Metabolic network-based predictions of toxicant-induced metabolite changes in the laboratory rat



Male Sprague-Dawley rats, 10 weeks of age, were purchased from Charles River Laboratories (Wilmington, MA). The rats were fed with Formulab Diet 5001 (Purina LabDiet; Purina Miles, Richmond, IN) and were given water ad libitum in an environmentally controlled room, set at 23 °C and on a 12:12-h light-dark cycle. All experiments were conducted in accordance with the Guide for the Care and Use of Laboratory Animals of the United States Department of Agriculture and the National Institutes of Health, and all protocols were approved by the Vanderbilt University Institutional Animal Care and Use Committee, and the U.S. Army Medical Research and Materiel Command Animal Care and Use Review Office. The investigators adhered to the Animal Welfare Act Regulations and other Federal statutes relating to animals and experiments involving animals.

Experimental Design

Surgery for implanting the catheters was performed 7 days before each experiment as previously described45. Rats were anesthetized with isoflurane. For studies to determine the appropriate APAP dose and exposure time and those to measure changes in gene expression and plasma metabolite profiles, the right external jugular vein was cannulated with sterile silicone catheters (0.51 mm inner diameter [ID]/0.94 mm outer diameter [OD]). For studies to measure metabolic flux, the carotid artery and the right external jugular vein were cannulated with sterile silicone catheters (0.51 mm ID/0.94 mm OD). The free end of the catheter was passed subcutaneously to the back of the neck where it was fixed. The catheter was occluded with a metal plug following a flush of heparinized saline (200 U heparin/ml). After surgery, rats were housed individually.

Preliminary studies for determining appropriate dose and exposure time

Two days before each study, the rats were moved from their regular housing cages to metabolic cages (Harvard Apparatus, Holliston, MA). To determine the appropriate dose and exposure time of APAP, they were treated with either vehicle (6 ml/kg of 50% polyethylene glycol, n = 6) or either 1 g/kg (n = 6) or 2 g/kg (n = 7) of APAP at 7 a.m. by gavage. Blood and accumulated urine were collected at 7 a.m. and 5 p.m. daily for 3 days.

Studies for measuring changes in gene expression and plasma metabolite profiles

We chose 2 g/kg as the appropriate APAP dose and two exposure times, one short (5 h, n = 8) and the other long (10 h, n = 8), based on the results of the dose-response study (Fig. 1a,b). Following blood collection, animals were given either vehicle or APAP by gavage at 7 a.m. and moved to new housing cages, where they could access water ad libitum but were not given food. At 12 p.m. (5-h group) or 5 p.m. (10-h group), after blood was collected from each group, animals were anesthetized by an intravenous injection of sodium pentobarbital through the jugular vein catheter and a laparotomy was performed immediately. The liver was dissected and frozen using Wollenberger tongs precooled in liquid nitrogen. The collected plasma was kept in a −80 °C freezer prior to analyses.

Studies for measuring metabolite flux

For flux measurements, at 7 a.m. on the day of the study, rats were administered either APAP (2 g/kg, n = 8) or vehicle (50% polyethylene glycol, 6 ml/kg, n = 8) by oral gavage, and food and water were subsequently removed. At 12:50 p.m., they were anesthetized with isoflurane, and following collection of 200 µl of arterial blood through the carotid artery catheter to determine the natural isotopic abundance of circulating glucose, a bolus of [2H]2water (99.9%) was delivered subcutaneously to enrich total body water to 4.5%. At 1 p.m. (i.e., 6 h after dosing), after they had recovered from anesthesia, the rats were placed in bedded containers without food or water and connected to sampling and infusion lines. A prime-constant infusion of [6,6–2H2]glucose (80 mg/kg prime + 0.8 mg/kg/min infusion) was administered into the systemic circulation through the jugular vein catheter for the duration of the study. Sodium [13C3]propionate (99%) was delivered as a prime-constant infusion (110 mg/kg + 5.5 mg/kg/min infusion) starting 120 min after the [2H]2water bolus. All infusates were prepared in a 4.5% [2H]2water-saline solution unless otherwise specified. Stable isotopes were obtained from Cambridge Isotope Laboratories (Tewksbury, MA). Blood glucose was monitored (AccuCheck, Roche Diagnostics, Indianapolis, IN) and donor erythrocytes were infused to maintain hematocrit throughout the study. Three blood samples (300 μl each) were collected over a 20-min period following 100 min of [13C3] propionate infusion. Arterial blood samples were centrifuged in EDTA-coated tubes for plasma isolation, and the three 100 μl plasma samples were stored at −20 °C prior to glucose derivatization and gas chromatography-mass spectrometry (GC-MS) analysis. Rats were rapidly euthanized through the carotid artery catheter immediately after the final steady-state sample.

Preparation of glucose derivatives

Plasma samples were divided into three aliquots and derivatized separately to obtain di-O-isopropylidene propionate, aldonitrile pentapropionate, and methyloxime pentapropionate derivatives of glucose. For di-O-isopropylidene propionate preparation, proteins were precipitated from 20 µl of plasma using 300 µl of cold acetone, and the protein-free supernatant was evaporated to dryness in screw-cap culture tubes. Derivatization proceeded as described previously46 to produce glucose 1,2,5,6-di-isopropylidene propionate. For aldonitrile and methyloxime derivatization, proteins were precipitated from 10 µl of plasma using 300 µl of cold acetone, and the protein-free supernatants were evaporated to dryness in microcentrifuge tubes. Derivatizations then proceeded as described previously46 to produce glucose aldonitrile pentapropionate and glucose methyloxime pentapropionate. All derivatives were evaporated to dryness, dissolved in 100 µl of ethyl acetate, and transferred to GC injection vials with 250-µl glass inserts for GC-MS analysis.

Measurement of tissue injury markers in blood

Plasma levels of ALT and AST were measured using ALT and AST activity assay kits (Sigma-Aldrich, St Louis, MO), respectively.

GC-MS analysis

GC-MS analysis was performed using an Agilent 7890 A GC system with an HP-5 ms (30 m × 0.25 mm × 0.25 μm; Agilent J&W Scientific) capillary column interfaced with an Agilent 5975 C Mass Spectrometer. Samples were injected into a 270 °C injection port in splitless mode. Helium flow was maintained at 0.88 ml∙min−1. For analysis of di-O-isopropylidene and aldonitrile derivatives, the column temperature was held at 80 °C for 1 min, ramped up at 20 °C∙min−1 to 280 °C and held for 4 min, then ramped up at 40 °C∙min−1 to 325 °C. For methyloxime derivatives, the same program was used except the ramp up to 280 °C was 10 °C∙min−1. After a 5 min solvent delay, the mass spectrometer collected data in scan mode from m/z 300 to 320 for di-O-isopropylidene derivatives, m/z 100 to 500 for aldonitrile derivatives, and m/z 144 to 260 for methyloxime derivatives. Each derivative peak was integrated using a custom MATLAB function47 to obtain mass isotopomer distributions (MIDs) for six specific ion ranges: aldonitrile – m/z 173–177, 259–265, 284–288, 370–374; methyloxime – m/z 145–149; di-O-isopropylidene – m/z 301–308. To assess uncertainty, the root mean squared error was calculated by comparing the baseline MID of unlabeled glucose samples with the theoretical MID computed from the known abundances of naturally occurring isotopes.

2H/13C metabolic flux analysis (MFA)

The in vivo MFA methodology employed in these studies has previously been described in detail48. Briefly, a reaction network was constructed using the INCA software package49 ( This network defined the carbon and hydrogen transitions for biochemical reactions linking hepatic glucose production and associated intermediary metabolic reactions. The flux through each reaction was estimated relative to citrate synthase (fixed at 100) by minimizing the sum of squared residuals between the simulated and experimentally determined MIDs of the six fragment ions previously described. Flux estimation was repeated 25 times from random initial values. Goodness-of-fit was assessed by the chi-square test, and 95% confidence intervals were computed by evaluating the sensitivity of the sum-of-squared residuals to variations in flux values50. The average SSR of each experimental group fell within the 99% confidence interval of the corresponding chi-square distribution with 22 degrees of freedom (i.e., the regressions were overdetermined by 22 measurements). Control SSR: 29.65 ± 7.05; APAP SSR: 41.87 ± 2.47. 99% CI = [8.6, 42.8]. Relative fluxes were converted to absolute values using the known [6,6-2H2]glucose infusion rate and rat weights. Flux estimates for the steady-state samples were averaged to obtain a representative set of values for each rat.

Metabolomic analysis

Sample preparation was carried out at Metabolon Inc. (Durham, NC), in a manner similar to a previous study51. Briefly, individual samples were subjected to methanol extraction and then split into aliquots for analysis by ultrahigh performance liquid chromatography/MS (UHPLC/MS). The global biochemical profiling analysis comprised four unique arms, consisting of reverse-phase chromatography positive ionization methods optimized for hydrophilic compounds (LC/MS Pos Polar) and hydrophobic compounds (LC/MS Pos Lipid), reverse-phase chromatography with negative ionization conditions (LC/MS Neg), as well as a hydrophilic interaction liquid chromatography (HILIC) method coupled to negative ionization (LC/MS Polar)52. All methods alternated between full scan MS and data-dependent MSn scans. The scan range varied slightly between methods but generally covered 70–1000 m/z.

Metabolites were identified by automated comparison of the ion features in the experimental samples to a reference library of chemical standard entries that included retention time, molecular weight (m/z), preferred adducts, and in-source fragments as well as associated MS spectra, and curated by visual inspection for quality control using software developed at Metabolon. Identification of known chemical entities was based on comparison to metabolomic library entries of purified standards53.

Two types of statistical analyses were performed: 1) significance tests and 2) classification analysis. Standard statistical analyses were performed in ArrayStudio on log‐transformed data. The R program (http://cran.r‐ was used for non-standard analyses. Following log transformation and imputation of missing values, if any, with the minimum observed value for each compound, Welch’s two-sample t-test was used to identify biochemicals that differed significantly (p < 0.05) between experimental groups. An estimate of the FDR (q‐value) was calculated to take into account the multiple comparisons that normally occur in metabolomics‐based studies.

RNA isolation and sequencing

Frozen whole liver was powdered in liquid nitrogen. Total RNA was isolated from the liver using TRIzol Reagent (Thermo Fisher Scientific, Waltham, MA) and the direct-zol RNA MiniPrep kit (Zymo Research, Irvine, CA). The isolated RNA samples were then submitted to the Vanderbilt University Medical Center VANTAGE Core (Nashville, TN) for RNA quality determination and sequencing. Total RNA quality was assessed using a 2100 Bioanalyzer (Agilent, Santa Clara, CA). At least 200 ng of DNase-treated total RNA with high RNA integrity was used to generate poly-A-enriched mRNA libraries, using KAPA Stranded mRNA sample kits with indexed adaptors (Roche, Indianapolis, IN). Library quality was assessed using the 2100 Bioanalyzer (Agilent), and libraries were quantitated using KAPA library Quantification kits (Roche). Pooled libraries were subjected to 75-bp single-end sequencing according to the manufacturer’s protocol (Illumina HiSeq3000, San Diego, CA). Bcl2fastq2 Conversion Software (Illumina) was used to generate de-multiplexed Fastq files.

Analysis of RNA-seq data

We analyzed RNA-seq data with Kallisto, a recently developed RNA-seq data analysis tool for read alignment and quantification. Kallisto pseudo-aligns reads to a reference, producing a list of transcripts that are compatible with each read while avoiding alignment of individual bases54. In this study, we pseudo-aligned the reads to the rat transcriptome downloaded from the Kallisto web-site ( Kallisto achieves a level of accuracy similar to that of other methods but is orders of magnitude faster; this allows calculation of the uncertainty of transcript abundance estimates, via the bootstrap technique of repeating analyses after resampling with replacement from the data. Here we employed bootstrapping by repeating analyses 100 times with resampling for each data set. Considering that the average number of reads per data set is 35 million (25 to 51 million single-end reads), using other software tools to perform the same bootstrap analysis becomes prohibitively expensive.

To identify DEGs from transcript abundance data quantified by Kallisto, we used the companion tool Sleuth, which uses the results of the bootstrap analysis during transcript quantitation to estimate the technical variance directly for each sample55. Many software tools for differential gene expression analysis of RNA-seq experiments assume that the technical variance of gene counts follows a Poisson distribution, in which the variance equals the mean56. However, for many genes, the technical variance can be much higher than the expected Poisson variance57. A distinct advantage of Sleuth is that it models biological and technical variances explicitly with a response error model.

To understand the biological significance of the lists of genes whose expression levels were altered by APAP exposure, we used the DEGs derived from Kallisto-Sleuth analyses and identified significantly altered DEGs that were mapped to the rat GENRE, and used KEGG pathways to identify molecular pathways that were significantly enriched. We used the online tool Database for Annotation, Visualization, and Integrated Discovery (DAVID)27 to perform this task.

Rat GENRE and model curation

We reconstructed a functional rat GENRE (iRno), using orthology annotations from genes in the Human Metabolic Reaction 2 (HMR2) database58, and manually reconciled several reactions by referring to the experimental literature and annotation databases16. The developed model, which contains 2,324 genes and 5,620 metabolites in 8,268 reactions connected by GPR rules, was validated for simulating 327 liver-specific metabolite functions successfully representing liver metabolism16. In this work, we further updated the iRno by incorporating new reactions or modifying some of the existing reactions based on experimental evidence (Supplementary Table S8). For example, although the pyruvate kinase reaction (EC: was reported as a reversible reaction in the original model, the Gibb’s free energy of the reaction under physiological conditions suggests that it is irreversible59. Similarly, for the heme:oxygen oxidoreductase reaction (EC:, we corrected the substrates and stoichiometry of the reaction components for consistency60. Using the metabolite evidence from the global metabolic profiling of plasma samples in the current study, we added 90 transport and 105 exchange reactions to the original model to increase the number of metabolites mapping to the data. We provide the updated iRno model with these modifications in Supplementary Table S9.

Boundary conditions for iRno in the fasting state

Our experimental design involved subjecting rats to APAP treatment under fasting conditions to maintain similar weight loss in the control and treatment groups. During fasting, the liver takes up gluconeogenic substrates, such as amino acids (AAs), lactate, and glycerol, to produce blood glucose, urea, and ketone bodies and takes up free fatty acids (FAs) for energy maintenance. Thus, the input fluxes (uptake rates) to our model are those of 1) AAs, 2) lactate, and 3) FAs and glycerol. The output fluxes (secretion rates) are those of 1) glucose derived from glycogenolysis and gluconeogenesis, as well as 2) urea and ketone bodies.

Multiple studies have used rats fasted overnight to deplete glycogen and measured input (AAs and lactate) and output (urea and ketone bodies) fluxes, using liver perfusion and in situ MFA61,62,63,64,65,66. In these studies, sham-treated control animals subjected to fasting conditions showed significant uptake rates of AAs and lactate with subsequent production of glucose, urea, and ketone bodies62. Furthermore, most of these studies measured the uptake/secretion rates from rats fasted for about 24 h and evaluated the metabolic state of the liver under ex vivo perfusion conditions61,62,64,65,66. We noted considerable inconsistency in the uptake rates reported in these studies (Supplementary Table S5). In contrast, the study by Izamis et al.63 measured uptake/secretion rates from rats fasted overnight and used metabolite concentrations and flow rates in the major vessels entering and leaving the liver under in situ conditions to evaluate the metabolic state of the liver. These conditions were similar to those in our experimental design. Thus, we used the majority of the approximated uptake and secretion rates derived from the study by Izamis et al. to constrain the respective input and output conditions for simulating metabolite alterations. In doing so, we strictly enforced the values for all of the uptake rates by constraining the lower and upper bounds in our model, while we constrained the values for the secretion rates only in terms of the lower bounds. We provide a detailed summary of the uptake and secretions rates from these studies in Supplementary Table S5.

Transcriptionally inferred metabolic biomarker response (TIMBR) algorithm for metabolite predictions

TIMBR is a novel method used for predicting toxicant-induced perturbations in metabolites by integrating gene expression changes into GENREs16. Briefly, it converts log2 fold changes of all DEGs into weights (W) for each of the GPRs in the GENRE. These reaction weights are then transformed into larger (or smaller) weights to represent relative levels of expression between the control and toxicant-treated conditions. TIMBR then calculates the global network demand required for producing a metabolite (Xmet) by minimizing the weighted sum of fluxes across all reactions for each condition and metabolite, so as to satisfy the associated mass balance and an optimal fraction of maximum network capability (vopt) to produce a metabolite as follows16 (see ref.16 for details):



where W denotes the vector representing the reaction weights, v is a vector of reaction fluxes, and S is the stoichiometric matrix. We integrated the aforementioned boundary conditions for uptake and secretion rates into the algorithm by fixing the respective lower (vlb) and upper bounds (vub) of the exchangeable reactions (vex) in the model (Eq. 2). Similarly, we integrated measurements from the 13C-labeled tracer studies for some of the central carbon metabolism fluxes into the TIMBR algorithm by constraining the lower and upper bounds of the respective reactions in the model (vmfa) (Eq. 3).

Using this method, we determined the relative production scores for all metabolites (Xraw) from control (Xcontrol) and toxicant-treated (Xtreatment) conditions (Eq. 4), and then calculated the TIMBR production scores (Xs) as the z-transformed scores across all exchangeable metabolites (Eq. 5).



The schematic in Fig. 6 depicts the overall integration strategy. More detailed descriptions of the TIMBR algorithm and the corresponding codes are available in the original publication16.

We used the experimental log2 fold changes of significantly altered (FDR < 0.10) plasma metabolites from the global metabolic profiling data (Supplementary Table S3) and then compared the corresponding iRno model predictions under no MFA and MFA conditions 5 or 10 h after APAP treatment (Supplementary Tables S4 and S7). Here, the model predictions of altered metabolite levels were considered as having increased or decreased based on TIMBR production score cut-off values of greater than 0.1 and less than −0.1, respectively. Metabolites with scores that were between −0.1 and 0.1 were considered as unchanged.

Data and code availability

Normalized gene expression data from the RNA-seq analysis and genes mapped to the iRno model are provided in Supplementary Table S1. The results of KEGG pathway enrichment analysis using the mapped genes are provided in Supplementary Table S2. The results from global metabolic profiling are provided in Supplementary Table S3. Metabolites mapped to iRno model are provided in Supplementary Table S4. The physiological boundary constraints required to simulate the metabolite predictions are provided in Supplementary Table S5. TIMBR predictions under random gene expression changes and addition of noise to the gene expression changes are provided in Supplementary Table S6. TIMBR predictions (Figs 7 and S1) are provided in Supplementary Table S7. Details of the modifications made to the iRno model are provided in Supplementary Table S8. An Excel file of the updated iRno model is provided in Supplementary Table S9. Additional information required to reproduce the figures can be obtained via the code made available as part of this publication at Detailed explanations for TIMBR algorithm are available as part of the original TIMBR publication16 at

Source link


Please enter your comment!
Please enter your name here