rs11130233	G/T	−0.992	4.7E−116	−1.099	2.2E−04	0.98
rs12051272	adiponectin	CDH13	CDH13	T/G	1.370	4.3E−80	0.000	2.8E−01	rs12051272	T/G	1.370	4.3E−80	0.000	2.8E−01	1.00
rs6805251	HDL cholesterol	GSK3B	GSK3B	C/T	−0.696	1.5E−43	0.067	7.4E−01	rs334533	C/T	−0.718	9.2E−47	−0.786	1.3E−04	0.95
rs8077889	triglycerides	MPP3	MPP3	C/A	−0.696	9.2E−33	−0.174	7.4E−01	rs55768269	T/C	−0.695	5.6E−33	−0.569	2.8E−01	0.99
rs4148008	HDL cholesterol	ABCA8	ABCA8	G/C	−0.538	5.9E−26	0.415	2.1E−02	rs1156340	T/C	−0.598	1.8E−31	−1.019	2.5E−08	0.92
rs12489828	waist-hip ratio	NT5DC2	NT5DC2	G/T	0.503	4.0E−24	0.089	7.6E−01	rs6778735	T/C	0.510	1.8E−24	0.398	1.9E−01	0.97
rs138777	total cholesterol	TOM1	HMGXB4	A/G	−0.495	1.7E−22	0.141	8.4E−01	rs9306298	T/C	−0.498	1.1E−22	0.636	3.7E−01	0.99
rs2254287	LDL cholesterol	B3GALT4	HSD17B8	G/C	−0.417	3.5E−17	0.000	2.1E−01	rs2254287	G/C	−0.417	3.5E−17	0.000	2.1E−01	1.00
rs12679556	waist-hip ratio	MSC	EYA1	G/T	0.463	1.1E−16	−0.202	6.2E−01	rs4738141	G/A	0.469	3.4E−17	−0.677	9.7E−02	0.98
rs12748152	HDL cholesterol	NR0B2-PIGV	PIGV	T/C	−0.698	4.0E−14	−0.054	8.5E−01	rs6656815	A/G	−0.732	2.7E−15	−0.687	1.9E−02	0.90
rs10919388	waist-hip ratio	GORAB	PRRX1	C/A	−0.448	4.5E−14	−0.212	3.3E−01	rs6427242	G/C	−0.448	3.7E−14	−0.245	2.5E−01	0.93
rs8077889	triglycerides	MPP3	DUSP3	C/A	0.430	6.6E−13	0.000	5.6E−01	rs2342310	C/T	0.430	6.6E−13	0.000	5.6E−01	1.00
rs11136341	LDL cholesterol	PLEC1	PLEC	G/A	−0.358	9.5E−13	0.084	6.5E−01	rs10107388	C/T	−0.379	4.9E−14	−0.455	1.5E−02	0.92
rs2590838	adiponectin	GNL3	NEK4	G/A	0.346	8.4E−12	−0.018	9.9E−01	rs35212380	C/G	0.346	7.8E−12	−0.415	6.7E−01	1.00
rs439401	triglycerides	APOE-TOMM40	APOE	T/C	0.384	2.1E−11	0.000	6.7E−01	rs439401	T/C	0.384	2.1E−11	0.000	6.7E−01	1.00
rs181362	HDL cholesterol	UBE2L3	YDJC	T/C	0.327	7.6E−11	−0.254	1.3E−01	rs11089620	G/C	0.368	3.1E−13	−0.621	2.3E−04	0.91
rs7134375	HDL cholesterol	PDE3A	PDE3A	C/A	0.342	1.6E−10	0.000	3.5E−01	rs7134375	C/A	0.342	1.6E−10	0.000	3.5E−01	1.00
rs11869286	HDL cholesterol	STARD3	STARD3	G/C	0.333	1.1E−09	0.000	9.8E−01	rs11869286	G/C	0.333	1.1E−09	0.000	9.8E−01	1.00
rs9400239	body mass index	FOXO3	FOXO3	C/T	0.323	3.0E−09	0.008	9.7E−01	rs3800228	G/T	0.337	5.9E−10	0.350	5.5E−02	0.92
rs12748152	HDL cholesterol	NR0B2-PIGV	ARID1A	T/C	−0.521	2.2E−08	−0.091	7.8E−01	rs34217609	C/T	−0.515	1.1E−08	−0.388	2.2E−01	0.92
rs3812316	triglycerides	MLXIPL	BCL7B	C/G	−0.411	6.5E−08	−0.059	8.2E−01	rs799166	C/G	−0.431	2.1E−08	−0.395	1.4E−01	0.92
rs4846914	HDL cholesterol	GALNT2	GALNT2	G/A	−0.267	4.1E−07	0.221	4.4E−01	rs4631704	C/T	−0.279	1.3E−07	0.503	8.4E−02	0.97
rs12145743	HDL cholesterol	PMVK-HDGF	RRNAD1	T/G	−0.262	7.9E−07	−0.115	8.1E−01	rs3806415	C/T	−0.263	7.8E−07	−0.139	7.7E−01	0.99
rs10501320	proinsulin	MADD	ACP2	G/C	0.309	1.3E−06	0.024	9.7E−01	rs11039149	A/G	0.310	1.2E−06	0.259	6.5E−01	0.99
rs6784615	waist-hip ratio	NISCH-STAB1	NISCH	T/C	0.649	1.6E−06	0.000	6.6E−01	rs728408	A/G	0.649	1.5E−06	0.000	7.4E−01	1.00
rs780094	glucose, triglycerides	GCKR	EMILIN1	T/C	0.240	3.3E−06	0.000	7.2E−01	rs780094	T/C	0.240	3.3E−06	0.000	7.2E−01	1.00
rs10761731	triglycerides	JMJD1C	NRBF2	A/T	−0.231	5.8E−06	0.000	5.8E−02	rs10761739	G/C	−0.233	5.2E−06	0.000	4.3E−02	1.00
rs1532624	HDL cholesterol	CETP	CETP	C/A	0.237	5.8E−06	−0.549	2.1E−01	rs4784741	C/T	0.246	2.4E−06	−0.777	7.4E−02	0.99
rs4311394	adiponectin	ARL15	FST	A/G	0.287	6.2E−06	−0.707	3.1E−01	rs59061738	A/G	0.292	3.7E−06	−1.011	1.5E−01	0.99

This table contains the 29 GWAS-relevant eQTLs for which the pairwise LD r² > 0.9 between GWAS index and lead eSNP, for which p_initial < 10⁻⁵ for the association between GWAS SNP and gene expression level, and which had not been previously described in large-scale genome-wide association studies that had interrogated available databases/resources for cis-eQTLs searches. The GWAS traits were limited to the most common cardio-metabolic traits, including body mass index, waist-hip ratio adjusted for body mass index, triglycerides, HDL cholesterol, LDL cholesterol, total cholesterol, glucose, proinsulin, adiponectin, and risk of type 2 diabetes. A full list of the 944 adipose eQTLs identfied at GWAS loci for cardio-metabolic diseases and traits, including the 140 GWAS coincident eQTLs with pairwise LD r² > 0.8, are shown in Table S8. A1/A2: the effect/non-effect alleles for association tests in METSIM eQTL analyses. A1 for the GWAS variant was the allele previously reported to increase the cardio-metabolic risk. LD r² between the GWAS variant and lead eSNP was calculated from 770 MESTIM individuals.

To consider potential pleiotropic effects of the GWAS variants on gene expression and cardio-metabolic traits, we performed two additional tests. We applied the SMR Mendelian randomization method to loci for 61 traits and identified 46 genes (p_SMR < 2.1 × 10⁻⁶; p_HEIDI ≥ 0.05) (Table S9), 19 of which were identified via LD (r² > 0.8) and conditional analyses. The SMR analysis also detected nine genes (C18orf8, CABLES1, CADM1, CDK6, CENPW, LMOD1, MFAP2, MT1M, and STARD10) that were not identified via the conditional analysis approach.

We applied a causal-inference test⁵⁰ and identified 15 variant-transcript pairs for seven traits that showed evidence of causal mediation (p_CIT < 0.05) (Table S10 and Figure S3). None of the target genes identified by the causal-inference test overlap with the genes identified by the SMR analysis.

The coincident GWAS and eQTL signals suggest candidate genes that might influence cardio-metabolic risk. For example, at the ARL15 locus associated with adiponectin and HDL-C^{14, 38}, the GWAS index variant rs6450176 was associated with expression of FST (p = 3.3 × 10⁻⁹), located ∼500 kb away, and exhibited strong LD (r² = 0.99) with the lead eSNP rs59061738, associated with FST expression (p = 1.8 × 10⁻⁹, Figure 1A and Table 1). Reciprocal conditional analyses provided additional evidence that this eQTL was coincident with the GWAS signal (p_cond = 0.43 for rs6450176 and 0.17 for rs59061738). FST expression levels were negatively associated with HDL-C (p = 1.2 × 10⁻³) and adiponectin (p = 6.0 × 10⁻⁶, Figure 1B). Encoded by FST, the protein follistatin has been shown to promote adipocyte differentiation and reduce fat mass and insulin resistance.^{70, 71} The same GWAS variant showed no eQTL for the nearest gene, ARL15 (p = 0.43), and the expression level of ARL15 was not associated with HDL-C (p = 0.85) or adiponectin (p = 0.084). These data suggest that the associations with the levels of adiponectin and HDL-C at this GWAS locus might be mediated at least in part through the altered expression of FST.

Figure 1.

Example Subcutaneous Adipose eQTL Genes at GWAS Loci

(A) The adipose eQTL of FST is coincident with the adiponectin GWAS locus ARL15.⁶⁹ Regional association of variants with expression level of FST is shown with the GWAS variant rs6450176 plotted as the index (purple diamond). LD is colored on the basis of the METSIM population.

(B) Association between adiponectin level and FST expression level in METSIM.

(C) The adipose eQTL of RQCD1 is coincident with the USP37 GWAS locus for BMI.¹⁷ Regional association of variants with expression level of RQCD1 is shown with the eSNP rs4674320 (r² = 1.0 with BMI index SNP rs492400) plotted as the index.

(D and E) Association between (D) BMI and RQCD1 expression level in humans from the METSIM study and (E) body fat and Rqcd1 expression level in mice from the HMDP study.

Association between Gene Expression Level and Phenotypic Traits

We investigated the effects of gene expression levels on cardio-metabolic risk by evaluating the association between the expression levels of all 43,145 probe sets and 23 cardio-metabolic-related traits (Table S11). At FDR < 1% (equivalent p < 6.5 × 10⁻⁴), we observed 48,365 significant associations between probe set and trait, and these corresponded to 29,920 gene-trait associations and 7,643 genes that were associated with 1 to 16 cardio-metabolic traits (Table S12). 10,819 probe sets (6,064 genes) were associated with BMI, 6,640 probe sets (3,940 genes) with Matsuda index, and 4,933 probe sets (3,039 genes) with insulin levels.

We next examined associations between the identified eQTL genes and cardio-metabolic traits measured in extensively phenotyped participants of the METSIM study. Among the genes for the 140 GWAS-relevant eQTLs, the expression levels of 49 were associated with the corresponding GWAS traits at p < 0.05 (Figure 2; also Figure S4 and Table S8). For example, TBX15 at the TBX15-WARS2 locus was associated with waist-hip ratio adjusted for BMI (p = 1.3 × 10⁻⁸) and 12 other traits (p = 4.3 × 10⁻²¹ to 2.6 × 10⁻⁴), and expression of GPR146 at the lipid locus GPR146 was associated with HDL-C (p = 3.3 × 10⁻¹³), triglycerides (p = 3.9 × 10⁻¹³), and ten other traits (p = 2.3 × 10⁻⁴⁵ to 6.6 × 10⁻⁵, Figure 2). These gene expression and trait associations at GWAS loci further suggest plausible roles of these genes in mediating variant effects on cardio-metabolic disorders.

Figure 2.

Heatmap of Effect Sizes for Significant Associations between Gene Expression Level and Cardio-Metabolic-Trait Levels at GWAS Loci with Coincident eQTLs

Rows show 23 selected cardio-metabolic traits, and columns show the eQTL genes (and reported GWAS trait at the coincident locus). Negative values (blue) indicate that increased gene expression level was associated (p < 0.05) with decreased trait level, whereas positive values (orange) indicate that increased gene expression level was associated with increased trait level.

Human GWAS eQTL Genes in Mice

We next sought additional evidence for the involvement of GWAS-relevant eQTL genes in regulating cardio-metabolic traits by comparing results with those from a diverse panel of 120 inbred mouse strains known as the Hybrid Mouse Diversity Panel (HMDP).⁷² In the panel, we identified microarray probes for 107 of the 140 mouse orthologs and tested them for association between adipose expression level and metabolic traits, including plasma lipids, insulin, glucose, and body-fat composition. We observed a significant correlation between 70 genes and one of the metabolic traits (p < 5.2 × 10⁻⁵; 0.05/963 tests for nine traits × 107 genes) (Figure S5). Of these genes, 25 showed expression levels correlated with a similar metabolic trait in the same direction (Table S13). 13.1% of 10,771 orthologous genes had expression-trait correlations in the same direction in the adipose tissue of both species for waist-hip ratio, total cholesterol, insulin, and glucose. For example, GWAS variants associated with the homeostatic model of insulin resistance (HOMA-IR) are associated with decreased expression level of IRS1 (p = 2.0 × 10⁻⁵⁰). IRS1 expression is negatively correlated with HOMA-IR (r = −0.40, p = 3.0 × 10⁻³⁰) in METSIM and the HMDP (r = −0.50, p = 1 × 10⁻¹³) (Figure S6).

The HMDP data support a biological contribution of 25 eQTL genes at GWAS loci; 18 of these genes did not have significant associations in humans (p > 0.05) but had significant associations in mice, possibly because of the heterogeneous environmental effects in humans and controlled experimental environment in mice (Table S13). For example, the chromosome 2q35 variants reported by the GIANT consortium to be associated with BMI¹⁷ span ∼300 kb and nine genes. In METSIM, the BMI risk allele rs492400 was significantly associated with increased expression of RQCD1 (p = 8.9 × 10⁻⁴³, Figure 1C), although RQCD1 expression was only nominally correlated with BMI (r = 0.09, p = 0.01, Figure 1D). In the HMDP, Rqcd1 expression was more strongly positively correlated with subcutaneous fat weight (r = 0.45, p = 1.2 × 10⁻¹², Figure 1E), providing further support for the role of RQCD1 in BMI. Consistent correlation between gene expression and clinical traits in mice for cis-regulated genes at human GWAS loci supports the fruitful use of mouse models to further understand the roles of candidate genes for cardio-metabolic traits.

trans-eQTLs

Associations of variants with expression of distant genes in trans allow the identification of regulatory networks in adipose tissue. We observed trans-eQTLs at a Bonferroni-corrected significance threshold (p < 1.51 × 10⁻¹³) for 90 genes (Table S3). To better identify regulatory networks in human samples, we searched for variants that were associated with at least five distal genes at a more liberal threshold of p < 5 × 10⁻⁸. These variants were located on chromosomes 3, 7, 11, and 16 in five independent loci (pairwise LD r² = 0) that we termed trans-eQTL hotspots (Table S14, Figure 3). Four of the hotspot loci were associated with local (cis) gene expression at p < 1 × 10⁻³ (Figure S7).

Figure 3.

trans-eQTL Hotspots in Subcutaenous Adipose Tissue

(A) KLF14 trans-eQTL hotspot and target genes in subcutaneous adipose tissue. Representation of the location of 55 distal genes for which expression level is associated with rs12154627 at the KLF14 locus.

(B) Representation of the location of 65 distal genes for which expression level is associated with one of four trans-eQTL hotspots: SLC25A38 locus (black), ZNF800 locus (blue), CIITA locus (green), and HBB locus (red).

Arrowheads point to the trans-eQTL loci, and curves indicate the associations with the four sets of target genes.

40 of the 55 trans-mediated genes at the KLF14 locus were significantly associated with metabolic phenotypes at 1% FDR, representing a significant enrichment with respect to all the genes assayed on the microarray (Chi-square p value = 8.4 × 10⁻⁶) (Figure S8). The trans-regulated genes were co-expressed with pairwise correlations of −0.48 to 0.67 (p = 8.9 × 10⁻⁴ to 1.6 × 10⁻¹⁰⁰); however, they were not enriched for known biological pathways according to the analysis involving the DAVID database.⁷³ Some of the distal genes have been previously studied in relation to adipose biology, but their association with the KLF14 locus was not known. Among the trans-regulated genes, PLIN5 was associated with 13 of the metabolic phenotypes (Figure S8). PLIN5 encodes Perilipin 5, a lipid-droplet-associated protein that helps maintain a balance between lipolysis and lipogenesis and has been shown to play a role in fatty-acid oxidation in adipose tissue.⁷⁴ The expanded KLF14-mediated trans network of genes suggests that these targets might also influence cardio-metabolic traits.

Variants located near CIITA (class II, major histocompatibility transactivator), which is known to activate the MHC class II genes at the HLA locus⁷⁵ (Figure 3B), were associated with six MHC class II genes (HLA-DPA1, HLA-DMA, HLA-DPB1, HLA-DOA, HLA-DRA, and HLA-DMB) located on chromosome 6 and with CD74 on chromosome 5. CD74 encodes for a protein that associates with the class II MHC complex and regulates antigen presentation.⁷⁶ The variants were also suggestively associated with CIITA expression (p = 1.0 × 10⁻³, Figure S7), suggesting that variation in its expression might be responsible for the trans-eQTL signals. Our eQTL results captured the known biology of CIITA regulation of MHC class II genes. Further, expression of CIITA and the seven trans-mediated genes was significantly associated with 14 of the metabolic traits (|β|= 0.12 to 0.37, p = 6.4 × 10⁻⁴ to 7.7 × 10⁻²⁶), suggesting a role for this network of genes in adipose tissue (Figure S8).

The second largest trans-eQTL hotspot was located on chromosome 3 and was associated with the expression of 44 genes (Figure 3B). This locus has been reported in other tissues, including liver and omental fat,⁷⁷ but the target genes identified in our study do not overlap with the previous ones, suggesting a unique trans-regulatory network in subcutaneous adipose tissue.¹¹ Variants in this locus showed the strongest local association with the expression of SLC25A38, located ∼1 Mb away (p = 1.7 × 10⁻⁷, Figure S7). Missense mutations in SLC25A38, which encodes a mitochondrial solute carrier, lead to congenital sideroblastic anemia characterized by defective erythropoiesis and mitochondrial iron overload.⁷⁸ The function of SLC25A38 in adipose tissue remains unclear. The signal observed at this locus might represent cell types other than adipocytes in light of the cellular heterogeneity of adipose tissue. The adipose expression level of SLC25A38 was nominally associated with 24 of the trans-mediated genes (|β| = 0.07 to 0.29, p = 6.8 × 10⁻¹⁶ to 4.7 × 10⁻²) and with 14 cardio-metabolic-related traits (|β| = 0.08 to 0.33, p = 6.7 × 10⁻²¹ to 3.1 × 10⁻²) (Figure S7). For example, SLC25A38 is negatively associated with BMI in METSIM participants (β = −0.33, p = 6.7 × 10⁻²¹). On the basis of structural similarity with other SLC25 proteins, SLC25A38 is predicted to transport glycine into the mitochondrial matrix for condensation with succinyl-coenzyme A to form 5-aminolevulinic acid, which is exported to the cytoplasm for haem synthesis.^{79, 80} Only a few of the trans-mediated genes have been shown to play important roles in adipocytes. For example, ESRRA encodes estrogen-related receptor alpha, which modulates the expression of adipogenesis genes during adipocyte differentiation.⁸¹ SPTLC1 encodes one of the subunits of serine palmitoyltransferase, the key enzyme in sphingolipid biosynthesis.⁸² These bioactive lipids are altered with obesity and can regulate inflammatory gene expression in adipocytes.⁸³ Another target gene, GOT2, encodes glutamic-oxaloacetic transaminase, which functions in the inner mitochondrial membrane and has been shown to play a key role in amino acid metabolism.⁸⁴ Although the function in adipose of the genes associated with SLC25A38 variants remains unclear, our results suggest a metabolic role.

We also observed a trans-eQTL hotspot, located on chromosome 7 but ∼3.3 Mb away from KLF14, associated with the expression of nine genes. The variants at this locus were independent from those at KLF14 and had distinct target genes (Figure 3B). The lead variant, rs62621812 in ZNF800, encodes a missense change (p.Pro103Ser) and is also associated with that gene’s expression level (p = 2.8 × 10⁻¹⁶), providing evidence suggesting that this gene regulates the target genes. ZNF800 encodes a C2H2 zinc finger protein and is a putative transcription factor.⁸⁵ ZNF800 expression level is associated with the Matsuda Index (β = 0.12, p = 1.3 × 10⁻³), suggesting a role in insulin sensitivity.

A fifth trans-eQTL, located on chromosome 11, was associated with expression of five trans genes. The peak variant, rs10742583, is located ∼2 kb upstream of HBB, which encodes beta hemoglobin, and is in complete LD (r² = 1.0) with a HBB synonymous variant, rs713040, although rs10742583 was not associated with the expression level of any local gene (p < 2.4 × 10⁻⁴). A functional role for HBB in adipose biology is not clear. Although HBB is distinctly and highly expressed in whole blood, the trans-mediated target genes are expressed in a range of tissues.⁵⁶ This trans-eQTL signal might reflect cellular heterogeneity, especially blood cells present in adipose tissue.

To validate these trans-eQTL hotspots, we asked whether eQTLs were also observed in the MuTHER study. Variant information was available for three of the loci, near CIITA, SLC25A38, and KLF14. All seven genes that mapped to the CIITA locus, 35 of the 44 genes that mapped to the SLC25A38 locus, and 41 of the 55 genes that mapped to the KLF14 locus had corresponding probes available in the MuTHER study. Although only two of the 35 target genes for the SLC25A38 locus had evidence of association (FDR < 1%) in the MuTHER study, many KLF14 and CIITA target genes were replicated (Table S14). For the KLF14 trans-eQTL hotspot, originally described in MuTHER,²⁰ 25 of the 41 target genes showed consistent direction of effect and nominal associations (p < 0.05) in the MuTHER study, whereas none of the genes showed an opposite effect. For the CIITA locus, six of the seven target genes were validated in MuTHER with a consistent direction of effect (p < 0.05), indicating strong support for these two trans-eQTL hotspots.

Discussion

We report an analysis of genetic variants, adipose gene expression, and cardio-metabolic traits in a deeply phenotyped cross-sectional sample of 770 participants in the METSIM study and compare the results to those of studies performed in 120 diverse inbred strains of mice. These results help to prioritize potential target genes at GWAS loci, and they identify several trans-regulatory networks associated with metabolic syndrome and its component traits. In addition, the results reveal concordance between mice and humans in terms of correlations between gene expression and traits.

A main challenge with regard to gaining biological insights from genetic associations is to identify which genes and pathways explain the associations. Systems genetics aims to address this challenge by integrating genetic variants with molecular phenotypes to more comprehensively define the relationship between genotype and phenotype. Because a large fraction of the variation underlying common diseases appears to be regulatory,^{4, 86, 87} eQTLs that coincide with GWAS loci can link trait-associated loci with molecular perturbations. Overall, using reciprocal conditional analyses, we provided evidence for 140 genes at 105 loci that may be involved in metabolic traits (Table 1 and Table S8). This analysis of gene expression in a single tissue might not be relevant for the mechanisms of some of the cardio-metabolic traits. Although our analysis provides indirect evidence of an association, because the colocalization of an eQTL with a disease locus could be coincidental, the cardio-metabolic phenotypes available in the METSIM cohort provided additional evidence for 49 genes for which expression level was associated with the relevant trait. In addition, a subset of the 140 genes was also linked to GWAS traits via the summary-data-based Mendelian-randomization and causal-inference tests. Of note, we limited our analysis to lead eQTL variants and lead GWAS variants in high pairwise LD (r² ≥ 0.8). This threshold might be conservative if the GWAS lead variant imperfectly tags a causal variant and might underestimate the number of GWAS loci that are coincident with eQTLs. Consistent with this possibility, we found evidence of association for 579 genes at 415 loci for which conditioning on the GWAS variant partially attenuated the eQTL effect (Table S8). Furthermore, secondary independent associations for both eQTLs and GWAS signals have been shown to play an important role in altering the expression of a gene in a locus;²² however, additional secondary and tertiary eQTL signals were not separately analyzed in this study. These genes provide a larger set of possible targets for GWAS loci.

Employing genotype imputation with a dense reference panel and analyzing 43,145 transcripts on a microarray to measure expression, we were able to identify more eQTLs than were identified in other subcutaneous adipose studies with similar sample sizes.^{10, 11, 37} Using these eQTLs, we implicated 50 genes that were distinct from the initial locus annotations used as signposts for the GWAS loci (Table S8). For example, rs11231693 is associated with waist-hip ratio adjusted for BMI¹⁶ and is located in the intron of MACROD1. Our results showed a strong association with the expression level of VEGFB (p = 2.2 × 10⁻⁶⁷), located 140 kb from the variant. For ten loci, we identified multiple genes associated with the risk variants. For example, BMI-associated variants in the INO80E locus¹⁷ are also the strongest variants associated with the expression of five nearby genes: YPEL3, INO80E, TMEM219, TBX6, and HIRIP3 (p = 1.4 × 10⁻⁸ to 1.8 × 10⁻⁴), suggesting that variation in one or more of these genes might influence BMI. Finally, at the GCKR, LCAT, and LDLR genes that harbor common and rare coding variants that can affect plasma lipid levels, we observed eQTL effects on nearby genes, including EMILIN1 at the GCKR locus (p = 3.3 × 10⁻⁶), GFOD2 and NUTF2 at the LCAT locus (p = 1.9 × 10⁻⁶ and 1.6 × 10⁻⁴, respectively), and YIPF2 at the LDLR locus (p = 1.5 × 10⁻⁶). Some coincident associations between GWAS variants and eQTLs might not contribute strongly to trait variation.

Recent studies demonstrated the evolutionary conservation of regulatory networks that increase susceptibility to atherosclerosis and other cardio-metabolic traits in disease-relevant tissues in humans and mice.⁸⁸ Genetically diverse mouse populations allow for the control of confounding factors, such as environmental exposure that are difficult to assess in humans. The HMDP has also helped researchers to understand contributions of genetic factors to cardio-metabolic traits.⁷² The results of our cross-species analysis showed consistent association between traits and the expression of 25 genes in humans and mice, providing support for further study of these genes in mouse models (Figure S5 and Table S13). Of course, the possibility exists that mechanisms might differ between species or that the wrong tissues are being compared. In addition to prioritizing candidate genes, studies of natural variation in gene expression in mice should be useful in elucidating mechanisms relating to the gene-by-gene, gene-by-environment, and gene-by-sex interactions.^{51, 52, 89} For example, although it is clear that genetic background contributes to weight gain and related traits in response to dietary challenge,⁹⁰ the inability to accurately ascertain environmental factors over a lifetime in humans has made molecular dissection difficult. In contrast, studies of natural variation in mice have revealed that gene-by-environment interactions can contribute substantially to the overall variance in traits such as obesity and that the underlying genes and pathways can be identified.^{51, 52}

Whereas numerous studies have successfully mapped cis associations, few studies have reported trans-eQTLs because of small effect sizes, multiple testing thresholds, and computational burden.^{22, 37, 91} These issues are consistent with the low reciprocal replication rates of trans-eQTLs identified in MuTHER and METSIM cohorts. However, in-depth characterization of the architecture of trans regulation of gene expression is useful in attempting to understand complex biological mechanisms, as shown by the expanded network of trans-regulated genes at the KLF14 locus, which has been associated with T2D and several metabolic traits. The 55 genes that map to this locus demonstrate the extensive downstream effects of an individual association signal and provide additional genes that might influence disease manifestation. The identification of trans-regulatory networks also offers opportunities for understanding fundamental biology. For example, ZNF800 is a putative transcription factor, and although little is known about the function of the encoded protein, our results implicate a novel role for ZNF800 in adipose biology. In addition, some trans-eQTL associations might be influenced by the cellular composition of the adipose tissue; for example, adipose tissue of obese individuals tends to have higher macrophage content.^{90, 91} Although we failed to observe any association between the trans-acting loci and expression of macrophage markers such as ABCG1 and CD68 (data not shown), we cannot rule out a role for cell-type heterogeneity. We also observed weaker cis associations compared to trans associations in the trans-eQTL hotspots. This may be an artifact of data pre-processing or inaccuracy in measuring low levels of expression. Future studies involving single-cell analyses might clarify the role of adipose tissue by identifying cell-type-specific eQTLs. Our study highlights the power of a systems-genetics approach in dissecting complex traits and identifying causal genes and pathways.

Conflicts of Interest

A.H., C.T., T.K., and P.S.G. are employees and shareholders of Bristol-Myers Squibb.

Published: March 2, 2017

Accession Numbers

Newly deposited human expression data can be obtained from the Gene Expression Omnibus with the accession number GEO: GSE70353.

Web Resources

• Efficient and Parallelizable Association Container Toolbox (EPACTS), http://genome.sph.umich.edu/wiki/EPACTS

• GWAS Catalog, https://www.ebi.ac.uk/gwas/

• The Haplotype Reference Consortium, http://www.haplotype-reference-consortium.org

• METSIM and HMDP, https://systems.genetics.ucla.edu/

Supplemental Data

Table S9. Results of the Summary-Data-Based Mendelian-Randomization Analysis

The source of the GWAS summary results are given in Table S5.

Table S10. Results of the Causal-Inference Test

The causal-inference test was performed for the GWAS index SNP, expression of the cis-eQTL gene, and inverse-normalized traits measured in the METSIM population.

Table S14. Trans-eQTL Hotspots Identified in METSIM Adipose Tissue and Their Look ups in the MuTHER Cohort

Red indicates a cis-regulated gene in the locus.