Multi-omic Integration Reveals Cell-type-specific Regulatory Networks of Insulin Resistance in Distinct Ancestry Populations

Summary

Our knowledge of the cell-type-specific mechanisms of insulin resistance remains limited. To dissect the cell-type-specific molecular signatures of insulin resistance, we performed a multiscale gene network analysis of adipose and muscle tissues in African and European ancestry populations. In adipose tissues, a comparative analysis revealed ethnically conserved cell-type signatures and two adipocyte subtype-enriched modules with opposite insulin sensitivity responses. The modules enriched for adipose stem and progenitor cells and immune cells showed negative correlations with insulin sensitivity. In muscle tissues, the modules enriched for stem cells and fibro-adipogenic progenitors responded to insulin sensitivity oppositely. The adipocyte and muscle fiber-enriched modules shared cellular respiration-related genes but had tissue-specific rearrangements of gene regulations in response to insulin sensitivity. Integration of the gene coexpression and causal networks further pinpointed key drivers of insulin resistance. Together, this study revealed the cell-type-specific transcriptomic networks and signaling maps underlying insulin resistance in major glucose-responsive tissues. A record of this paper’s Transparent Peer Review process is included in the Supplemental Information.

Graphical Abstract

Electronic table of contents (eTOC) blurb:

Peng et al. performed a multiscale gene network analysis of adipose and muscle tissues in African and European ancestry populations and identified conserved cell-type-specific subnetworks associated with insulin resistance. The results provide a cell-type-specific landscape of molecular interactions and regulators of insulin resistance in glucose-responsive tissues.

Introduction

Reduced insulin sensitivity, or insulin resistance, is an intermediate phenotype and early marker of Type 2 Diabetes (T2D) risk^1,2. Many insulin-resistant individuals develop T2D and are at increased risk for dyslipidemia, hypertension, and atherosclerotic cardiovascular disease³. Recent studies started recognizing the role of intrinsic cellular defects in insulin resistance. At the molecular level, insulin resistance is associated with dysregulations of various interconnected cellular processes, such as inflammation, oxidative stress, lipotoxicity, endoplasmic reticulum stress, and mitochondrial dysfunction; the majority affecting pathways outside of the canonical insulin signaling pathway^4,5. Physiologic and genomic studies revealed that insulin resistance results from derangement in gene expressions of insulin-responsive tissues involved in glucose homeostasis, including the adipose, muscle, and liver⁶. The adipose tissue acts as a major endocrine organ with active hormone secretion and plays an essential role in storing lipids for energy homeostasis⁷. The skeletal muscle is the largest insulin-sensitive organ that accounts for approximately 80% of postprandial glucose disposal in humans⁸. Dysregulations of the adipose and muscle tissues, such as macrophage infiltration into the adipose and ectopic lipid accumulation in the skeletal muscle, influence glycogen synthesis, de novo lipogenesis, hyperlipidemia, and gluconeogenesis, in the liver, leading to the pathogenesis of insulin resistance and T2D⁹. Integrating multi-omic data from these tissues through systems biology approaches in human population cohorts will help expand our knowledge of the biological processes including genetically regulated mechanisms of insulin resistance.

Both adipose and muscle tissues are composed of a complex mixture of different cell types, which are responsive to dynamic regulations of metabolic signals. In the adipose tissue, adipocytes play a dominant role in storing and releasing lipids for energy balance. Adipocytes as well as the stromal vascular fraction of adipose tissue cells, including preadipocytes, fibroblasts, vascular cells, and immune cells, provide systemic metabolic regulations by secreting a plethora of adipocytokines (e.g., leptin, adiponectin, TNFα, and IL-6)¹⁰. In the muscle tissue, the muscle fiber cells provide basic support for physical movement and glucose metabolism. Other muscle resident cells, such as muscle stem cells and non-myogenic cells (e.g., immune cells and endothelial cells), are essential for immune response, cell death and regeneration¹¹. However, due to the difficulty in dissecting the mixture of cells from the biopsy of complex tissues, the bulk tissue transcriptome sequencing-based approaches have limited power in defining cell-type-specific transcriptomic networks associated with insulin resistance in adipose and muscle tissues^12,13. Recent single-cell RNA-seq (scRNA-seq) studies directly measured gene expressions in single cells and revealed cell-type-specific regulations of the adipose and muscle tissues in obesity and T2D. Different subpopulations of adipocytes were identified to be associated with obesity in the adipose tissue, and a group of skeletal muscle fibro-adipogenic progenitors was involved in muscle degeneration in T2D patients^14,15. The application of scRNA-seq is however limited by expensive cost, small cohort size, extensive gene dropout effect, and the difficulty in harvesting certain cell types (e.g., fragile adipocytes and elongated muscle fibers)¹⁶. Alternative single-cell transcriptome technology like single-nucleus RNA-seq (snRNA-seq) partially circumvents some of the limitations of scRNA-seq but has its own limitations¹⁷. Existing sc/snRNA-seq studies only profiled a few individuals with the discrete statuses of obesity and diabetes. Thus, the cell-type-specific signatures of molecular networks of insulin resistance have not been revealed in the adipose and muscle tissues.

Recent advances in network biology provide a scalable and effective avenue to decode molecular interactions and determine the cell-type specificity in bulk tissue transcriptomic data from human populations. As gene activities in each cell-type are regulated and coordinated as a system, this cellular system can be mathematically modeled in a coexpression network, where highly co-regulated genes are embedded in highly connected modules and interact with each other through network “hubs”¹⁸. Thus, instead of considering each gene in isolation, the gene co-expression network portrays the landscape of gene activities and regulations of different cell types in the context of network modules. To reveal the fine structure of gene coexpression networks, we recently developed Multiscale Embedded Gene Coexpression Network Analysis (MEGENA)¹⁹. Compared with the large and sparse modules generated by the traditional weighted gene co-expression network analysis (WGCNA)²⁰, empirical evaluations using simulated and real-world large-scale gene expression data established that MEGENA is capable to resolve fine hierarchical network structures and produce more compact and coherent modules^19,21–23. The high resolution of hierarchical network structures at different compactness scales enables MEGENA to efficiently infer cell-type-specific modules and further identify network drivers in large-scale bulk tissue transcriptomes of complex human diseases^23–26. While coexpression networks can identify the global patterns of interacting genes directly or indirectly involved in a biological system, they are not capable of defining and identifying causal relationships. By integrating genetic, transcriptomic, and other high throughput data into Bayesian probabilistic causal networks, we have previously discovered novel causal relationships in complex human diseases such as diabetes and obesity^27–30, Inflammatory bowel disease³¹, cancer³², and Alzheimer’s disease^23,33,34, and have elucidated potential mechanisms of these causal relationships. Integration of gene coexpression and causal networks (termed multiscale network analysis (MNA)) can further reveal module-specific drivers^23,33.

In this study, we applied the well-established MNA to study the cell-type-specific transcriptomic networks and their regulations in glucose-responsive tissues of human populations. Our analyses were motivated by the following scientific questions: 1) What are the landscapes of cell-type-specific gene signatures, gene coexpression and regulatory networks, in response to insulin resistance in the adipose and muscle tissues? 2) How are cell-type-specific signatures preserved in different human populations and across different glucose-responsive tissues? 3) How cell-type-specific gene modules are regulated by genetic variants such as expression quantitative trait loci (eQTLs)? As large-scale genetic analyses in African and European ancestry cohorts suggested heterogeneous genetic regulatory mechanisms in the prevalence of insulin resistance among ethnic groups³⁵, we hypothesize that major trends in cell-type specificity for insulin resistance would be reflected in well-curated cohorts from diverse populations and would be helpful in the cross-validation of our findings. Therefore, we performed multi-omic network analyses of adipose and muscle tissues in different ethnic populations: the African American Genetics of Metabolism and Expression cohort (AAGMEx, n = 256)³⁶, the Metabolic Syndrome in Men cohort (METSIM, n = 770)³⁷, and the Arkansas European-American (AREA, n = 99) cohort³⁸. By network preservation analysis and integration of related sc/snRNA-seq datasets, we systematically identified insulin resistance-associated gene signatures, cell-type-specific gene modules and drivers in major glucose-responsive tissues of human populations.

Results

A pipeline of cell-type-specific network analysis in adipose and muscle tissues

Our analyses include two glucose-responsive tissues from three independent human cohorts: the adipose and muscle tissues from AAGMEx and AREA, and the adipose tissue from METSIM (Figure 1 and Table S1). Participants in the AAGMEx cohort were African Americans, while the participants in METSIM and AREA cohorts were European ancestry individuals (Finnish European and European Americans, respectively). In these cohorts, two numeric indices were calculated to quantify insulin sensitivity: the S_I from insulin modified-frequently sampled intravenous glucose tolerance test (FSIGT) and the Matsuda index from 75g-oral glucose tolerance test (OGTT). Matsuda index is a well-validated dynamic measurement of insulin sensitivity comparable to Rd (rate of disappearance of plasma glucose) measured by insulin clamp³⁹. In AAGMEx, Matsuda index showed a significant positive correlation with S_I (r = 0.57, p = 1.1 × 10⁻²⁰)³⁶. As both S_I and Matsuda index are available in AAGMEx and AREA cohorts, and only Matsuda index is available in the METSIM cohort, we defined insulin sensitivity-associated modules (IS-modules) as the modules enriched for Matsuda index- and/or S_I-correlated genes (termed insulin sensitivity-correlated genes).

Figure 1. The schematic study design and analysis workflow to construct and compare gene networks to define molecular mechanisms of insulin resistance.

The network modeling pipeline includes tissue sampling from different human cohorts, physiological and omics profiling, coexpression and Bayesian network construction, cell-type landscape analysis, and network driver discovery.

Our network analysis pipeline consists of the following procedures (Figure 1): first, we performed MEGENA to resolve network structures and identify the cell-type landscape. Three coexpression networks were constructed: the adipose and muscle networks of AAGMEx, and the adipose network of METSIM. The AREA cohort, which contains adipose and muscle tissue expression data on a smaller number of samples, was primarily used for the validation of network preservations. The IS-modules and cell-type-specific modules were identified by enrichment analysis of the Matsuda index-correlated genes and the cell-type marker genes from sc/snRNA-seq datasets, respectively. Cell-type-specific modules and IS-modules were compared across different ethnic groups and glucose-responsive tissues. Ethnically conserved and specific modules were identified by the network preservation analyses. Second, by combining genotypic and gene expression data, eQTL enrichment and Bayesian probabilistic causal network analysis were performed to detect causal associations and network driver genes in the adipose and muscle tissues. Finally, we conducted in vitro experiments in relevant cell models to validate the cell-type-specific signatures and regulatory roles of network driver genes.

It is noted that MEGENA utilizes the same naming system for the modules in each of the three coexpression networks and thus many module names are shared by multiple networks. To avoid any confusion, we clarify the tissue and cohort sources when discussing modules.

Cell-type specificity of coexpression modules in the adipose tissue of African Americans

We performed MEGENA on the adipose tissue transcriptomic data of AAGMEX and identified gene modules of various compactness scales, leading to a hierarchy of parent and child modules (Figure 2A). The coexpression network in African Americans harbored 16,198 genes and 536 modules. Module eigengene (the first principal component of module gene expression profiles) was used to calculate the correlation between a module and the gluco-metabolic phenotype. Among the 536 gene modules, 263 were significantly correlated with insulin sensitivity (Matsuda index or S_I) (adjusted p-value (Padj) < 0.01, Spearman correlation; Figure S1A and Table S2). Fisher’s exact test (FET) identified 167 IS-modules enriched (Padj < 0.01) for Matsuda index- or S_I-correlated genes. The top-ranked IS-modules included three major hierarchical modules: M6 and M9, and their child modules were negatively correlated with insulin sensitivity, while M5 and its child modules were positively correlated (Figure 2A). Pathway enrichment analysis revealed important biological functions associated with the top-ranked IS-modules (Figure 2B). For example, the modules negatively correlated with insulin sensitivity were involved in immune response and hormone response, while the modules positively correlated with insulin sensitivity were associated with mitochondrial and metabolic pathways.

Figure 2. Characterization of coexpression network modules in adipose tissues of African and European ancestry individuals.

(A-C) The network modules and cell-type specificity of the AAGMEx cohort. (A) Sunburst plot showing the hierarchical structures of the modules and the correlation between module eigengenes and Matsuda index. The color scale shows the Spearman correlation coefficients. (B) Bar graph showing the biological pathways most enriched in the top 15 IS-modules in the adipose network. The pathways most significantly enriched in each module are shown (FET false discovery rate (FDR) < 0.05). The bar color intensity is proportional to the Spearman correlation coefficients between module eigengenes and Matsuda index. (C) Heatmap showing enrichment of cell-type specific marker genes in the adipose gene modules (FET FDR < 0.01) and histogram on the top shows the Spearman correlation coefficients between module eigengenes and Matsuda index. The red asterisks indicate major IS-modules for downstream characterization. (D-F) The network modules and cell-type specificity of the METSIM cohort.

Single-nucleus RNA-sequencing data helps identify the cell-type-specific transcriptomic architecture of most cell types in the adipose tissue, including highly fragile mature adipocytes. Using the enrichment analysis of these gene modules and cell-type marker gene signatures from the snRNA-Seq dataset (n = 57,599 nuclei)¹⁵, we identified the cell-type-specific network modules in the adipose tissue of AAGMEx (Figure 2C and Table S3). Among the top-ranked IS-modules, M5 and M9 specifically expressed adipocyte marker genes, but in the opposite directions of correlations with insulin sensitivity. M5 contained 129 adipocyte markers which were positively correlated with insulin sensitivity while M9 contained 91 adipocyte markers which were negatively correlated with insulin sensitivity (Table S4), suggesting that these modules captured the two major trends of adipocyte markers in transcriptional response to insulin resistance. M5 and M9 contained the adiponectin gene encoding ADIPOQ and the leptin encoding gene LEP, respectively, indicating that these modules are involved in different sets of adipocytokine activities. Apart from adipocyte modules, M6 preferentially expressed the marker genes from the adipose stem and progenitor cells (ASPCs) and immune cells, and was negatively correlated with insulin sensitivity. To avoid potential bias from a single snRNA-Seq reference, we also analyzed the cell-type-specific modules from the PanglaoDB database which compiled marker genes from 178 cell types of 1368 scRNA-seq samples⁴⁰. Consistent with the module cell-type inference from the snRNA-Seq dataset, cell-type enrichment with the PanglaoDB database also showed that M5 was enriched for the adipocyte markers, and M6 was enriched for immune cell marker genes of macrophages and dendritic cells (Figure S1B).

Apart from enrichment analysis of cell-type marker genes, we determined the cell-type identity of AAGMEx adipose network modules by directly mapping these modules in single nuclei. We performed the gene set enrichment analysis (GSEA) to identify the nuclei that preferentially expressed the module genes. Among the total 57,599 nuclei in snRNA-seq, genes from M5, M9, and M6 were preferentially expressed in 10,784, 5993, and 13,237 nuclei, respectively (Padj < 0.05 and enrichment score > 0, 1,000 permutations by GSEA) (Figure S1C–E). Among them, 99.5% of the M5-enriched nuclei and 94.3% of the M9 module-enriched nuclei were adipocytes, suggesting M5 and M9 genes specifically expressed in adipocytes. For the M6-enriched nuclei, 53.8%, 12.5%, and 11.0% were from macrophages, T cells, and ASPCs, respectively. The nuclei expressing M9 and M6 module genes were preferentially derived from obese individuals (Figure S1F and Table S5). As M9 and M6 were negatively correlated with insulin sensitivity, their preferential expressions in obese individuals may suggest that the expressions of these module genes are associated with obesity-induced insulin resistance.

Cell-type specificity of coexpression modules in adipose tissue of European ancestry individuals

To validate our findings from African ancestry individuals described above, we further analyzed cell-type-specific transcriptomic signatures of insulin resistance in adipose tissue of European ancestry individuals. The MEGENA network in the adipose tissues of the METSIM participants identified 506 hierarchical modules from 18,041 genes (Figure 2D). Among these network modules, 167 were IS-modules enriched for the Matsuda index-correlated genes (Figure S2A and Table S6). The top-ranked modules were distributed into several hierarchical clusters with different transcriptional responses to insulin resistance. The module M12 and its child modules were positively correlated with the insulin sensitivity index and mainly participated in mitochondrial and metabolic pathways (Figure 2E). In contrast, modules M15 and M19 were negatively correlated with the insulin sensitivity index and involved in immune response and extracellular matrix processes.

Integration of the METSIM adipose gene modules and the snRNA-seq data¹⁵ identified the modules enriched for the marker genes of major cell types in the adipose tissue (Figure 2F and Table S7). Similar to the AAGMEx cohort, the adipocyte-enriched modules from the METSIM population showed two opposite directions of insulin sensitivity correlations: M12 and its child modules were positively correlated with insulin sensitivity, while M19 exhibited a significant negative (inverse) correlation. M12 contained 124 adipocyte markers with positive correlations of insulin sensitivity, and M19 contained 20 adipocyte markers with negative correlations (Table S8). M12 and M19 also contained cytokine ADIPOQ and LEP, respectively, suggesting opposite transcriptional responses of the two adipocytokines to insulin resistance. Apart from adipocyte modules, the ASPC and immune cell-enriched modules showed negative correlations with insulin sensitivity. The cell-type specificity was also confirmed by the enrichment analysis of cell-type markers from the PanglaoDB database. For example, M12 was significantly enriched for the adipocyte marker genes and M15 was enriched for the cell-type markers of immune cells from the PanglaoDB database (Figure S2B).

We further applied GSEA to determine the cell-type specificity of the top-ranked adipose IS-modules in METSIM by directly mapping them to single nuclei. We identified 11,319, 2970, and 10,921 nuclei of the snRNA-seq dataset which preferentially expressed M12, M19, and M15-module genes, respectively (Figure S2C–E). Among them, 97.9% and 74.7% of M12 and M19-enriched nuclei belonged to adipocyte cells, respectively. The majority of M15-enriched nuclei came from macrophages (44.3%), T cells (23.1%), and NK cells (11.7%). Consistent with the negative correlations with insulin sensitivity, the M19 and M15-enriched nuclei were more likely observed in obese individuals (Figure S2F and Table S9).

The preservation and cell-type signatures of adipose gene modules across ethnic groups

As many adipose modules of the AAGMEx and METSIM cohorts shared similar functions and cell-type specificity, we evaluated the conservation of the adipose coexpression networks in two ethnic groups. Module preservation analysis by network connectivity and density⁴¹ identified 32 adipose IS-modules in AAGMEx preserved in the adipose network of METSIM (Z score > 10; Figure 3A and Table S10). Module gene enrichment by FET showed that 93 IS-modules in AAGMEx significantly overlapped with 109 IS-modules in METSIM (Table S11). Unsupervised hierarchical clustering of the top-ranked modules from the module overlap metrics of FET significance score (−log10(FDR)) identified three mutually conserved module clusters that corresponded to the major IS-modules in each cohort (Figure 3B). The first conserved module cluster was negatively correlated with insulin sensitivity, including AAGMEx M6 and METSIM M15. These two modules shared 726 immune-related genes involved in macrophage activation, inflammatory response, and cytokine secretion (Figure S3A). The second conserved module cluster contained several modules positively correlated with insulin sensitivity, such as AAGMEx M5 and METSIM M12. These modules were adipocyte-specific and shared 405 genes enriched for the GO processes necessary for adipocyte metabolic functions, such as lipid oxidation, glucose metabolic process, and amino acid metabolism (Figure S3B). The third conserved module cluster had three modules, including AAGMEx M9 and METSIM M19. These modules also had the adipocyte identity but were negatively correlated with insulin sensitivity. Among 61 genes shared by the two modules, 4 genes (UCHL1, LEP, OXT, NPY5R) are related to eating behavior and 7 genes (AKR1C3, SPARC, CALM1, LOX, OXT, ADM, RHOXF1) participate in steroid hormone pathways (Figure S3C).

Figure 3. Preservation of the insulin sensitivity-associated gene modules in adipose tissues of different ethnicities.

(A) Scatter plot showing network preservation (Z score) of the AAGMEx adipose gene modules in the adipose tissue of METSIM Europeans. (B) Heatmap plot showing unsupervised hierarchical clustering of top 20 IS-modules in the adipose from AAGMEX and METSIM cohorts. The color scale indicates the log-transformed FET significance score of the module overlap matrix. The color bars on the top and right of the heatmap show the Spearman correlation coefficients between module eigengenes and Matsuda index. The red asterisks indicate the major IS-modules from the two cohorts. (C) Proportion bar plot showing the distribution of adipocyte nuclei in different subclusters in the donors with different BMI categories. (D-E) Bar plot showing enrichment of module-specific nuclei in the adipocyte subclusters in AAGMEx (D) and METSIM (E). For each module, GSEA was used to determine module-specific nuclei of subcutaneous adipocytes. FET was used for the enrichment test.

As the module preservation analysis revealed that adipocytes contain two distinct module clusters with opposite directions in correlation with insulin sensitivity (Figure 3B), we further investigated whether the module clusters were from different adipocyte subtypes. Louvain algorithm-based clustering analysis of single nuclei of subcutaneous adipocytes (n = 14,396) identified 10 subclusters of adipocytes (Figure 3C and S4A). Based on the distribution of BMI categories, the majority of the nuclei in subcluster 5 were from extreme obesity individuals (BMI: 40–50). GSEA identified 678 and 959 nuclei enriched for conserved modules AAGMEx M9 and METSIM M19 (Figure S4B–C), respectively, which were preferentially distributed in the adipocyte subcluster 5 (Figure 3D–E). In contrast, 525 and 729 nuclei were enriched for conserved modules AAGMEx M5 and METSIM M12, respectively, and mainly distributed in adipocyte subcluster 1. These observations suggest that the two conserved module clusters originated from distinct subtypes of adipocytes, which were associated with different degrees of obesity.

Based on the established relationship of ligand-receptor pairs⁴², we performed a statistical test to identify the significant interactions or communications between the top preserved network modules. Both AAGMEx and METSIM coexpression networks in the adipose tissue displayed mutual cellular communications between adipocytes and immune cells. For example, strong cellular communications were detected from M6 to M9 and M5 to M6 in the adipose tissue of AAGMEx (Figure S4D). Similarly, the modules of the adipose tissue in METSIM showed strong cellular communications from M19 to M15 and M15 to M19. In the METSIM adipose gene coexpression network, the TGF-β, BMP, and WNT signaling factors in M19 bind to their corresponding receptor genes (e.g., TGFBR1, ACVR2B, and FZD2) in M15 (Table S12). Conversely, the semaphorins (SEMA3F and SEMA4A) and cytokines (CCL4 and CCL5) secreted from M15 were predicted to bind to their receptors (PLXNA1, ACKR2) in the adipose module M19 in METSIM.

We also validated the module preservation in a smaller European ancestry cohort (AREA with 99 adipose samples). Among the 32 preserved adipose IS-modules of AAGMEx, 23 were also preserved in the adipose gene coexpression network of the AREA cohort, suggesting these modules are strongly preserved or trans-ethnic modules involved in insulin resistance in both African and European ancestry individuals (Figure S5A). Apart from preserved adipose modules, we also identified a few coexpression modules specific to each population. For example, 5 IS-modules from AAGMEx showed a Z score < 2 compared with the METSIM cohort, and 8 METSIM IS-modules showed a Z score < 2 compared with the AAGMEx, suggesting they are potential ethnic-specific IS-modules. The lowest preservation of AAGMEx adipose modules in European ancestry individuals were observed for M316 (Z score = 0.27 and 2.3 in METSIM and AREA cohort, respectively) and M382 (Z score = 1.3 and 2.8 in METSIM and AREA cohort, respectively). M316 contains 44 genes with JPH2 as a hub gene (Figure S5B), and M382 contains 30 genes with hub genes UQCRHL and UQCRC2 (Figure S5C). These two IS-modules showed positive and negative correlations with insulin sensitivity, respectively, but they were not enriched for any known biological pathways.

Cell-type specificity of gene coexpression modules in the muscle tissue

Given that most insulin-mediated glucose disposal occurs in muscle, dysregulated gene expression in skeletal muscle likely plays an important role in insulin resistance. Thus, we performed MEGENA on the muscle transcriptomic profile data in AAGMEx. The muscle gene coexpression network in AAGMEx included 523 modules from 14,335 genes (Figure 4A). Among the 87 modules significantly correlated with insulin sensitivity (Matsuda index or S_I), 63 were enriched for the S_I- or Matsuda index-correlated genes (Padj <0.01), including the top-ranked IS-modules M5, M95, M97, and M8 (Figure S6A and Table S13). The majority of the top-ranked IS-modules in the muscle tissue, except for M8, were positively correlated with the Matsuda index and involved in the ribosome and protein translation process, while M8 mainly functioned in the mitochondria process (Figure 4B).

Figure 4. Characterization of the modules in the muscle tissue coexpression network for the African Americans in the AAGMEx cohort.

(A) Sunburst plot showing the hierarchical structures of the modules and the correlation of module eigengenes with S_I and Matsuda index. The color scale shows the Spearman correlation coefficients. (B) Bar graph showing the biological pathways most enriched in the top 20 IS-modules in the muscle network. The pathway most significantly enriched in each module is shown (FET FDR<0.05) and the bar color intensity is proportional to the Spearman correlation coefficients between module eigengenes and Matsuda index. (C) Heatmap showing enrichment of cell-type specific marker genes in the muscle gene modules (FET FDR < 0.01) and histogram on the top shows the Spearman correlation coefficients between module eigengenes and Matsuda index. The red asterisks indicate the major candidate modules for downstream characterization.

To determine the cell-type specificity of the co-expressed gene modules in the muscle tissue, we analyzed single-cell transcriptome profiles (n = 22,000 cells) of the human muscle tissue⁴³ and identified marker genes of different muscle cell types. The integration of the muscle modules and the cell-type marker genes revealed different cell-type specificity for the modules responsive to insulin sensitivity (Figure 4C and Table S14). For example, M97, which was positively correlated with insulin sensitivity, was significantly enriched for the marker genes of muscle stem cells; M91 and its child modules (e.g., M319), which were negatively correlated with insulin sensitivity, preferentially expressed the fibro-adipogenic progenitor markers; M8 showed a negative correlation with insulin sensitivity, and was significantly enriched for marker genes from muscle fiber cells.

We further performed the preservation analysis of muscle modules from the AAGMEx cohort against the AREA cohort. The cell-type-specific modules in AAGMEx were strongly preserved in the AREA cohort (Figure S6B and Table S15): M91 and M8 were highly preserved in the AREA muscle transcriptomic data with Z score = 27.9 and 17.1, respectively, while M97 was moderately preserved with Z score = 7.7. We also found that 17 IS-modules in the muscle of AAGMEx were not preserved (Z score < 2) in the muscle tissue of AREA European-Americans. For example, the Z score was 1.25 for module M362 (194 genes, rank 4^th) (Figure S6C).

Concordant and tissue-specific regulations of module genes in the adipose and muscle tissues

The matched adipose and muscle transcriptomic data from the same set of individuals in AAGMEx allowed a direct comparison of gene activities across the two insulin-responsive tissues. We performed a cross-tissue module enrichment analysis of insulin sensitivity-correlated genes (ISGs). For the adipose tissue of AAGMEx, 3,429 and 3,185 genes were positively and negatively correlated with the Matsuda index, respectively (Spearman correlation Padj < 0.05). These adipose ISGs were significantly enriched in 45 and 38 gene modules of the muscle network, respectively (Figure 5A and Table S16). Among the cell-type-specific modules, M8 and M97 in the muscle network were enriched for the positive ISGs in the adipose, and M91 was enriched for the negative adipose ISGs. For modules M97 and M91, their insulin sensitivity correlations were consistent with the directions of adipose ISGs. However, while the eigengene of the muscle module M8 was negatively correlated with the Matsuda index, positive ISGs of the adipose tissue were enriched in M8, suggesting opposite characteristics of gene response to insulin sensitivity in the two tissues.

Figure 5. Enrichment of AAGMEx muscle gene modules for the adipose ISGs and cell-type-specific marker genes.

(A) The muscle modules enriched for the adipose ISGs in AAGMEx. The pink and blue colors indicate the muscle modules enriched for the positive and negative adipose ISGs, respectively (FET FDR < 0.05). The x-axis shows the Spearman correlation coefficients between module eigengenes and Matsuda index, and the y-axis shows the significant levels of enrichment. The black rectangles indicate three cell-type specific modules for enlarged network plots in (B-D). Network plots of three cell-type specific modules in AAGMEx muscle coexpression network: (B) M8; (C) M97; (D) M91. The cell-type marker genes are labeled for each module (M8, muscle fibers; M97, muscle stem cells; M91, fibro-adipogenic progenitors). The pink and blue colors of node labels indicate the positive and negative correlations with Matsuda index in the muscle tissue, respectively (Spearman correlation p < 0.05). The pie chart of the network nodes shows the Matsuda correlation trends in adipose and muscle tissues: whole blue, both negative correlations; whole pink, both positive correlations; half blue half pink, opposite correlation directions.

We further analyzed the cell-type marker genes in response to insulin sensitivity in the adipose and muscle tissues. The muscle module M8 contained 21 muscle fiber marker genes that were negatively correlated with the Matsuda index (Figure 5B). Meanwhile, these cell-type marker genes showed a difference in response to insulin sensitivity in the adipose: 13 genes in the adipose showed positive correlations with Matsuda Index, and 6 genes were not correlated with the Matsuda index. At the network level, the muscle module M8 was most similar to the adipose module M5 and they shared 205 genes, among which 47 genes showed opposite directions of insulin sensitivity correlation in the adipose and muscle tissues, and only 6 had the same insulin sensitivity correlation. Pathway analysis revealed that the 47 genes with opposite insulin responses were most enriched for the cellular respiration process. Specifically, 16 genes associated with the cellular respiration process were positively and negatively correlated with insulin sensitivity in the adipose and the muscle, respectively (Figure S7), suggesting tissue-specific differences in the dysregulation of mitochondrial functions in response to insulin resistance. In contrast to the muscle module M8, the cell marker genes of M97 and M91 showed the same trend of insulin sensitivity response in the adipose tissue. M97 contained 20 muscle stem cell marker genes associated with insulin sensitivity, and 13 were positively correlated with the Matsuda index in both adipose and muscle tissues (Figure 5C). M91 contained 72 marker genes of fibro-adipogenic progenitors associated with insulin sensitivity, and the majority (n = 49) of them were negatively correlated with the Matsuda index in both adipose and muscle tissues (Figure 5D).

Genetic regulations of cell-type-specific gene modules

Genetic variants (e.g., SNPs) contribute to altered gene expression in complex diseases including diabetes and its intermediate phenotypes. Here we investigate how genetic variants potentially regulate gene modules underlying insulin resistance. Cis-eQTL analysis determined the association between the gene expression level and the genotype of local SNPs (within ± 1Mb of a gene) in the adipose and muscle tissues of AAGMEx participants. We applied the enrichment test to integrate cis-eQTL genes and coexpressed gene modules. Eight modules in the adipose gene coexpression network and 24 modules in the muscle gene network were enriched (FET Padj < 0.05) for the respective cis-eQTL genes (Figure 6A–B). For example, the adipocyte-specific gene modules M5 and M9 in the adipose gene network were enriched for the cis-eQTL genes in the adipose tissue (Table S2), while the fiber-specific module M8 in the muscle gene network showed the strongest enrichment for the cis-eQTL genes (Table S13).

Figure 6. Enrichment of genetically regulated genes in gene modules and driver genes’ network neighborhoods.

Scatter plot showing cis-eQTL gene enrichment in (A) adipose and (B) muscle modules in AAGMEx. Spearman correlation coefficients between module eigengenes and Matsuda index are shown on x-axis and the FET enrichment significance of cis-eQTL genes are plotted on y-axis. (C) Scatter plot showing cis-eQTL gene enrichment in network neighborhood (2-layers) of network driver genes in adipose modules in AAGMEx. FET odds ratios and enrichment significance of cis-eQTL regulated genes are plotted on the x- and y-axis, respectively. (D) Driver-centric circular network plot showing eQTL enrichment in network neighborhood (2-layer) of adipose driver gene LEP. In the network plots, the orange circles are the eQTL genes and the blue triangles indicate the genes without eQTLs. The eQTL genes with network connectivity above four are labeled. (E) Scatter plot showing cis-eQTL gene enrichment in network neighborhood (2-layers) of network driver genes in muscle modules in AAGMEx. (F) Driver-centric circular network plot showing eQTL enrichment in network neighborhood (2-layer) of muscle key driver gene CKMT2.

Bayesian probabilistic causal network analysis is a powerful approach to detecting causal associations by combining genotypic and gene expression data in segregating populations^23,44. To identify potential network regulators among co-expressed genes, we constructed a Bayesian probabilistic causal network by integrating genetic variants (SNPs), gene expression, and known transcription factor-target relationship (see Methods). By projecting the module genes from MENEGA onto the Bayesian network, potential regulators (or termed “network drivers”) were identified as the nodes whose network neighborhoods were enriched for the module genes^33,45,46. Bayesian network analysis identified 284 network driver genes in the adipose tissue in AAGMEx. Seven of the 284 network drivers, including leptin encoding gene LEP, were enriched for the cis-eQTL genes in the MEGENA network neighborhood (FET Padj < 0.05 for 2 network layers of driver genes) (Figure 6C–D; Table S17), suggesting these network drivers were strongly connected with genetic variant-mediated regulations. In the AAGMEx muscle tissue, we identified 197 network drivers. Among them, 14 driver genes were enriched for the cis-eQTL genes in the network neighborhood, including the most significant enrichment around the creatine kinase mitochondrial 2 (CKMT2, Padj = 4.74 X 10⁻⁷), a sarcomeric mitochondrial gene crucial to energy metabolism (Figure 6E–F; Table S18). Altogether, these network analyses revealed cell-type-specific network modules and their potential regulators in the adipose and muscle tissues.

GWASs have successfully identified many loci associated with complex diseases including cardiometabolic diseases and other phenotypes linked to insulin resistance. However, most of these disease-associated genetic variants (SNPs) are noncoding and in the linkage disequilibrium blocks, which challenges the ability to pinpoint disease genes solely based on GWAS findings. To address the limitation, previous studies developed approaches to combine layers of functional genetic information (e.g. eQTL, enhancer-gene linking epigenetic map, promoter capture Hi-C based chromatin interaction map, and ATAC-seq based open chromatin map) to link disease-associated SNP to Gene (S2G)^47,48. A recent study applied a combined S2G score (cS2G) to nominate target genes by testing 50 S2G linking strategies for SNPs associated with a large number of complex diseases and traits⁴⁹. We determined if network driver genes for insulin sensitivity-associated modules in our study were also target genes (with cS2G score >0.5) of causal SNPs associated with insulin resistance-related phenotypes in the GWAS catalogues and UK biobank studies (Table S19). We found that 188 and 20 network driver genes of insulin sensitivity-associated modules in adipose and muscle tissues, respectively, were target genes of cardio- and gluco-metabolic disease-associated SNPs from GWAS. For example, NQO1, a driver gene of adipose tissues in AAGMEx and METSIM cohorts, is a target gene of a T2D-associated SNP rs2032912. Similarly, MRPL33, a driver gene of AAGMEx muscle tissue, is a target gene of fasting blood glucose-associated SNP rs3736594. Thus, the network driver genes of the insulin sensitivity-associated network modules in our study can capture causal mechanisms revealed by GWAS of cardiometabolic diseases.

Experimental validation of cell-type-specific expressions and biological functions of network regulators

By integrating coexpression and Bayesian probabilistic causal networks, we identified 114 genes as MEGENA module hub genes or Bayesian network drivers in conserved IS-modules of the adipose tissues in African American and European ancestry (Table S20). Among the top-ranked genes correlated with insulin sensitivity in adipose IS-modules, we selected five genes representative of different cell types for in vitro experimental validations: HADH (Hydroxyacyl-CoA dehydrogenase), ACADM (acyl-CoA dehydrogenase medium chain), and ALDH6A1 (aldehyde dehydrogenase 6 family member A1) were from the adipocyte-specific module M5 in AAGMEx adipose coexpression network; NPL (N-acetylneuraminate pyruvate lyase) and LAPTM5 (lysosomal protein transmembrane) were from the immune-specific module M6. We performed qRT-PCR assays to evaluate their cell-type specificity in the adipocyte- and stromal vascular- fraction (AF and SVF) of the human subcutaneous adipose tissue. Compared to SVF, the expression of ACADM, ALDH6A1, and HADH was 4.7, 3.5, and 11.7 fold higher, respectively in AF (Figure 7A). In contrast, the expression of LAPTM5 and NPL was 26.9 and 14.2 fold higher, respectively in SVF compared to AF (Figure 7B). We further tested the expressions of these genes in respective cell models of in vitro human adipocyte and macrophage differentiation. Expressions of ACADM, ALDH6A1, and HADH in human adipose-derived stem cells (hADSCs) gradually increased throughout the stages of differentiation of mature adipocytes, leading to a 7.07, 6.07, and 20.4 fold increase compared to undifferentiated preadipocytes (Figure 7C). THP1 is a human monocytic leukemia cell line and is a well-validated model for the study of monocyte-macrophage differentiation^50,51. Compared to undifferentiated THP1 monocytes, expressions of LAPTM5 and NPL strongly increased (5.13, and 6.56 fold for LAPTM5 and NPL, respectively) at 24hrs of differentiation of THP1 macrophages (Figure 7D). Consistent with the cell-type predictions of the network modules, these results confirmed the preferential expressions of ACADM, ALDH6A1, and HADH in adipocytes, and the preferential expressions of LAPTM5 and NPL in macrophages of SVF.

Figure 7. Experimental validations of network module genes and regulators.

Relative expressions of ACADM, ALDH6A1 and HADH (A), and LAPTM5 and NPL (B) in the AF and SVF of subcutaneous adipose tissues in Africa Americans. The bar plot shows the mean and standard deviation of relative expression data normalized to RPLP0 (36B4) endogenous control gene. The circles indicate three technical replicates of the pool RNA samples from 5–6 individual donors. The p-values were calculated by Student’s t-test. (C) Expression of ACADM, ALDH6A1, and HADH at 0, 2, 7, and 14 days of in vitro differentiation of hADSC. Relative mRNA expression data were normalized to RPLP0 for three biological replicates. (D) Expression of LAPTM5 and NPL in undifferentiated THP1 monocytes and at 6, 24, 48, and 72 hours of Phorbol ester (PMA/TPA, 10ng/ml) induced macrophage differentiation. Relative mRNA expression data were normalized to PPIA endogenous control gene for three biological replicates. (E) NPL expression determined by qRT-PCR in knockdown and control THP1 cells and in PMA induced (10ng/ml, 48 hours) macrophage differentiation state. Relative mRNA expression data with biological triplicates were normalized to PPIA. (F) Volcano plot showing the differentially expressed genes (FDR < 0.05, Fold change > 1.2). (G) Top five enriched pathways for the up- and down-regulated DEGs in differentiated THP1 macrophages (48 hours PMA treatment) expressing NPL-shRNA. FET was used for the enrichment test. (H) Network showing the DEGs from NPL knockdown THP1 cells in 2-layer neighbors of the NPL gene in the MEGENA network of AAGMEx adipose tissue. Genes downregulated and upregulated in NPL-shRNA transduced THP1 cells are shown in blue and red, respectively.

The NPL gene encodes an enzyme that regulates cellular concentrations of sialic acid (N-acetyl-neuraminic acid) by mediating the reversible conversion of sialic acid into N-acetylmannosamine and pyruvate⁵². As the functions of NPL gene in obesity and gluco-metabolic phenotypes have not been studied, we knocked down NPL in THP1 cells to understand its role in modulating the human monocyte-macrophage expression network. Transduction of THP1 cells by NPL-specific lentiviral shRNA stably knocked down its expression at baseline monocytes and in the PMA-induced macrophage state (Figure 7E). Global transcriptomic analysis by RNA-seq validated the downregulation of NPL, and the comparison of NPL knockdown cells with control-shRNA treated cells further identified 1,183 differentially expressed genes (DEGs) (Figure 7F and Table S21). Genes downregulated by the NPL knockdown were significantly enriched for cytokine production, while upregulated genes were enriched for extracellular structure organization (Figure 7G and Table S22). These DEGs were enriched in several adipose IS-modules, including M6 in AAGMEx. The network neighborhood genes of NPL in the AAGMEx adipose MEGENA network were also significantly enriched for these DEGs induced by knockdown of NPL in THP1 cells, including 25 DEGs in the 2-layer network neighborhood around NPL (Figure 7H). When SGBS adipocytes (on day 4 of differentiation) were treated with macrophage conditioned media (MCM) derived from NPL-shRNA-THP1 cells, peroxisome proliferator activated receptor gamma (PPARG) was less repressed and LEP was less activated (Figure S8), suggesting knockdown of NPL influenced macrophage-induced gene expressions of adipocytes.

Discussion

Although insulin resistance has been extensively studied for decades, little has been done to elucidate its cell-type-specific organizations and gene expression regulations. In this study, we applied a multi-omic network approach and identified a comprehensive cell-type landscape of IS-modules in adipose and muscle tissues. Distinct from the scRNA-seq experiments that suffer from expensive costs and small cohort size, our approach directly models gene interactions and cell-type specificity from network modules of bulk tissues. Taking advantage of well-curated human populations from different ancestry, our study provided a systematic cross-population and cross-tissue overview of IS-modules. Many top-ranked IS-modules were preserved across different ethnic cohorts and shared cell-type-specific signatures. Meanwhile, distinct regulations of cellular respiration-related genes were observed in adipocytes and muscle fibers in response to insulin resistance. A small number of putatively cohort-specific IS-modules of unknown biological significance were also detected. By integrating Bayesian and MEGENA networks, we identified driver genes for the cell-type-specific IS-modules and potential regulations by genetic variants. A subset of these driver genes were targets for cardio- and gluco-metabolic disease-associated SNPs from GWAS and indicate causal mechanisms of cardiometabolic diseases. Overall, our results provide a comprehensive population-level understanding of the organizations and cell-type-specific regulations of gene expressions underlying insulin resistance.

We employed three different approaches to validate the robustness of the key findings: First, the preservation analyses, including module preservation and enrichment tests, showed that the top-ranked adipose and muscle modules could be captured in human populations of two different ethnicities. Independent coexpression networks from ethnically diverse groups identified similar cell-type signatures in the top-ranked IS-modules, suggesting the robustness of network modules underlying insulin resistance. Second, the cell-type-specific signatures were supported by various scRNA-seq datasets. Different sources of single-cell datasets, including the PanglaoDB single-cell database and sc/snRNA-seq experiments on human adipose and muscle tissues, collectively confirmed cell-type signatures of the co-expressed modules. Third, the in vitro experiments validated expression patterns of selected genes from the cell-type-specific modules. Consistent with network predictions, qRT-PCR experiments showed that HADH, ACADM, and ALDH6A1 from the adipocyte-enriched module (AAGMEx M5) were preferentially expressed in adipocyte fraction and induced during the stages of adipocyte differentiation. Meanwhile, NPL and LAPTM5 from the immune-related module (AAGMEx M6) were mainly detected in stromal vascular fraction and upregulated during macrophage differentiation. Additionally, knockdown of NPL in macrophages influenced macrophage-induced gene expressions of adipocytes. These observations demonstrated that the network approach identified cell-type stratified and functionally meaningful modules, which reflected the rearrangement of gene expressions associated with insulin resistance.

Our study revealed different behaviors of network modules in response to insulin resistance in glucose-responsive tissues. We found that a subset of adipocyte-enriched modules was positively correlated with insulin sensitivity (Matsuda index and S_I), while the adipocyte progenitor and immune-enriched modules were negatively correlated with insulin sensitivity. A recent deconvolution analysis, which was based on bulk adipose tissue RNA-seq data of 331 individuals and snRNA-seq experiments of 13 individuals, reported a similar trend in the cell-type association with BMI¹⁵. These results support our observations on the cell-type proportion changes in insulin resistant individuals. Mature white adipocytes are generally considered monotypic cells with uniform function. Interestingly, our analyses showed that adipocytes were enriched for two types of modules (e.g., M5 and M9 in AAGMEx, M12 and M19 in METSIM) with different responses to insulin sensitivity. These modules contained adiponectin and leptin encoding genes, ADIPOQ and LEP, respectively. Compared with the ADIPOQ-containing modules, the LEP modules were negatively correlated with insulin sensitivity index and preferentially upregulated in the cells of obese individuals. Leptin is an adipocyte-derived peptide hormone that functions as an afferent signal in a negative feedback loop that controls feeding and maintains energy homeostasis. Leptin may regulate adipose tissue metabolism directly through autocrine signaling or indirectly mediated by sympathetic neurons that innervate adipocytes at neuroadipose junctions¹⁰. Consistently, genes in the LEP-containing modules were enriched in a specific adipocyte subcluster of obese individuals and function in eating behavior. In line with our findings, single-cell immunostaining of primary cultured and freshly isolated cells in a recent study suggested the existence of white adipocyte subtypes specialized for the production of ADIPOQ and LEP, and are putatively under the regulation of distinct transcription factors⁵³. A spatial transcriptomic analysis of human adipose tissue identified three different types of mature adipocytes with distinct transcriptional profiles and spatial arrangements⁵⁴. The adipocyte subtype with marker gene leptin or Adipo^LEP was insulin resistant, while the adipocyte subtype with top marker gene perilipin or Adipo^PLIN strongly expressed ADIPOQ and was insulin sensitive. Our experiments also showed that conditioned media from THP1 macrophages induced the expression of LEP but repressed ADIPOQ expression, suggesting the two adipokines are involved in different regulatory mechanisms in immune cell interactions. Network driver genes co-expressed with the two adipokines may play an important role in such immune-adipocyte signaling pathways, which warranted future studies for functional investigation.

Our study also identified cell-type-specific modules in response to insulin resistance in the muscle tissue. The muscle stem cell modules were positively correlated with insulin sensitivity, while the modules enriched for muscle fibers and fibro-adipogenic progenitors showed negative correlations. Consistent with our findings, a recent study reported that fibro-adipogenic progenitor cells increased in diabetes patients and contributed to degenerative remodeling of the extracellular matrix in muscle tissues¹⁴. A previous study using a diverse panel of mouse strains suggests that insulin resistance in adipose and muscle tissues can occur independently, and tissue-specific mechanisms are involved in the progression of metabolic diseases⁵⁵. In our study, marker genes from muscle stem cells and fibro-adipogenic progenitors showed a consistent insulin sensitivity response in both adipose and muscle tissues. In contrast, cellular respiration-related genes shared by adipocyte and muscle fiber modules exhibited an opposite pattern for insulin sensitivity correlation, suggesting the distinct pattern in the rearrangement of gene expression networks in different cell types of insulin-responsive tissues.

The integration of network modules and genotype information revealed that the adipocyte modules (M9 and M55) and the muscle fiber modules (M8) were enriched for the cis-eQTL genes in the AAGMEx cohort. Several network driver genes, including LEP in adipose module M9 and CKMT2 in muscle module M8, showed significant enrichment for the cis-eQTL genes in their network neighbors. The regulatory SNPs in the network neighbors may configure the networks by modulating gene expression and are likely causal determinants for insulin resistance. LEP is an extensively characterized adipokine involved in insulin resistance-related metabolism¹⁰. CKMT2 encodes a sarcomeric mitochondrial creatine kinase isoenzyme highly expressed in oxidative myocytes and plays a central role in energy transduction in tissues with large, fluctuating energy demands⁵⁶. So far, how genetic variations modulate dynamic changes of LEP and CKMT2 expression levels, especially in obesity-related metabolic disorders, is little understood. A recent study reported that adipose-specific and quantitative LEP transcript expression was controlled by redundant cis-acting enhancer elements and trans factors interacting with the proximal promoter together with a long noncoding RNA⁵⁷. As both adipocyte and muscle fiber modules were enriched with eQTLs, it would be interesting for future studies to investigate how SNPs are involved in regulating these module genes and determining insulin resistance and metabolic disorders.

Our previous work applied a computationally less expensive blockWise module detection approach of WGCNA for exploratory analysis of the AAGMEx cohort and identified 5 and 3 co-expressed modules associated with insulin sensitivity in adipose and muscle, respectively³⁶. Compared with those sparse and exclusive WGCNA modules, our current approach not only captured module-related pathways from WGCNA (Table S23) but also identified additional molecular processes and biological pathways that were enriched among a large number of IS-modules. This is consistent with previous findings that MEGENA was capable to produce more compact and coherent modules^19,21. The advantages of the analytical framework of MEGENA not only expanded our knowledge of the repertoire of cellular processes involved in insulin resistance, but also allowed more comprehensive downstream analyses such as cell-type-specific module identification, module preservation analysis, and network driver discovery.

While the network approach was effective to decipher the cell-type-specific modules from human populations, the network modules showed a limited resolution to distinguish cell subtypes, such as different subpopulations of immune cells. This may be because the genes from different cell subtypes displayed similar correlations with insulin resistance, which hindered the module identification by the co-expression network approach. Although the top-ranked insulin-responsive modules were highly conserved across two ethnic groups, we also observed specific modules in the coexpression networks from different cohorts. However, there are only a few insulin sensitivity-associated modules specific to each ethnic group, and most of these ethnic-specific modules were not enriched for genes in known biological pathways. Although the differences in the network modules may suggest ethnically-predominant mechanisms, we could not exclude the possibility caused by variations in sample sizes and the platforms for gene expression measurement. It would be important for future studies to validate the unique network modules and hub genes between different ethnic groups. Finally, our sc/snRNA-seq data analysis did not account for the influence of the ethnicity of donors. Considering the strong preservation of cell types in human tissues (as well determined by our network preservation analysis) and the sc/snRNA-seq markers were also supported by a broad single-cell database (PanglaoDB), we believe the ethnic background of the donors in sc/snRNA-seq data may have a minor effect on our cell type interpretation.

STAR METHODS

RESOURCE AVAILABILITY Lead contact

Further information and request for resources and reagents should be directed to and will be fulfilled by the lead contact, Swapan K. Das (sdas@wakehealth.edu).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• The multi-omic datasets reported in this study are publicly available. The Accession numbers are listed in the key resource table. The full MEGENA and Bayesian networks can be downloaded through Zenodo (DOI:10.5281/zenodo.7331429).

• No original code was produced in this paper. Code for bioinformatics analyses on coexpression network and single-cell RNA-seq analysis is available through the download of the specified (key resources table/STAR Methods), publicly available packages.

• Any additional information required to reanalyze the data reported in this paper is available from the Lead Contact upon request.

KEY RESOURCES TABLE

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Biological samples
Human tissues	This paper	N/A
Human DNA	This paper	N/A
Human RNA	This paper	N/A
Human Plasma	This paper	N/A
Chemicals, peptides, and recombinant proteins
miRNeasy Mini Kit	Qiagen	217004
Ultraspec RNA total RNA extraction reagent	Biotecx laboratories	BL-10100
HumanHT-12 v4 Expression BeadChip	Illumina, San Diego	N/A
Affymetrix U219 microarray	Affymetrix	N/A
RNAeasy Lipid Tissue Mini kit	Qiagen	74804
Collagenase Type 2	Worthington, Lakewood, NJ	CLS-2
BGISEQ Eukaryotic Transcriptome resequencing BGISEQ-500 platform	BGI	N/A
Gentra Puregene blood DNA isolation kit	Qiagen.	158489
Infinium HumanOmni5Exome-4 v1.1 DNA Analysis BeadChip	Illumina	N/A
Illumina Human OmniExpress BeadChip array	Illumina	N/A
RNAqueous kit	Ambion, Inc	AM1912
RPMI-1640 Medium (ATCC® 30-2001™),	ATCC	ATCC® 30-2001™
Bench Mark FBS	Gemini, CA	100–106
L-glutamine and antibiotics	Omega Scientific	PG-30
Phorbol 12-myristate 13-acetate	Sigma	P1585
Human NPL gene-specific shRNA lentiviral (lv) particle	Santcruz biotechnology Inc.	sc-88481-V
Polybrene	Santacruz	sc-134220
Control shRNA lentiviral Particles-A	Santacruz	sc-108080
Puromycin	Gibco, Thermo Fisher Scientific	A1113803
RNAeasy Total RNA Mini Kit	Qiagene	74104
QuantiTect reverse transcription kit	Qiagen	205313
Power SYBR green master mix	Applied Biosystems Thermofisher	43 687 02
Illumina HiSeq, PE 2x150	Illumina done at GENEWIZ, LLC.	N/A
DMEM/Ham’s F-12 (1:1 v/v) adipocyte basal medium	Zenbio, Inc	BM-1
Adipocyte differentiation medium	Zenbio, Inc	DM-2
Adipocyte maintenance medium	Zenbio, Inc	AM-1
Deposited data
AAGMEx cohort adipose and muscle expression data	36	GEO id #GSE95674 and #GSE95675 in super series #GSE95676
METSIM cohort adipose expression data set	37	GEO (id #. GSE70353)
AREA cohort adipose expression data set	38	GEO (accession number GSE65221)
AREA cohort muscle expression data set	This paper	DOI:10.5281/zenodo.7331429
NPL gene knockdown in THP1 cell expression data	This paper	GSE196888
Human Adipose tissue snRNA-seq dataset	15	Single Cell Portal (Study #SCP1376)
Human Muscle tissue scRNA-seq dataset	43	GSE143704
Experimental models: Cell lines
THP-1	ATCC, obtained from the Wake Forest Cell and Viral Vector Core Laboratory	TIB-202
Simpson-Golabi-Behmel syndrome (SGBS) preadipocytes	Developed by Dr Martin Wabitsch, University Medical Center Ulm, Germany	N/A
hADSC	Coriell Cell Repositories(Camden, NJ)	AG17928 and AG172929
Oligonucleotides
ACADM-RTF: TCATTGTGGAAGCAGATACCC	Integrated DNA Technology (IDT)	N/A
ACADM-RTR: CAGCTCCGTCACCAATTAAAAC	IDT	N/A
ALDH6A1-RTF: GGGCATCCAATTCTACACTCAG	IDT	N/A
ALDH6A1-RTR: AGGGAGATTACTCAGGATGGAG	IDT	N/A
HADH-RTF: GCTTCTAGATTATGTCGGACTGG	IDT	N/A
HADH-RTR: TGGGCTGATGTAATGGGTTC	IDT	N/A
LAPTM5-RTF: GCCCACCTATCTCAACTTCAAG	IDT	N/A
LAPTM5-RTR: GATGAAGGCGATGGAAAAGATG	IDT	N/A
NPL-RTF: TCACATTCCTGCCTTGACAG	IDT	N/A
NPL-RTR: ACATTGCCCGAAGTCTAAGAG	IDT	N/A
RPLP0(36B4)-RTF: CGACCTGGAAGTCCAACTAC	IDT	N/A
RPLP0(36B4)-RTR: ATCTGCTGCATCTGCTTG	IDT	N/A
PPIA-RTF: TCCTGGCATCTTGTCCAT	IDT	N/A
PPIA-RTR: TGCTGGTCTTGCCATTCCT	IDT	N/A
Software and algorithms
GenomeStudio V2011.1.	Illumina	https://support.illumina.com/array/array_software/genomestudio/downloads.html
SOAPnuke	58	https://github.com/BGI-flexlab/SOAPnuke
HISAT	59	http://www.ccb.jhu.edu/software/hisat/index.shtml
Bowtie2	60	http://bowtie-bio.sourceforge.net/bowtie2/index.shtml
RSEM	61	https://deweylab.github.io/RSEM/
DESeq2	62	https://bioconductor.org/packages/release/bioc/html/DESeq2.html
KING	63	http://people.virginia.edu/~wc9c/KING/
ADMIXTURE	64	https://dalexander.github.io/admixture/index.html
MEGENA (v1.4)	19	https://github.com/songw01/MEGENA
MatrixEQTL	65	https://cran.r-project.org/web/packages/MatrixEQTL/index.html
WGCNA (v1.69)	41	https://cran.r-project.org/web/packages/WGCNA/index.html
Seurat (v3.9.9)	66	https://cran.r-project.org/web/packages/Seurat/index.html
fgsea (v1.16)	67	https://github.com/ctlab/fgsea
Trimmomatic v.0.36	68	http://www.usadellab.org/cms/index.php?page=trimmomatic
STAR aligner v.2.5.2b	69	https://github.com/alexdobin/STAR
Subread package v.1.5.2	70	http://subread.sourceforge.net/
RIMBANet	71	https://icahn.mssm.edu/research/genomics/about/resources
clusterProfiler (v3.18)	72	https://bioconductor.org/packages/release/bioc/html/clusterProfiler.html
Other
Amicon™ Ultra-15 Centrifugal Filter Unit (3 KDa)	Millipore	UFC900324

EXPERIMENTAL MODEL AND SUBJECT DETAILS

Human Subjects

Analyses presented in this study used multi-omic data on African and European ancestry individuals from three independent cohorts from previously published studies. Characteristics of these cohorts are described briefly below.

African American Genetics of Metabolism and Expression (AAGMEx) cohort

Gluco-metabolic phenotype, gene expression, and genotype data available from 256 unrelated and non-diabetic individuals in the AAGMEx cohort were used as the primary discovery cohort^36,73,74. All participants provided written informed consent under protocols approved by the Institutional Review Boards at Wake Forest School of Medicine. Cohort participants were healthy, self-reported African American men and women residing in North Carolina, aged 18–60 years, with a body mass index (BMI) between 18 and 42 kg/m². A standard 75-g oral glucose tolerance test (OGTT) was used to evaluate insulin sensitivity and exclude individuals with diabetes. Detailed evaluation of insulin sensitivity was performed by insulin modified (0.03 U/kg) frequently sampled intravenous glucose tolerance test (FSIGT). Fasting blood samples were drawn for DNA isolation and biochemical analyses. Abdominal subcutaneous adipose near the umbilicus and vastus lateralis skeletal muscle biopsies were obtained under local anesthesia. Tissue biopsies were collected by Bergstrom needle from participants after an overnight fast. Tissues were immediately rinsed in sterile saline, quick-frozen in liquid nitrogen, and stored at −80°C. Clinical, anthropometric, and physiological characteristics of the AAGMEx cohort have been described.

Metabolic Syndrome in Men (METSIM) cohort

We analyzed phenotypic and genomic data from 770 males who were part of the METSIM cohort⁷⁵. 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. To evaluate insulin sensitivity METSIM participants underwent a 75-g OGTT, and the blood samples were drawn at 0, 30, and 120 min. Among the 770 participants 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.

Arkansas European-American (AREA) Cohort

European-American or African-American men and women who were generally in good health, between 19 and 60 years of age, and had a BMI between 19 and 45 kg/m² were recruited from Arkansas. Methods for subject recruitment, physical examination, physiological experiments, and the method for obtaining biopsy samples were described previously³⁸. Similar to AAGMEx, in this cohort insulin sensitivity was evaluated by 75g-OGTT and FSIGT; and biopsies were obtained under local anesthesia using a Bergstrom needle from abdominal subcutaneous fat and vastus lateralis muscle. Among the 168 recruited individuals, genome-wide adipose and muscle tissue expression data for 99 nondiabetic European Americans were used for validation and replication analyses in this study.

METHOD DETAILS

Laboratory measures and physiologic phenotypes

Details of clinical laboratory measures have been described. In brief, for the AAGMEx and AREA cohort, plasma glucose levels were analyzed by glucose oxidase methods at a CLIA-certified commercial laboratory (LabCorp). Plasma insulin was measured using an immuno-chemiluminometric assay (Invitron Limited, Monmouth, UK). Plasma glucose and insulin data from five OGTT time points (0, 30, 60, 90, and 120 min) were used to calculate the Matsuda insulin sensitivity index (http://mmatsuda.diabetes-smc.jp/MIndex.html). The MINMOD Millennium program was used to analyze FSIGT data to determine insulin sensitivity index (S_I) by Minimal model analysis⁷⁶. As reported earlier, in the METSIM cohort plasma glucose was measured by enzymatic hexokinase photometric assay (Konelab Systems reagents; Thermo Fischer Scientific), and insulin was determined by immunoassay (ADVIA Centaur Insulin IRI no. 02230141; Siemens Medical Solutions Diagnostics). Insulin sensitivity in the METSIM cohort was evaluated by calculating the Matsuda index from three OGTT data points (0, 30, and 120 min)³⁷.

Multi-omic Data

A. AAGMEx cohort adipose and muscle expression data

Extraction of total RNA from adipose and muscle was performed using miRNeasy Mini Kit (Qiagen) and Ultraspec RNA total RNA extraction reagent (Biotecx laboratories), respectively. Quantities of RNA samples were determined by ultraviolet spectrophotometry (Nanodrop, Thermo Scientific), and the quality of RNA was determined by electrophoresis (Experion nucleic acid analyzer, BioRad Laboratories, Inc.). Genome-wide expression data were generated using HumanHT-12 v4 Expression BeadChip (Illumina, San Diego, CA; Data submitted to Gene Expression Omnibus: GEO id #GSE95674 and #GSE95675 in super series #GSE95676)^36,73,74. Participants were block-randomized by age, gender, and BMI, and their RNA samples were assigned to a group, totaling 12 samples for hybridization per BeadChip. Chips were scanned in the Illumina HiScan Reader using Illumina iScan Control Software. Genome-wide gene expression data (probe level) were extracted using Illumina GenomeStudio V2011.1. Expression levels were log₂ transformed, robust multi-array average normalized (RMA, includes quantile normalization), and batch-corrected using ComBat. The HumanHT-12 v4 Expression BeadChip includes 47,231 probes annotated to transcripts; however, data on transcript probes encompassing common SNPs (based on ReAnnotator, or SnpInProbe annotation, and UCSC SNPv141) and transcripts that were not significantly expressed (p-value<0.05) in ⩾25% of the samples were excluded.

B. METSIM cohort adipose expression data set

Total RNA from METSIM participants was isolated from adipose tissue via the QIAGEN miRNeasy kit. RNA integrity numbers (RINs) were assessed with the Agilent Bioanalyzer 2100 instrument, and 770 samples with RIN > 7.0 were used for transcriptional profiling. Genome-wide transcript expression profiling with the Affymetrix U219 microarray was performed at the Department of Applied Genomics at Bristol-Myers Squibb, and data were submitted to GEO (id #. GSE70353)³⁷. The microarray image data were processed with the Affymetrix GCOS algorithm via the robust multiarray average (RMA) method to determine the specific hybridizing signal for each gene. 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 based on the RefSeq (version 59) and Ensembl (version 72) databases were removed. The final dataset included 43,145 probe sets.

C. AREA cohort adipose and muscle expression data set

Because the METSIM cohort includes only males, gene expression data is limited to adipose tissue, and insulin sensitivity was not evaluated by FSIGT, we used adipose and muscle tissue expression data on AREA cohort participants for replication and comparison with the AAGMEx cohort. Total RNA was isolated from whole adipose tissue using the RNAeasy Lipid Tissue Mini kit (QIAGEN Inc-USA, Valencia, CA). The quantity and quality of the isolated total RNA samples were determined by ultraviolet spectrophotometry (Nanodrop, Thermo Scientific, Pittsburgh, PA) and electrophoresis (Experion nucleic acid analyzer, BioRad Laboratories, Inc., Hercules, CA), respectively. High-quality RNA with RIN (RNA integrity number) > 8 was used for genome-wide transcriptome analysis. Genome-wide transcriptome analyses were done using HumanHT-12 v4 Expression BeadChip (Illumina, San Diego, CA) whole-genome gene expression arrays according to the vendor-recommended standard protocol. Participants were randomized and their RNA samples assigned to a group, totaling 12 samples for hybridization per BeadChip. Chips were scanned in the Illumina BeadArray Reader. Raw expression intensity was background subtracted and normalized by the average normalization algorithm as implemented in GenomeStudio Gene Expression Module v1.0 application software (Illumina). Normalized data were used for further analysis. Expression data have been deposited in GEO (accession number GSE65221)^38,77. Genome-wide expression profiling of muscle samples in the AREA cohort was performed at Beijing Genome Institute, Inc (BGI Hongkong Tech Solution NGS Lab) by RNA-sequencing. RNA samples were isolated from 40–100mg frozen tissue biopsy samples with ULTRASPEC^® RNA Total RNA isolation reagent. RNA samples were independently quality checked by BGI using Agilent 2100 Fragment Analyzer. Twelve samples not meeting QC criteria (RIN>7, 28S/18S ratio>1, low 5S peak) were re-extracted by RNeasy Plus Universal Mini Kit (Qiagen, 73404). All RNA samples meeting QC criteria were finally used for sequencing library preparation. Samples were sequenced by BGISEQ-500 platform to generate paired-end 100bp reads (PE100). The SOAPnuke software (BGI) was used to filter reads and clean reads stored in FASTQ format⁵⁸. Clean reads were mapped to the reference genome (hg19) using HISAT (Hierarchical Indexing for Spliced Alignment of Transcripts)⁵⁹. Clean reads were mapped to reference transcripts using Bowtie2⁶⁰, then expression levels for each sample (gene and isoform expression levels) from RNA-Seq data were calculated with RSEM⁶¹. On average, 65.32 million reads per sample were generated. All sample shows uniformity in the mapping result, and the average mapping ratio with the reference genome was 94.93%. The average mapping ratio with genes is 65.55%. Expression levels of a total of 18,688 genes were detected in this set of muscle samples. Expression levels were normalized based on RNA-Seq read counts with DESeq2, which utilizes the median of ratios method, and was used for all downstream analyses⁶².

D. AAGMEx cohort genotype data set

DNA was isolated from whole blood using the Gentra Puregene blood kit (Qiagen). DNA samples were measured by NanoDrop and concentrations were adjusted for genome-wide genotyping. Infinium HumanOmni5Exome-4 v1.1 DNA Analysis BeadChip (Illumina) and Infinium LCG Quad Assay kits were used to genotype DNA samples (400 ng per subject) based on the manufacturer’s instructions. The Illumina HiScan System was used to scan the BeadChips. Genotype data were examined to verify sample and SNP quality. Samples were excluded if they had a call rate <90% or excess heterozygosity (F <−0.10). Genetic markers were considered high quality if call rates were >95% without departure from expected Hardy-Weinberg proportions (P > 1 × 10⁻⁶). Identity-by-descent statistics computed by the program KING⁶³ were examined and did not reveal unexpected duplicates or first- or second-degree relatives. HapMap Phase 3 CEU, YRI, and CHB samples were merged with study samples and admixture estimates were computed using the software ADMIXTURE⁶⁴. Samples with >50% European ancestry proportion were excluded. Genotype assays of 4,210,443 SNPs passed technical quality filters. The genotype of 2,296,925 autosomal SNP assays (representing 2,210,735 unique high-quality genotyped SNPs with MAF > 0.01 and HWE-p value > 1 X 10⁻⁶) was used in eQTL analysis⁷³. After combining available phenotype, transcript, and genotype data sets, most analyses in AAGMEx adipose and muscle tissue include 251 and 248 samples, respectively.

E. METSIM cohort genotype data set

METSIM samples were genotyped using the Illumina Human OmniExpress BeadChip array and the Illumina Human CoreExome array. After quality control and genotype imputation of the 681,789 directly genotyped variants, the METSIM study used 7,677,146 variants (MAF ⩾ 0.01) for eQTL analysis³⁷. Based on current European Union law on human subjects privacy regarding genotype data, METSIM Data Use and Share Committee was unable to share the genotype data, but shared the precomputed eQTL summary statistics which was used for this study.

Data Analysis methods

A. MEGENA coexpression network analysis

MEGENA was performed by the R package “MEGENA” (v1.4) according to the recommended pipelines¹⁹. The normalized expressions of all transcripts were used for MEGENA. For each gene, the transcript with the most variance was used for network construction. Pearson correlation coefficients (PCCs) were computed for all gene pairs. Significant PCCs were filtered by a false discover rate (FDR) cutoff of 0.05 with 10 permutation analyses. The ranked significant PCCs were iteratively tested for planarity to grow a Planar Filtered Network (PFN) using the PMFG algorithm. The resulting PFN was analyzed by Multiscale Clustering Analysis (MCA) to identify coexpression modules at different network scale topologies. Hub genes, which are highly connected in each cluster, were identified by the Multiscale Hub Analysis (MHA).

B. eQTL analysis

For AAGMEx datasets, eQTLs were identified by integrating the transcript expression and the SNP genotype data. The MatrixEQTL package was used to identify SNPs significantly associated with gene expression traits⁶⁵. Significant SNPs (eQTLs) were classified into cis- and trans-acting elements according to whether they are located within 1 Mb from the gene or not. Benjamini & Hochberg’s (BH) corrected p-value threshold of 0.05 was used to define eQTLs. Detail of the method for eQTL analysis used in the METSIM cohort is published elsewhere³⁷. In summary, results from factored spectrally transformed linear mixed models (FaST-LMM) eQTL analysis in the METSIM cohort were used for this study. In METSIM 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. Variants in cis-region with association p < 2.46 X10⁻⁴ (corresponding to 1% FDR) were considered significant, whereas a conservative Bonferroni-corrected p < 1.51 × 10⁻¹³ was used for significant trans-eQTLs. The cis-eQTL regulated genes were considered as the cis-eQTL genes.

C. Bayesian probabilistic causal network inference

To construct the Bayesian probabilistic causal Network (BN), the known tissue-specific transcription factor (TF)-target relationships from the ENCODE project and the genetic information from eQTL analysis (described above) were used as prior information to assist the inference of regulatory relationships between genes. In the causal network of TF-targets, the edges are directional: 1) the TFs can be the parent nodes of their target genes, but the targets are not allowed to be parent nodes of their TFs, and 2) For cis- and trans-eQTL regulated genes which are associated with the same SNP (i.e., eQTL), the trans-associated gene(s) cannot be the parent node of the cis-associated gene(s). Then the BN was built based on the structure priors combined from both TF-target and the eQTL-based genetic relationships. A Monte Carlo Markov Chain (MCMC) simulation-based procedure was applied to construct the BN⁴⁴. This MCMC process started with different random structures and 1,000 networks were generated. The links that appeared in more than 30% of the networks were used to define a final consensus network following previous practices^44,71. To get a final network structure that is a directed acyclic graph, the weakly supported links involved in a loop were removed in the consensus network. To identify the driver genes that are predicted to modulate downstream nodes, the module genes from MENEGA were projected onto the BN. Driver genes were identified as the nodes whose network neighborhoods were enriched for the module genes by Key Driver Analysis (KDA)^33,45,46.

D. Module preservation analysis

Two paralleled testing methods were applied to study the preservation of two MEGENA coexpression networks. The first method is the network-based statistics calculated by the modulePreservation function developed for WGCNA⁴¹. As modules from MEGENA are hierarchical with multiple layers, the modules were tested by the modulePreservation layer by layer. For each preservation test, 100 permutations were applied. Following the original software guideline, the main network-based statistics Zsummary.pres was used for preservation interpretation. The second preservation analysis method is the FET. For each module, FET was performed to test the enrichment of module genes against the modules from the other network. Then the BH adjusted enrichment p-value was used for preservation interpretation. Unsupervised hierarchical clustering of top-ranked modules was performed by the “ComplexHeatmap” package in R software⁷⁸. The FET significance score (−log10(FDR)) of the module overlap matrix was used as the input for clustering analysis.

E. Single-cell RNA-seq analysis

For adipose tissue, the normalized gene expressions and cell type annotations of the snRNA-seq dataset were obtained from the Single Cell Portal (Study #SCP1376)¹⁵. The nuclei collected from human subcutaneous adipose tissue (n = 57,599) were retained for downstream analysis. For muscle tissue, the normalized gene expressions and cell type annotations of the scRNA-seq dataset (n = 22,000) were obtained from the GEO database, under accession number GSE143704⁴³. To identify marker genes in the two tissues, the “FindAllMarkers” function from Seurat (v3.9.9) was applied to identify differentially expressed genes using a Wilcoxon Rank Sum test⁶⁶. Only significantly upregulated genes (FDR < 0.05) with 0.25 log fold change and 0.25 minimum expression fraction were regarded as marker genes. To identify the cell-type specificity of network modules, the FET from the R package clusterProfiler (v3.18)⁷² was performed to compare the overlapping between the cell-type maker genes with the module genes, and calculate the enrichment p values based on the deviation from a null hypothesis. BH method was applied to adjust multi-testing p-values to calculate the Padj.

To identify the module enrichment in each single nuclei, the GSEA was performed by the R package fgsea (v1.16)⁶⁷. Conceptually, the GSEA enrichment score, which corresponded to a weighted Kolmogorov-Smirnov-like statistic, was calculated by walking down the ranked list, increasing a running-sum statistic when encountering a gene in the module, and otherwise decreasing it⁷⁹. As the enrichment score calculated the maximum deviation from zero encountered in the random walk, it reflected the degree to which the module gene set overrepresented at the extremes (top or bottom) of the entire ranked list. The statistical significance (nominal p value) of the enrichment score was estimated by 1,000 permutations.

The Seurat pipeline was applied to integrate and identify subclusters of adipocytes⁶⁶. Briefly, subcutaneous adipocytes from individual donors were extracted from Single Cell Portal (Study #SCP1376)¹⁵ and normalized by the “LogNormalize” function. For the gene expression matrix from each individual, 30 principle components were calculated by the “RunPCA” function using the top 2,000 variable genes. Then the gene expression matrices of different individuals were integrated by the reciprocal PCA (RPCA) method, followed by gene expression scaling and a linear dimensional reduction which generated 30 principle components. The “FindNeighbors” function was used to construct a K-nearest neighbor (KNN) graph based on the euclidean distance with the principal components. “FindClusters” function was then applied to optimize modularity by the Louvain algorithm. The resolution parameter for the clustering granularity was set to 0.5. Finally, the UMAP method was used for non-linear dimensional reduction and cluster visualization.

F. Module communication analysis

Module communication analysis was performed similarly to the cell communication analysis of single-cell studies. To assess whether two modules have significant communications, the genes from two modules were searched against each other for the ligand-receptor pairs in the CellChatDB database⁴², which contains manually curated literature-supported ligand-receptor pairs. Meanwhile, one million permutations were performed from the randomly chosen genes to calculate the background frequency of ligand-receptor pairs. The significance of module communication was determined by the permutation test, which compared the observed ligand-receptor frequency with the background values.

Methods for functional validation

A. Isolation of adipocyte and stromal vascular fractions from adipose tissue

The adipocyte fraction (AF) was separated from the stromal-vascular fraction (SVF) after collagenase digestion (Collagenase Type 2, Worthington, Lakewood, NJ) of freshly collected adipose biopsy samples using the method of Rodbell⁸⁰. Total RNA was isolated from adipocyte fractions and the stromal vascular fraction of adipose by using the RNAeasy Lipid Tissue Mini kit (QIAGEN Inc., Valencia, CA) and RNAqueous kit (Ambion, Inc., Austin, TX), respectively. Total RNA from 5–6 AAGMEx participants was mixed to make separate pools of AF and SVF RNA. Pooled RNA samples were reverse transcribed using Qiagen QuantiTect reverse transcription kit (Qiagen) based on the manufacturer’s protocol. The qRT-PCR analysis of cDNA prepared by this protocol shows very high expression of ADIPOQ (adiponectin) and CD68 (tissue macrophage-specific marker) in AF and SVF, respectively, and indicates the quality of isolated AF and SVF.

B. THP1 cell culture and experiments

THP1 cells (American Type Culture Collection, Manassas, VA), a model for human macrophages⁵⁰ were obtained from the Wake Forest Cell and Viral Vector Core Laboratory and cultured in RPMI-1640 Medium (ATCC^® 30–2001^™), supplemented with 10% fetal bovine serum (Bench Mark FBS, Gemini, West Sacramento, CA), L-glutamine, and antibiotics. Cells were cultured at 37°C in a humidified incubator maintaining a 5% CO₂ atmosphere. Cells were treated with Phorbol 12-myristate 13-acetate (PMA or TPA, P1585, Sigma; 10ng/ml) to differentiate THP1 monocytes to macrophages. THP1 monocytes grow as suspension culture, but PMA-stimulated differentiated THP1 attaches to the plastic surface of cell culture plates and shows distinctive macrophage cell morphology. Cells were harvested for RNA isolation at different time points after PMA treatment (0, 6, 24, 48, and 72 hrs.) to understand the transcript expression of selected genes at different stages of macrophage differentiation. Most experiments used RNA from 48hr PMA treated THP1 cells. To obtain macrophage conditioned media (MCM) we seeded 4×10⁶ THP1 cells in a T75 flask and treated with PMA (10ng/ml). After 48hrs of PMA treatment attached THP1 macrophages were washed with DPBS, and cultured in 15ml RPMI-1640 Medium supplemented with 0.5% FBS for additional 24hrs to obtain MCM. MCMs were collected from the flask, cleared by centrifuging at 150g for 5min, and then concentrated (to 2ml) using Amicon^™ Ultra-15 Centrifugal Filter Unit (3 KDa, UFC900324, Millipore) by centrifuging at 3500rpm for 45 min at 4 °C⁸¹. Concentrated MCMs were further filtered through 0.2 μM filter and were used in experiments to test the interaction between macrophages and adipocytes.

C. Transduction of THP1 cells to knock down target genes

For stable RNA interference, the NPL gene was silenced by infecting the THP1 cells with gene-specific lentiviral (lv) shRNA (sc-88481-V, Santcruz biotechnology Inc., Dallas, TX) in the presence of polybrene (sc-134220, Santacruz; 8 μg/ml) according to an optimized method based on the manufacturer’s protocol³⁶. A Control shRNA lentiviral Particles-A (sc-108080, Santacruz) was used as negative control (encodes a scrambled shRNA sequence that will not lead to the specific degradation of any known cellular mRNA). Cells successfully transduced and stably expressing shRNA were selected using 2 μg/ml of puromycin (A1113803, Gibco, Thermo Fisher Scientific, USA). Total RNA from gene specific-shRNA and control-shRNA expressing THP1 cells were isolated at baseline and 48hr PMA treated condition using RNAeasy kit (Qiagene). RNA samples were reverse transcribed using Qiagen QuantiTect reverse transcription kit (Qiagen) based on the manufacturer’s protocol. To confirm shRNA-mediated downregulation, expression of NPL was measured in cDNA by quantitative real-time PCR (qRT-PCR) using Power SYBR green chemistry (Applied Biosystems, Inc., Foster City, CA). Similar to published studies, the expression of target genes was normalized to the expression of endogenous control gene cyclophilin A (PPIA)³⁶. Two independent experiments with 2–3 biological replicates for each treatment condition were performed.

D. RNA sequencing for in vitro experiments

RNA isolated from NPL-shRNA and control-shRNA expressing THP1 cells at 48hr PMA treated condition from an experiment (in biological triplicate) was used for comparison by RNA-seq analysis to identify target genes and pathways for these genes. RNA library preparations with Poly A selection, and HiSeq sequencing reactions were conducted at GENEWIZ, LLC. (South Plainfield, NJ, USA) using standardized protocols. RNA samples were quantified using Qubit 2.0 Fluorometer (Life Technologies, Carlsbad, CA, USA) and RNA integrity was checked using Agilent TapeStation 4200 (Agilent Technologies, Palo Alto, CA, USA). RNA sequencing libraries were prepared using the NEBNext Ultra RNA Library Prep Kit II for Illumina following the manufacturer’s instructions (NEB, Ipswich, MA, USA). Briefly, mRNAs were first enriched with Oligo(dT) beads. Enriched mRNAs were fragmented for 15 minutes at 94 °C. First-strand and second-strand cDNAs were subsequently synthesized. cDNA fragments were end-repaired and adenylated at 3’ends, and universal adapters were ligated to cDNA fragments, followed by index addition and library enrichment by limited-cycle PCR. The sequencing libraries were validated on the Agilent TapeStation (Agilent Technologies, Palo Alto, CA, USA), and quantified by using Qubit 2.0 Fluorometer (Invitrogen, Carlsbad, CA) as well as by quantitative PCR (KAPA Biosystems, Wilmington, MA, USA). The sequencing libraries were pooled and clustered on 1 lane of a flowcell. After clustering, the flowcell was loaded on the Illumina HiSeq instrument (4000 or equivalent) according to the manufacturer’s instructions. The samples were sequenced using a 2×150bp Paired-End (PE) configuration. Image analysis and base calling were conducted by the HiSeq Control Software (HCS). Raw sequence data (.bcl files) generated from Illumina HiSeq was converted into fastq files and de-multiplexed using Illumina’s bcl2fastq 2.17 software. One mismatch was allowed for index sequence identification.

E. RNA-seq data analysis for in vitro experiments

After investigating the quality of the raw data, sequence reads were trimmed to remove possible adapter sequences and nucleotides with poor quality using Trimmomatic v.0.36⁶⁸. The trimmed reads were mapped to the Homo sapiens GRCh38 reference genome available on ENSEMBL using the STAR aligner v.2.5.2b⁶⁹. The STAR aligner is a splice aligner that detects splice junctions and incorporates them to help align the entire read sequences. BAM files were generated as a result of this step. Unique gene hit counts were calculated by using feature Counts from the Subread package v.1.5.2⁷⁰. Unique reads that fall within exon regions were counted. After extraction of gene hit counts, the gene hit counts tables were used for downstream differential expression analysis. Using DESeq2, a comparison of gene expression between the groups of samples was performed⁶². The Wald test was used to generate p-values and log₂ fold changes. Following the previous study²³, genes with adjusted p-values < 0.05 and absolute fold change > 1.2 were called differentially expressed genes for each comparison. A pathway enrichment analysis was performed on the statistically significant set of genes. Validation of differentially expressed genes identified in the first RNA-seq experiment was performed by a second independent RNA-seq experiment and by quantitative real-time PCR (qRT-PCR) of selected genes.

F. Adipocytes culture and experiments

We studied the expression of selected genes at different stages of differentiation in human adipose stroma-derived stem cells (hADSCs). The hADSCs (AG17928 and AG172929) derived from abdominal subcutaneous tissue samples donated by two different women were obtained from Coriell Cell Repositories (Camden, NJ). Most experiments on adipocytes however used Simpson-Golabi-Behmel syndrome (SGBS) pre-adipocytes, a well-characterized human adipocyte cell model that is more amenable to in vitro studies than hADSC’s and has an expression profile after differentiation that closely mimics mature adipocytes⁸². We grew hADSC and SGBS cells under standard culture conditions in DMEM/Ham’s F-12 (1:1 v/v) adipocyte basal medium (BM-1, Zenbio, Inc; Research Triangle Park, NC) supplemented with 10% FBS and antibiotics. Cells were differentiated to adipocytes using adipocyte differentiation medium (DM-2, Zen Bio) for seven days and maintained for an additional seven days in adipocyte maintenance medium (AM-1, ZenBio) for complete maturation of adipocytes following the vendor-recommended protocol⁷⁴. To investigate the interaction of adipocytes with macrophages, and to test if those cell-cell interactions are modified by knocking down of key regulatory genes identified in our adipose tissue network analyses, differentiating SGBS adipocytes were treated with MCMs harvested from THP1 macrophages expressing control-shRNA and compared with MCMs from THP1 cells expressing NPL-shRNA. MCMs were added to adipocyte culture media (20 μl concentered MCM per ml of DM-2 and AM-1 media) and SGBS cells were treated with this culture media throughout the differentiation phase. THP1 culture media harvested and processed similarly from a cell-free blank culture flask was used as a negative control for MCM treatment. Total RNA was isolated using RNAeasy kit (Qiagen). RNA samples were reverse-transcribed using QuantiTect reverse transcription kit (Qiagen) based on the manufacturer’s protocol. To determine expression levels at different stages of adipocyte differentiation and in other experiments with SGBS cells, expression was measured in cDNA by quantitative real-time PCR (qRT-PCR) using Power SYBR green chemistry (Applied Biosystems, Inc., Foster City, CA). Similar to published studies, the expression of target genes was normalized to the expression of an endogenous control gene, 36B4 (RPLP0)⁷⁴. Two independent experiments with three biological replicates for each condition were performed.

QUANTIFICATION AND STATISTICAL ANALYSIS

Spearman correlation was calculated between each transcript and the trait values. BH method was applied to adjust multi-testing p-values to calculate the Padj. For correlation calculation, the S_I and Matsuda index were log-transformed, and the BMI value was transformed by square root. The R package clusterProfiler (v3.18)⁷², which relies on FET, was used for enrichment tests of various purposes, including module pathway annotations and cell type enrichment. BH method was used to adjust multi-testing p values. The package fgsea (v1.16) was applied for GSEA to identify the significantly enriched module genes in each single nuclei. Permutation analysis (n = 1000) was used to calculate the significance p values. Further detail on the computational and statistical methods are described in the Data Analysis method section above.

Supplementary Material

3

Table S1. Demographics of study population and genomic resources.

4

Table S2. Adipose tissue coexpressed gene modules in AAGMEx cohort.

5

Table S3. Cell-type enrichment of MEGENA modules in adipose tissue of AAGMEx cohort.

6

Table S4. The adipocyte marker genes in human snRNA-seq and their correlations with logMatsuda in AAGMEx cohort.

7

Table S5. BMI category enrichment of module-specific snRNA-seq nuclei in AAGMEx adipose. For each cell type, FET was performed to analyze enrichment of module-enriched cells for different categories of BMI.

8

Table S6. Adipose tissue coexpressed gene modules in METSIM cohort.

9

Table S7. Cell-type enrichment of MEGENA modules in adipose tissue of METSIM cohort.

10

Table S8. The adipocyte marker genes in human snRNA-seq and their correlations with logMatsuda in METSIM cohort.

11

Table S9. BMI category enrichment of module-specific snRNA-seq nuclei in METSIM adipose. For each cell type, FET was performed to analyze enrichment of module-enriched cells for different categories of BMI.

12

Table S10. Preservation of IS-modules of AAGMEx adipose tissue in METSIM and AREA cohort. Preservation analysis was based on network connectivity and density by the modulePreservation function of WGCNA package.

13

Table S11. Significant overlapping of AAGMEx modules and METSIM modules by FET. Only significant comparisons are shown (Padj < 0.05).

14

Table S12. Module interactions by ligand-receptor pairs in three cell-type-specific modules of adipose tissues in AAGMEx and METSIM cohorts.

15

Table S13. Muscle tissue coexpressed gene modules in AAGMEx cohort.

16

Table S14. Cell-type enrichment of MEGENA modules in muscle tissue of AAGMEx cohort.

17

Table S15. Preservation of IS-modules of AAGMEx muscle tissue in the AREA cohort. Preservation analysis was based on network connectivity and density by the modulePreservation function of WGCNA package.

18

Table S16. Enrichment of AAGMEx muscle modules for the adipose insulin sensitivity-correlated genes.

19

Table S17. Enrichment of Bayesian network drivers for the cis-eQTL genes in AAGMEx adipose.

20

Table S18. Enrichment of Bayesian network drivers for the cis-eQTL genes in AAGMEx muscle.

21

Table S19. Adipose and muscle tissue Bayesian network drivers identified as target gene for cardiometabolic trait-associated SNPs by SNP-to-gene (S2G) linking strategies in A) UK biobank and B) GWAS catalogue.

22

Table S20. Genes identified as MEGENA hubs or Bayesian network drivers for conserved adipose tissue IR-modules in both AAGMEx African Americans and METSIM Europeans. Genes are ranked by Spearman correlation with S_I and red color labels the genes for experimental validations.

23

Table S21. RNA-seq analysis of genes differentially expressed in THP1 macrophages expressing NPL-shRNA compared to control-shRNA. THP1 monocytes were differentiated to macrophage by phorbol ester (PMA or TPA, 10ng/ml, 48 hrs.) treatment. Differentially expressed transcripts with FDR<0.05 and fold change ≥1.2 are shown. Normalized counts for transcripts based on Deseq2 analysis for each biological replicate in RNA-seq analysis is shown.

24

Table S22. Biological pathways enriched for genes differentially expressed in THP1 macrophages expressing NPL-shRNA compared to control-shRNA. Selected pathways (q-value<0.05) enriched among upregulated and downregulated genes are shown.

25

Table S23. Comparison of WGCNA modules and MEGENA modules by FET enrichment analysis. The insulin sensitivity-associated modules of WGCNA were obtained from Sharma et al., 2016.

Highlights:

• Coexpression networks revealed insulin resistance (IR)-related modules across ethnicity

• Cell-type-specificity of IR-related modules was determined by single-cell data

• Network comparison showed tissue-specific gene network rewiring in IR

• Integration of the gene coexpression and causal networks identified key drivers of IR