Systems genetics identifies a co-regulated module of liver microRNAs associated with plasma LDL cholesterol in murine diet-induced dyslipidemia

Abstract

Chronically altered levels of circulating lipids, termed dyslipidemia, is a significant risk factor for a number of metabolic and cardiovascular morbidities. MicroRNAs (miRNAs) have emerged as important regulators of lipid balance, have been implicated in dyslipidemia, and have been proposed as candidate therapeutic targets in lipid-related disorders including atherosclerosis. A major limitation of most murine studies of miRNAs in lipid metabolic disorders is that they have been performed in just one (or very few) inbred strains, such as C57BL/6. Moreover, although individual miRNAs have been associated with lipid phenotypes, it is well understood that miRNAs likely work together in functional modules. To address these limitations, we implemented a systems genetics strategy using the Diversity Outbred (DO) mouse population. Specifically, we performed gene and miRNA expression profiling in the livers from ~300 genetically distinct DO mice after 18 wk on either a high-fat/high-cholesterol diet or a high-protein diet. Large-scale correlative analysis of these data with a wide range of cardio-metabolic end points revealed a co-regulated module of miRNAs significantly associated with circulating low-density lipoprotein cholesterol (LDL-C) levels. The hubs of this module were identified as miR-199a, miR-181b, miR-27a, miR-21_-_1, and miR-24. In sum, we demonstrate that a high-fat/high-cholesterol diet robustly rewires the miRNA regulatory network, and we identify a small group of co-regulated miRNAs that may exert coordinated effects to control circulating LDL-C.

dyslipidemia, or the state of having chronically altered lipid levels in the blood, is a major risk factor for developing atherosclerosis and cardiovascular disease (31, 55). The liver is the primary organ regulating plasma lipid levels, and dysfunction in certain hepatic processes has been shown to be a main contributor to dyslipidemia (17, 55). Thus, understanding the underlying molecular mechanisms in the liver that cause or respond to dyslipidemia is important for ultimately identifying novel therapeutic targets.

MicroRNAs (miRNAs), which are small noncoding RNAs that fine-tune gene expression primarily at the posttranscriptional level, have emerged as key players in many processes, including those involved with lipid homeostasis. Several hepatic miRNAs have been associated with atherosclerosis and hyperlipidemia, including microRNA (miR)-27 (44, 49), miR-122 (11), miR-148a (15), miR-33 (18, 39, 40), and miR-30c (47).

A number of these miRNAs, including miR-33 and miR-30c, have been identified as potential therapeutic targets for atherosclerosis and hyperlipidemia (7, 19). However, there are at least two limitations shared by most of these studies. First, the vast majority of the studies of miRNAs in lipid-related disorders have been performed in C57BL/6 mice, an inbred mouse strain. Although these have produced promising results, the relevance of the findings to genetically diverse outbred populations (like humans) is unclear. Second, all of these studies involve a focused effort to understand an individual miRNA and its contextually relevant target genes. While this is a reasonable approach, it is not conducive to understanding the roles of miRNAs within a network of other miRNAs and genes. This is an important limitation since there is evidence to support the idea that miRNAs often work in cooperative groups to regulate gene expression (22, 24, 51).

One approach to addressing both of these limitations is to utilize a systems genetics strategy wherein transcript levels are quantified in tissues of interest, integrated with underlying genetic information, and related to clinical traits of interest. This has been successfully performed in a number of studies focused on gene networks in glucose and lipid metabolism in humans (34), mice (52), and flies (4). Subsequent studies have consistently demonstrated that the candidates identified with these approaches are critical mediators of these processes (12, 29, 54).

A resource that is ideal for such a study is the Diversity Outbred (DO) mouse population. The DO mice were created by strategic outbreeding of eight parental strains of mice, the same ones used to generate the Collaborative Cross (CC), which are distinct lines of mice that are maintained as recombinant inbred strains (8). While DO mice are similar to CC mice in that they represent mosaics of the eight founder lines, they are different from the CC in that each DO mouse has a unique, nonreproducible genome with dramatically increased levels of accumulated recombination (9). With each mouse harboring around 45 million variants in its genome, there is a high degree of allelic and phenotypic variation within the population. The extent of genotypic and phenotypic diversity, as well as the high frequency of recombination events, is very useful for identifying genetic contributions to traits of interest with high resolution (48).

In the present study, we utilized a cohort of almost 300 DO mice to interrogate the hepatic network of miRNAs associated with circulating lipid levels in diet-induced dyslipidemia. We identify a key co-regulated module of miRNAs that is strongly associated with LDL cholesterol (LDL-C), which is a significant risk factor for many downstream morbidities, including atherosclerosis and metabolic disease.

MATERIALS AND METHODS

Animals, diets, and phenotyping.

Details on the origin, housing, husbandry, and treatment of the DO animals, diet compositions, and measurement of total cholesterol, triglycerides, glucose, and insulin have been provided previously (46). For HDL precipitation, 100 μl aliquots of blood plasma samples were diluted 1:4 with PBS, combined with 9 μl of heparin-MnCl₂, and centrifuged. Supernatant was removed and combined with working cholesterol reagent (mixture of reagent, HDCBS, cholesterol oxidase, cholesterol esterase, and horseradish peroxidase) (35). Samples were run in triplicate on 96-well flat-bottom plates, and absorbance was read at 515 nm using BioTek plate reader and Gen5 software. Absorbance values were averaged across triplicates, and concentrations were calculated. To determine on very-low-density lipoprotein cholesterol/low-density lipoprotein cholesterol (VLDL-C/LDL-C) levels, high-density lipoprotein cholesterol (HDL-C) was subtracted from total cholesterol. Markers of liver inflammation, alanine aminotransferase (ALT) and aspartate aminotransferase (AST), were measured using a Biolis 24i Analyzer (Carolina Liquid Chemistries, Winston Salem, NC).

RNA extraction.

Livers were flash-frozen in liquid nitrogen and subsequently stored at −80°C until their use. Total RNA was isolated by automated instrumentation from ~25 mg of liver tissue per sample using Norgen Total RNA Purification Kit (Norgen, Ontario, Canada; catalog no. 24300). Quant-iT RiboGreen from ThermoFisher Scientific (Waltham, MA; catalog no. R11490) was used to measure RNA concentration by fluorimetry. RNA integrity was determined by Bioanalyzer from Bio-Rad Technologies. Only samples with RNA quality indicator ≥7.5 were used for microarray and sequencing.

Microarray.

High-quality RNA was available from livers of 268 of the 288 DO mice and was used for microarray gene expression analysis. The RNA was hybridized to Affymetrix Mouse Gene 2.1 ST 96-Array Plate using the GeneTitan Affymetrix instrument according to standard manufacturer’s protocol. Robust multiarray average (RMA) method was used to estimate normalized expression levels of transcripts (median polish and sketch-quantile normalization). Affymetrix Expression Console software was used for quality control assessment, and as a result, six of the mice were removed for not passing tests, leaving 262 samples to be analyzed. All probes containing known single nucleotide polymorphisms (SNPs) from the eight founder inbred mouse strains of the DO mouse population were masked (165,204 probes) during normalization by downloading the SNPs from the Sanger sequencing website (http://www.sanger.ac.uk/science/data/mouse-genomes-project) and overlapping them with probe sequences. We removed all control probes (190 probes), reporter probes (82 probes), and normalization probes (6,683 probes) from the probe sets before running weighted gene coexpression network analysis (WGCNA). Probes were filtered using an expression threshold of a minimum RMA of 4 in at least one-quarter of the samples, which left 15,105 probes. Differential expression analysis was performed by Student’s t-test, and P values were corrected using the Bonferroni method. Microarray data are available on the Gene Expression Omnibus (GEO) repository, accession number GSE99561.

Small RNA-sequencing.

High-quality RNA was available from livers of 269 of 288 of the DO mice was used for small RNA sequencing (smRNA-seq). Libraries were created using New England Biosciences NEBNext Multiplex Small RNA Library Prep Set for Illumina, and 50 bp single-read sequencing was carried out on the Illumina HiSeq platform resulting in an average of over 16 million reads per sample. miRquant 2.0 (20) was used to trim off adapter sequences, align reads to the mouse genome, and quantify miRNAs and their isoforms (termed isomiRs). A previous study in CC mice has shown that miRNAs do not contain variants across founder strains within their seed regions, so reads were aligned to the mm9 mouse genome (41). Reads were normalized to reads per millions mapped to miRNAs (RPMMMs). An expression threshold of at least 50 RPMMMs in about one-quarter of all samples was set to filter out the lowly expressed miRNAs, which resulted in a set of 246 robustly expressed miRNAs. The results were consistent when repeated with an expression threshold of at least 100 RPMMMs in at least a quarter of all samples. Hierarchical clustering of the samples’ expression profiles was performed based on several data set characteristics (library prep date, plate, etc.) to ensure there were no batch effects. Differential expression of miRNAs was performed by Student’s t-test, and P values were corrected by the Benjamini-Hochberg method. Small RNA-seq data are available on the GEO repository, accession number GSE99561.

WGCNA.

The WGCNA R package was used to identify the co-regulated modules (CRMs) for miRNAs and genes (25). Only expression data from the high-fat/cholic acid (HFCA)-fed mouse samples were used to identify the CRMs.

For the miRNA network analysis, we matched the smRNA-seq samples with the mice for which phenotypic data was measured, and were left with 256 DO mice, of which 135 were HFCA-fed mice. The RPMMMs were transformed to log2(x+1) scale. The soft threshold was chosen by running the pickSoftThreshold function to find the best fit to a scale-free topology, and beta was set to 14 because it fit with an R² value ≥ 0.8, and connectivity measures suggested the possibility of identifying hubs. An adjacency matrix was created using Pearson correlations. From the adjacency matrix, the topological overlap measure (TOM) was calculated by the signed method. The dissimilarity measure was calculated by 1-TOM, and this was used to create a dendrogram according to the Ward’s hierarchical clustering method. We use Ward’s method instead of the default average method because it considers the variance in expression between miRNAs before choosing to put them in a clade together. Thus, miRNAs within the same clade have the lowest variance in expression possible, which is meaningful for identifying clusters of co-regulated miRNAs. The hybrid tree-cutting algorithm was used to form the modules, which were left unmerged. Module eigenmiRs (MEms) were calculated with the moduleEigengenes function and were correlated with each phenotype measured in the mice using the biweight midcorrelation. Module significance was also calculated by biweight midcorrelation method. Modules with the highest correlation or inverse correlation (|coefficient| > 0.4) were taken as those of interest. miRNAs that were found to have the highest aggregated TOMs within the respective CRMs, the highest kWithin (intramodular connectivity measure) and the highest Pearson correlations to their MEms were identified as hub miRNAs. Network files were exported to Cytoscape (43) for visualization.

For WGCNA on mRNAs, we matched microarray samples with the mice we have measured phenotypes for, and we were left with 249 DO mice, of which 134 were HFCA-fed mice. We parsed out the top 5,000 most variably expressed mRNAs within the HFCA samples from the 15,105. We found that a large number of the mRNAs were not separated into CRMs but were allocated to an undefined group, which is where genes are assigned when they cannot be placed into any module. We gradually reduced the number of mRNAs from 5,000 to 3,000. Using the top 3,000 most variably expressed mRNAs within the HFCA samples allowed for each gene to be assigned to a specific CRM. Soft threshold beta was set to 9 because it fit a scale-free topology with an R² value > 0.8, and connectivity measures suggested the possibility of identifying hub genes. An adjacency matrix was calculated with Pearson correlations, and the TOM was calculated by the signed method. The 1-TOM distance measure was used to create a dendrogram according to the Ward’s method of hierarchical clustering. The hybrid tree-cutting algorithm was used to form the modules, which were left unmerged. Module eigengenes were calculated and correlated with each phenotype measured in the mice by the biweight midcorrelation. The same method was used to calculate correlations between gene CRMs and the brown miRNA CRM. Modules with the highest correlation or inverse correlation (|coefficient| > 0.4) were recognized as those of potential interest.

Pathway enrichment analysis.

Enrichr was used to perform pathway enrichment analysis and gene ontology analysis on the genes within each gene CRM (5, 23).

miRNA target site enrichment analysis.

miRhub (42) was used to perform target site enrichment analysis for miRNAs. Briefly, miRhub employs a Monte Carlo simulation strategy to determine which miRNAs, if any, have an overrepresentation of predicted target sites at a specified level of conservation in a set of input genes. We ran miRhub on genes up- and downregulated in the liver from HFCA-fed mice and required positional conservation of predicted target sites in at least two mammalian species in addition to mouse.

RESULTS

Effects of a HFCA, dyslipidemia-inducing diet on plasma lipoprotein cholesterol levels in the DO mouse population.

In a previous study (46), we demonstrated the value of a specific multiparental mouse population, the DO resource, for mapping quantitative trait loci (QTL) and identifying candidate genes and potential therapeutic targets for dyslipidemia and atherosclerosis. An initial cohort of 288 DO mice comprising 144 sibling pairs was fed either an HFCA-containing diet, which induces dyslipidemia, or a calorie-matched high-protein diet (HP) for 18 wk. We identified QTLs for atherosclerotic lesion size, prediet circulating triglycerides, and postdiet circulating total cholesterol. In the present study (Fig. 1), we have analyzed plasma samples from the same mouse cohort for several additional cardio-metabolic end points, with a primary focus on VLDL/LDL-C and HDL-C before and after the 18 wk diet exposure. We found that the average levels of ALT, AST, VLDL/LDL-C, but not HDL-C are significantly elevated in HFCA-fed DO mice relative to the HP-fed DO mice (Fig. 2, A–D). Notably, all but three of the HFCA-fed mice had higher circulating VLDL/LDL-C levels than the average level among HP-fed mice. Moreover, VLDL/LDL-C levels were highly variable among the HFCA-fed mice, ranging from <10 mg/dl to > 300 mg/dl (Fig. 2C). This finding indicates that effects of the HFCA diet on plasma VLDL/LDL-C is highly dependent on genetic background.

Fig. 1.

Summary diagram of study design: 288 Diversity Outbred (DO) mice, each having a different composite of the 8 founder mouse strain genomes, were fed either high-protein (HP) or high-fat/cholic acid (HFCA) diet. Cardio-metabolic end points were measured before and after diet intervention. RNA was isolated from the livers of each of the mice and used for microarray analysis to measure gene expression and small RNA sequencing.

Fig. 2.

Plasma very-low-density lipoprotein/low-density lipoprotein (VLDL/LDL), aspartate aminotransferase (AST), and alanine aminotransferase (ALT) are dramatically affected by HFCA diet. Box plot of postdiet AST (mg/dl) (A), postdiet ALT (mg/dl) (B), postdiet plasma VLDL/LDL-C (mg/dl) (C), and postdiet plasma high-density lipoprotein cholesterol (HDL-C) (mg/dl) (D) concentration in HP-fed DO mice and HFCA-fed DO mice. Each dot represents 1 mouse in the respective diet. Hinges of boxplots represent the 1st and 3rd quartile of expression.

Effects of the HFCA diet on liver miRNA expression in the DO cohort.

Given the importance of the liver in maintaining cholesterol and lipoprotein homeostasis, and the growing appreciation for miRNAs in the control of cholesterol metabolism, we reasoned that hepatic miRNAs may associate with the observed variation in lipid phenotypes, particularly VLDL/LDL-C, across the mice in the DO cohort. To test this hypothesis, we performed smRNA-seq on liver tissue from 269 of the same DO mice at an average depth of 16 million reads per sample (range 7,662,595–30,8861,518 reads). The reads were mapped to the mouse genome (mm9), and miRNAs and their isoforms (referred to as isomiRs) were annotated and quantified with miRquant 2.0 (20). Detailed information on the mapping statistics is provided in Supplemental Table S1. (The online version of this article contains supplemental material.)

To normalize miRNA expression, we used the RPMMM method. After we filtered out those miRNAs with low levels of expression across the majority of samples, 246 miRNAs remained (materials and methods). More than three-quarters of these were significantly altered [false discovery rate (FDR) < 0.05] in the livers of the mice fed the dyslipidemic HFCA diet relative to those fed the HP diet (Fig. 3). Specifically, 44 miRNAs were significantly upregulated (fold change ≥ 1.5), and 39 miRNAs were significantly downregulated in the HFCA-fed mice (fold change ≤ −1.5). These differentially expressed miRNAs included several that have previously been implicated in the development and/or progression of atherosclerosis or hyperlipidemia. One such miRNA is miR-34a, which is upregulated in the plasma of ApoE knockout (KO) C57BL/6 mice (16), a well-established animal model of atherosclerosis; in the liver tissue of high-fat diet-fed C57BL/6 mice (10, 13), a model of hepatic steatosis and obesity; as well as in the plaques of humans with coronary artery disease (37). We found that liver miR-34a expression levels in most HFCA diet-fed mice were significantly greater than in the HP diet-fed mice (Fig. 4A). Notably, miR-34a expression in all but two (98.7%) of the HFCA samples was above the average expression in the HP-fed mice. In addition, analysis of miR-34a expression within the sibling pairs shows the majority of the pairs follow the same trend of an increase in miR-34a expression as a result of HFCA feeding (Fig. 4B). We also identified other miRNAs significantly altered by HFCA diet, including several that have not been previously associated with regulation of lipid levels, such as miR-874, which was downregulated by almost twofold, and miR-1247-5p, which was upregulated by over fourfold (Fig. 4, C and D).

Fig. 3.

Diet alters microRNA (miRNA) expression. Volcano plot of differentially expressed liver miRNAs between HFCA-fed and HP-fed DO mice after those that had low expression were filtered out. Each dot represents 1 miRNA. Red dots are miRNAs that are upregulated in HFCA-fed mice relative to HP-fed mice with a fold-change of 1.5 or more and a false discovery rate (FDR) ≤ 0.05. Blue dots are miRNAs that are downregulated in HFCA-fed mice relative to HP-fed mice with a fold-change of 1.5 or more and an FDR ≤ 0.05. Horizontal dashed line denotes FDR = 0.05. Vertical dashed lines denote fold change of −1.5 (left) and 1.5 (right).

Fig. 4.

Diet alters expression of specific miRNAs. A: box plot of miR-34a-5p. B: miR-34a-5p expression with lines connecting DO mouse sibling pairs in either diet group. C–E: box plots of miR-874-3p, miR-1247-5p, and miR-30c-2-5p expression in HP-fed and HFCA-fed DO mice. For all box plots, each dot represents 1 mouse in the respective diet. Hinges of boxplots represent the 1st and 3rd quartile of expression. F: miR-30c-2-5p expression with lines connecting DO mouse sibling pairs in either diet group.

The HFCA diet did not universally alter expression for every miRNA, including miR-30c, which has been shown to be significantly downregulated in the liver of the ApoE C57BL/6 KO model. Overexpression of miR-30c mimic can mitigate hyperlipidemia and regress atherosclerosis in both wild-type and ApoE KO C57BL/6 mice (47). However, in our DO cohort, while there is an expected trend toward lower liver miR-30c expression in the HFCA-fed mice, the difference is not substantial and the expression distributions among HFCA- and HP-fed mice are largely overlapping (Fig. 4E). The response of the miR-30c levels to the HFCA diet is much more mixed than what is seen for miR-34a (Fig. 4F). These data suggest that while the dyslipidemia-inducing diet has a robust effect on liver miR-34a, the effect on miR-30c may be more dependent on genetic composition than diet. The extensive variation in hepatic expression that we observed for miRNAs across the DO mice that were fed the same HFCA diet is likely due to interactions between the underlying genetics and diet, and it provided a unique opportunity to identify groups or modules of miRNAs that exhibit highly similar genotype-dependent responses to the HFCA diet.

Identification of co-regulated miRNA modules and correlation with lipid phenotypes.

We next sought to identify groups, or CRMs, of hepatic miRNAs across the HFCA-fed DO mice. We applied the WGCNA (25) to the 246 miRNAs robustly expressed in the liver (materials and methods). This analysis identified five main miRNA CRMs (mCRMs) with anywhere from 32 to 60 miRNAs in each CRM (Fig. 5A). We then used the biweight midcorrelation (bicor) analysis to assess the extent of correlation between each of these modules and the various end points that were measured in these mice (materials and methods). We found that one particular module, which we will refer to as the brown module, comprises 34 miRNAs/isomiRs and is highly correlated with important end points, most notably postdiet circulating VLDL/LDL-C (bicor coefficient 0.49) (Fig. 5B). Furthermore, all of the top 20 correlations between any of the 246 miRNAs and any of the end points involve miRNAs from the brown module, and almost all are associated with the diet-induced change in plasma VLDL/LDL-C (Table 1).

Fig. 5.

miRNA coexpression analysis identifies module associated with metabolic traits. A: miRNA modules formed by weighted gene coexpression network analysis (WGCNA). Only HFCA mice were used during analysis. Reads per millions mapped to microRNAs (RPMMMs) were converted using log2(x+1) and used to calculate Pearson correlations. Dendrogram was created using the 1-topological overlap measure (TOM), and Ward’s method of hierarchical clustering. Modules were formed with the hybrid tree-cutting function in the WGCNA software package. Heat map of miRNA module eigenmiRs correlated to cardio-metabolic end points measured in the HFCA-fed DO mice (B) and HP-fed DO mice (C). EigenmiRs were calculated using the WGCNA function moduleEigengenes and correlated using the biweight midcorrelation to normalized end point values. The intensity of orange or blue denotes how close the correlation coefficient is to 1 or −1, respectively. Top numbers are biweight midcorrelation coefficients; bottom numbers are P values. D: Cytoscape visualization of brown miRNA co-regulated modules (mCRM). Each node represents 1 miRNA. Each edge represents high cocorrelation. The dashed circle highlights the hub miRNAs in this module as determined by number of connections and weight of connections.

Table 1.

Top correlations between miRNAs and cardio-metabolic end points

miRNA	Associated Phenotype	Cor_coef	P Value
mmu-mir-24-2-5p	delta VLDL/LDL-C	0.66	2.88E-09
mmu-mir-24-1-3p	delta VLDL/LDL-C	0.63	2.48E-08
mmu-mir-24-2-3p	delta VLDL/LDL-C	0.63	2.48E-08
mmu-let-7i-5p	delta VLDL/LDL-C	0.61	5.35E-08
mmu-let-7i-5p	postdiet ALT	0.61	5.61E-13
mmu-mir-29a-3p_-_1	delta VLDL/LDL-C	0.60	9.75E-08
mmu-let-7i-5p	postdiet AST	0.59	7.19E-07
mmu-mir-181b-1-5pd	delta VLDL/LDL-C	0.58	3.95E-07
mmu-mir-214-3p_-_1	delta VLDL/LDL-C	0.58	4.18E-07
mmu-mir-21-5p	delta VLDL/LDL-C	0.58	4.19E-07
mmu-mir-501-3p	postdiet ALT	0.58	1.42E-11
mmu-mir-143-3p	delta VLDL/LDL-C	0.58	5.18E-07
mmu-mir-99b-3p	delta VLDL/LDL-C	0.58	5.32E-07
mmu-mir-199a-2-3p	delta VLDL/LDL-C	0.57	5.94E-07
mmu-mir-199a-1-3p	delta VLDL/LDL-C	0.57	6.22E-07
mmu-mir-199b-3p	delta VLDL/LDL-C	0.57	6.22E-07
mmu-mir-181b-2-5p	delta VLDL/LDL-C	0.56	9.75E-07
mmu-mir-99b-5p	delta VLDL/LDL-C	0.56	1.02E-06
mmu-let-7e-5p	delta VLDL/LDL-C	0.56	1.04E-06
mmu-mir-21-5p_-_1	delta VLDL/LDL-C	0.56	1.33E-06
mmu-mir-24-2-5p	postdiet LDL-C	0.55	6.37E-12
mmu-mir-214-3p	delta VLDL/LDL-C	0.54	2.77E-06
mmu-mir-21-5p	postdiet ALT	0.54	5.40E-10
mmu-mir-146b-5p	postdiet AST	0.54	7.10E-06
mmu-mir-146b-5p	postdiet ALT	0.54	5.85E-10
mmu-mir-342-3p	postdiet ALT	0.54	6.13E-10
mmu-mir-143-3p	postdiet ALT	0.54	7.01E-10
mmu-mir-146a-5p	postdiet ALT	0.53	1.49E-09
mmu-mir-21-5p	postdiet LDL-C	0.52	1.07E-10
mmu-mir-501-3p	postdiet AST	0.52	1.67E-05
mmu-mir-27a-3p	delta VLDL/LDL-C	0.52	9.24E-06
mmu-mir-24-2-5p	postdiet ALT	0.52	4.08E-09
mmu-mir-146a-5p	delta VLDL/LDL-C	0.52	1.12E-05
mmu-mir-99a-5p	delta VLDL/LDL-C	0.51	1.14E-05
mmu-mir-191-5p_+_1	postdiet ALT	0.51	4.83E-09
mmu-let-7e-5p	postdiet AST	0.51	2.56E-05
mmu-mir-1839-5p	postdiet ALT	0.51	7.11E-09
mmu-mir-1839-5p_+_1	postdiet ALT	0.51	8.68E-09
mmu-mir-199a-1-5p	delta VLDL/LDL-C	0.50	1.93E-05
mmu-mir-199a-2-5p	delta VLDL/LDL-C	0.50	1.94E-05
mmu-mir-532-5p	postdiet ALT	0.50	1.27E-08
mmu-mir-191-5p	delta VLDL/LDL-C	0.50	2.39E-05
mmu-mir-322-3p	postdiet ALT	0.50	2.12E-08
mmu-mir-21-5p	postdiet AST	0.49	5.07E-05
mmu-let-7i-5p	postdiet LDL-C	0.49	1.65E-09
mmu-mir-140-5p	delta VLDL/LDL-C	0.49	3.78E-05
mmu-let-7e-5p	postdiet ALT	0.49	3.95E-08
mmu-mir-191-5p_+_1	delta VLDL/LDL-C	0.49	4.09E-05
mmu-mir-322-3p	delta VLDL/LDL-C	0.48	4.28E-05
mmu-mir-425-5p	delta VLDL/LDL-C	0.48	4.77E-05
mmu-mir-200b-3p	postdiet AST	0.48	8.57E-05
mmu-mir-10a-5p_+_1	postdiet ALT	0.48	7.24E-08
mmu-mir-143-3p	postdiet LDL-C	0.48	7.54E-09
mmu-mir-674-3p	delta VLDL/LDL-C	0.48	6.14E-05
mmu-mir-21-5p_-_1	postdiet AST	0.47	0.0001225
mmu-mir-99b-5p	postdiet ALT	0.47	1.43E-07
mmu-mir-21-5p_-_1	postdiet ALT	0.47	1.43E-07
mmu-mir-143-3p	postdiet AST	0.47	0.00015041
mmu-mir-146b-5p	delta VLDL/LDL-C	0.47	8.95E-05
mmu-mir-21-5p_-_1	postdiet LDL-C	0.46	2.32E-08
mmu-mir-10a-5p_+_1	delta VLDL/LDL-C	0.46	0.00010479
mmu-mir-186-5p	delta VLDL/LDL-C	0.46	0.00010562
mmu-mir-10a-5p	postdiet AST	0.46	0.00017707
mmu-mir-24-2-5p	delta cholesterol	0.46	2.84E-08
mmu-mir-24-2-5p	postdiet cholesterol	0.46	3.15E-08
mmu-mir-532-5p	postdiet AST	0.45	0.00023692
mmu-mir-140-3p_+_1	delta VLDL/LDL-C	0.45	0.00014825
mmu-mir-122-5p	postdiet ALT	−0.45	4.22E-07
mmu-mir-122-5p	delta VLDL/LDL-C	−0.56	1.24E-06

Biweight midcorrelation coefficients of microRNAs (miRNAs) as they are related to the cardio-metabolic end points. Only correlation values ≥0.45 and ≤−0.45 are shown.

As a comparison, we performed the WGCNA analysis with the HP-fed mice using the liver expression data for the same 246 miRNAs. Of the five mCRMs that were identified, none were strongly correlated with any of the cardio-metabolic end points that were measured in the mice (Fig. 5C). Also, none of the mCRMs in the HP-only analysis exhibited substantial overlap with any of the mCRMs identified in the HFCA analysis, with an average of only ~8 shared miRNAs (which represents an average of just ~20% shared between any one HP mCRM and any one HFCA mCRM). This indicates that the HFCA diet leads to a very specific, robust rewiring of the regulatory networks governing miRNA expression.

Identification of miRNA “hubs” in the brown module correlated with postdiet plasma VLDL/LDL-C.

The strength of connection between miRNAs in an mCRM is indicative of the extent of cocorrelation. The miRNAs with the highest connectivity scores are defined as “hubs.” To define the miRNA hubs in the brown mCRM, we ranked the miRNAs according to their TOM scores and intramodular connectivity measures, which are the WGCNA metrics for interconnectedness. The miRNAs in the 75th percentile were identified as hubs in the brown mCRM: miR-199a, miR-181b, miR-27a, miR-21_-_1, and miR-24. Each of these miRNAs was very highly correlated with the module eigenmiR, or first principal component, of the brown module (Fig. 5D, Table 2), with miR-199a being the most highly correlated (Pearson coefficient = 0.93).

Table 2.

miRNAs with the highest interconnectedness are identified as hubs

Brown mCRM miRNAs	kWithin	kWithin (Scaled)	Cor to MEbrown	Summed TOM
mmu-mir-199a-2-3p	7.0038	1.0000	0.93	8.0038
mmu-mir-199a-1-3p	6.9947	0.9987	0.93	7.9947
mmu-mir-199b-3p	6.9947	0.9987	0.93	7.9947
mmu-mir-181b-1-5p	6.2731	0.8957	0.92	7.2731
mmu-mir-181b-2–5p	6.1845	0.8830	0.92	7.1845
mmu-mir-199a-2-5p	6.0431	0.8628	0.92	7.0431
mmu-mir-199a-1-5p	6.0431	0.8628	0.92	7.0431
mmu-mir-24-2-3p	6.0018	0.8569	0.89	7.0018
mmu-mir-24-1-3p	6.0018	0.8569	0.89	7.0018
mmu-mir-27a-3p	5.0559	0.7219	0.88	6.0559
mmu-mir-21-5p_-_1	5.0528	0.7214	0.89	6.0528
mmu-let-7e-5p	4.9871	0.7121	0.88	5.9871
mmu-mir-200a-3p	4.8912	0.6984	0.86	5.8912
mmu-mir-214-3p_-_1	4.8437	0.6916	0.86	5.8437
mmu-mir-214-3p	4.6433	0.6630	0.84	5.6433
mmu-let-7i-5p	4.4421	0.6342	0.86	5.4421
mmu-mir-200c-3p	4.2868	0.6121	0.81	5.2868
mmu-mir-200b-3p	4.2822	0.6114	0.81	5.2822
mmu-mir-142-5p_-_2	4.2114	0.6013	0.82	5.2114
mmu-mir-24-2-5p	4.2033	0.6001	0.86	5.2033
mmu-mir-146b-5p	3.9551	0.5647	0.84	4.9551
mmu-mir-21-5p	3.8489	0.5495	0.81	4.8489
mmu-mir-29a-3p_-_1	3.7823	0.5400	0.80	4.7823
mmu-mir-872-5p	3.6169	0.5164	0.78	4.6169
mmu-mir-99b-3p	3.4688	0.4953	0.77	4.4688
mmu-mir-99b-5p	3.4553	0.4933	0.76	4.4553
mmu-mir-146a-5p	3.3812	0.4828	0.78	4.3812
mmu-mir-143–3p	3.3189	0.4739	0.78	4.3189
mmu-mir-125a-5p	3.1855	0.4548	0.69	4.1855
mmu-mir-99a-5p	3.1538	0.4503	0.74	4.1538
mmu-mir-342-3p	3.0082	0.4295	0.72	4.0082
mmu-mir-322-3p	3.0044	0.4290	0.75	4.0044
mmu-mir-501-3p	2.9523	0.4215	0.73	3.9523
mmu-mir-532-5p	1.6634	0.2375	0.53	2.6634

Members of the brown microRNA co-regulated modules (mCRM) ranked according to intramodular connectivity measure (kWithin). Column 2 lists kWithin scaled by the module. Column 3 lists Pearson correlations to the module eigenmiR, and the last column lists the total summed topological overlap measure (TOM) for each miRNA.

Analysis of gene expression data and identification of several gene CRMs associated with postdiet LDL-C and inversely correlated with the brown miRNA module.

To determine the effects of the HFCA diet on gene expression, we performed microarray analysis on 262 liver samples from the DO mice (34,390 mRNAs detected, 15,105 mRNAs with RMA ≥ 4 in at least one-quarter of the samples). Differential expression analysis revealed that 4,236 genes were significantly (corrected P value < 1.20 × 10⁻⁶) upregulated, of which 401 exhibited a fold-change >2, and 3,603 genes significantly (corrected P value < 1.20 × 10⁻⁶) downregulated, of which 140 exhibited a fold-change <−2, in the liver tissue from HFCA-fed mice compared with HP-fed mice (Fig. 6A). We found that lipid processing and metabolism genes such as lipoprotein lipase (Lpl) are upregulated (+14.8-fold), whereas cholesterol biosynthesis genes such as squalene epoxidase (Sqle) are downregulated (−19.5-fold). These results are expected in response to HFCA, which leads to a dramatic increase in dietary lipid and cholesterol, eliciting a suppression of endogenous lipid/cholesterol synthesis and increase in lipid metabolic activity.

Fig. 6.

Differential expression analysis for gene expression. A: volcano plot of differentially expressed liver genes between HFCA-fed and HP-fed DO mice. Each dot represents 1 probe. Red dots are probes that are upregulated (n = 401) in HFCA-fed mice relative to HP-fed mice with a fold-change of 2 or more and a P value ≤ 1.20e-06 (Bonferroni correction). Blue dots are probes that are downregulated (n = 140) in HFCA-fed mice relative to HP-fed mice with a fold-change of 2 or more and an P value ≤ 1.20e-06. Horizontal dashed line denotes –log10(1.20e-06). Vertical dashed lines denote fold change of −2 (left) and 2 (right). B, C: correlation plots illustrating the inverse relationship between miR-27a and Hmgcr, Ldlr, Acly, and Lpin1 expression (correlations calculated with bicor).

Using the tool miRhub we demonstrated that the 3,603 genes significantly downregulated in the liver from HFCA-fed mice are significantly enriched for predicted conserved target sites for the brown module hub miRNAs miR-27a (empirical P value = 0.008), miR-199a (P = 0.015), miR-24 (P = 0.020), and miR-181b (P = 0.045), each of which is significantly upregulated in the liver in response to HFCA. This finding held for miR-27a (P = 0.017) upon miRhub analysis of the subset of significantly downregulated genes that are altered by more than twofold (n = 140). We and others have shown previously that the miR-27 family is involved in the control of lipid balance in part through regulation of genes in the lipid synthesis and uptake pathways. We confirmed that the previously validated targets of miR-27, Hmgcr (42) and Ldlr (1), which encode proteins critical for cholesterol biosynthesis and LDL-C uptake, respectively, are indeed among the 140 genes significantly downregulated in the liver from HFCA-fed mice and are significantly inversely correlated with miR-27a levels (Fig. 6B). We additionally identified a predicted conserved target site for miR-27 in the 3′-untranslated regions of Acly and Lpin1, each of which is reduced in the liver by more than twofold in HFCA-fed mice and is inversely correlated with miR-27a levels (Fig. 6C). Acly catalyzes an early step in fatty acid synthesis, and Lpin1 mediates one of the final steps in triglyceride biosynthesis in the liver. Analysis of published Argonaute (Ago) CLIP-seq data from human Huh7 hepatoma cells provided strong experimental support for a regulatory interaction between miR-27 and Acly and, to a slightly lesser extent, miR-27 and Lpin1 (27), thereby adding to the evidence that miR-27 is a critical diet-responsive regulator of lipid homeostasis.

As we did with miRNA expression data, we next analyzed the gene expression data via WGCNA. Using the top 3,000 most variable genes within the HFCA-fed mice, we identified 20 groups, or gene CRMs (gCRMs), each of which contains highly cocorrelated mRNAs that are likely co-regulated (Fig. 7A). Among these there are two gCRMs, pink and midnight blue, that are relatively highly correlated with postdiet plasma LDL-C (pink bicor coefficient = 0.5, midnight blue bicor coefficient = 0.58), and three gCRMs, magenta, tan, and turquoise, which are strongly inversely correlated with postdiet LDL-C (bicor coefficients −0.56, −0.56, and −0.51, respectively) (Fig. 7B). As expected, individual miRNAs from the brown mCRM (e.g., miR-199a) exhibited strong inverse correlation with the genes in the magenta, light cyan, tan, and turquoise gCRMs, whereas miRNAs in other mCRMs did not (e.g., miR-151), suggestive of a unique association in diet-induced dyslipidemia between miRNAs in the brown mCRM and genes in these four gCRMs (Fig. 7C). Genes in these gCRMs that are candidate targets of one or more of the miRNAs in the brown mCRM are listed in Supplemental Table S2.

Fig. 7.

Gene coexpression analysis identifies gene co-regulated modules (gCRMs) that are correlated with the brown mCRM. A: gCRMs formed with WGCNA. Only HFCA mice were used during analysis. The top 3,000 most variable genes and the hybrid tree-cutting function in the WGCNA software package were used to form modules. B: heat map of correlations between mRNA (gene) modules and brown miRNA module, and list of cardio-metabolic end points measured in the DO mice. Eigengenes and eigenmiRs were calculated by the WGCNA function and correlated with the biweight midcorrelation to normalized end point values. The intensity of orange or blue denotes how close the correlation coefficient is to 1 or −1, respectively. Numbers in parentheses are Student P values. C: aggregate correlation values of miRNAs to gene module members. Biweight midcorrelations were calculated between each individual miRNA and each gene. Values were averaged (mean) across gene module members. Dashed lines denote significant correlation values (−0.198, 0.198) as determined by 97.5% quantile of 1,000 permutations.

DISCUSSION

To our knowledge, this is the first large-scale study aimed at using a genetically diverse mouse population to identify hepatic miRNAs associated with a variety of different cardio-metabolic end points pertinent to diet-induced dyslipidemia and metabolic dysfunction. Our study utilizes a population of outbred mice to delineate the miRNA regulatory networks associated with diet-induced metabolic dysfunction. This systems-level approach yielded at least four novel findings. First, we determined that certain metabolic phenotypes in DO mice, notably plasma VLDL/LDL-C, are far more variable in response to a dyslipidemia-inducing diet than other phenotypes. Second, we identified specific miRNAs (e.g., miR-34a) that exhibit a dramatic response to the dyslipidemic diet irrespective of host genotype; these are miRNAs for which the diet had the most dominant effect. Third, we showed that some miRNAs (e.g., miR-30c), though previously identified as dramatically altered in diet-induced dyslipidemia, were not altered by HFCA in most DO mice, indicating the prominent contribution of genetics to the responses of certain liver miRNAs to diet. Fourthly, we identified a co-regulated module of miRNAs that is strongly associated with circulating levels of VLDL/LDL-C, a significant risk factor for atherosclerosis and other related cardiovascular conditions. Moreover, we predicted that miRNAs in this module together target about one-third of the genes that are inversely correlated with VLDL/LDL-C.

We found that some miRNAs, such as miR-34a, are altered by high-fat/high-cholesterol diet across most strains in a manner that is consistent with expectation based on previous studies in C57BL/6 mice. However, others such as miR-30c, exhibited much more genotype-dependent behaviors, indicating that previous implications of miR-30c as a potential therapeutic in dyslipidemia and atherosclerosis based on studies in C57BL/6 mice are not necessarily generalizable to different genetic backgrounds. This finding underscores the importance and value of performing studies across diverse genetic backgrounds to more broadly define the variability in miRNA responses to diet or other perturbations. Multiparental resources such as the DO and CC appear to be especially promising in this regard.

It is worth noting that another prominent atherosclerosis-related miRNA, miR-33 (18, 28, 30, 38–40), was not included in the set of miRNAs considered in this analysis because the levels at which it was detected did not reach our threshold for robust expression. This may be because the library preparation protocol we used in this study is biased against the detection of some miRNAs such as miR-33 [most likely due to adapter ligation bias, as has been reported previously (2)]. Due to the importance of miR-33 in lipid biology and atherosclerosis, it may be worth reanalyzing these samples in the future with alternate detection methods such as RT-qPCR as well as the Bioo Scientific NextFlex V3 library preparation protocol, which is intended to mitigate adapter ligation bias (2).

To identify those mCRMs that are most affected by HFCA diet interactions with genetic composition, we used only the HFCA-fed mouse samples in the WGCNA process. Since all of these mice were fed the same impactful HFCA diet, and each harbored a distinct genome, we reasoned that the use of just these samples would best accentuate the CRMs affected most by genetics and the dyslipidemic diet. Indeed, running the analysis with just the HP-fed mouse samples resulted in mCRMs that are poorly correlated with the cardio-metabolic end points. Furthermore, the compositions of the HP mCRMs were very different from the HFCA mCRMs, with the mean overlap between modules being ~8. Taken together, these data demonstrate that HFCA leads to a rearrangement of the regulatory networks controlling miRNA expression.

To identify hub members of gene and miRNA CRMs, we calculated the total TOM for each member within the respective module, ranked them, and found those with the top aggregated TOMs and intramodular connectivity measures. A hub’s expression is believed to influence the expression of the other members within their module, so their TOMs and scaled intramodular connectivities are relatively high because of their strong interconnectedness to the other members in the CRM. Hubs are also most similar to the eigenvalues of the modules itself, so correlations to the eigenmiRs and eigengenes should be very high. In regards to the brown mCRM, we identified miR-199a, miR-181b, miR-27a, miR-21_-_1, and miR-24 as hubs because of their high TOMs, scaled kWithin, and their high correlations with the brown mCRM eigenmiR. In this context, these are miRNAs that are purported to have the most regulatory influence in the module. miRNA hubs may modulate the expression of regulatory genes, such as those encoding transcription and/or biogenesis factors, which subsequently effect the expression and/or stability of other members of the module.

Several of the hub miRNAs we identified have been studied in the context of hyperlipidemia and related metabolic diseases. miR-27 is involved in the regulation of lipid synthesis and metabolic pathways (49) and has been implicated in the etiology of viral hepatitis-induced steatosis (45) and atherosclerosis (6). Although miR-21_-_1 (a 5′-shifted isomiR of miR-21) is not well studied, the canonical miR-21 is more highly expressed in the liver tissue of the DO mice and is also a member of the brown mCRM. Notably, hepatic miR-21 is a key partner of miR-27 in the negative regulation of factors mediating cholesterol synthesis (42) and lipid metabolism (21) and was very recently proposed as a driver of metabolic disorders associated with diet-induced obesity as well as fatty liver disease (3). Our findings underscore the importance of these three miRNAs in contributing to diet-induced dyslipidemia. miR-24 also has been connected previously to the development of hyperlipidemia. It is known to be upregulated in the livers of C57BL/6 mice fed a high-fat diet, and to directly target Insig1, which leads to hepatic lipid accumulation and hyperlipidemia (32). In the DO mice, miR-24 is upregulated by ~1.4-fold in the HFCA-fed mice relative to the HP-fed mice. In our gene expression microarray data, Insig1 is downregulated by ~2.4-fold. Hepatic miR-199a is not as well studied in the context of dyslipidemia; however, it has been linked to a lipid-related condition as it was shown to be elevated in livers of humans with nonalcoholic fatty liver disease (26). Hepatic expression of miR-181b in dyslipidemia is even less studied as it is more known for its role in liver fibrosis (53), and hepatocarcinogenesis in mice (50) and rats (14).

We found that the brown mCRM is inversely correlated with four of the gCRMs. Moreover, we observed that for each of the gCRMs ~30–40% of their genes are predicted to harbor conserved target sites for one or more brown mCRM miRNAs. The brown mCRM members may function within a larger regulatory network to cooperatively regulate gene expression pertinent to the control of circulating VLDL/LDL-C levels.

The brown mCRM and four inversely correlated gCRMs are associated with postdiet circulating VLDL/LDL-C levels, AST and ALT levels in the HFCA-fed DO mice, but not strongly associated with any of the other end points. It is noteworthy that the same mCRM was not associated with end points such as atherosclerotic lesion size. One reason for this seemingly discordant result is the fact that long-term feeding of HFCA diet consistently results in early atheroma formation where lesions consist primarily of macrophages and immune cells (33, 36). Furthermore, we found DO mice to be generally resistant to atherosclerosis on the HFCA diet, but all were hyperlipidemic (46). This indicated that factors other than hepatic gene expression and hyperlipidemia, perhaps those operating at the vessel wall itself, may be responsible for inhibition of lesion formation.

In utilizing the DO mice in a systems genetics study, we have identified not only miRNAs that may potentially function cooperatively within a network of other miRNAs and genes, but also hub miRNAs that may contribute substantially to the observed increase in VLDL/LDL-C levels or may contribute to the liver’s adaptive response to HFCA. These miRNAs are prime candidates for future loss- and gain-of-function studies in diverse genetic strains in the context of diet-induced dyslipidemia.

GRANTS

This research was supported in part by National Institutes of Health Grants 5R01HL-128572 (B. J. Bennett), P30DK-056350 (B. J. Bennett and D. Pomp), and R01DK-105965 (P. Sethupathy), a pilot grant from the Nutrition Research Institute (B. J. Bennett and P. Sethupathy), and a National Science Foundation Graduate Research Fellowship Program DGE-1144081 (A. R. Coffey).

DISCLOSURES

No conflicts of interest (financial or otherwise) are declared by the authors.

AUTHOR CONTRIBUTIONS

A.R.C., D.P., B.J.B., and P.S. conceived and designed research; A.R.C., T.L.S., M.K., and B.J.B. analyzed data; A.R.C., B.J.B., and P.S. interpreted results of experiments; A.R.C. prepared figures; A.R.C. drafted manuscript; A.R.C., D.P., B.J.B., and P.S. edited and revised manuscript; A.R.C., B.J.B., and P.S. approved final version of manuscript; T.L.S., J.A., and K.H. performed experiments.