Investigating the Mechanistic Basis for Hepatic Toxicity Induced by an Experimental Chemokine Receptor 5 (CCR5) Antagonist Using a Compendium of Gene Expression Profiles

Abstract

A compendium of hepatic gene expression signatures was used to identify a mechanistic basis for the hepatic toxicity of an experimental CCR5 antagonist (MrkA). Development of MrkA, a potential HIV therapeutic, was discontinued due to hepatotoxicity in preclinical studies. Rats were treated with MrkA at 3 dose levels (50, 250, and 500 mg/kg) for 1, 3, or 7 days. Hepatic toxicity (vacuolation, consistent with steatosis, and elevated serum transaminase levels) was observed at 250 and 500 mg/kg, but not at 50 mg/kg. Hepatic gene expression profiles were compared to a compendium of hepatic expression profiles. MrkA was similar to 3 β-oxidation inhibitors (valproate, cyclopropane carboxylate, pivalate), 8 PPARα agonists (fenofibrate, bezafibrate and 6 fibrate analogues), and 3 other diverse compounds (diethylnitrosamine, microcystin LR & actinomycin D). These data indicate MrkA to be a mitochondrial inhibitor, and activation of PPARα-regulated transcription was thought to be due to an accumulation of endogenous ligands. While mitochondrial inhibition was likely responsible for steatosis, canonical pathway analysis revealed that progression to liver injury may be mediated by activation of the innate immune system primarily through NF-kB pathways. These results demonstrate the utility of a gene expression response compendium in developing transcriptional biomarkers and identifying the mechanistic basis for toxicity.

Introduction

Hepatotoxicity remains the most common reason for non-approval, warning labels and market recalls for pharmaceuticals (Senior, 2001; Temple, 2001; Lee, 2003). Drugs that have been withdrawn from the market due to unacceptable risk of hepatotoxicity include troglitazone and bromfenac, and drugs that have warnings or are second line due to liver toxicity include isoniazid, tolcapone, trovafloxacin, labetalol, diclofenac, valproate, ketoconazole, and acetaminophen among others. Despite its prevalence and often severe impact on patient health and the pharmaceutical industry, accurate prediction of hepatotoxicity is still difficult at best.

Analysis of available data indicates that liver toxicity is one of the leading target organ effects observed in preclinical testing. For example, of 31 compounds terminated due to toxicity in humans, 31% (9 compounds) were due to liver toxicity (Ballet, 1997, 2001). According to Ostapwicz et al., drug-induced toxicity remains one of the leading causes of acute liver failure in humans (2002). In a study of 308 acute liver failure patients, 52% were putatively due to adverse drug reactions, with 75% of drug-induced liver failures due to a single drug, acetaminophen, while the remainder included bromfenac, troglitazone, isoniazid, and other drugs. Even more problematic is that about one-third of drugs that produce liver toxicity in humans were negative for liver toxicity in animals (Lumley, 1990; Ballet, 1997, 2001). In this regard, one study showed that 55% of compounds that were in development and caused liver changes in animals were negative for liver damage in clinical trials (false positives) while 35% of compounds that were negative for liver damage in animals caused liver toxicity in human trials (false negatives) (Ballet, 2001).

Reducing or ameliorating the risk for hepatic toxicity in the clinic requires a greater predictive value from preclinical animal models, which in turn requires an understanding of the mechanistic basis of toxicity. While the concordance between preclinical species and humans for several forms of toxicity is reasonable (Olson et al., 2000), the ability to identify specific mechanisms in rodents that may constitute occult risks is poor. In particular, mitochondrial impairment may or may not produce frank toxicity in rodents, but it is recognized as a risk factor for idiosyncratic responses in humans. For example, the experimental anxiolytic panadiplon did not produce toxicity in typical preclinical species but subsequently produced hepatic toxicity in clinical trials. Subsequent studies revealed the rabbit to be sensitive to toxicity by panadiplon, and also revealed the mechanistic basis to be mitochondrial impairment (reviewed in Ulrich et al., 2001). However, mitochondrial impairment alone likely does not produce liver toxicity, since mitochondrial β-oxidation inhibitors such as aspirin and valproate are typically well tolerated. Thus, toxicity appears to require the presence of additional risk factors as forms of secondary stress.

Identification of additional risk factors is thus important, particularly when such risk factors are not obvious from histology or other traditional biomarkers. One approach that has been proposed to help identify such risk factors and thus improve the predictive value of animal toxicity testing is gene expression profiling (Ulrich and Friend, 2002). Many have suggested the use of expression profiles to identify mechanisms of toxicity and to classify or predict toxicity, and some have demonstrated its utility (Amin et al., 2002; de Longueville et al., 2004; Hamadeh et al., 2001; Hamadeh et al., 2002a; Hamadeh et al., 2002b; Hu et al., 2000; Steiner et al., 2004; Thomas et al., 2001; Waring et al., 2003; Waring et al., 2002). Expression profiling can link together various compounds with a similar mechanistic basis for toxicity, including mitochondrial impairment (Jolly et al., 2004).

In this study, we examined the effects of a selective chemokine (C-C motif) receptor 5 (CCR5) antagonist, MrkA, on hepatic gene expression in the rat. MrkA (1, N-[3-[4-(3-benzyl-1-ethyl-1H-pyrazol-5-yl)piperidin-1-yl]methyl- 4-(3-fluorophenyl)cyclopentyl]-N-methylvaline, Figure 7) was under development for HIV therapy but was discontinued due to toxicity in preclinical species. Previous studies have shown liver toxicity in rats as evidenced by hepatic steatosis and elevated serum transaminases, and metabolomic and proteomic investigations demonstrated mitochondrial impairment as a likely mechanistic basis (Meneses-Lorente et al., 2004; Mortishire-Smith et al., 2004). Since mitochondrial dysfunction is of particular concern for HIV therapeutics (Miro et al. 2005), the aim of our investigation was to determine if expression profiling could reveal clues regarding this mechanism and in particular to detect potential pathways by which mitochondrial impairment and steatosis can progress to produce liver toxicity. Additionally, this study gave us the opportunity to test the utility of a liver-specific gene expression database for this purpose. For this, we compared the expression profile of MrkA in rat liver to a previously constructed database of expression profiles induced by a large variety of hepatotoxicants (i.e., a compendium of expression profiles).

Materials and Methods

Animal Treatments

Male rats [Crl: CD^® (SD) IGS BR] (Charles River Laboratory) were obtained at 8–10 weeks of age and were acclimatized for at least 1 week. Animals were housed individually in suspended wire-mesh cages on a 12-hour light/dark cycle and were given ad libitum access to water and Lab Diet Certified Rodent Diet #5002 (PMI International, Inc).

The test compound (MrkA, Figure 7) was dissolved/suspended in the vehicle (0.25% methylcellulose, viscosity = 400 centipoise at 2% in H₂0, Sigma) at a concentration that would allow daily doses of 10 mL suspended compound/kg body weight. Animals were dosed daily by oral gavage at 3 dose levels (50, 250 & 500 mg/kg) for 1, 3, and 7 days (n = 3 animals/dose level • time point). Control animals were treated daily with vehicle for the same amount of time (n = 6 animals/time point).

After an overnight fast, animals were euthanized by carbon dioxide inhalation 24 hours after the last dose. Following euthanasia, blood was collected by cardiac puncture. The left lateral lobe of the liver was snap-frozen in liquid nitrogen while the remainder of the liver was fixed in 10% neutral buffered formalin. Standard clinical chemistry (sodium, potassium, chloride, calcium, phosphorus, alkaline phosphatase, total bilirubin, bile acids, γ-glutamyltransferase (GGT), aspartate aminotransferase (AST), alanine amino-transferase (ALT), sorbitol dehydrogenase, creatinine kinase (CK), urea nitrogen, creatinine, total protein, albumin, globulin, triglycerides, cholesterol, and glucose) and hematology parameters were measured, and pathological changes were identified by examining H&E-stained paraffin sections.

A compendium of liver gene expression responses to treatment with 65 chemicals was assembled from multiple studies (Table 1, Waring et al., 2003; Cornwell et al., 2004; Jolly et al., 2004). In addition to the 2 short chain fatty acids described in Jolly et al., 4 other short chain fatty acids (4-pentenoic acid, 20 mg/kg; n-octanoic acid, 700 mg/kg; pivalate, 500 mg/kg; propionic acid, 700 mg/kg) were examined in the same paradigm. Also included in this compendium from the present study were animals treated with an experimental mixed PPARα/PPARγ agonist (MRL6; 5, 30, and 75 mg/kg). For all compounds in this compendium, Sprague–Dawley rats (6–15 weeks of age, 225–300g) were treated once daily at a minimum of one dose level for 3 days. For some compounds, other dose levels or time points were available.

The in-life portion of this study and the RNA extraction procedure (see below) were completed at MPI Research (Mattawan, MI). MPI Research is fully accredited by the Association for Assessment and Accreditation of Laboratory Animal Care (AAALAC) International. The animal study protocol was reviewed and approved by the MPI Institutional Animal Care and Use Committee (IACUC) and the Merck Research Laboratories IACUC; and was conducted according to The Guide for the Care and Use of Laboratory Animals (National Research Council, 1996).

RNA Extraction

RNA was extracted from liver by a method that combined TRIzol RNA extraction (Invitrogen Life Technologies, Carlsbad, CA, USA) with the RNeasy RNA extraction kit (Qiagen, Valencia, CA, USA). Tissue was incubated in TRIzol reagent (1 mL/100 mg tissue) for 15 seconds at room temperature, homogenized with a Polytron homogenizer followed by a ~5-minute room temperature incubation. After the addition of 100 μL chloroform, 500 μL of homogenate was mixed by shaking for 15 seconds, incubated at room temperature for 2–3 minutes and centrifuged at 10,000 ×g for 10 minutes at 2–8°C. The supernatant was used as the input material for the RNeasy RNA extraction kit, and RNA was isolated according to the manufacturer’s protocol. Following isolation, RNA quantity, purity and quality were determined using a SpectraMax Plus384 (Molecular Devices, Sunnyvale, CA, USA) spectrophotometer and a 2100 Bioanalyzer (Agilent Technologies, Palo Alto, CA, USA).

Expression Profiling

Expression profiling was done using custom arrays consisting of 22,505 60mer oligonucleotides (plus control sequences) representing rat genes (available in the Gene Expression Omnibus (GEO): 〈http://www.ncbi.nlm.nih.gov/projects/geo/query/acc.cgi? acc=GPL2050〉). The arrays were synthesized by Agilent Technologies (Palo Alto, CA, USA) using an inkjet printing method (Hughes et al., 2001). The 60mer probes were selected as described in Waring et al. (2003). Cy3 or Cy5 labeled cRNA was created from total RNA using reverse transcription (RT) followed by in vitro transcription (IVT) and a 2-step label incorporation method (Hughes et al., 2001). All treated individual samples were hybridized against a pool of RNA from time matched (concurrent) control animals (e.g., animals treated for 3 days were hybridized against a pool of animals treated for 3 days with vehicle). All hybridizations (Hughes et al., 2001) were performed in duplicate, with fluor reversal (Cy3 or Cy5) in the second hybridization.

The resultant fluor-reversed pairs were combined to give a single ratio measurement for each gene for each treated liver. The ratio of individual animal expression to control pool expression was used for all data analysis. In addition, a subset of the vehicle treated animals was hybridized against the control pool in a manner identical to the compound treated samples. Arrays were scanned using a DNA microarray scanner (Model G2565AA; Agilent Technologies, Palo Alto, CA, USA), and feature intensities (background subtracted) were determined using feature extraction software developed at Rosetta (Qhyb, Marton et al. 1998). All hybridizations described herein passed a set of quality control criteria that evaluate feature quality/spot success rate (at least 90% of spots must pass metrics assessing handling or array synthesis problems), ratio reproducibility (intrachip standard deviation of the log₁₀ (ratio) <0.0607 and the standard deviation of the log₁₀ (ratio) <0.0792 within chip pairs), ratio accuracy (mean observed intrachip ratio not biased by more than 50% from expected ratio) and ratio sensitivity (at least 50% of transcripts spiked in at 1 part in 100,000 and at 1.5 copies/cell vs. 0.5 copies/cell (1:3) are within 50% of the expected ratio) based on the performance of 10 control transcripts spiked-in at known concentrations and ratios and present on each chip in approximately 30 locations dispersed evenly across the chip.

Analysis of RNA Expression

Genes that were significantly modulated between groups where no study-related changes would be expected (i.e., vehicle-treated groups) were identified by two methods. One method identified genes that were changed greater than 2-fold (p ≤ 0.01, Resolver error model) from the control pool in at least 1 control animal. In this case, where the expression of a gene in an individual animal is two fold or greater from a pool of RNA made up (in part) from RNA from that individual (along with the other animals from that group), the fold change between the individual animal in question and the other animals in the same group must be greater than two fold. This magnitude of variability may indicate that the expression of the gene in question is not tightly regulated; thus, analysis of genes in this category may lead to unreliable results. The other method treated the expression data from vehicle treated (i.e., control) animals from each day as an ANOVA group and identified genes that were differentially expressed between these groups (p ≤ 0.01, error weighted). Again, significant changes in the relative expression of genes between time points with only vehicle treatment identifies genes that are not tightly regulated and are possibly unreliable. These two gene sets were united (resulting in a set of 924 sequences) and excluded from all further analyses.

ANOVA was performed to identify genes that were changed by MrkA. For each time point, an ANOVA was done for each dose group vs. the control group from the appropriate day. Genes whose expression in treated animals was significantly different from control (p < 0.001, error weighted) were analyzed further. Genes whose expression appeared to trend with dose were also identified. For directly related trends, a gene was included if its expression was never lower at a given dose compared to lower doses at the same time point. For inversely related genes, a gene was included if its expression was never higher at a given dose compared to lower doses at the same time point. Genes that were included in the intersection of the ANOVA and trend lists will be referred to as “dose related” genes. Cluster diagrams of various gene sets were created using several cluster algorithms (e.g., K-means, divisive) and various similarity metrics (e.g., Euclidean distance, cosine correlation).

The expression profiles of all replicate animals within a treatment group (i.e., same dose and time point) were combined using an error-weighted average to give a single expression profile for each treatment group. These expression profiles from animals treated with a MrkA (as well as from animals treated only with the vehicle) were compared to the compendium of expression profiles described above and in Table 1 using the ROAST feature of Rosetta Resolver 4.0. ROAST is a tool for determining the correlation (and the significance level of that correlation) between two expression profiles. The gene set used for the comparison consisted of the union of the dose-related genes (method of identifying dose-related genes described above) from all time points. Significant correlations between profiles were defined as having a p-value less than the lowest p-value observed in a comparison of a vehicle-treated profile vs. any other profile.

Canonical pathways analysis identified the pathways from the Ingenuity Pathways Analysis (Ingenuity Systems, 〈http://www.ingenuity.com/〉) library of canonical pathways that were most significantly enriched in the data set. Subsets of the dose-related genes that were associated with a canonical pathway in the Ingenuity Pathways Knowledge Base were considered for the analysis. Fischer’s exact test was used to calculate a p-value determining the probability that the association between the genes in the dataset and the canonical pathway is explained by chance alone.

Rosetta Resolver Expression Data Analysis System 4.0 (Rosetta Biosoftware, an independent business unit of Rosetta Inpharmatics LLC, Seattle, WA, USA) was used for ANOVA, comparing profiles (Rosetta Array Search Tool (ROAST) analysis feature), creating and comparing sequence sets, and creating cluster diagrams. Custom PERL scripts were created to identify genes that trend with dose.

Results

Toxicology

Hepatic toxicity was observed for MrkA as evidenced by hepatocellular vacuolation and elevated serum transaminases. Hepatocellular vacuolation ranging from trace to moderate was observed in all animals treated at 500 mg/kg/day of MrkA for 1 and 3 days (Figure 1C and Figure 2C). No animals at this dose level survived to day 7. At the 250 mg/kg/day dose level, trace to moderate vacuolation was observed in all animals at all time points (Figure 1C and Figure 2B). In the 50 mg/kg/day dose level group, only trace subacute inflammation was observed at any time point; this was also observed at other dose levels (data not shown). Significant increases in serum transaminase levels (AST and ALT) accompanied vacuolation at the 500 mg/kg/day level (Figure 1B). Significantly increased liver to body weight ratios were observed for at least one time point in animals dosed at 250 and 500 mg/kg/day (Figure 1A), which were due to increased liver weight, not decreased body weight (Figure 1B). These results are consistent with the outcomes observed in previous studies of MrkA (Meneses-Lorente et al., 2004; Mortishire-Smith et al., 2004).

Expression Profiling

Hepatic total RNA from individual rats treated with MrkA (or vehicle) at various doses for various lengths of time was profiled against a pool of hepatic total RNA from animals treated with vehicle for the same length of time (i.e., time-matched controls) using a two-color system. The expression of genes whose expression was modulated at day seven at the lowest dose compared to the controls is shown in Figure 3 (1880 transcripts, ANOVA p ≤ 0.001). Using genes with a putative dose-related gene expression pattern in animals treated with MrkA (4357 transcripts, Figure 4), the expression profiles of animals treated with MrkA were compared to the expression profile of animals treated with compendium compounds (Table 1) to determine the degree of similarity (i.e., p-value). This degree of similarity was then identified as significant or not significant based upon the similarities observed in comparisons involving vehicle-treated animals (see Materials & Methods). It is expected that these similarities are more likely due to steatosis (whose occurrence more closely resembles the dosing paradigm, Figure 1C) than transaminase changes (which are not as similar to the dosing paradigm, Figure 1B).

A similarity score was calculated, where the numerator was the number of significant comparisons (see Materials and Methods section for the definition of significant) between MrkA and a given compound (e.g., valproate) and the denominator was the total number of comparisons between MrkA and the given compound. No distinction was made for different doses or time points (e.g. all dose levels of MrkA and the compound of interest were given equal weight in the numerator and denominator). This scoring system revealed significant similarities between MrkA-treatment and treatment with any one of 14 other compounds in the compendium (Figure 5). This list included 8 PPAR α agonists (fenofibrate, bezafibrate, and 6 fibrate analogues MRL1–6), 3 known inhibitors of mitochondrial β-oxidation (cyclopropane carboxylate (CPCA), valproate, pivalate), and 3 non-related hepatotoxins (diethylnitrosamine (DEN), actinomycin D, microcystin LR).

Two subsets of dose-related genes were identified from Figure 4. The genes in cluster six represented the genes whose expression was very different in MrkA treated animals versus PPARα treated animals but was similar to DEN, actinomycin D and microcystin. The genes in clusters 1–3, 8, and 10–12 represented genes whose expression was very similar between MrkA-treated animals and PPARα-treated animals. These 2 sets were used to generate 2 sets of relevant canonical pathways through the use of Ingenuity Pathways Analysis tools (see Materials and Methods) shown in Figure 6.

Discussion

In this study we found MrkA, a selective chemokine receptor 5 (CCR5) antagonist, to produce hepatic toxicity characterized by hepatocellular vacuolation, hepatocellular hypertrophy, increased serum transaminase levels and mortality in rats (Figure 1 and Figure 2). Our findings are consistent with previous studies in monkeys and rodents that demonstrated hepatocellular vacuolation and increases in serum transaminase levels (Meneses-Lorente et al., 2004; Mortishire-Smith et al., 2004) indicative of liver toxicity. Histological findings from our study and previously reported Oil Red-O staining (of liver sections from rates treated with MrkA, Mortishire-Smith et al., 2004) are consistent with microvesicular steatosis, and the observed lethal dose level was consistent across studies. In the study reported here we further examined the effects of MrkA on the hepatic transcriptome using custom oligonucleotide microarrays.

Genes regulated by MrkA treatment were first identified; then, the expression of those genes was compared to a compendium of hepatic expression profiles from animals treated with various compounds known to induce hepatic toxicity through a variety of mechanisms (Table 1). Significant similarities between several compendium compounds and MrkA were identified, and the frequency of occurrence of these significant overlaps was used to derive a similarity score.

Using the similarity score method, it was possible to identify which compounds were similar to the test compound, an approach that is more revealing than merely identifying the closest match or matches. If no compound in the compendium had a biological/transcriptional effect similar to MrkA, then no compound would show a significant similarity metric. Indeed, of the 65 compounds in the compendium used in our analysis, 14 had gene expression patterns more similar to MrkA than would be expected by chance (Figure 5). Of these, 8 were PPARα-agonists (including the fibric acid analogues MRL1–MRL6, fenofibrate and bezafibrate), 3 were known mitochondrial β-oxidation inhibitors (CPCA, pivalate, valproate) and 3 were hepatotoxicants from various apparently unrelated classes (DEN, a DNA damaging agent; microcystin, a mitochondrial permeability transition inducer; actinomycin D, a protein synthesis inhibitor).

A superficial interpretation of the similarity score would be that MrkA, in addition to being a CCR5 antagonist, is also primarily a PPARα agonist. All 8 PPARα agonists in the compendium showed similarity scores higher than any other compounds, and the PPARα agonists with the lowest scores are mixed PPAR agonists (i.e., MRL6 is a mixed PPARα/γ agonist and bezafibrate is a PPAR pan-agonist). Compounds with known inhibitory effects on mitochondrial β-oxidation (CPCA, pivalate and valproate) had the lowest similarity scores of the compounds with significant expression profile similarities. The greatest similarities were driven primarily by the relatively large and highly significant PPARα signature; however, it is unlikely that MrkA directly activates this transcription factor, since PPARα transactivation assays indicated that MrkA is not a PPARα ligand (data not shown).

The β-oxidation signature of MrkA (discussed later) produced a small but significant similarity with known mitochondrial β-oxidation inhibitors. CPCA (Duncombe and Rising, 1972; Buxton et al., 1983; Ulrich et al., 1998; Jolly et al., 2004;) and valproate (Baillie, 1992; Dixon et al., 1990; Kesterson et al., 1984; Kibayashi et al., 1999; Ponchaut et al., 1992; Spahr et al., 2001; Trost and Lemasters, 1996; Yao et al., 1994) have been studied extensively regarding their inhibition of β-oxidation and induction of fatty liver changes.

Pivalate decreases hepatic and cardiac carnitine levels (reviewed in Brass 2002), which would be expected to impair β-oxidation; however, only one study demonstrated decreased β-oxidation and only in the heart (liver was not studied, Broderick et al., 1995). Part of the similarity to PPAR agonists may have been driven by mitochondrial effects, as many PPAR ligands have also been observed to impair mitochondrial function independent of their receptor binding activities (Scatena et al., 2004a, 2004b). Previous studies on MrkA using metabolomic (Mortishire-Smith et al., 2004) and proteomic (Meneses-Lorente et al., 2004) tools also indicated this compound to be a mitochondrial β-oxidation inhibitor, consistent with our transcriptional analysis. Further, MrkA was shown to have histopathological and metabolic characteristics similar to MCAD deficiency (a deficiency in β-oxidation) (Mortishire-Smith et al., 2004).

It is thus likely that the similarity score outcome was driven proximally by inhibition of mitochondrial function, and we propose that the dominant PPARα signature for MrkA was driven by the resultant accumulation of endogenous ligands due to inhibition of β-oxidation. Such an accumulation is indicated by histological changes, as increased fatty acid levels due to steatosis have been observed to induce peroxisome proliferation in studies on other compounds (Desvergne and Wahli, 1999; Spaniol et al., 2003). Also, Meneses-Lorente et.al. observed an induction of several proteins involved in the production of acetyl-CoA and suggested that this was a homeostatic mechanism in reaction to decreased β-oxidation by MrkA (2004). Our data are consistent with this and concur with observations by others that metabolic changes associated with PPAR agonism can occur when β-oxidation is inhibited.

For example, Patel et al. observed decreased pyruvate dehydrogenase activity at some concentrations of propionate exposure in perfused rat livers (1983). This may be a result of the observed induction of pyruvate dehydrogenase kinase 4 (PDK4) by PPARα activation (Cornwell et al., 2004; Holness et al., 2003; Kramer et al., 2003). The protein inductions observed by Meneses-Lorente et al., (2004) and the increased transcript levels of several acyl-CoA dehydrogenases observed by Kibayahsi et al. (1999) could also be explained by increased PPARα agonism (Cornwell et al., 2004; Kramer et al., 2003).

While our data concur with previous observations (Meneses-Lorente et al., 2004; Mortishire-Smith et al., 2004) that MrkA inhibits mitochondrial β-oxidation, this activity alone does not explain the hepatic toxicity. While mitochondrial inhibitors, including the positive classifiers in this study, produce steatosis they typically do not directly cause liver damage. In fact, progression to liver damage for mitochondrial inhibitors such as valproate or aspirin is idiosyncratic and relatively rare. To understand what other effects of MrkA might directly contribute to liver injury, we more closely examined its similarity to unrelated compounds.

Three compounds in our compendium that are not characterized as mitochondrial fatty acid oxidation inhibitors (DEN, actinomycin D and microcystin LR) showed a strong similarity to MrkA. DEN, a genotoxic carcinogen, is typically used in a single dose paradigm with analysis of the results coming weeks or month later; however, there are several reports that chronic treatment results in hepatocellular vacuolation and steatosis (Braunbeck et al., 1992; Dunsford et al., 1989; Goldfarb et al., 1983; Ha et al., 2001). Most evidence points to mitochondrial permeability transition (MPT) and induction of apoptosis as the primary mechanism of microcystin LR-induced hepatotoxicity (Ding and Nam Ong, 2003; Ding et al., 2001); although, there are reports of this compound inducing lipid changes and/or steatosis (Guzman and Solter, 1999; Sturgeon and Towner 1999). Interestingly, microcystin LR is more toxic in fasted animals compared to fed animals (reviewed in Sturgeon and Towner, 1999). This effect was reported in two studies with MrkA as well (Meneses-Lorente et al., 2004; Mortishire-Smith et al., 2004). We did not find literature to support effects of actinomycin D on mitochondrial β-oxidation or hepatic steatosis (Akahori et al., 1999; Bader et al., 1974; Czauderna et al. 2000; Goldblatt et al., 1969; Kleeberg et al., 1979; Kuehl, 1969; McGeachin et al., 1978). Thus, the basis for the transcriptional similarity of MrkA to these compounds does not appear to center on mitochondrial effects.

Canonical pathway analysis of the gene expression signatures shared by MrkA and non-PPARα agonists (Figure 6A) revealed a potential mechanism by which mitochondrial inhibition can lead to liver injury. The most significantly affected pathway is NF-κB signaling, followed by B-cell signaling and a variety of other inflammatory pathways including IL-2 and IL-6. It is becoming increasingly apparent that hepatic steatosis increases the sensitivity to damage by activation of the innate immune system (a possible on-target effect of a CCR5 inhibitor). For example, fatty livers in ob/ob mice showed increased NF-κB activation and increased IL-6 following LPS stimulation (Romics Jr et al. 2004), where NF-κB was proinflammatory and PPARα activation was thought to be anti-inflammatory. The increased PPARα signaling we observed for MrkA and the non-PPAR compounds is consistent with this report.

In dietary-induced steatohepatitis (Dela Pena et al., 2005), NF-κB was also found to be proinflammatory independent of TNF-α. In humans, disease severity for nonalcoholic and alcoholic steatohepatitis appears to correlate with active NF-κB expression (Ribeiro et al., 2004). For MrkA, NF-κB may drive the progression from steatosis to steatohepatitis and necrosis in a manner analogous to that observed in ob/ob mice (Yang et al., 2001). The severity of the injury may also depend on the degree of mitochondrial inhibition, since mitochondrial inhibition elevates the levels of reactive oxygen species (ROS, reviewed in Begriche et al., 2006), and ROS can drive NF-κB-regulated gene transcription by inactivation of inhibitory κBs (reviewed by Wang et al., 2002).

What is not known is the extent to which this progression involves the intended therapeutic target, CCR5, since a CCR5 deficiency in mice can promote fulminant liver failure (Ajuebor et al., 2005) and hepatitis in a manner that appears to involve TNF-α (Moreno et al., 2005). Thus it is clear that, while the primary effect of MrkA is inhibition of mitochondrial β-oxidation, the effects leading to actual liver injury likely include an activation of the innate immune system driven at least in part by steatosis and the subsequent activation of NF-κB.

This study also demonstrates the utility of a gene response compendium for providing clues and knowledge regarding mechanism; however, our similarity scoring approach, though useful, did have apparent false negatives. Similarity scoring showed a relationship of MrkA to short-chain fatty acid β-oxidation inhibitors CPCA, valproate and pivalate, but the compendium also included profiles for butyrate, propionate, 4-pentenoate and the medium chain fatty acid n-octanoate (Table 1). Along with affecting pyruvate oxidation, propionate is a known inhibitor of β-oxidation (Brass and Beyerinck, 1988; Brass et al., 1986; Fedotcheva et al., 1991, 1993; Jesse et al., 1986). There are reports that 4-pentenoate (pent-4-enoate) inhibits fatty acid oxidation and induces microvesicular steatosis (Billington et al., 1978; Dixon et al., 1990; Kesterson et al., 1984; Ponchaut et al., 1992; Sugimoto et al., 1990; Tang et al., 1995).

While it is not clear that butyrate or n-octanoate have any significant effect on fatty acid metabolism, it does appear that propionate and 4-pentenoate are false negatives in this analysis. These results could be explained by the fact that, in these experiments, 4-pentenoate and propionate produced very weak gene expression signatures (Figure 4), perhaps due to insufficient exposure in these experiments. Conversely, the octanoate signature (Figure 4) appears to be similar to the MrkA signature by visual inspection despite having a similarity score of zero (Figure 5). This illustrates the risk of relying on visual inspection to identify similarity: similarities gleaned from visual inspection rely on the proper ordering of the genes while statistical methods like ROAST^® do not.

It is possible that a reordering of the genes would reveal differences in the gene expression profiles that were not previously apparent from visual inspection, while it is equally possible that octanoate may indeed have previously undiscovered effects on β-oxidation. Perhaps another explanation of the false negatives for 4-pentenoate and propionate is based upon rejecting the assumption that inhibition of β-oxidation is via a single mechanism and results in a single expression profile. Our data is not in agreement with that assumption.

In fact, our data suggest the existence of at least 2 distinct methods of β-oxidation inhibition with 2 independent transcript profiles. Consistent with this hypothesis is that the fatty acids in our compendium and MrkA can be divided into 2 somewhat different classes based upon their chemical structure at the α-carbon. Pivalate, CPCA, valproate and MrkA are all branched at the α-carbon while propionate, butyrate, n-octanoate and 4-pentenoate are all straight-chain fatty acids with no branching (Figure 7). This may indicate that α-branched carboxylic acids inhibit β-oxidation by one mechanism while 4-pentenoate and propionate affect fatty acid metabolism by a different mechanism.

One significant challenge encountered when using a compendium is how to identify the appropriate set of genes to use in comparing expression profiles; this set may number in the thousands to as few as a dozen (Thomas et al., 2001). Our initial method used the signature of the test compound (MrkA) to compare profiles, as an unsupervised use of data should allow for the identification of similarities representing mechanisms of toxicity not previously considered. In contrast, an argument can be made for identifying a toxicity or mechanism-specific set of genes based upon the signatures of profiles in the compendium of responses. Repeating our comparisons of MrkA to the compendium using a signature determined by several variations of the latter method gave very similar results (data not shown).

Two of these alternate methods of signature gene selection, based upon what is known about CPCA, pivalate, valproate, 4-pentenoate, propionate, octanoate and butyrate, resulted in a set of similarity scores comparable to the results already discussed. Of the 14 compound similarities originally identified (Figure 5), pivalate was excluded by one of the alternate methods while the other 13 were identified by both methods. Only one other compound not in the original 14 (diclofenac) was identified as being similar to MrkA by both alternative methods. There are reports that classify diclofenac as a β-oxidation inhibitor as well (Baldwin et al., 1998).

The predictive value of the compendium approach, and therefore the utility in identifying risk factors such as mitochondrial impairment or activation of cytokine pathways, is also apparent from this study in the expression profiles from the low-dose animals (50 mpk, Figure 3). The signature for MrkA that is responsible for its classification as a mitochondrial β-oxidation inhibitor was observed in animals at this dose level, particularly at the later time points, even though this dose did not produce steatosis or other indications of toxicity. This observation is important, since it indicates the potential to detect occult effects that are not observed using traditional toxicological endpoints.

There are limitations to using a compendium for the identification of a mechanistic basis for toxicity. While it can be proposed that MrkA is a β-oxidation inhibitor based on transcriptional analysis alone, this effect could not be confirmed in the absence of other data including histology (steatosis suggests mitochondrial inhibition) and mitochondrial function assays. Other data, such as PPAR transactivation data, metabonomic data and/or proteomic data, can help further define the specificity of this effect. Also, the high PPARα similarity scores illustrates the potential for a secondary effect with a strong or dominant expression signature to overshadow or mask the primary mechanism, as is illustrated in Figure 4.

For most clusters in this figure (i.e., 1–3, 8, and 10–12), it is difficult to identify any difference between MrkA and the PPARα agonists. The main difference is observed almost completely in cluster 6 (with some subtle differences apparent in clusters 4 and 9). In this case, an approach that used the compound with the closest match to the test compound or even a k-nearest neighbor approach would likely have misdiagnosed the basis of toxicity as PPARα activation, which is typically not considered to produce hepatic toxicity of this nature. The genes whose expression levels were changed similarly between MrkA and PPARα agonists are in pathways known to be part of the PPARα response, which includes genes regulating mitochondrial and metabolic function (Figure 6B, e.g., pyruvate metabolism, citrate cycle, fatty acid metabolism, glycolysis/gluconeogenesis, reviewed in Cornwell et al., 2004).

In this report, we have used gene expression profiling to identify a plausible mechanistic basis for the hepatic toxicity of an experimental compound. The proximal effect was an inhibition of mitochondrial function, and inhibited fatty acid oxidation likely led to the observed accumulation of lipids as evidenced by steatosis. Accumulated lipids in turn produced a response mediated by PPARα, and an activation of the innate immune system ensued. Evidence for such an effect was observed at the transcriptional level even at a dose level that did not produce detectable toxicity, demonstrating that expression profiling can detect occult adverse effects that may not be found through more traditional methods. Further, we have demonstrated the utility of using a compendium of expression profiles to help identify such a mechanistic basis using a similarity scoring metric. With refinement to reduce false positive and false negative discoveries, this approach will enhance not only the timing in which we can identify liabilities of potential new drugs, but also the accuracy with which we can identify putative mechanisms and risk factors.