Multilayered Genetic and Omics Dissection of Mitochondrial Activity in a Mouse Reference Population

SUMMARY

The manner by which genotype and environment affect complex phenotypes is one of the fundamental questions in biology. In this study, we quantified the transcriptome—a subset of the metabolome—and, using targeted proteomics, quantified a subset of the liver proteome from 40 strains of the BXD mouse genetic reference population on two diverse diets. We discovered dozens of transcript, protein, and metabolite QTLs, several of which linked to metabolic phenotypes. Most prominently, Dhtkd1 was identified as a primary regulator of 2-aminoadipate, explaining variance in fasted glucose and diabetes status in both mice and humans. These integrated molecular profiles also allowed further characterization of complex pathways, particularly the mitochondrial unfolded protein response (UPR^mt). UPR^mt shows strikingly variant responses at the transcript and protein level that are remarkably conserved among C. elegans, mice, and humans. Overall, these examples demonstrate the value of an integrated multilayered omics approach to characterize complex metabolic phenotypes.

INTRODUCTION

The central dogma of molecular biology states that genetic information encoded in DNA is first transcribed by RNA polymerase, then translated by ribosomes into proteins. However, the DNA sequence of a gene provides little information for predicting when, where, and to what extent its associated RNA and protein products will be expressed. Since the advent of microarray technology, comprehensive gene expression patterns—i.e., the transcriptome— can be precisely and comprehensively quantified across large populations. Unfortunately, transcript levels generally have only modest correlation with the levels of corresponding proteins (Ghazalpour et al., 2011; Gygi et al., 1999; Schwan-häusser et al., 2011), and genetic variants similarly affecting both the transcript and peptide levels of a gene are relatively uncommon (Albert et al., 2014; Skelly et al., 2013). As proteins in most cases are more directly responsible than transcripts in the regulation of cellular pathways—and ultimately phenotypic traits—there is a critical need for efficient, large-scale, and accurately quantitative proteomics methods to complement transcriptomic data sets.

Over the past decade, the development of discovery mass spectrometry (“shotgun”) has allowed the first large-scale studies on quantitative proteomics. In this approach, protein extracts are cleaved into short peptide sequences, which are then chromatographically separated and analyzed by tandem mass spectrometry. This allows the untargeted discovery of thousands of peptides, but if the number of unique peptide fragments in a sample significantly exceeds the number of available sequencing cycles (as in whole proteome extracts), any individual peptide will be inconsistently sampled across repeat analyses. This reduces the technical reproducibility but moreover means that the number of peptides consistently quantified across all (or most) samples decreases as the study size increases (Karpievitch et al., 2012). Consequently, discovery mass spectrometry strategy has yielded mixed results in large population studies (Ghazalpour et al., 2011; Holdt et al., 2013), particularly as specific peptides of interest cannot be targeted, and the most consistently identified peptides are biased toward those of higher abundance (Callister et al., 2006). To overcome these hurdles, selected reaction monitoring (SRM) was developed, which perfects technical reproducibility and allows consistent multiplexed quantitation of target proteins by deploying a mass spectrometric measurement assay that is specific for each targeted peptide (Lange et al., 2008a). Thus, hundreds of target peptides can be consistently and accurately quantified across large populations of samples. Recent studies in yeast have shown that the proteins and transcripts of genes are typically controlled by different, distinct mechanisms (Albert et al., 2014; Picotti et al., 2013). However, these hypotheses have not been well tested in mammalian genetic reference populations (GRPs) through multilayered transcriptomic and proteomic strategies.

Large GRPs are frequently used to determine to which extent phenotypic variation is driven by genetic variants (i.e., heritability), and to subsequently identify genes driving such variation. These genes can be identified by genome-wide association studies (GWASs) or by quantitative trait locus (QTL) mapping, approaches that have been applied to various species and have led to the successful identification of dozens of major allelic variants (Andreux et al., 2012; De Luca et al., 2003; Deeb et al., 1998; Yvert et al., 2003). In mammals, the murine BXD family is the largest and most well-studied GRP, consisting of ~150 recombinant inbred strains descended from C57BL/6J (B6) and DBA/2J (D2) (Andreux et al., 2012). Using 40 strains of this population on both chow (CD) and high fat (HFD) diets, we have obtained major metabolic phenotypes and established a multilayered data set focused on 192 metabolic genes expressed in the liver. For all genes, we know the sequence variants, transcript levels, and protein levels in all cohorts. These data are further supplemented with targeted metabolite analysis in the liver and serum, generating the first large-scale multilayered quantitative picture of any cellular process in the BXD population.

RESULTS

Protein Targeting across a Genetically and Environmentally Diverse Murine Population

We first selected 192 metabolic proteins for study, with particular focus on genes regulating mitochondria and general energy metabolism. For each gene, synthetic peptides were generated based on established assays (Picotti et al., 2010) (Figure 1A) to accurately quantify each protein across all cohorts. To validate peptide measurements, we compared the coefficients of variation (CV) among technical and biological peptide replicates. Both technical and biological replicates showed a high degree of reproducibility (CV ~ 0.09 and 0.12, respectively), indicating the quantification of targeted peptides is sufficiently accurate across large, diverse populations. More importantly, the results for biological replicates reveal nearly equally high overall reproducibility (r > 0.98; Figure 1B), suggesting low biological variance within each cohort, particularly relative to cross-cohort variance (CV ~ 0.33); 82.1% of the quantified peptides are highly variable (CV > 0.20), indicating that the biological error (i.e., variation within a cohort) is much smaller than the variation induced by differences in genotypes and diet. We observed nearly complete quantification of all 192 proteins across all cohorts, with only ~60 missing peptide counts out of ~15,000 measurements (i.e., 99.6% completion; Figure 1C, top). This completion is similar to the level of completeness achieved by microarrays (Figure 1C, bottom) and is a major contrast to shotgun-acquired proteomics, which typically have completeness of ~70% (Karpievitch et al., 2012).

Figure 1. SRM-Based Protein Quantification and Covariation Network.

(A) SRM assay development for targeted proteomic measurements; 309 peptides corresponding to 192 genes were designed and synthesized via SPOT synthesis. Fragment ion spectra were generated on a triple-quadrupole MS with the SRM-triggered MS2 mode, and then ions were selected based on their relative intensities. Dot colors indicate different amino acids. The following abbreviation was used: m/z, mass-to-charge ratio. Mouse liver homogenate combined with the heavy reference proteome was analyzed with SRM on a triple-quadrupole MS. Different dot colors represent different peptides (Quadrupole 1) or product ions (Quadrupole 3).

(B) Biological replicates had near-perfect reproducibility (r² ~ 0.98), shown here for all three biological replicates of the BXD60 CD cohort.

(C) Two-way cluster analysis of protein (top) and transcript (bottom) abundances in all 77 cohorts (40 CD, 37 HFD). Columns are clustered based on samples, and rows are clustered based on gene-product abundances. Protein and transcript abundances are colored in a red-blue scale. Red, high abundance; blue, low abundance; white, missing data.

(D) Protein association network based on robust Spearman correlation measures for all protein pairs. Statistically significant and strong positive associations (p < 0.01 and r > 0.6) are edges. The largest correlation clusters are labeled “a,” “b,” and “c.” Nodes are labeled with protein names and colored according to their biological process, as reported by DAVID (Huang et al., 2009).

It has been well established that transcriptomic networks of many metabolic processes covary quite well, e.g., within the electron transport chain or within the citric acid cycle (Ihmels et al., 2002). At the protein level, proteins that function in common biological processes or that localize to the same functional modules are also reported to be subject to a similar regulatory process and generally covary (Foster et al., 2006). To validate and identify which of these 192 proteins vary synchronously, we computed the robust Spearman correlation network for all protein pairs using the full SRM data set (Figure 1D). The resulting network contained 82 correlated nodes (proteins) with 211 edges in 3 main enrichment clusters. As expected, proteins within ontologies are highly correlated such as for mitochondria (Figure 1Da) and lipid metabolism (Figure 1Db). Within Figure 1Da are five of the six measured proteins involved in mitochondrial unfolded protein response (UPR^mt) (HSPD1, HSPE1, HSPA9, CLPP, and LONP1, indicated in red), along with three of the four measured NADH dehydrogenase genes (NDUFA1, NDUFB3 and NDUFS6, in blue) and four of the eight measured mitochondrial-encoded proteins (MT-CYB, MT-CO2, MT-CO3, and MT-ND3, in black). Meanwhile, proteins involved in carbohydrate metabolism are enriched in the same cluster (Figure 1Dc). These results show that functionally related proteins tend to be coordinately regulated at a protein level and that coregulation of protein abundance is strongly maintained for certain processes. To validate the biological significance of these function-based covariation clusters, we further investigated one: the UPR^mt network (elaborated on later).

Protein and mRNA Gene Products Generally Do Not Correlate

With the general protein measurements validated, we generated a global overview of how genotype and diet influence differential transcript and protein expression (Figure 2A; numerical values listed in Table S1 available online). At the genetic level, transcripts and proteins map to an equivalent number of significant QTLs: we detected 65 significant transcript QTLs (eQTLs—blue lines at the center) and 57 significant protein QTLs (pQTLs—red lines at the center). However, though the total number of eQTLs and pQTLs are roughly equivalent, the predominant type of regulation was very different: 74% of the eQTLs are cis-mapping (ratio of solid to dashed blue lines), versus only 31% of pQTLs (ratio of solid to dashed red lines). This indicates a closer connection between a transcript and its gene than a protein and its gene. In general, trans-mapping proteins and transcripts mapped evenly across the genome, not yielding any clear “hot spots” for these metabolic genes (sample magnification of QTLs mapping to chromosome 5 is shown in Figure 2B).

Figure 2. mRNA and Protein Overview.

(A) Circos plot of mRNA and protein data for all 192 genes, labeled on outer edge. Genes are represented by two bars: light for CD and dark for HFD. Genes are arranged by relative chromosome position; the chromosome length is according to number of genes measured. Blue bars indicate the transcript relative expression CD versus HFD; orange bars indicate the protein relative expression CD versus HFD. Bars with more unequal heights indicate diet has a larger impact. Green bars indicate the correlation between the transcript and protein for each gene within each diet. Fuchsia bars represent the strength of the peak eQTL. Yellow bars represent the strength of the peak pQTL. The two bars are overlaid with transparency. The dashed green line represents the simplified significance cutoff (LRS ≥ 18). The inner ring indicates the chromosome location. The blue central lines represent significant eQTLs, and the red central lines represent significant pQTLs. The central solid lines represent cis-QTLs, and the central dashed lines represent trans-QTLs. QTL lines stem from the LRS bar graph and terminate on the inner side of the chromosome ring at the approximate QTL location.

(B) Magnified view of eQTLs and pQTLs mapping to chromosome 5.

(C) In CD ~25% (left), and in HFD ~30% (right), of transcripts correlate nominally significantly with their protein. The lower strip charts show correlation distribution. Spearman correlation values corresponding to nominal significance (p < 0.05) and corrected significance (p < 0.0002) are displayed on the axis.

(D) Venn diagram of genes that are differentially regulated between CD and HFD as transcripts (blue), proteins (red), both (purple), or neither (gray).

(E) Volcano plot for mRNA showing the magnitude of dietary effect versus significance. ~45% vary with nominal significance (p < 0.05) between the dietary conditions. Approximately 19% vary with corrected significance (raw p < 0.0003). Some extreme genes are labeled.

(F) Plot of the effect of diet on transcripts versus the effect of diet on proteins. In general, transcripts and proteins are similarly affected by diet.

Related to Tables S1 and S2.

Within each diet, we found proteins and transcripts to be nominally correlated for ~25% of genes (i.e., Spearman correlation p value < 0.05; Figure 2C, diets considered separately), similar to findings in other species and populations (Foss et al., 2007; Ghazalpour et al., 2011; Schwanhäusser et al., 2011; Skelly et al., 2013). Of the correlated transcript-peptide pairs (46 in CD, 55 in HFD), 31 correlate significantly in both cohorts, whereas 70 pairs correlate in only one dietary cohort. Correspondingly, while ~50% of genes were affected by diet, transcripts were more frequently influenced (84 of 189) than proteins (37 of 192), with only 21 genes affected at both the transcript and protein levels (Figure 2D). Genes that were the most strongly affected by diet at the transcript level (Figure 2E) tend to be similarly affected at the protein level (Figure 2F)—e.g., Cyp3a11 is higher in CD—though exceptions are frequent, e.g., Srebf1 mRNA is induced by HFD, but unaffected at the protein level, while in another counterexample, the HFD cohorts have more ETFDH protein, but less of the transcript. Thus, while transcripts and proteins are moderately covarying estimations of their gene’s activity and typically have covarying responses to external factors (e.g., diet), these trends are too weak to support the measurement of any one particular transcript to serve as a proxy for the protein (or vice versa) without prior knowledge.

Most Transcript and Protein QTLs Do Not Overlap

Of the 192 target genes, 79 map to a significant eQTL or pQTL in at least one dietary condition. A strong majority of significantly mapped QTLs, ~80%, are unique to either the transcript level or protein level (Figure 3A). At the transcript level, 28 genes map to cis-eQTLs, 20 of which are in both diets, while 17 transcripts map as trans-eQTLs, none of which are observed in both diets. Together, we observed 65 significant eQTLs stemming from 45 distinct genes (Figure 3B, left; the 20 distinct genes that map cis in both diets contribute 40 significant eQTLs). The range of transcript variance within a diet was a strong predictive factor for observing an eQTL. Transcripts in the least variable quartile (range < 1.5-fold from the lowest to the highest expressing BXD strain) contained only 10% of the significant QTLs. In contrast, the second quartile (range 1.5- to 1.65-fold) contained 17%, the third quartile (range 1.65- to 2.0-fold) contained 29%, and the top quartile (range ≥ 2.0-fold) contained 41% of the significant QTLs.

Figure 3. QTL Overview.

(A) Venn diagram separating all genes with distinct QTLs based on provenance.

(B) Venn diagrams separating eQTLs or pQTLs by dietary source and regulatory mechanism. Overlapping regions indicates genes giving an eQTL or pQTL in both diets; white numbers are counted twice for QTL count, but once for distinct gene count (e.g., there are 48 significant cis-eQTLs, which stem from 28 distinct genes).

(C) In both diets, DHTKD1 and Dhtkd1 share a common cis-QTL, are unaffected by diet, and strongly correlate.

(D) NNT and Nnt display a similar pattern.

(E) Car3 has only one significant pQTL despite an absence of dietary effect and a strong transcript-protein correlation.

(F) PMPCB does not map to a significant pQTL; however, Pmpcb maps to a significant cis-eQTL in both diets, despite a strong transcriptional upregulation by HFD. The transcript and protein levels do not correlate.

(G) Mup3 and MUP3 do not map to significant QTLs, despite having high levels of variation and a strong transcript-protein correlation.

Related to Figure S1.

At the protein level, 57 significant pQTLs stem from 48 distinct proteins (Figure 3B, right). In striking contrast to transcript regulation, only 13 distinct proteins map to cis-pQTLs, five in both diets, while 36 proteins map as trans-pQTLs. No trans-pQTLs are consistent across diets, although three proteins significantly map to separate trans-pQTLs in both diets (Cd44, Acss2, and Ndufs6), and one (Hmgcs1) maps to both cis- and trans-pQTLs. Together with transcript data, we could moderately predict pQTLs: of the 28 genes with cis-eQTLs, 12 correlated well (r > 0.40) between the transcript and protein. Of these 12 genes, five (Dhtkd1, Nnt, Tymp, Gclm, and Bckdhb) map significantly as pQTLs and eQTLs in both diets (Figures 3C, 3D, S2A, and S2B; Bckdhb in Figure 4B), and four (Acox1, Mfn1, Mri1, and Pm20d1) map as pQTLs and eQTLs in at least one diet. Interestingly, three proteins mapped as cis-pQTLs and correlated with their transcript, but did not have associated eQTLs (Hmgcs1, Car3, and Acsf2), suggesting the genetic variant driving these changes manifests more prominently at the protein level than at the transcript level (e.g., Figure 3E).

Figure 4. Metabolic Consequences of BCKDHB and DHTKD1.

(A) BCKDHB is the E1b subunit of the BCKD complex, which irreversibly converts several BCKAs.

(B) Bckdhb mRNA has the strongest eQTLs of the 192 target genes examined in the liver and has among the strongest pQTLs. Neither transcript nor protein is affected by diet.

(C) The BCAA/alanine ratio is significantly increased in animals with the dysfunctional C57BL/6 allele in CD serum and liver measurements and HFD liver measurements, in line with it acting as a risk for MSUD. Red line represents median of strains with the C57BL/6 allele, and purple lines represent the median of strains with the DBA/2 allele.

(D) DHTKD1 is the E1 subunit of the dehydrogenase complex that catalyzes the irreversible conversion of 2-oxoadipate to glutaryl-CoA.

(E) Left: serum 2-AA levels are strongly related to DHTKD1, as are liver levels (not shown). Right: other upstream metabolites in the DHTKD1 pathway correlate strongly with one another, e.g., 2-AA and 2-AA semialdehyde, shown here in liver.

(F) Serum 2-AA maps as an mQTL in both diets to proximal chromosome 2, the location of DHTKD1.

Diet also plays a clear role in the consistent identification of cis-QTLs. For the 20 genes with cis-eQTLs identified in both diets, HFD significantly affects the levels of only one (Pmpcb; Figure 3F). Conversely, for the eight genes mapping to an eQTL in only one diet, the levels of only three are affected (Ndufs2, Acads, and Aldh4a1). Similar trends are observed at the protein level: of the five cis-pQTLs found in both diets, none are affected by diet, while for the eight genes with cis-pQTLs in only one diet, only the levels of one (Hmgcs1) are affected. This relationship is similar for trans-eQTLs: of the 17 genes with trans-eQTLs, only four are influenced by diet (whereas ~50% of total transcripts are influenced by diet). However, diet does not appear to strongly influence trans-pQTLs, where 11 of the 39 are affected by diet, a similar proportion to the overall effect, recalling that ~20% of proteins have levels influenced by diet (Figure 2D). Together this suggests diet can change the genetic factors regulating these genes without changing their overall levels.

While high expression variability is associated with stronger QTLs, it is not a definitive predictor, e.g., Mup3 is highly variable, strongly affected by diet, and consistently expressed between mRNA and protein, but does not map to a QTL in any measurement (Figure 3G). Therefore, while the ability to predict peptide levels based on transcript measurements on a systems scale is quite powerful (the ~25% to ~37% of correlated transcript-peptide pairs is much better than the ~5% expected by chance), the probability to fail of any one particular gene is quite high. These chances can be adjusted somewhat—perturbations dramatically affecting transcript levels are more likely to manifest at the protein level and vice versa—but even so, prior research must be established before gene expression can be confidently considered a proxy for protein levels in targeted genetic studies. This also indicates that applying quantitative proteomic data to pathways established at the transcript level can indicate new links that were previously obscured.

Functional Relationships of pQTLs to Phenotypes

To characterize the cellular function and potential physiological relevance of the pQTLs, we first collated all Entrez (Maglott et al., 2005) and UniProt (Magrane and Consortium, 2011) entries for genes with the most significant pQTLs (LRS ≥ 20) (Table 1). As the BXDs have extensive historical phenotype and metabolite data available on GeneNetwork (Wang et al., 2003), we performed a phenome-wide association study to determine if any collected phenotype data mapped at least suggestively (likelihood ratio statistic [LRS] ≥ 12) as clinical QTLs (cQTL) to the same loci. A handful of phenotypic connections in the BXDs were supported by literature, including a link between Nnt and insulin (Wong et al., 2013) and between Car3 and subcutaneous adipose mass (Mitterberger et al., 2012). However, for the majority of pQTLs, no established cQTLs mapped to the same loci. We thus selected the two genes with the most significant and novel pQTLs in both diets for follow-up analysis and validation— Bckdhb and Dhtkd1.

Table 1.

pQTLs Link to Phenotypes

Protein	Diet	LRS	QTL Area [Mb]	Gene	Function	Phenotype
DHTKD1	CD	43	Chr2: 3–8	cis	lysine catabolism	Reye syndrome, glucose regulation
DHTKD1	HFD	41	Chr2: 3–8	cis	lysine catabolism	Reye syndrome, glucose regulation
NNT	CD	37	Chr13: 119–120	cis	NADPH regeneration	insulin secretion
GCLM	HFD	34	Chr3:118–124	cis	glutathione synthesis	detoxification, ischemia, anemia
NNT	HFD	31	Chr13: 119–120	cis	NADPH regeneration	insulin secretion
PRPS1	CD	31	Chr6: 61–66	73	nucleotide synthesis	arts syndrome
GPX1	CD	30	Chr11: 119–122	70	glutathione peroxidase	detoxification, cardiomyopathy
PM20D1	HFD	29	Chr1: 134–137	cis	zinc binding (putative)	tuberculosis sensitivity
MT-ND3	CD	28	Chr3: 11–18	21	complex I subunit	Leigh syndrome
CMAS	CD	28	Chr7: 95–105	70	sialic acid adhesion	fibrosarcoma, dermatomyositis
MRI1	CD	28	Chr8: 88–90	cis	tyrosine phosphorylation	cell invasion regulation
NDUFS6	CD	27	Chr3: 11–18	21	complex I subunit	lactic acidosis
PUM2	HFD	27	Chr10: 106–109	7	RNA binding	development, cell differentiation
BCKDHB	CD	26	Chr9: 83–85	cis	BCAA degradation	maple syrup urinary disease
BCKDHB	HFD	24	Chr9: 83–83	cis	BCAA degradation	maple syrup urinary disease
TYMP	CD	24	Chr15: 89–90	cis	thymidine phosphorylation	angiogenesis, mtDNA regulation
RPL18A	CD	24	Chr3: 25–31	33	ribosome 60S subunit	hepatitis susceptibility
CAR3	CD	24	Chr3: 14–21	cis	carbon dioxide hydration	laryngeal disease, adipogenesis
DAK	HFD	24	Chr4: 128–130	34	dihydroxyacetone kinase, FMN	glycerolipid metabolism
NDUFA4	HFD	24	Chr11: 33–40	5	complex I subunit	glucose regulation
ACADS	CD	23	Chr5: 67–76	42	beta oxidation, short chain	SCAD deficiency
HMGCS1	HFD	23	Chr4: 80–84	10	cholesterol synthesis	cholesterolemia
FASN	CD	23	Chr5: 137–139	38	fatty acid synthesis	obesity
GCLM	CD	23	Chr3: 118–124	cis	glutathione synthesis	ischemia, anemia
ABCE1	CD	23	Chr13: 114–121	28	ribonuclease L inhibitor	viral load, RNA stability
ABCB8	CD	23	Chr5: 23–25	cis	mito membrane transporter	iron homeostasis
ARG1	HFD	22	Chr11: 64–66	13	urea cycle	argininemia, hyperammonemia
DCN	HFD	22	Chr11: 51–52	11	collagen assembly	cell growth and proliferation
AKR7A2	HFD	22	Chr1: 23–24	cis	succinic semialdehyde redox	detox of aldehydes and ketones
AFP	HFD	21	Chr6: 86–88	29	fetal serum albumin (putative)	copper, fatty acid, bilirubin binding
COPB1	CD	21	Chr16: 8–15	56	Golgi complex transporter	lipid storage, LDL levels
ACSS2	HFD	21	Chr11: 115–121	101	acetate activation	lipid synthesis, energy generation
TYMP	HFD	21	Chr15: 89–90	cis	thymidine phosphorylation	angiogenesis, mtDNA regulation
COX6C	CD	20	Chr7: 141–144	14	cytochrome C electron transfer	energy generation
MYBBP1A	HFD	20	Chr14: 112–116	1	regulation of ribosomal DNA	nucleolar stress response
GSTK1	HFD	20	Chr4: 67–70	3	glutathione regulation	cellular detoxification
EPAS1	HFD	20	Chr10: 106–109	7	O2-responsive transcription factor	erythryocytosis, hypoxia response
ACSF2	CD	20	Chr11: 116–118	cis	fatty acid metabolism	tuberculosis sensitivity
SRSF7	HFD	20	Chr15: 40–48	11	spliceosome, mRNA export	immune response

A list of all genes with highly significant (LRS ≥ 20) pQTLs. For each gene, the significant peak QTL width considers a 5 LRS drop-off on either side of the peak. A count of all genes under the wide peak is displayed, unless it is a cis-pQTL as the candidate gene can be easily identified (itself). The cellular function and overt physiological effect described by literature for each gene is given.

Bckdhb is a subunit of the branched chain amino α-keto acid dehydrogenase (BCKD, EC 2.7.11.4) complex (Figure 4A) and maps to two of the most significant eQTLs and among the most significant consistent pQTLs (Figure 4B; Tables 1 and S1). Bckdhb transcript and protein levels are highly variable across strains and are unaffected by diet (Figure 4B), but no cQTL mapped to this locus. In humans, variant alleles of BCKDHB are known to cause Type IB maple syrup urinary disease (MSUD), an inborn error of metabolism characterized by a buildup of branched chain amino acids (BCAAs) and branched chain α-keto acids (BCKAs). MSUD manifests in newborns and is associated with neurological problems and delayed development, and if untreated it can be fatal (Chuang and Shih, 2001). This pathway is a multi-tissue system driven particularly by the muscle and liver, and consequently it is diagnosed by significant increases of BCAAs and BCKAs in both the serum and urine, a decrease in alanine, and/or by detection of L-alloisoleucine (Chuang and Shih, 2001; Haymond et al., 1978). Genetically engineered mouse models of MSUD have similar phenotypic progression as humans, including the same metabolite buildups and related phenotypes such as movement difficulties and reduced body growth (Koepp et al., 1974; Wu et al., 2004). We thus examined the BXDs with respect to these traits. While L-alloisoleucine was not detected, strains with the less functional B6 allele of Bckdhb had a marked increase in the BCAA/alanine ratio in the serum and liver (Figure 4C, left and right, respectively). However, no association with movement, growth or body weight was present in the BXDs (data not shown), suggesting that this Bckdhb variant carries only the subclinical metabolic effects of MSUD. This observation is also in line with the human disease, where MSUD symptoms are only apparent when BCKD activity is < ~30% of normal (Skvorak, 2009). In contrast, strains with the B6 allele of Bckdhb have ~50% the enzymatic activity compared to those with the D2 allele. However, this raises the possibility that a sustained BCAA loading test in the BXDs may lead to clinical manifestation of the intermittent form of MSUD.

To continue scanning for physiological links with the proteins examined, we also investigated the strongest pQTL in both diets. This gene, Dhtkd1, encodes a mitochondrial protein operating as the E1 subunit of a dehydrogenase complex involved in lysine metabolism (EC 1.2.4.2; Figure 4D). In humans, variants in DHTKD1 have been linked to urinary buildup of 2-aminoadipate (2-AA) and 2-oxoadipate, yet the few studies on the physiological consequences of these variants are conflicted (Danhauser et al., 2012; Houten et al., 2013; Xu et al., 2012). After measuring the metabolites involved in this pathway in the BXD liver and serum, we observed a striking correlation between DHTKD1 levels and 2-AA (Figure 4E, left) as well as between metabolites in the pathway itself (Figure 4E, right and Figure S2A). Moreover, 2-AA mapped significantly in both diets as an mQTL to the locus of Dhtkd1 itself, indicating the BXDs share an inborn error of metabolism similar to humans (Figure 4F). Due to the strength and consistency of this correlation across diets, we hypothesized this variant may lead to phenotypic consequences and inform for human validation.

DHTKD1: A Regulator of Glucose Homeostasis

Inhibition of DHTKD1 in human liver cells is known to diminish mitochondrial activity (Xu et al., 2013), and external administration of 2-AA improves insulin secretion in mice (Wang et al., 2013). We thus examined whether the natural genetic variants in Dhtkd1 may also influence glucose homeostasis in the BXDs. The BXDs have varying responses to HFD-induced diabesity and include some resistant strains, but most strains are affected, leading to general increases in liver weight, serum cholesterol, glucose, and overall body weight (Figures 5A and S2B). Strikingly, 2-AA is negatively associated with liver mass and fasted glucose levels in both diets, and with serum cholesterol and insulin in CD (Figure 5B). Other metabolites in this pathway were also connected to the same phenotypes and with negative correlation [e.g., saccharopine (Figure 5C) and α-ketoglutarate (Figure S2C)]. Interestingly, 2-AA levels are significantly decreased in the HFD cohorts (Figure 5D), despite the genetic regulation of 2-AA by Dhtkd1 being equivalent between CD and HFD cohorts (Figures 4E and 4F). However, this decrease is not directly linked to diet, but instead with the animals’ progression toward diabetes. Based on HOMA insulin resistance indices (Lee et al., 2008), most HFD-fed BXD animals (108 of 180) became glucose intolerant and insulin resistant over the six-month period of the dietary challenge (Figure S2D). Conversely, only 18% of CD animals (29 of 165) have poor insulin sensitivity. Strikingly, the changes in 2-AA levels go in tandem with the HOMA index: animals with worse insulin sensitivity have lower 2-AA, and animals with good insulin sensitivity have equivalent levels of 2-AA, with no effects linked directly to diet (Figure 5E). With this in mind, we further examined 2-AA in two independent population studies. In the Hybrid Mouse Diversity Panel, a recent profile of liver metabolomes in 96 cohorts in the fasted state contained 2-AA measurements (Ghazalpour et al., 2014). In this diverse population, which includes~30 BXD strains (with only one strain overlapping with those in our study), 2-AA is diminished in cohorts with higher fasted glucose (Figure 5F), and the two are again negatively correlated (Figure S2E). Likewise, we observed the same trend in a human population study (Cohorte Lausannoise [CoLaus] study [Firmann et al., 2008]), in which we recently analyzed the spot urine samples of 835 middle-aged or elderly individuals (Rueedi et al., 2014). In this cohort, diabetic patients (Figure S2E) also displayed a marked decrease in 2-AA levels (Figure 5F). Together, these findings indicate that Dhtkd1 plays a major role in the regulation of 2-AA in both mice and humans and that the modulation of this pathway may influence the development of diabetes.

Figure 5. Physiological Consequences of DHTKD1 Variants.

(A) Liver size, serum cholesterol, and fasted glucose levels increase after HFD across the BXDs. Error bars represent mean + SEM.

(B) In CD livers, 2-AA is associated negatively with liver mass, serum cholesterol, insulin, and glucose. In HFD livers, 2-AA is associated negatively with liver mass and fasting glucose.

(C) Liver levels of saccharopine, a metabolite upstream of 2-AA, are also associated negatively with the same phenotypes, although only in CD.

(D) 2-AA levels in the liver are significantly decreased in HFD-fed BXD cohorts.

(E) However, when correcting for diabetes status (HOMA-IR > 10 as diabetic, or HOMA-IR < 5 as healthy), there was no difference between CD and HFD—instead, only between diabetic and nondiabetic cohorts.

(F) Using publicly available data from a recent metabolomic profiling of livers from the HMDP (Ghazalpour et al., 2014), we can observe that the inverse relationship between 2-AA and fasted glucose is highly consistent in mouse populations. p = 0.01 if the four high nonoutliers in the low group are suppressed.

(G) Also in a human population study with urine metabolomics, diabetic patients had markedly lower levels of 2-AA.

Related to Figure S2.

The Mitochondrial Unfolded Protein Response

The UPR^mt is a mitochondrial stress response pathway that is activated by proteostatic stress, such as by accumulation of unassembled or unfolded proteins in the mitochondria (Zhao et al., 2002), by the presence of an imbalance between mitochondrial and nuclear encoded proteins (Houtkooper et al., 2013), or by electron transport chain defects (Durieux et al., 2011; Runkel et al., 2013). The activation of UPR^mt in turn leads to the transcription and translation of nuclear-encoded protective genes such as mitochondrial chaperones and proteases to reestablish mitochondrial proteostasis (reviewed in Haynes et al., 2013; Jovaisaite et al., 2014; Wolff et al., 2014). The bulk of research on UPR^mt has taken place using C. elegans and mammalian cell lines, thus little is known about when or how UPR^mt occurs in vivo in mammals. Furthermore, as the UPR^mt is a stress response tied to maintaining mitochondrial protein balance, we hypothesized that its protein correlation networks may be different than those generally examined at the transcriptional level.

In the worm, two “classical” approaches have been typically used to induce UPR^mt: the loss of function of cco-1, a nuclear encoded component of the electron transport chain (Durieux et al., 2011), or the loss of function of spg-7, a mitochondrial protein quality-control protease (Yoneda et al., 2004). We confirmed that the knockdown of either gene by RNAi triggers the UPR^mt response in C. elegans, by strong induction of the mitochondrial chaperone hsp-6 and of the proteases lonp-1 and clpp-1 (Figure 6A). Moreover, we linked this UPR^mt activation to specific phenotypes—a major reduction in size and mobility, as well as a decrease in oxygen consumption—which are consequences of mitochondrial stress (Figure 6B). However, it has not been previously shown whether this coordinated regulation of UPR^mt genes is conserved in mammals in vivo.

Figure 6. The Mitochondrial Unfolded Protein Response.

(A) UPR^mt induction in C. elegans triggered by interference with ETC (RNAi of cco-1) or mitochondrial proteostasis (RNAi for spg-7). These triggers result in upregulation of UPR^mt effectors hsp-6, clpp-1, and lonp-1 and a reduction in ubl-5. The orthologous mouse genes are indicated below the respective C. elegans gene symbol. Error bars represent mean + SEM.

(B) UPR^mt induction in C. elegans decreases movement, size, and oxygen consumption.

(C) UPR^mt genes and proteins form a network of coordinately expressed mRNAs and proteins in vivo in mice, which is stronger at the protein than at the mRNA level.

(D) Cox5b and Spg7 (orthologs of C. elegans cco-1 and spg-7) are generally negatively associated with the levels of all UPR^mt genes in CD cohorts, particularly at the protein level, in line with observations in the worm.

(E) While the levels of Cox5b and Spg7 are not affected by diet, expression is consistent by strain across the two diets only for Cox5b.

(F) The UPR^mt network in HFD livers is similar to that observed in CD, but somewhat weaker. Ubl5 remains a striking negative correlate at the mRNA level.

(G) In HFD, Cox5b remains a negative correlate of UPRmt transcripts and proteins, while Spg7 does not.

(H) The features of the UPR^mt network are also conserved in 427 human liver biopsies (Schadt et al., 2008), 405 lung biopsies (Ding et al., 2004), 180 lymphoblast lines (Monks et al., 2004), and 43 hippocampi (Berchtold et al., 2008).

(I) In humans, SPG7 is a consistent negative correlate of the UPR^mt transcripts.

In the BXDs, we investigated the expression of six members of the UPR^mt pathway, which are well conserved from C. elegans: mitochondrial chaperones (Hspd1, Hspe1, Hspa9), proteases (Clpp, Lonp1), and a transcriptional regulator involved in UPR^mt (Ubl5). These UPR^mt genes are also coordinately regulated at both mRNA and protein level in the BXDs, but with much stronger connections among proteins (Figure 6C). Moreover, the UPR^mt network correlates negatively with Cox5b and Spg7 (mouse orthologs of worm cco-1 and spg-7, respectively), indicating that low abundance of these genes amplifies UPR^mt in mammals as in C. elegans (Figure 6D). The network is also influenced in part by diet. While Cox5b expression patterns are similar between CD and HFD, Spg7 covariation is disjointed between the dietary cohorts (Figure 6E). This may explain why, despite a similar overall UPR^mt response in both diets (Figures 6D and 6F), Spg7 trends positively in HFD cohorts, while Cox5b remains consistent (Figure 6G). Using four large transcriptional studies of human tissue biopsies, we observe similar transcriptional links, particularly including a strong network between all the UPR^mt genes (Figure 6H). In humans, SPG7 is a consistent negative correlate of this network, in contrast to COX5B, which generally has positive covariation with the UPR^mt response (Figure 6I). Thus, while many of the overall regulators of UPR^mt remain coregulated across species—worm, mouse, and human—particular nuances of its activation pathways appear to be variable dependent on species, environment, tissue, and likely other factors.

DISCUSSION

Due to major differences in transcript and protein regulation, it has become increasingly clear that systems proteomics is essential for the analysis of complex systems such as metabolism (Khan et al., 2013; Skelly et al., 2013). Traditionally this was attempted by semiquantitative immunoassays such as ELISA or western blotting, yet these techniques allow only a handful of proteins to be measured in parallel and are moreover limited by the scarcity of quantitative immunoassays and the dubious quality of many antibodies (Marx, 2013). Recent shotgun proteomics experiments on diverse populations have provided fundamental data on protein variance in populations of yeast (Albert et al., 2014; Skelly et al., 2013), mammals (Ghazalpour et al., 2011), and humans (Hwang et al., 2010). However, the inability to measure target proteins has limited the application of this approach in the study of disease pathways. New developments in mass spectrometry have led to SRM, a targeted quantitative systems proteomics technique in which choice proteins can be quantified based on a priori information. This technique has been recently applied in a moderate-sized yeast population study (Picotti et al., 2010), but until now it has not been applied to study genetic regulation in multicellular species or for a choice pathway.

In this study, we quantified 192 metabolism genes at the transcript and protein levels in livers from 77 cohorts of the BXD GRP under two different dietary conditions. Along with sequence variants, basic metabolomics, and phenotype data, this combined multilayered population data set enabled us to accurately estimate abundance changes in gene products due to genotype, diet, and gene-by-environment interactions (GXE). By complementing these layers, we were able to tentatively link dozens of pQTLs with phenotypes. Of the ~50 of significant pQTLs identified, we established links between four genes leading to phenotypic consequence: two novel and two confirmatory. For these genes, Nnt, Car3, Dhtkd1, and Bckdhb, coding differences between B6 and D2 alleles were linked to robust differences at both the transcript level and protein level, leading to shared cis-regulatory eQTLs and pQTLs and then to phenotypic consequences. While Nnt has been previously attributed as causal to differences in insulin secretion in the BXDs (Wong et al., 2013), the other pQTLs are novel, including Car3, which was previously linked to adiposity only in knockout mice (Mitterberger et al., 2012). For Bckdhb, strains carrying the B6 allele have a ~50% reduction in enzyme levels, which leads to a buildup of BCAA levels and a decrease in alanine levels, as in the intermittent form of MSUD. In humans, MSUD is a rare inborn error of metabolism, and while BXDs do not suffer any overt physiological consequences of BCKDHB deficiency in the basal state, certain dietary stresses may magnify the symptoms and make the BXDs appropriate to model MSUD.

For Dhtkd1, the BXD strains map to highly significant and overlapping eQTLs and pQTLs with an approximately bimodal distribution with 2-fold variance between strains with the B6 (low) or the D2 allele (high) of the gene and with no effect from diet. Furthermore, the metabolite 2-AA maps as a significant mQTL to the Dhtkd1 locus in both diets, despite a marked decrease of 2-AA levels in HFD cohorts. Interestingly the decrease in 2-AA is not directly due to the HFD, but rather it is due to the increased penetrance of insulin resistance after HFD. Moreover, prior works show that external administration of 2-AA can improve insulin sensitivity and glucose response in mice (Wang et al., 2013), indicating that variants in Dhtkd1 and 2-AA levels may be driving the pathogenesis of diabetes, rather than being mere passive biomarkers of the disease. Strikingly, we replicated this association between 2-AA and high glucose levels in a recent and completely independent metabolomic data set collected in the Hybrid Mouse Diversity Panel (Ghazalpour et al., 2014). Furthermore, and clinically relevant, we demonstrated in a human population-based sample, Co-Laus, that diabetic patients likewise have lower urinary levels of 2-AA (Rueedi et al., 2014). Thus, 2-AA levels are consistently linked to the diabetes status in both mice and humans, and in the BXDs we can unequivocally attribute a causal part of this variance to variants in Dhtkd1. The BXD strains were necessary to bring the hypothesis to power a human study and also confirm that the mouse link has relevance for human disease. Together, these findings further validate the BXD mouse population as a model for human metabolic diseases (Andreux et al., 2012), and they indicate a critical need for multilayered measurements to effectively analyze complex systems, particularly within the context of the study of GXE.

It is worth stressing that novel regulatory mechanisms can be found either through QTL analysis and their equivalent from GWAS, SNP analysis, or through network analyses, which are a complementary and powerful approach to dissect complex traits. The network approach is particularly viable when backed by high-depth multilayered data sets such as illustrated by our example of UPR^mt. UPR^mt is a reparative pathway activated by mitochondrial proteotoxic stress that has been primarily studied in the C. elegans and in cultured cells, but little is known about whether it occurs in vivo in mammals. We examined six genes that are known to be major regulators of C. elegans UPR^mt and that are conserved in mammals. These six genes form a robust coexpression network in both diets at the transcriptional and proteomic levels, with the proteomic connections being stronger, befitting the role of UPR^mt as a sensor and regulator of protein stress. One observation that stood out in the analysis of the UPR^mt was the striking “contradiction” between the Ubl5 transcript and UBL5 protein correlations to the UPR^mt network. Ubl5 is a transcriptional regulator known to induce UPR^mt; yet, in both worms and mice, its transcript levels decrease when UPR^mt is activated. Conversely, the UBL5 protein is increased with UPR^mt activation in the BXDs, an observation also previously reported in C. elegans (Benedetti et al., 2006). This discordance in protein and transcript regulation suggests the existence of posttranscriptional mechanisms or a negative feedback loop, which could not be detected at the transcript or protein level alone. While there remains a great deal of this pathway left to be explored, it is clear that accurate, systems-scale protein measurements are essential to effectively model complex protein response networks like UPR^mt.

Today, there is an unprecedented capacity for accurate measurement quantification of nearly all aspects of biology. Surprisingly, perhaps, proteins remain one of the most difficult aspects of biology to precisely measure on a systems scale, though it has been long clear that transcripts only serve as a weak proxy for protein levels (Gygi et al., 1999). Our data show that the integration of systems proteomics data sets with different layered systems measurements provides unprecedented insights into the mechanistic regulation of complex systems, such as exemplified by metabolism, which can eventually lead to the improved diagnosis and treatment of metabolic disease.

EXPERIMENTAL PROCEDURES

Animals

Forty strains of the BXD population—40 on CD, 37 on HFD—with ~10 male animals from each strain were separated into 2 cohorts of 5 for each diet. Food access was ad libitum for CD—Harlan 2018 (6% kCal/fat, 74% carbohydrate, 20% protein)—and for HFD (Harlan 06414 [60% kCal/fat, 20% carbohydrate, 20% protein]). HFD cohorts received the diet from week 8 until sacrifice. Each cohort was communally housed until week 23, after which animals were single caged until tissues collection at 29 weeks after an overnight fast. Tissues were collected from 183 CD and 168 HFD animals, with at least 3 biological replicates for all cohorts. All research was approved by the Swiss cantonal veterinary authorities of Vaud under licenses 2257.0 and 2257.1.

Sample Preparation and Analysis

For liver analyses, three ~100 mg pieces were taken from cold storage for each individual and then weighed and sorted for mRNA, protein, and metabolite measurements. For microarray, mRNA from three to five individuals per cohort were prepared separately and then pooled equally after nanodrop quantification and run on the Affymetrix Mouse Gene 1.0 ST array platform. For proteomics, protein was prepared from one to three biological replicates per cohort. For metabolite measurements, all individuals in all cohorts were measured individually after polar extraction of samples.

SRM Assay Development and Protein Quantification

Generation of peptide library and development ofSRM assays were performed as described (Picotti et al., 2010). Identical SRM assays for all 192 target proteins were run on all 77 cohorts. SRM traces were manually checked according to established criteria (Lange et al., 2008a). For relative quantification of each protein across the set of different cohorts, the raw intensity of transitions of the native and (¹³C₆, ¹⁵N₄)-Arginine, (¹³C₆, ¹⁵N₂)-Lysine peptides were considered. The technical reproducibility of SRM-based quantification was validated by measuring the individual samples with three independent mass spectrometry injections.

General Informatic Analyses

Outliers were Windsorized prior to all analyses. SNP analysis was also performed using GeneNetwork, which contains the full sequence data of C57BL/6J and DBA/2J, as well as high density genotype information for all the BXD lines. The correlations are Pearson’s r or Spearman’s rho, depending on the absence or presence of outliers. Student’s t test was used for comparing two groups in normalized data (all protein and mRNA are normalized). Bonferroni’s correction for multiple testing was performed, and cutoffs for both nominal significance (p < 0.05) and corrected significance (p < 0.05/n tests) are displayed. Except for QTL plots, graphs and analyses were performed in R.

Metabolite Measurements

Murine metabolite measurements for this study were performed on serum and livers from the BXDs using ultra performance liquid chromatrography-mass spectrometry and flow-injection time-of-flight mass spectrometry, respectively. Human urinalysis was performed on 835 Caucasian adults in the CoLaus study, including 43 diabetics and 792 nondiabetics, males and females. The CoLaus study was approved by the Institutional Ethics Committee of the University of Lausanne. All study participants gave written consent. Full details of the human NMR metabolomics data are in a separate parallel publication (Rueedi et al., 2014).

Data Repository

All BXD data can be found at http://www.GeneNetwork.org. To download the clinical phenotype data, change the type to “Phenotypes” and enter “LISP2” to find all associated results. Select all and export to recover the data, or analyze it online. Microarray data and protein measurements can be found on the same resource; change the type to “Liver Proteome” or “Liver mRNA” and navigate to the named EPFL data sets under CD or HFD (e.g., EPFL/LISP BXD CD Liver Affy Mouse Gene 1.0 ST [Apr13] RMA). To download these data, please click the “INFO” button at the right-hand side of the search page and download the raw data set in the upper right box of the screen.

Extended Experimental Procedures and Supplemental References are available in the Supplemental Information.