Ethics statement

All animal work was approved by the Institutional Animal Care and Use Committee (IACUC) at the University of Tennessee Health Science Center (IACUC protocol 12–148.0-B; Legacy protocol 680).

Generation of BXD recombinant inbred strains

BXD strains were generated from crosses between B6 and D2 strains resulting in a panel of recombinant inbred strains with either the parental B or D allele fixed at each locus [31]. There are now roughly 160 BXD lines—about 80 strains available from The Jackson Laboratory and ~80 new lines from the University of Tennessee Health Science Center [32]. Animals were maintained on an equal light dark cycle with ad libitum access to food and water.

GeneNetwork data sets

Detailed information for data sets used in the analysis, including tissue acquisition, RNA extraction, array hybridization methods, data processing, and normalization methods, can be found in the “Info” page at www.genenetwork.org. All mouse expression data sets used in this study were generated through collaborative efforts [33,34] and can be publically accessed from www.genenetwork.org. Representative expression data sets include: VU BXD Midbrain Agilent SurePrint G3 Mouse GE (May12) Quantile [35], Hippocampus Consortium M430v2 (Jun06) [34], VCU BXD PFC Sal M430 2.0 (Dec06) RMA [36], INIA Amygdala Cohort Affy MoGene 1.0 ST (Mar11) RMA, GSE16780 UCLA Hybrid MDP Liver Affy HT M430A (Sep11) [37], and UT-VGX Hepatocytes Affy Mouse Gene 1.0 ST Gene Level (Oct14) RMA. Human cortical expression data sets containing AD and control cases were also used, including GSE15222 Human Brain All Cases Myers (Apr09) RankInv Database [38].

Gene expression measurements

Expression of cytosolic and membrane-bound GSTs was measured across the BXD family using Affymetrix and Illumina microarray platforms in brain and liver. Representative probe sets (S1 Table) were selected for each gene based on probe set specificity and mean expression. SNPs and other sequence variants, such as indels (insertions and deletions) or CNVs (copy number variants), overlapping probes have an important impact on the false discovery rate of expression differences and on identification of local expression quantitative trait loci (cis-eQTLs) [39–41]. No representative probe set target sequences overlap SNPs or other sequence variants that could interfere with GST measurements in this study.

Heritability calculations

Broad sense heritability [42] was estimated by comparing the genetic variation between strains to the environmental variance within strains. Biological replicates for each strain are required for broad sense heritability calculations and were only performed for expression data sets for which replicates were available. For recombinant inbred lines, which are homozygous at all loci, the following formula was used: 0.5V_A / (0.5V_A + V_E) or V_A / (V_A +2V_E) where V_A = additive genetic variance (variances of strain means) and V_E = environmental variance (V_E = V_T—V_A, V_T = total variance among subjects) [43].

QTL mapping

Genome-wide interval mapping of gene expression patterns across the BXDs using different data sets was performed in GeneNetwork (www.genenetwork.org), as previously described [44]. Genome-wide empirical P-values are estimated by permuting trait data for each transcript randomly 5000 or more times, to determine the genome-wide significance threshold (P value < 0.05). We estimated QTL intervals for downstream analysis based on a 1.5 LOD confidence interval.

Analysis of allele-specific expression by RNA-seq

We downloaded paired-end RNA-seq data from the European Nucleotide Archive (accession number ERP000591) for lung, liver, and hippocampus of B6xD2 F1 female hybrids generated by crossing B6 females with D2 males [45]. The data consist of transcriptome sequence from six biological replicates. We aligned RNA-seq reads to both the B6 reference genome (mm10 assembly) and the SNP-substituted D2 genome using “Splice Transcripts Alignment to a Reference” tool (STAR, version 2.3.1a) [46] with the following parameters: “—outFilterMultimapNmax 10—outFilterMismatchNmax 12”. We used SAMtools (version 0.1.19) “pileup” function [47] and an in-house Python script (https://github.com/ashutoshkpandey/ASE_prealignment/blob/master/Allele_specific_SAM.py) to assign reads overlapping SNPs to either the B6 or D2 parental allele using both parental genomes. We calculated allelic ratios for each SNP as the ratio of number of reads assigned to the reference allele (B6) to the total number of aligned reads (B6+D2). For each SNP we used an interquartile range (IQR) method to identify outlier allelic ratios from the set of six F1 replicates. Outlier ratios were located outside the [Q1–1.5(IQR) and Q3 + 1.5(IQR)] range where Q1 and Q3 represent first and third quartiles and IQR is calculated as Q3 –Q1. Reads from replicates showing concordant allelic ratios were merged and allelic ratios were recalculated. We used the chi-square goodness of fit test to determine allelic imbalances for a given SNP. For a SNP showing an allelic imbalance, the ratio will deviate from 0.5. We defined genes as having an allele-specific expression difference if they contained one or more SNPs with an allelic imbalance at an FDR threshold of less than 0.1 [48]. We also required the expression fold difference to be >1.25.

Global protein quantification of Gsto1

Data from global protein quantification of mouse hippocampus using tandem mass tag technology was mined for Gsto1 peptides. Hippocampal tissue was collected from three B6 and three D2 mice of two different ages (2 to 3 months or 12 months of age). Isobaric labeling was used to differentiate sample peptides and 10 fractions were collected for each sample and analyzed by low pH reverse phase LC-MS/MS. A total of 7,074 proteins were identified in this study, of which 7,014 were quantified. 28 spectral counts and 8 peptides were identified for Gsto1. A t-test (p < 0.05) was used to determine significance.

Partial correlation analysis

Pearson product-moment correlations were computed in GeneNetwork between Gsto1 (probe set 1416531_at) and all other probe sets in the Hippocampus Consortium M430v2.0 data set after controlling for the genetic variation near Gsto1 (locus rs13483647; Chr 19 at 46.66 Mb) and Ina (locus rs13483649; Chr 19 at 47.22 Mb). The residual values after partial correlation for Gsto1 are stored in the BXD Published Phenotypes database as trait ID 17328. One hundred forty highly correlated and relevant genes were selected based on the following criteria: partial correlation P < 0.001 (r ≥ |0.38|) and a literature correlation of r > 0.5. Literature correlations are available on GeneNetwork and were originally computed using the Semantic Gene Organizer (SGO) software to automatically extract gene-gene relations from titles and abstracts in MEDLINE citations [49]. The Chilibot system (http://chilibot.net), Alzheimer Disease & Frontotemporal Dementia Mutation Database (http://www.molgen.ua.ac.be/ADMutations/), and the ALZGENE database (www.alzgene.org) were used to scan candidate genes for relations with AD. Results were also checked manually.

Identification of downstream targets and phenotypes

The following analytical steps were taken to identify transcripts that are putative downstream targets. First, transcripts with expression level greater than 8 units modulated by a suggestive trans-eQTL (LOD > 2) located in a 4 Mb region overlapping the location of the target GST gene in at least one tissue were selected. Second, the correlation coefficient between the target GST and the expression of each trans-regulated gene was computed (p < 0.01 is considered a significant correlation).

To identify phenotypes that are likely to be downstream of the variation in GST genes we queried a phenotype database of nearly 5000 published phenotypes to identify those significantly correlated with the expression of each GST gene. We also identified phenotypes whose expression was significantly regulated by genomic variants situated within 2 Mb of each target GST gene locus.

Variation in expression of several GST family members is under genetic control

Genome-wide expression profiles are available for ~20 peripheral tissues or cell types and ~10 individual brain regions for the BXD population (www.genenetwork.org). We surveyed all available data but chose to focus on several representative data sets from brain (hippocampus, prefrontal cortex, amygdala, and midbrain) and peripheral tissue (liver and liver hepatocytes) for detailed systems genetics analysis. Representative data sets are robust for expression QTL mapping (eQTL), have been extensively error-checked, and include a large number of BXD strains.

GST expression levels vary tremendously across B6, D2, and the BXD population both within and across tissues (Table 1, S1 Table). Variation tends to be greater in liver compared to individual brain regions. Some GSTs, such as Gsta2 and Gsta3, are predominately expressed in liver, whereas others have higher expression in brain, such as Gstm5 and Mgst3. The expression of many GSTs is moderately to highly heritable (H² > 0.3) suggesting that some of the variation in expression of individual GSTs is controlled by genetic factors. In agreement with this observation, several GSTs demonstrate genetic modulation of expression by local (cis eQTLs) or distant (trans eQTLs) loci (S1 Table). In general, distant eQTLs are not well conserved across tissues but cis eQTLs are often highly reproducible. We surveyed the genetic regulation of GST family members from all seven classes and found consistent cis modulation of expression across multiple tissues for Gsta4, Gstt2, Gstz1, Gsto1, and Mgst3 (Table 2). No significant and consistent cis or trans modulation of expression was detected for mu, pi, or kappa class members or for the sigma class, which is represented by the highly specialized Hpgds gene that converts prostaglandin H2 to prostaglandin D2. With the exception of Gsto1, all cis-modulated genes (Gsta4, Gstt2, Gstz1, and Mgst3) have higher expression in the BXD population that is associated with inheritance of the D allele from the D2 parental strain. For all cis-modulated GST family members we leveraged multiple informatics resources to identify the genetic architecture underlying expression variation, explore potential functional consequences of that variation, and evaluate biological function through coexpression analysis.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Summary of GST gene expression.

https://doi.org/10.1371/journal.pone.0148230.t001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Summary of cis modulation of expression.

https://doi.org/10.1371/journal.pone.0148230.t002

Gsta4

Expression is variable across the BXD set (Fig 1) ranging from a modest level in the midbrain (fold change of 1.23) to high levels in prefrontal cortex, hippocampus, and liver (fold changes of 1.99, 1.80, and 11.43, respectively). This variation is highly heritable (H² = 0.59) and is strongly modulated by a significant cis eQTL in 10 tissues (Table 2), including midbrain (LOD = 10.5), hippocampus (LOD = 14.2), prefrontal cortex (LOD = 7.5), liver (LOD = 4), and liver hepatocytes (LOD = 4.2) (Fig 1). The 1.5 LOD confidence interval, which defines the boundaries of the QTL, is ~3 Mb (77 to 80 Mb on Chr 9), precisely overlapping the location of the gene (Chr 9 at 78 Mb). Higher expression of Gsta4 in the BXD family is driven by inheritance of the D allele (Fig 1). Expression of Gsta4 is also cis-modulated in an F2 intercross between C57BL/6J and C3H/HeJ (BHF2, Table 2) and higher expression is driven by inheritance of the C3H/HeJ allele. This result is consistent with a mutation within or near Gsta4 that occurred in the B6 parental strain.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Summary of expression and QTL mapping across BXD strains for Gsta4.

The top five panels contain bar plots representing the expression of Gsta4 for each BXD individual in five different tissues. Average log₂ expression is shown on the y-axis and unique strains are shown on the X-axis (BXD1 = 1). Red and blue indicate inheritance of the paternal D2 (D) or maternal B6 (B) allele of Gsta4 in each strain, respectively. If only a single individual was used for expression measurements, error bars are not shown. For genetic reference populations, mapping power is derived from the number of individuals as opposed to the number of biological replicates. Higher expression is associated with inheritance of the D allele. Bottom two panels show the genetic mapping results in each tissue. Association strength (LOD) is shown on the Y-axis and plotted across the genome on the X-axis (by chromosome) for hippocampus (blue), midbrain (red), prefrontal cortex (green), liver (dark purple), and hepatocytes (purple). Genome-wide significance is determined by permutation (n = 5000) with the threshold for significance indicated as black (significant, p <0.05) and grey (suggestive, p < 0.3) horizontal lines. Expression of Gsta4 is modulated by variants within or near its own locus on Chr 9, a cis eQTL.

https://doi.org/10.1371/journal.pone.0148230.g001

Gsta4 is situated within a relatively variant-rich region on Chr 9; 96 SNPs and 22 small insertion/deletions (indels) are segregating between the B6 and D2 parental strains. None of these variants are in coding exons, known regulatory regions, or splice sites. Instead, one or more of these variants alter expression through undefined regulatory regions or alternative mechanisms. Although the causal variant has not yet been identified, the cis eQTL regulating Gsta4 expression in multiple tissues across the BXD panel has been validated in liver, lung, and hippocampus by allele-specific expression (ASE) assays using RNA sequencing in B6xD2 F1 hybrid mice (Table 3).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Summary of GST ASE.

https://doi.org/10.1371/journal.pone.0148230.t003

Gstt2

Among the BXD population expression is variable across brain regions, with average expression levels in the midbrain, and low levels of expression in the hippocampus (Fig 2). In midbrain, Gstt2 expression variation among BXD strains is modest at nearly 1.3-fold. In hippocampus, expression is more variable at over 1.6-fold. The expression of Gstt2 is significantly modulated by a cis eQTL in seven tissues (Table 2) from the BXD population, including the midbrain (LOD = 5.0; 1.5 LOD confidence interval on Chr 10 from 74 to 76 Mb) and hippocampus (LOD = 4.1; 1.5 LOD confidence interval on Chr 10 from 69 to 79 Mb) (Fig 2). Inheritance of the D allele confers higher expression in all regions. Although expression across the BXD set is higher in the liver, there is no significant cis modulation of expression detected. Gstt2 is located in a variant poor region and, in both midbrain and hippocampus, it is the only gene within each respective 1.5 LOD confidence interval with demonstrable cis modulation of expression. Two intronic variants in Gstt2 are segregating among the BXD population but their effect on expression is unknown. Due to the location of these variants in introns, validation by ASE analysis was not possible.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Summary of expression and QTL mapping across BXD strains for Gstt2.

Top two panels (bar plots) show expression of Gstt2 in each BXD strain in hippocampus and midbrain. Average log₂ expression is shown on the y-axis and unique strains are shown on the X-axis. Red and blue indicate inheritance of the paternal D or maternal B allele of Gsta4 in each strain, respectively. Black indicates a heterozygous (likely erroneous) genotype call. Higher expression is associated with inheritance of the D allele. Bottom panel shows the genetic mapping results in each tissue. Association strength (LOD) is shown on the Y-axis and plotted across the genome on the X-axis (by chromosome) for hippocampus (blue) and midbrain (red). Genome-wide significance is determined by permutation (n = 5000) with the threshold for significance indicated as black (significant, p <0.05) and grey (suggestive, p < 0.3) horizontal lines. Expression of Gstt2 is modulated by variants within or near its own locus on Chr 10, a cis eQTL.

https://doi.org/10.1371/journal.pone.0148230.g002

Gstz1

Expression is high with a modest level of variation (1.3-fold) in midbrain while expression in the hippocampus is average and more variable with a fold change of over 1.5 (Fig 3). Cis-modulation of expression is observed in four brain tissues in the BXD set (Table 2), including midbrain (LOD = 17.4; 1.5 LOD confidence interval from 88 to 90 Mb on Chr 12) and hippocampus (LOD = 2.9) (Fig 3). In addition, a significant cis eQTL is also detected in crosses between A/J (A) and B6 (B), B6 and C3H/HeJ (H), and CAST/EiJ (CT) and B6 (Table 2). Higher expression is associated with inheritance of the D allele in the BXD set. In contrast, the parental allele associated with higher expression varies in other crosses, such that expression is higher for inheritance of A or CT alleles compared to the B allele and higher for inheritance of the B allele compared to the H allele. This pattern might be expected if several functional variants are segregating among inbred mouse strains.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Summary of expression and QTL mapping across BXD strains for Gstz1.

Top two panels (bar plots) show expression of Gstz1 in each BXD strain in hippocampus and midbrain. Average log₂ expression is shown on the y-axis and unique strains are shown on the X-axis. Red and blue indicate inheritance of the paternal D or maternal B allele of Gstz1 in each strain, respectively. Higher expression is associated with inheritance of the D allele. Bottom panel shows the genetic mapping results in each tissue. Association strength (LOD) is shown on the Y-axis and plotted across the genome on the X-axis (by chromosome) for hippocampus (blue) and midbrain (red). Genome-wide significance is determined by permutation (n = 5000) with the threshold for significance indicated as black (significant, p <0.05) and grey (suggestive, p < 0.3) horizontal lines. Expression of Gstz1 is modulated by variants within or near its own locus on Chr 12, a cis eQTL.

https://doi.org/10.1371/journal.pone.0148230.g003

Gstz1 is polymorphic between the B6 and D2 parental strains with four SNPs in the 3’ UTR and one in an intron, however no allelic imbalance was detected for this gene in lung and liver using ASE. There was a trend towards allelic imbalance in the hippocampus with higher expression driven by the D allele but this was not significant (p = 0.15) (Table 3). Gstz1 is located within a region that is relatively gene and variant sparse. In midbrain and hippocampus the only other cis-modulated gene is Tmed8, located on Chr 12 at 88.51 Mb, however, the function of Tmed8 is not well defined.

Gsto1

In most tissues, Gsto1 is well expressed and variable across the BXD population (Fig 4). Nearly 60% of the variation between individual BXD strains can be explained by a genetic component (hippocampus, 1416531_at, h² = 0.56) and inheritance of the B allele is always associated with higher expression (Fig 5). A cis-eQTL for Gsto1 was consistently detected in 20 tissues from the BXD set (Table 2) The 1.5 LOD confidence interval for the cis-eQTL is between 47.1 and 48.0 Mb at markers rs3705264 and rs8257607 (hippocampus), precisely aligned with the position of Gsto1. Molecular validation of the cis-eQTL was performed by ASE assay in lung, liver, and hippocampus (Table 3) with the B allele having a significantly increased number of reads relative to the D allele in all tissues. Variation at the gene and mRNA level is also associated with a modest but significant 1.5–fold increase (p = 0.0013) in Gsto1 protein levels in the hippocampus of B6 relative to D2 as measured by global protein quantification using tandem mass tag technology. Cis-modulation of Gsto1 expression is also detected in crosses between A/J and C57BL/6J, Inbred Long-Sleep (ILS) and Short-Sleep (ISS), and a mouse diversity panel.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Summary of Gsto1 expression across BXD strains.

Panels (bar plots) show expression of Gsto1 in each BXD strain in multiple tissues. Average log₂ expression is shown on the y-axis and unique strains are shown on the X-axis. Red and blue indicate inheritance of the paternal D or maternal B allele of Gsto1 in each strain, respectively. If only a single individual was used for expression measurements, error bars are not shown. For genetic reference populations, mapping power is derived from the number of individuals as opposed to the number of biological replicates. Higher expression is associated with inheritance of the B allele.

https://doi.org/10.1371/journal.pone.0148230.g004

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Summary of QTL mapping of Gsto1 expression across BXD strains.

Genetic mapping results are shown for brain and peripheral tissue. Association strength (LOD) is shown on the Y-axis and plotted across the genome on the X-axis (by chromosome) in each QTL map. Genome-wide significance is determined by permutation (n = 5000) with the threshold for significance indicated as black (significant, p <0.05) and grey (suggestive, p < 0.3) horizontal lines. Expression of Gsto1 is modulated by variants within or near its own locus on Chr 19, a cis eQTL.

https://doi.org/10.1371/journal.pone.0148230.g005

Many variants within the Gsto1 gene are segregating among the BXD population that could account for the variation at this locus, including a 5’ UTR SNP, three SNPs located in the 3’ UTR, three synonymous variants in exons, and 47 intronic SNPs, one of which is located in a splice region. Unlike previously discussed GSTs, Gsto1 resides within a gene- and variant-rich region that contains 10 other well expressed genes modulated by significant cis-eQTLs (LOD > 3.3 or LRS > 15) located within 4 Mb of the Gsto1 locus (Fig 6). An important consideration in the analysis of Gsto1 biological networks and downstream traits is the effect of linkage disequilibrium at this locus. For example, internexin neuronal intermediate filament protein alpha (Ina) is located within 1 Mb of Gsto1 at 47.01 Mb and is associated with a strong cis-eQTL (1448992_at; LOD = 20). Ina is a major structural component of the cytoskeleton that is expressed primarily in neurons and involved in axonal architecture. Overexpression of Ina leads to neuronal death and degeneration in cortex, thalamus, and cerebellum [67] and is a signature of neuronal interfilament inclusion disease in humans. INA immunoreactivity is also observed in other neurodegenerative disorders, such as AD [68]. Phenotypes and genes that map into this locus, correlations between Gsto1 and Ina, and correlations between other genes and Gsto1 could result from shared biological function or result from the proximity of other cis-modulated genes that are in linkage disequilibrium at this locus.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Summary of recombination and linkage disequilibrium near the Gsto1 locus.

(A) The association score (LOD) (y-axis) for Gsto1 expression in the hippocampus is plotted as a solid blue line across the genome (x-axis) with horizontal lines indicating significance thresholds for significant (red) and suggestive (grey) LOD values. The physical position of Gsto1 is indicated by the orange triangle and the locations of other cis-modulated genes are shown as light purple triangles. Haplotypes are shown above the linkage map for BXD strains with strain number shown to the right and left of the haplotypes. Vertical black lines designate marker position with marker names below the haplotype map. Red and green blocks indicate a chromosomal region inherited from the maternal B6 or paternal D2 strain, respectively. Blue areas are heterozygous and grey areas are undefined meaning that more markers would be needed to pinpoint the exact recombination breakpoint. As expected for a recombinant inbred population, this region is primarily inherited as an entire haplotype block from either parental strain. (B) Pairwise correlations between markers in the region are shown. Intensity reflects correlation strength. Markers in this region are tightly linked.

https://doi.org/10.1371/journal.pone.0148230.g006

Mgst3

Expression is more robust in brain compared to liver (Fig 7). Variation across the BXD population is moderate in midbrain and amygdala (~1.5-fold), greater than 2-fold in hippocampus and prefrontal cortex, and greater than 7-fold in liver and hepatocytes (Fig 7). Mgst3 is modulated by a strong and significant cis eQTL in 18 different tissues from the BXD set (Table 2, Fig 8). Cis modulation of expression is also detected in F2 crosses between B6 and D2 and CAST/EiJ and B6 (CTB6F2). In most data sets the 1.5 LOD confidence interval is 2 Mb and inheritance of the D allele confers higher expression. Inheritance of the B allele confers higher expression in the CTB6F2 cross.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 7. Summary of Mgst3 expression across BXD strains.

Panels (bar plots) show expression of Mgst3 in each BXD strain in multiple tissues. Average log₂ expression is shown on the y-axis and unique strains are shown on the X-axis. Red and blue indicate inheritance of the paternal D or maternal B allele in each strain, respectively. If only a single individual was used for expression measurements, error bars are not shown. For genetic reference populations, mapping power is derived from the number of individuals as opposed to the number of biological replicates. Higher expression is associated with inheritance of the D allele.

https://doi.org/10.1371/journal.pone.0148230.g007

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 8. Summary of QTL mapping of Mgst3 expression across BXD strains.

Genetic mapping results are shown for brain and peripheral tissue. Association strength (LOD) is shown on the Y-axis and plotted across the genome on the X-axis (by chromosome) in each QTL map. Genome-wide significance is determined by permutation (n = 5000) with the threshold for significance indicated as black (significant, p <0.05) and grey (suggestive, p < 0.3) horizontal lines. Expression of Mgst3 is modulated by variants within or near its own locus (distal Chr 1), a cis eQTL.

https://doi.org/10.1371/journal.pone.0148230.g008

Mgst3 is highly polymorphic between B6 and D2 with 239 intronic and 56 UTR SNPs or small indels. Similar to the other cis-modulated GSTs, one or more non-coding variants underlie the expression variation at this locus as the cis eQTL for Mgst3 has been validated by ASE in both liver and hippocampus (Table 3). Mgst3 resides in a variant rich but relatively gene sparse region, although several neighboring genes (within 2 Mb of Mgst3) are also significantly cis-modulated (LOD > 2) in multiple data sets. This includes two genes involved in developmental processes, Pbx1 (pre B cell leukemia homeobox 1) and Tmco1 (transmembrane coiled-coil domains protein 1) that are cis modulated in hippocampus and midbrain and Pogo transposable element with KRAB domain (Pogk) that is cis-modulated in midbrain and amygdala. Immunoglobulin-like domain containing receptor 2 (Ildr2), an endoplasmic reticulum protein involved in lipid homeostasis, is cis-modulated in hippocampus and amygdala. No neighboring genes within 2 Mb of Mgst3 are cis-modulated across BXD liver tissue, although transcriptional adaptor 1 (Tada1), which is part of the STAGA histone acetyltransferase complex, is cis-modulated in liver hepatocytes and amygdala. Although several genes are cis-modulated in one or two tissues, Mgst3 is the only gene in this interval consistently cis-modulated across all BXD brain and liver datasets surveyed.

GST superfamily expression is correlated in both brain and peripheral tissue

In addition to exploring the genetic regulation of individual GSTs, we also explored interactions among all family members and found that there is extensive covariation in hippocampus and liver. Covariation can be driven by shared biological function, transcriptional regulation, or genetic regulation. Close physical proximity on chromosomes can drive coexpression through common transcriptional control. Paralogs of the alpha, mu, omega, and theta class are located in gene clusters on Chrs 9 (78 Mb), 3 (107.7Mb), 19 (47.9 Mb), and 10 (75.2 Mb), respectively. Theta class members Gstt1-3 demonstrate robust expression covariation and Mu class members Gstm1, Gstm5, Gstm7, and Gstm2 are also well correlated (Fig 9). In contrast, Alpha and Omega paralogs do not display extensive covariation. No common genetic mechanism—such as a single regulatory locus or quantitative trait loci (QTL)—was detected that modulates the expression of multiple GSTs, suggesting that coexpression in these tissues is primarily driven by shared biological function and close physical proximity.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 9. GST coexpression networks in brain and liver.

Top and bottom panels show GST coexpression networks in hippocampus and liver, respectively. Positive correlations are indicated by warm line (edge) colors and negative correlations are indicated by cool edge colors. In addition, correlations greater than |0.7| are indicated as bold lines and those less than |0.5| are indicated by dashed lines. All correlations are greater than |0.3|. The expression of many GST genes is positively correlated.

https://doi.org/10.1371/journal.pone.0148230.g009

In hippocampus, nearly all members of the GST network are connected by at least two edges (|r| > 0.3) with the exception of Gsta2 and genes with low expression in hippocampus (Ptgds, Mgst2, and Gstm4; Fig 9). Correlations in the hippocampal GST network are mostly positive with the notable exception of most edges connecting Gstm3. Several GSTs, including Gstk1, Gstt1-3, and Mgst1, are network hubs with connectivity greater than seven nodes. The expression of Gstm2, Gstt1, Gstk1, Mgst1, and Gsta4 is tightly correlated and these genes are highly interconnected. The first principle component (PC1) of this module explains nearly 60% of the covariation in expression and was used as a signature to find additional highly correlated genes in hippocampus. The top 500 genes in this GST network are significantly enriched for localization to the mitochondrion and many metabolic and antioxidant related process (S14 and S15 Tables). Taken together these results suggests that clustering of physically unlinked GST family members in brain is most likely due to underlying biology and the role of these genes in energy production and metabolic homeostasis.

Expression covariation of GSTs is even more pronounced in liver (Fig 9). Network connections are overwhelmingly positive and there is a much higher degree of network connectivity when compared to brain. Gstm5 is unconnected from the network at a threshold of |r| > 0.3 and genes with low expression—Mgst3 and Ptgds—are not well connected. Interestingly, positive and negative correlations are not well conserved between the hippocampus and liver network and may result from differences in expression level, which tends to be higher in liver, or tissue heterogeneity, which is greater in brain. However, comparison of tightly correlated but physically unlinked GSTs in liver revealed the same underlying network biology. Gstm1, Gsta2, Gstm6, Gstt1, Gstk1, and Gstz1 cluster tightly and PC1 captures over 70% of the variance in coexpression. The top 500 correlates of the GST network signature in liver are also significantly enriched for localization to the mitochondrion and metabolic terms (S16 and S17 Tables). Despite the strong overlap of biological function associated with both GST networks there is only about a 9% overlap between the top correlates in liver in brain, suggesting tissue specific expression of correlated metabolic genes in these tissues.

Minimal obvious impact of Gstt2 and Gstz1 variants

The cis eQTLs in Gstt2 and Gstz1 could not be confirmed by ASE due to the lack of an informative coding variant and relatively low hippocampal expression, respectively. Both GSTs reside in relatively gene or variant sparse regions and are the only genes within each interval with local modulation of expression. Coding mutations—missense, stop, or deletions—could alter protein function without a concomitant change in gene expression and, consequently, would not be identified as cis eQTLs. A search for high impact coding variants in all genes located within each respective QTL confidence interval was conducted for Gstt2 and Gstz1 using inbred strain genomic sequence data available and annotated by the Wellcome Trust Sanger Institute (http://www.sanger.ac.uk/sanger/Mouse_SnpViewer/rel-1303; [45]). This analysis did not reveal any genes within the confidence interval with potentially damaging coding variants, except for a frameshift mutation in a predicted gene model Gm6803 located in the Gstz1 interval. Gstt2 and Gstz1 are therefore good candidates for modulating any traits that map back to the vicinity of each gene. However, these genes appear to exert only a subtle effect on phenotype with few transcripts or higher order traits mapping back to either locus. Variation in Gstt2 or Gstz1 may be buffered by the activities of other GSTs and genes with overlapping functions, or an effect may only be evident after a challenge to the system.

Compensation or buffering is an important consideration for Gstt2 as there are multiple highly related Theta class members in the mouse genome. Gstt1 and Gstt2, for example, share greater than 50% sequence homology in both rodents and humans. However, in humans there are only two theta class members, compared to four in mice, and the locus is very complex; GSTT1 is frequently deleted [75–78] and an inverted duplication of GSTT2 leads to the transcription of pseudogenes whose function is not well defined [76,79]. While several studies have reported associations between GSTT2 and differential cancer susceptibility [79–81], the potential role of this gene is still unclear due to the structural complexity of this locus in humans. Several targeted mutations of murine Gstt2 are available, however, phenotypic abnormalities have yet to be reported (Mouse Genome Informatics, www.informatics.jax.org). The lack of a strong phenotype upon deletion is consistent with our lack of evidence for direct genetic control over more complex traits, however, multiple transcripts are modulated from the Gstt2 locus in the BXD population. These genes play a role in development and cell growth including, Apold1, Aebp2, Eif4a1, Hoxb, Mtpn, Fgf12, Deptor, Nfasc. Our genetic dissection of transcripts modulated by variation in Gstt2 in the BXD population suggests a plausible link between regulation of cell growth pathways and cancer GWAS [79–81] and further investigation is warranted.

The failure to associate robust trait variation to segregating mutations in Gstz1 is perhaps especially surprising as this gene has no paralogs and plays a distinct role in both rodents and humans in the catabolism of phenylalanine and tyrosine [82] and in the biotransformation of dichloroacetate (DCA, a chemotherapeutic agent that also inactivates Gstz1) to glyoxylate [83]. Indeed, deletion of this gene in mice leads to upregulation of alpha, mu, and pi class GSTs and is also associated with increased sensitivity to diets high in phenylalanine and tyrosine [84,85]. These strong effects are puzzling given our findings, however, naturally occurring variants in regulatory regions are expected to result in smaller expression changes than complete gene deletion and may therefore exert a more modest influence on phenotype. In humans there are common haplotypes in GSTZ1 that could affect gene function [86], although the implications on metabolic function or pharmacokinetics of drug metabolism remain unclear [11]. The BXD family would thus be ideally suited for targeted dietary or toxicological studies to evaluate potential consequences of naturally occurring variation in Gstz1.

Strong and consistent variation in Gsto1 levels is not associated with major phenotypic abnormalities

Both members of the Omega class are conserved between human and rodent but only Gsto1 demonstrates significant cis-modulation of expression in the BXD population. Interestingly, this variation has only modest effects on downstream trait variation. In humans, a locus on human chromosome 10q that contains both GSTO1 and GSTO2 has been implicated in risk and age-at-onset of AD [87–89]. Li and colleagues also detected significantly lower GSTO1 levels in the hippocampus of AD patients compared to controls, and alleles of GSTO1 are significantly associated with age-at-onset for AD and Parkinson’s disease (PD) [88]. Despite these findings, several follow-up studies have been published with mixed results [90–93]. GSTO1 is also thought to play a role in inflammation [94], arsenic and alpha-haloketone biotransformation [95,96], and reduction of S-(phenacyl)glutathiones, such as S-(4-nitrophenacyl)glutathione [96]. However, deletion of Gsto1 in mice does not produce an observable phenotype [97] and GSTO1 deficiency in a human breast cancer cell line does not alter sensitivity to arsenic or cytotoxic cancer drugs [95] Despite these discordant findings Gsto1 expression is highly correlated with many metabolic genes, some of which are involved in AD pathways. Among the 2384 unique gene correlates of Gsto1 expression in hippocampus are 128 genes that share a high co-incidence of literature citations—24 of which are known to be associated with AD or dementia, including well studied candidates Chmp2b, Optn, Dlst, Sod2, and Acat1 [98–107]. GSTO1 expression in the cortex of human AD and control cases is highly variable and a significant decrease in GSTO1 levels is evident among AD cases. In both human (AD and control cases) and mouse GSTO1 correlation networks we found significant enrichment for mitochondrial genes. Although probably not the main genetic driver, GSTO1 could interact with other variants, environmental exposure, and aging processes to influence AD risk and pathology, perhaps through changes in glutathione regulation, metabolic processes, or mitochondrial function. Mitochondrial dysfunction has long been implicated in a neurodegenerative disorders such as AD, PD, Huntington’s disease, and amyotrophic lateral sclerosis [108,109]. A direct link between GSTO1 and mitochondrial function has not yet been reported but is suggested in both our human and mouse brain expression networks and further investigation is warranted

Variants in Gsta4 exert a downstream effect on trait variation

We found that expression of Gsta4 is variable and modulated by local sequence variants in the BXD and other murine populations. In primates and rodents, Gsta4 has a high affinity for a major byproduct of lipid peroxidation, 4-hydroxynon-2-enal (4-HNE) [110]. Toxic by-products of lipid peroxidation such as 4-HNE act as signaling molecules and markers of peroxidation that are often elevated after tissue injury or in inflammatory and degenerative diseases, including Parkinson’s Disease [111]. There is also evidence that higher levels of Gsta4 may protect against neurodegeneration, especially after traumatic brain injury, which is associated with elevated levels of 4-HNE [112]. Non-coding variants segregate among human populations, but it remains to be seen how these impact gene expression [113]. However, variation at GSTA4 has been associated with epilepsy remission following pharmacological intervention [114], suggesting a possible role in drug metabolism in addition to a putative neuroprotective role. Gsta4 null mice show a range of tissue-specific phenotypes, including susceptibility to oxidative stress [110], CCl4 hepatotoxicity [115], and increased mitochondrial dysfunction and protein carbonylation in adipose tissue [116].

Based on the biological function of Gsta4 and the strong phenotype after deletion, we expected Gsta4 variation to exert a downstream effect on gene expression and higher order phenotypes in the BXD panel. Indeed, our results support both known and novel functions. For example, Cxcl2 (prefrontal cortex), Bet1 (liver), and Nosip (hippocampus) are modulated by variation in Gsta4 and are involved in response to lipid peroxidation. In Gsta4 knockout mice, exposure to an alcohol liquid diet results in increased lipid peroxidation and expression of multiple inflammatory markers, including Cxcl2 [117]. Bet1 expression has been shown to increase during ER stress [118], which can be induced by various insults, including accumulation of misfolded proteins and by lipid oxidation byproducts such as 4HNE [119]. Combination of superoxide and the signaling molecule nitric oxide (NO) can generate peroxynitrite anions that damage membranes and induce lipid peroxidation [120]. Of note, Nosip is a negative regulator of the main enzyme (nitric oxide synthase) responsible for production of NO [121].

Other downstream targets of Gsta4 hint at an overlapping but undefined role in response to oxidative stress and include genes involved in apoptosis (Dedd2), ubiquitination (Fbxo46), transport into the ER (Sec63) transcriptional regulation (Zfp59, Ino8e, Fam60a), immune function (Gm2a, Effs, Ill7rc) and protein targeting to membranes (Pigp). In addition to a role in immune function, Gm2a is also thought to bind and traffic a range of lipid targets, including phosphatidylcholine [122] and could be involved in repair pathways in response to lipid peroxidation.

Finally, several of the gene products that mapped back to the Gsta4 locus in hippocampus are thought to be involved in cognition and behavior (Aust2, Csmd2, Dstyk, Pam). Interestingly, a number of traits related to hippocampal activity and behavior were also partially modulated by variation in Gsta4, such as hippocampal mossy fiber pathway volume, dopaminergic activity (homovanillic acid levels) in the medial septal nucleus, and several measures of locomotor activity. The medial septal nucleus is innervated by dopaminergic projections from the ventral tegmental area [123] and projects to the hippocampal formation. Stimulation of this ascending pathway produces theta wave activity in the hippocampus of rodents [124] that is critical for information processing and memory storage during exploratory behavior [125,126]. These downstream traits suggest a novel role of Gsta4 in exploratory activity, perhaps mediated through the septal-hippocampal pathway.

Although these findings are exciting and warrant further investigation, a caveat of this analysis is linkage disequilibrium and the size of the haplotype blocks in the BXD population. A neighboring gene involved in fatty acid synthesis, Elovl5, is also modulated by a cis eQTL and could be responsible for variation in traits mapping to this locus on chromosome 9. Elovl5 deletion results in female infertility, increased levels of liver triglycerides, and hepatic steatosis [127] with no effect on body weight or other phenotypic abnormalities. In contrast Gsta4 knockout mice exhibit a range of phenotypic alterations including, reduction of 4-HNE conjugating activity and higher baseline levels of 4-HNE [110,115], lower litter size, increased incidence of bacterial infections, greater susceptibility to the herbicide paraquat, higher fat content in bone marrow [110], mitochondrial defects [116] and extended lifespan [128]. The higher burden of phenotypic abnormalities in the Gsta4 knockout mouse and the overlapping roles in lipid peroxidation pathways of downstream targets identified in our study strongly suggest that the majority of traits mapping back to the chromosome 9 locus are indeed modulated primarily by variation in Gsta4 and not Elovl5.

Variants in Mgst3 exert a downstream effect on trait variation

The only membrane-associated microsomal GST family member with consistent genetic modulation of expression among the BXD population is Mgst3. Although not as well studied as some of the other cis-modulated genes, variation has previously been linked to hippocampal size in the BXD population [129]. In addition, activation of stress response transcriptional activators such as PPARα or Nrf2 lead to increased expression of Mgst3 [130] and studies have documented increased Mgst3 expression in liver after caloric restriction [131] and in heart in response to opioidergic preconditioning that is protective against ischemia-induced injury [132]. Conversely, downregulation of Mgst3 has been observed in Alzheimer’s disease [133]. These studies suggest that Mgst3 has overlapping roles with other GSTs in oxidative stress pathways, unlike many of the other MAPEG family members primarily involved in inflammatory pathways [2].

Downstream trait analysis in the BXD family supports involvement of Mgst3 in detoxification and oxidative stress pathways and hints at novel roles in pathways involved in cancer, cell growth, and behavior. In addition to hippocampal size, multiple measures of locomotor activity are modulated from the Mgst3 locus, suggesting that this gene may also play a role in circuitry or metabolism of brain regions involved in exploratory behavior. In liver, nearly all transcripts whose expression is modulated by Mgst3 are involved in some aspect of metabolism or detoxification (Acadm, Cyp2b, Dtd1, Vrt), with the notable exception of Sdcbp which is a signaling molecule implicated in tumor growth and metastasis for melanoma [134], hepatoma [135], lung cancer [136], glioma [137], urothelial cell carcinoma [138], and breast cancer [139]. Hundreds of transcripts map back to the Mgst3 locus in brain, with 67 and 52 modulated in hippocampus and midbrain, respectively. Among these, only Csnk1a1 was significantly modulated by variation in Mgst3 in both regions. Csnk1a1 is known to regulate the Wnt signaling pathway [140], which is frequently disrupted in cancer and modulates cell growth, adhesion, and survival. Loss of function of Csnk1a1 may be related to poor prognosis in colon cancer [141,142]. In addition, alterations in Csnk1a1 function or expression have been implicated in prostate cancer [143], leukemia [144], myelodysplatic syndrome [145], and breast cancer [146]. The large set of genes that are partially controlled by variation in Mgst3 in hippocampus and midbrain are significantly enriched for signaling to p38 via RIT and RIN, RAF activation and L-ascorbate biosynthesis. The P38 signaling pathway is activated by environmental stress, oxidative stress, and cytokine signaling [147,148]. Activation of the Ras/Raf/MEK/ERK signaling pathway is associated with cell growth and survival, and is often altered in oncogenesis [149]. Taken together, systems genetics analysis across the BXD family supports a role for Mgst3 in oxidative stress and metabolic pathways and suggests a possible novel role in cell growth signaling pathways often disrupted in many cancers.

Conclusion

GSTs are part of a larger cell surveillance system involving GSH. This system depends on the actions of genes involved in GSH synthesis (e.g. glutamate cysteine ligase and glutathione synthetase), export of GSH conjugates from the cell by multidrug resistance-associated proteins, and a whole host of enzymes involved in scavenging reactive oxygen species (ROS) and removal of toxic byproducts of cellular respiration, such as aldehyde and alcohol dehydrogenase, catalase, glutathione peroxidase, and superoxide dismutase [2]. Understanding the genetic architecture of GSTs and other members of the GSH surveillance system in the BXD family is critically important for modeling human disease and pharmacokinetics in murine populations and for evaluating the biological impact of these genes. Multiple GSTs, including Gsta4, Mgst3, Gstt2, Gstz1, and Gsto1 are variable among the BXD population with some individuals carrying mutations that alter gene expression. Future studies can leverage this variation in the BXD cohort to test the influence of these variants in response to perturbations such as exposure to drugs or toxins, cancer, or neurodegeneration, or to evaluate novel associations in human GWAS. Here we used an integrative translational approach to characterize gene regulation and co-expression, and to identify potential downstream targets of GST genes. Our analysis provides new insight into the biological function of GSTs and possible mechanistic links to disease.