Results

QTLs Associated with Gonadal Fat Mass
Characteristics of the B6.ApoE−/− and C3H.ApoE−/− parents and the F2 BXH.ApoE−/− generation on a Western diet are summarized in Table 1. Gonadal fat mass differed significantly between the sexes in F2 (p < 10−4) and in the parental C3H.ApoE−/− (p < 0.05), but not in B6.ApoE−/− mice. Gonadal fat mass was the fat pad collection that represented the most animals and the most accurate collections and was thus chosen for further analysis. Broad sense heritability (h 2) calculated as (σ2 Total − σ2 Parental)/σ2 Total for the gonadal fat mass trait was 54% for females and 36% for males, which is in close agreement with previous reports [18,19] and demonstrates significant heritability of gonadal fat mass.
Table 1  Characteristics of the BXH.ApoE−/− Cross A total of 334 F2 mice were genotyped at an average 1.5 cM density using 1,032 single nucleotide polymorphisms (SNPs) spanning all 19 autosomes. QTL analysis for several clinical traits (clinical QTLs [cQTLs]), including the unadjusted raw values for gonadal fat mass, was performed using a single marker regression approach (justified by the high-density of markers, making interval mapping unnecessary). In order to test specifically for sex effects of linkage, we included additive, dominant, sex, sex-additive, and sex-dominant parameters in our calculations (see Materials and Methods). A stepwise regression procedure was used to determine whether the addition of the final two terms significantly improved the linear regression model, conditional on realizing a significant additive QTL effect. We performed permutation analyses over all gene expression traits, estimating false discovery rates (FDRs) at different logarithm of odds (LOD) score thresholds and assessing the overall rate of QTL detection. From these analyses we constructed receiver operating characteristic (ROC)-like curves to demonstrate that our straightforward method has significantly increased power to detect QTLs compared to QTL mapping methods that do not incorporate sex and genotype–sex interactions (Figure S1). It is clear from the ROC curves that the sex and sex-interaction terms add significantly to the detection of QTL for the gene expression traits. Using previously described conventions [20], QTL models without the final two interaction terms (sex*add and sex*dom) have a suggestive threshold of 3.0 (p < 1 × 10−3) and a significant threshold of 4.3 (p < 5 × 10−5, genome-wide p < 0.05). QTL models incorporating only the sex*add interaction term in addition to the additive terms have one extra degree of freedom that leads to a corresponding increase in the LOD score thresholds to 3.5 (suggestive) and 4.9 (significant) for the 0.001 and 0.00005 p-value thresholds, respectively. QTL models incorporating both sex*add and sex*dom interaction terms possess two extra degrees of freedom, with a corresponding increase in LOD score thresholds to 4.0 (suggestive) and 5.4 (significant) for the 0.001 and 0.00005 p-value thresholds, respectively.
One suggestive (Chromosome 1) and four significant (Chromosomes 3, 5, 11, and 19) cQTLs for the gonadal fat mass trait were identified (Figure 1A; Table 2). Four out of the five cQTLs showed statistically significant better fits with the full model incorporating the interaction terms sex*add and sex*dom, compared to the model including only the additive terms. Interestingly, the cQTL over Chromosome 11 did not improve, suggesting that the additional terms did not contribute to improved detection of this locus. Table 2 summarizes the position and LOD score of maximal linkage for each cQTL. While the focus of this study was the gonadal fat mass trait, it is noted that a genome scan for the adiposity trait resulted in cQTLs at the same locations, with very similar LOD scores and sex dependence (unpublished data).
Figure 1  Genome Scan for Gonadal Fat Mass
(A) Animals were genotyped at an average 1.5 cM density using 1,032 SNPs polymorphic between the parental strains. LOD scores computed using sex as an additive covariate (black) failed to detect significant linkage. A genome scan accounting for interactions between sex and QTL (red) showed evidence for suggestive linkage on Chromosome 1 and significant linkage on Chromosomes 3, 5, 11, and 19. Dashed and solid lines are thresholds for suggestive (p < 1 × 10−3) and significant linkage (p < 5 × 10−5), respectively.
(B) Genome scans for gonadal fat mass using different models over mouse Chromosome 5. Scans for fat mass using all animals with (black) and without (green) sex as an additive covariate failed to detect significant linkage. Females analyzed alone (magenta) showed evidence for suggestive linkage (p < 2 × 10−4). When both sexes were analyzed to account for sex effects (red), a significant QTL was realized (p < 10−6).
For clarity, only the model incorporating both the “sex*add” and “sex*dom” terms is shown in red, although additional models incorporating the terms separately were also computed.
Table 2  cQTLs for Gonadal Fat Mass Results from the various regression models used to determine linkage for the Chromosome 5 cQTL are depicted in Figure 1B. Analysis of all animals with and without sex as a covariate failed to demonstrate evidence of linkage on Chromosome 5. When females were analyzed alone, a suggestive LOD score of 3.7 was realized (p = 2 × 10−4); males analyzed alone did not demonstrate evidence for linkage. However, using all 334 animals and adding the interaction terms to the QTL model significantly improved sensitivity, and a cQTL with a maximum LOD of 7.56 (p = 1.7 × 10−6) was realized.
Given the improved detection of four of the five cQTLs when sex-additive and sex-dominant interaction terms were considered, we hypothesized that the main genotype effect of these cQTLs on the gonadal fat mass trait would differ between the sexes (Figure 2). Indeed, cQTLs located on Chromosomes 1, 3, and 5 showed opposing effects on fat mass, or sex antagonism. The effect of the cQTL on Chromosome 11 was in the same direction in both males and females, but was sex-biased toward a larger effect in females (R 2 = 0.091 in females versus R 2 = 0.046 in males), confirming the minimal sex specificity of this cQTL. The cQTL on Chromosome 19 showed a sex-specific effect in females, with no effect in males.
Figure 2  Effect of Genotype on Fat Mass
Homozygous B6 (BB), C3H (CC), or heterozygous (BC) genotype at all five QTL positions, separated by sex, are shown. The underlying genotypic effects of the QTLs on fat mass differ between the sexes. Coefficients of determination (R 2) are shown along with associated ANOVAs *p < 0.05, **p < 0.01, ***p < 0.001. Overall, all five cQTLs for gonadal fat mass were biased toward a larger effect in females. Assuming purely additive effects of each genotype, these cQTLs account for approximately 42% of the variation in female F2 mice and 13% in males, consistent with the narrow sense heritability estimates for this trait and again demonstrating significant differences in the regulation and heritability of the gonadal fat trait between the sexes.

Liver eQTLs
Livers from 312 F2 animals (156 female, 156 male) were profiled using oligonucleotide microarrays manufactured by Agilent Technologies (Palo Alto, California, United States), which included probes for 23,574 mouse transcripts. Individual transcript intensities were corrected for experimental variation and normalized and are reported as the mean log10 ratio (mlratio) of an individual experiment relative to a pool composed of 150 mice randomly selected from the F2 population [21]. Each measurement was fitted to an error model and assigned a significance measurement (type I error). A heat map of the 2,320 transcripts most differentially expressed (p < 0.05 in 10% or more of animals) relative to the pool is depicted in Figure 3. This selection of genes was not biased on a priori known differential expression between the sexes, linkage, or correlation with a clinical phenotype. This is noteworthy because hierarchical clustering of these transcripts against the 312 F2 mice shows an almost perfect clustering into male and female subgroups, emphasizing striking effects of sex on liver gene expression levels and suggesting that sex is controlling more variance in these transcripts' expression than any other parameter.
Figure 3  Heat Map of Liver Gene Expression
Over 2,300 of the most differentially expressed genes in liver hierarchically clustered by animals (x-axis) against transcript levels (y-axis). Expression is reported as mlratio of individual experiment against a common pool. Red is over- and green underrepresented relative to pool. The expression values of the 23,574 transcripts were treated as quantitative traits and fitted to the same linear regression models used to compute LOD scores for clinical traits (eQTLs). The FDR at each threshold was determined by permuting the data 100 times and taking the mean number of QTLs detected over all of the permuted datasets at a given threshold, and dividing this count by the number of QTLs detected at the same threshold in the observed data. At the threshold for significant linkage (p < 5 × 10−5, genome-wide p < 0.05, based on a single trait), the FDR was estimated at 3.4% for the standard QTL model not accounting for any sex terms, 3.1% for the QTL model accounting for additive sex effects, and 3.2% for the QTL model accounting for additive sex effects and allowing for the sex interaction terms to enter the model. A list of all detected suggestive (p < 1 × 10−3) and significant eQTLs (p < 5 × 10−5) detected in the BXH.ApoE−/− intercross is provided in Table S1.
Characteristics of the eQTLs at different significance levels are summarized in Table 3 and shown graphically in Figure 4. We detected 6,676 eQTLs representing 4,998 transcripts at the 5 × 10−5 significant level. Of these, 2,118 eQTLs were located within 20 Mb (roughly 10 cM) of the corresponding gene, likely representing eQTLs regulated by cis-acting variation within the gene itself. Of the 6,676 significant eQTLs, 1,166 (17%) demonstrated a sex bias and were subsequently significantly improved with the addition of the sex-additive and sex-dominant terms.
Table 3  Characteristics of Liver eQTLs
Figure 4  Properties of All Liver eQTLs
(A) Distribution of all significant liver eQTLs across the genome in 2-cM bins. A total of 6,676 significant eQTLs were realized, representing 4,998 liver transcripts. Hotspots of nonrandom eQTL colocalization are clearly evident.
(B) Distribution of eQTLs with significant sex-specific effects. A total of 1,166 eQTLs representing 1,044 transcripts show an eQTL hotspot on Chromosome 5.
(C) Properties of eQTLs at increasing significance levels. As the threshold for significant linkage increases (p-value decreases, or LOD score increases), the proportion of cis-eQTLs (black) increases. The fraction of all eQTLs with sex effects (red) and cis-eQTLs with sex effects (blue) remains relatively constant at increasing thresholds. The dashed line indicates the genome-wide significance threshold (p < 5 × 10−5; genome-wide p < 0.05).
(D) Properties of sex-specific eQTLs at increasing significance levels. For eQTLs with significant sex effects, as with all eQTLs, the proportion of cis-eQTLs (black) increases and trans (blue) decreases as the threshold for significance increases. At the genome-wide threshold for significance (dashed line), over 70% of eQTLs with significant sex effects are trans. The distribution of all 6,676 significant eQTLs (p < 5 × 10−5) across the genome in 2 cM bins is shown in Figure 4A. Evidence for eQTL hotspots is clear on Chromosomes 1, 2, 4, 5, 6, and 7 where significant fractions of the 6,676 eQTLs colocalize within 2 cM regions on each chromosome. Approximately 67% of eQTLs at this threshold are trans, and these eQTL hotspots consist primarily of trans-acting effects on transcriptional variation. Distribution of the 1,166 eQTLs with significant sex effects, of which 852 (73%) are trans, showed enrichment on Chromosome 5 at approximately 49 cM (Figure 4B) as assessed using the Fisher exact test (p = 8.7 × 10−25 after Bonferroni correction). At this locus there were 250 eQTLs, 140 of which exhibited genotype–sex interactions.
At increasing thresholds for linkage, a higher fraction of detected eQTLs were cis-acting (Figure 4C). The increased proportion of cis-eQTLs with increasing LOD score thresholds has been reported before [5,15] and confirms what is likely to be our increased power to detect first-order cis-acting variations affecting transcription. The proportion of eQTLs with significant sex effects remained relatively constant at all thresholds (Figure 4C). Furthermore, the majority of these sex-specific eQTLs (73%) are acting in trans on a given gene's expression (Figure 4D), and similar proportions of sex-specific eQTLs (26%) are cis compared to the proportion of all liver cis-eQTLs (32%). These data demonstrate the profound effects of sex on the genetic regulation of gene expression.

Cis-eQTLs Are Candidate Genes for the Gonadal Fat Mass Trait
Given the marked effects of sex on liver gene expression and the genetic regulation of gonadal fat mass, we reasoned that cis-eQTLs with significant “sex-additive” and “sex-dominant” interactions would be potential candidate genes for our trait. Of the 2,118 significant cis-eQTLs, 304 (14%) were improved by the sex-interaction terms. Cis-eQTLs overlapping the confidence interval for the fat mass cQTL are candidate genes for the trait. Those genes with significant sex interactions receive increased consideration as potential candidates (Tables 4 and 5).
Table 4  Properties of Genes Coincident with the Gonadal Fat Mass cQTL
Table 5  Transcripts Coincident with cQTLs with Significant Sex Effects and Correlated with Gonadal Fat Mass Are Strong Candidate Genes for the Trait

Thousands of Genes Show a Sex-Specific Correlation with Gonadal Fat Mass
For each of the 23,574 oligonucleotides represented on the array, we computed a linear regression analysis to test for association between the trait “gonadal fat mass” and each transcript abundance measure, incorporating the terms “gene,” “sex,” and “gene-by-sex,” where the “gene-by-sex” parameter tests for sex-specific correlation between a gene and the trait. As before, a stepwise regression procedure was used to determine if the addition of the interaction term significantly improved the model fit (see Materials and Methods). Multiple testing was addressed by controlling for the FDR. Distribution of the p-values obtained from these 23,574 correlations is shown in Figure 5A. At FDR = 0.01, 4,613 genes were significantly correlated with gonadal fat mass. Of these genes, 4,524 (98%) showed significant “gene-by-sex” effects, supporting the high degree of sex specificity in the genetic regulation of this trait. A complete list of all genes correlated with gonadal fat mass is provided in Table S2.
Figure 5  Properties of Transcripts Significantly Correlated with Gonadal Fat Mass
(A) Distribution of p-values for trait–gene correlations between transcripts and gonadal fat mass. At FDR = 0.01, 4,613 transcripts are significantly correlated with the trait.
(B) Number of eQTLs generated by the 4,613 genes significantly correlated with gonadal fat mass. Of these, 1,130 genes possessed at least one significant eQTL.
(C) Distribution of 1,478 eQTLs significantly correlated with gonadal fat mass across the genome in 2-cM bins.
(D) Identification of genomic regions enriched for eQTLs correlated with gonadal fat mass. The x-axis represents genome position in 2-cM bins, and the y-axis represents the −log10 Fisher exact test p-value for enrichment of eQTLs in overlapping 6-cM bins. The dashed line corresponds to p = 0.05 after correction for multiple comparisons. One significantly enriched region on Chromosome 19 is shown. The Chromosome 19 (40-cM) hotspot is coincident with a cQTL for gonadal fat mass and is highlighted in red. Of the 4,613 genes correlated with gonadal fat mass, 1,130 generate 1,478 significant eQTLs (Figure 5B). The colocalization of eQTLs for these correlated genes with the cQTL for the fat mass trait provides useful implications for the possible role of these genes. Whether the eQTLs are cis or trans determines what that role may be. Genes that show significant correlation with the gonadal fat mass trait and that have cis-eQTLs coincident with the fat mass cQTLs are potential candidate genes for the trait (i.e., they may contain a genetic variation in that gene that is the cause of the trait cQTL). Table 5 summarizes the genes that possess these properties for each cQTL, increasing evidence for these genes as potential candidates. As addressed below, given the complex multiorgan regulation of adipose tissue mass, it is unlikely that the genetic regulation of all five loci resides in the liver. However, some may involve the liver, and even for those that do not, the liver transcriptional variation may reflect that of the relevant tissue.
Genes that show significant correlation with gonadal fat mass and have trans-eQTLs coincident with the fat mass cQTL cannot be candidates directly responsible for the trait, as they are physically located elsewhere in the genome. However, they are potentially involved in the pathway(s) leading from the causative gene to the expression of the fat mass trait (i.e., their transcription is closely regulated by the causative gene at the locus). All of the five fat mass cQTLs have colocalizing trans-eQTLs for correlated genes. However, for a trait such as fat mass that is regulated by multiple tissues and organs, it is unlikely that all five fat mass cQTLs are primarily driven by liver gene expression. As an approach to this problem, we hypothesized that those cQTLs that are most closely associated with liver gene expression would show an overrepresentation of colocalized eQTL for correlated genes, while those loci primarily controlled by other tissues would not have shown this pattern.
To assess this, we first determined the distribution of these 1,478 eQTLs across the genome in 2-cM bins as shown in Figure 5C. In order to see if there exist any hotspots for these eQTLs, we tested eQTLs with p < 0.001 for enrichment along the genome in overlapping 2-cM bins against the distribution of all liver eQTLs (Figure 4A) using a Fisher exact test. Figure 5D shows the significance of enrichment reported as −log10 of the enrichment p-value across the genome. One locus on Chromosome 19 was significantly enriched for eQTLs of transcripts correlated with the gonadal fat mass trait. As anticipated, there was an overlap of a correlated eQTL hotspot and a fat mass cQTL, specifically Chromosome 19 at 40 cM. This suggests that the genetic regulation of fat mass for the Chromosome 19 locus is more closely tied to liver gene expression than are the other four fat mass cQTLs.
The effect of the trans-eQTLs at the Chromosome 19 locus on gene expression is summarized in Figure 6. Twenty-nine trans-eQTLs colocalize to Chromosome 9 at 40 cM, suggesting that 29 genes correlated with gonadal fat mass are regulated in trans by a polymorphism at this position. The proportion of gene expression levels controlled by this locus (approximated as the coefficients of determination R 2) differs between males and females for the majority of the transcripts (as in Figure 6A), and for Chromosome 19 (Figure 6B), females demonstrated greater genetic regulation of gene expression than males. This substantial female bias is significantly higher than would be expected to arise by chance for the Chromosome 19 locus (p < 0.001 by χ2). This locus corresponds to one of the four sex-biased cQTLs for gonadal fat mass reported in this study, and the significant sex specificity of both the cis and trans genetic regulation of liver genes correlated with fat mass supports the functional significance of this locus in this organ.
Figure 6  The Effects of Sex on Trans-eQTL Correlated with Gonadal Fat Mass
(A) Example of the effect of genotype at a trans locus on gene expression. Presence of homozygous B6 (BB), C3H (CC), or heterozygous (BC) genotype at a trans locus affects transcript MMT00016118 levels (reported as mlratio) in a sex-specific manner, with effects detectable only in females. Coefficients of determination (R 2, or proportion variance explained) are shown along with associated ANOVA p-values. Several trans-eQTLs correlated with gonadal fat mass localize to regions overlapping with cQTLs for this trait, specifically, to Chromosome 19, 40 cM.
(B) For Chromosome 19, the vast majority of these correlated trans-eQTLs are biased toward larger effects on gene expression in females (red lines). The effect of any given trans-eQTL is approximated as R 2 determined in a manner similar to that depicted in (A).