ToxSci Advance Access originally published online on June 8, 2007
Toxicological Sciences 2007 99(1):90-100; doi:10.1093/toxsci/kfm156
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
A Gene Expression Biomarker Provides Early Prediction and Mechanistic Assessment of Hepatic Tumor Induction by Nongenotoxic Chemicals
Iconix Biosciences, Inc., Mountain View, California 94043
1 To whom correspondence should be addressed at Investigative Toxicology, Non-Clinical Drug Safety, Roche Palo Alto LLC, 3431 Hillview Avenue, Palo Alto, CA 94304. Fax: (650) 855-5588. E-mail: mark.fielden{at}roche.com.
Received April 13, 2007; accepted June 4, 2007
 |
ABSTRACT |
|---|
 TOP ABSTRACT INTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONSUPPLEMENTARY DATAREFERENCES |
|---|
There are currently no accurate and well-validated short-term tests to identify nongenotoxic hepatic tumorigens, thus necessitating an expensive 2-year rodent bioassay before a risk assessment can begin. Using hepatic gene expression data from rats treated for 5 days with one of 100 structurally and mechanistically diverse nongenotoxic hepatocarcinogens and nonhepatocarcinogens, a novel multigenebiomarker (i.e., signature) was derived to predict the likelihood of nongenotoxic chemicals to induce liver tumors in longer term studies. Independent validation of the signature on 47 test chemicals indicates an assay sensitivity and specificity of 86% and 81%, respectively. Alternate short-term in vivo pathological and genomic biomarkers were evaluated in parallel for comparison, including liver weight, hepatocellular hypertrophy, hepatic necrosis, serum alanine aminotransferase activity, induction of cytochrome P450 genes, and repression of Tsc-22 or alpha2-macroglobulin messenger RNA. In contrast to these biomarkers, the gene expression–based signature was more accurate. Unlike existing tests, an understanding of potential modes of action for hepatic tumorigenicity can be derived by comparison of the signature profile of test chemicals to hepatic tumorigens of known mechanism, including regenerative proliferation, proliferation associated with xenobiotic receptor activation, peroxisome proliferation, and steroid hormone–mediated mechanisms. This signature is not only more accurate than current methods, but also facilitates the identification of mode of action to aid in the early assessment of human cancer risk.
Key Words: biomarker; carcinogenicity; nongenotoxic; toxicogenomics; microarray.
|
INTRODUCTION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONSUPPLEMENTARY DATAREFERENCES |
|---|
Chemically induced liver tumors in rodents are the most frequently observed neoplastic lesion following long-term exposure to both genotoxic and nongenotoxic chemicals of commercial, therapeutic, and environmental interest (Davies and Monro, 1995
; Gold et al., 2005) and there is significant concern over the long-term safety of humans exposed to such ubiquitous chemicals. Since DNA damage and mutation are believed to be initiating events for carcinogenesis, a number of in vitro and short-term in vivo tests for these events are commonly used to identify genotoxicants. Genotoxicity, however, does not always correlate with carcinogenicity and many nongenotoxic chemicals are carcinogenic. As a result, the 2-year rodent bioassay remains the standard test for assessing carcinogenicity. Because this bioassay costs millions of dollars per compound and many years to complete, it is not applied until very late in product development after considerable resources have already been invested. Positive findings often result in significant delay or denial of successful product registration and add greatly to development costs. Furthermore, many environmental pollutants, industrial chemicals, and food contaminants to which humans are exposed have not been adequately tested for carcinogenicity due to the resources required.
It has been proposed that a combination of mechanism-based in silico, in vitro, or short-term in vivo tests could replace the 2-year bioassay (Cohen, 2004; McDonald, 2004). However, retrospective evaluation of their predictivity has indicated they are not adequate for regulatory purposes (Jacobs, 2005) and their utility as early screening tools is questionable due to limited validation. Several medium-term bioassays in rats and transgenic mice have been developed as alternatives to long-term testing, however, their format and associated expense do not facilitate routine screening (Cohen et al., 2001; Ito et al., 2003). To further complicate matters, a lack of concordance between rodent predictions and human carcinogenicity has called into question the relevance of rodent assays to human risk assessment (Cohen et al., 2001) and there is growing recognition that mode of action (MOA) must be taken into account when making carcinogenicity risk assessments for low-dose human exposures based on high-dose rodent experiments (Holsapple, 2006). As a result, it is unlikely that existing in vitro or short-term in vivo tests can adequately replace the requirement for the 2-year rodent bioassay.
A pragmatic approach to human cancer risk assessment is to develop and validate short-term in vivo screening methods with higher predictivity that will provide an early warning of potential hazard in rodents, while generating MOA information that can be used to assess relevancy of the findings to humans. Such information can guide decision making and/or prioritization and differentiation of chemicals for further development or evaluation, and ultimately contribute to the assessment of human risk at an earlier stage of product development. To this end, we have applied a hepatic gene expression–based approach to identify chemical-induced mRNA transcripts from short-term in vivo studies in rats that are predictive of, and mechanistically linked to, long-term liver tumor formation. In order to validate the assay, the signature was tested against 47 independent chemicals that were not used to derive the signature. To gain an understanding of the MOA of a putative nongenotoxic hepatocarcinogen, a novel approach was taken to compare the signature profile of a test compound against reference chemicals of known mechanism.
|
MATERIALS AND METHODS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONSUPPLEMENTARY DATAREFERENCES |
|---|
Animals and treatments.
Male Sprague–Dawley (Crl:CD(SD)(IGS)BR) rats (weight matched, 7–8 weeks of age and averaging 200–250 g) were purchased from Charles River Laboratories (Portage, MI) and housed individually in hanging, stainless steel, wire-bottom cages in a temperature (66–77°F), light (12-h dark/light cycle), and humidity (30–70%) controlled room. Water and Certified Rodent Diet #5002 (PMI Feeds, Inc, Richmond, IN) were available ad libitum throughout the studies. Housing and treatment of the animals were in accordance with regulations outlined in the United States Department of Agriculture and Code of Federal Regulations Animal Welfare Act (nine CFR Parts 1, 2, and 3). Animals were assigned to groups such that mean body weights were within 10% of the mean vehicle control group. Test articles were administered either orally (10 ml of corn oil/kg body weight) or by intraperitoneal, intravenous, or subcutaneous injection (5 ml of saline/kg body weight) as indicated (Supplementary Table 1). Animals were dosed once daily starting on day 0 and necropsied 24 h after the last dose as indicated. Time-, vehicle-, and route-matched control rats were treated concurrently as previously described (Ganter et al., 2005
). Animals were administered chemical once daily at a maximum tolerated dose that was selected based on initial range finding studies to reduce body weight gain to 5–10% in the absence of severe clinical signs, thus ensuring sufficient exposure to ensure a robust transcriptional response reflective of the MOA of the chemical. This dosing regimen also has the advantage of simplifying the dose selection criteria for a hazard identification screen. Since this approach is not meant to replace the 2-year rodent bioassay, but rather facilitate hazard identification, doses were not chosen to match those found in the original 2-year bioassays.
Compound classification.
A chemical was classified as an hepatic tumorigens if (1) it was found to induce liver tumors in a 2-year carcinogenicity study in at least one strain or gender of rat, or (2) if it was reasonably expected to induce liver tumors based on a known class effect (e.g., peroxisome proliferator–activated receptor [PPAR-
] agonists, steroid hormones), or (3) believed to induce or promote liver tumors through a nongenotoxic mechanism despite positive findings in genotoxicity assays (e.g., phenobarbital, clofibrate). A chemical was classified as negative for hepatic tumorigenicity if (1) it was found not to induce liver tumors in a 2-year carcinogenicity study in both male and female rats or (2) is not expected to induce liver tumors based on an antiproliferative MOA. Nonhepatotumorigens with at least one positive finding in a genotoxicity assay are not likely to adversely affect the signature and so were not specifically excluded. Since the assay is restricted to hepatic gene expression, tumorigenicity in other organs was not considered in the classification. Data were obtained from the Carcinogenicity Potency Database (http://potency.berkeley.edu), the National Library of Medicine Chemical Carcinogenesis Research Information System (http://toxnet.nlm.nih.gov/cgi-bin/sis/htmlgen?CCRIS), the National Toxicology Program Database (http://ntp-server.niehs.nih.gov), the Physician's Desktop Reference, and Pubmed (http://www.pubmed.gov) (Supplementary Table 1). Positive findings in the literature were used without reinterpretation or reclassification with respect to tumor formation. Furthermore, no attempt was made to segregate chemicals based on the incidence or severity of tumor formation since the doses used in the current study are likely higher than those used in 2-year carcinogenicity studies, and biasing the training set toward only potent tumorigens may hinder the sensitivity of the biomarker toward weaker carcinogens that are still of regulatory concern.
Microarray expression profiling.
Gene expression profiling on the Amersham Codelink Uniset Rat 1 Bioarray (GE HealthCare, Piscataway, NJ), data processing, and quality control were performed as previously described (Ganter et al., 2005). Briefly, mRNA was extracted from 100 mg tissue pieces from the left lateral lobe using an automated MagNA Pure LC robot. cRNA synthesis, preparation, and purification were as described in the Codelink Manual v2.1 as supplied by Amersham using the Qiagen Biorobot 9600 (Valencia, CA). The cRNA was fragmented and hybridized on the Codelink Uniset Rat 1 Bioarray. Images were scanned on the Axon Genepix Scanner (Axon Instrument, Union City, CA). Log-transformed signal data for all probes were array-wise normalized used Array Qualifier (Novation Biosciences, Palo Alto, CA), a nonlinear centralization normalization procedure adapted for the CodeLink microarray platform. Log10 ratios for each experimental group are computed as the difference between the average of the logs of the normalized experimental signals and the average of the logs of the normalized control signals for each gene. Statistical significance of expression changes were calculated as previously described (Ganter et al., 2005). Genes were considered significantly changed at p < 0.05. Microarray data for all experiments are available at the National Center for Biotechnology Information Gene Expression Omnibus, web site (accession: GSE8251TBD upon acceptance).
Signature derivation.
To derive a signature, 5443 probes on the array were preselected based on having no missing values (e.g., invalid measurement or below signal threshold) in the positive class of the training set, and less than 5% missing values in the negative class of the training set. The training set was classified used the adjusted sparse linear programming (A-SPLP) algorithm (El Ghaoui et al., 2003; Natsoulis et al., 2005). Briefly, the algorithm finds an optimal linear combination of variables (i.e., gene expression measurements) that best separate the two classes of experiments in dimensional space, where m is equal to 5443. The general form of this linear-discriminant based classifier is defined by n variables, r1, r2, ... rn, and n associated constants or weights, w1, w2, ... wn, where ri is the log10 ratio of gene i, wi is the weight of gene i in the signature, and b is the bias term:
 |
Evaluation of the scalar product S for a test experiment across the n genes in the signature determines what side of the hyperplane in multidimensional space the test experiment lies, and thus the classification of the test experiment. Test experiments with scalar products greater than 0 are predicted as nongenotoxic hepatic tumorigens.
Signature performance is measured by sensitivity, which is defined as the percentage of true positive (TP) predictions out of all positive samples evaluated, and specificity, which is defined as the percentage of true negative (TN) predictions out of all negative samples evaluated. Signature accuracy, or concordance, is defined as the percentage of correct classifications out of all samples evaluated. A log odds ratio (LOR) is also used to conveniently summarize the performance of the signature, and is defined as the natural log of the ratio of the odds of predicting a subject to be positive when it is positive, versus the odds of predicting a subject to be positive when it is negative. It is determined by the following equation:
 |
Biomarker validation.
Percent change in relative liver weight, presence of hepatocellular hypertrophy, hepatic necrosis, or elevated serum alanine aminotransferase (ALT) activity was evaluated as predictors of nongenotoxic hepatic tumorigenicity. To increase the sensitivity of prediction, day 5 measurements were used to estimate sensitivity and specificity for each treatment. Relative liver weight was calculated by dividing the liver weight by the terminal body weight. Percent liver weight change was calculated by dividing the relative liver weight (liver weight/body weight) by the average relative liver weight of a large control population of day 5 animals (n > 400) matched to the test set for vehicle and route of administration. This has the advantage of assessing treatment effects relative to a stable estimate of liver weight in control animals. An increase in liver weight was determined to be significant if induced more than 120% relative to controls, and not significantly induced if increased less than 110%. Based on the incidence and severity of hepatocellular hypertrophy in a large set of vehicle control animals (n > 400), hypertrophy was determined to be positive if detected as at least a minimal change in at least two of the three treated rats. Similarly, hepatic necrosis was determined to be positive if detected as at least a minimal change in at least one of three treated rats. More restrictive thresholds for histopathological findings would be expected to decrease sensitivity in detecting significant changes with an increase in specificity. Serum ALT was considered significantly elevated by treatment if the average serum ALT level measured across the treatment group (n = 3) was greater than the mean plus two standard deviations (> 81 U/l) of a large population of control animals (n > 400) matched to the test set for vehicle and route of administration. For the purposes of calculating specificity, an average serum ALT value less than the mean and one standard deviation of the control population (64 U/l) was considered negative. To facilitate comparison to previous studies (Elcombe et al., 2002), equivocal cases were treated as negatives for the purpose of scoring. Treating equivocal cases as positive or removing them from consideration did not alter the overall conclusions. To evaluate Tsc-22 (NM_013043_PROBE1) and alpha2-macroglobulin (J02635_PROBE1) mRNA as predictors, a statistically significant (p < 0.05) repression of the mRNA as measured on the microarray was considered a positive prediction, while the absence of a significant repression was considered a negative prediction. To evaluate P450 induction as a predictor, a statistically significant (p < 0.05) induction of Cyp1a1 (X00469_PROBE1), or Cyp3a1 (M13646_PROBE1, X00469_PROBE1, NM_013105_PROBE1) or Cyp4a3 (M33936_PROBE1) was considered a positive prediction, while no statistically significant induction of these mRNAs was considered a negative prediction.
|
RESULTS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONSUPPLEMENTARY DATAREFERENCES |
|---|
Biomarker Derivation
To develop a multigene RNA-based signature to predict nongenotoxic chemical-induced hepatic tumors in rats, we first identified reference chemicals that are positive or negative for the induction of rat hepatic tumors based on public databases and limited to compounds for which gene expression data are available in the toxicogenomic reference database DrugMatrix (Ganter et al., 2005
). From this we identified 147 compounds in DrugMatrix for which information was available, and randomly chose 25 nongenotoxic hepatic tumorigens and 75 nonhepatocarcinogens to form the training set (Supplementary Table 1). The remaining 47 chemicals (21 positives and 26 negatives) were reserved for independent validation of the signature. Each chemical was evaluated in a short-term repeat dose in vivo study in male Sprague–Dawley rats (n = 3 per group). The livers from treated and control rats were collected 24 h following 1, 3, and 5 or 7 days of dosing and liver mRNA was extracted for gene expression profiling. Gene expression changes in treated animals were compared to matched controls and represented as average log10 ratios for each treatment group. Since the development of tumors is a long-term, multistep process, the signature was derived using gene expression data from the latest time point available in order to capture gene expression changes that were likely to be reflective of repeat dose treatment effects that contribute to carcinogenicity and not reflective of an initial stimulus following a single dose that may be transient in nature. Using the A-SPLP algorithm, we derived a linear classification model using the training set of 3-, 5-, or 7-day hepatic gene expression profiles for 100 chemicals. The adjustment of the SPLP algorithm allows a flexible trade-off between false positives and false negatives by differentially weighting the terms in the loss function related to the positive and the negative class, which helps adjust for the unbalanced class size. Based on split sample cross validation (60% training, 40% test, 20 random splits), the signature has an estimated sensitivity of 56%, a specificity of 94%, and a LOR of 2.92. Repeated randomization of compound annotation and classification of the 100 chemical training set resulted in a mean LOR of –0.007 ± 0.331 (n = 100), which is significantly different from the true training set (p < 0.01) indicating the estimated performance is not due to chance. Furthermore, other training sets of identical class size randomly selected from the original 147 chemicals produced signatures with similar training and test performance estimates indicating that the initial training set selection was representative (data not shown). All in vivo studies, including the microarray processing, were completed over a period of 4 years in our lab, thus reducing the probability that the signature is detecting confounding variables or artifacts in the protocol, such as experiment date or reagent supply or lot. This also provides a means to evaluate the signature performance on test samples generated over a lengthy time period.
The complete signature generated using 100% of the training data consists of 37 probes, each with an associated weight corresponding to its contribution to the classification power of the signature and indicative of the direction of change (Fig. 1). While several of the genes contributing to the signature have little or no current functional annotation, we were able to relate the function and directional change of many of the annotated signature genes to proliferative and antiapoptotic changes as expected for carcinogens, thus adding biological plausibility to the model. Since the algorithm is designed to select the shortest list of genes that will best classify the training set, it is likely that additional genes exist that are known to play a significant role in tumorigenicity, but classify the training set to a lesser degree.
|
Biomarker Validation
To validate the 37 gene signature on independent test samples, 1-, 3-, and 5-day hepatic gene expression profiles from the remaining 21 positive chemicals and 26 negative chemicals not used to develop the signature were evaluated. To account for chemical-specific toxicodynamics and toxicokinetics, the maximum scalar product observed over the three time points for each compound was used to classify each compound as positive or negative. This increases the sensitivity of the signature with little loss of specificity. The signature correctly predicted 18 of the 21 (85.7%) positive test chemicals as being nongenotoxic hepatic tumorigens (Table 1). Of the 26 negative test chemicals, 21 (80.8%) were correctly predicted negative. Although the training set is composed of mainly small molecule therapeutics (88%), the signature has broad utility across diverse structural and mechanistic classes, including industrial toxicants and environmental pollutants. For example, the signature correctly predicted the hepatotoxicants chloroform and carbon tetrachloride as positive, and 1,1-dichloroethene and propylene glycol as negative. Additionally, the signature was also able to correctly classify diethylstilbestrol (DES) and chloroform as hepatic tumorigens using male liver gene expression data, despite the fact that these compounds were found to be positive in female rats only. Further testing would be required to evaluate whether the signature is sufficiently sensitive to predict gender-specific hepatic tumorigens.
|
If genotoxic mechanisms play a role in tumorigenicity for chemicals classified in the training set as positive, then the signature would be expected to be sensitive to genotoxic hepatocarcinogens. To test this possibility, an additional four genotoxic hepatocarcinogens were evaluated by the signature (Table 2). The results indicate that aflatoxin B1, hydrazine, and N-nitrosodiethylamine were classified as negative, while 2-acetylaminofluorene was classified as weakly positive (SP = 0.03). This supports the concept that the signature is specific for chemicals that induce hepatic tumors through nongenotoxic mechanisms, as expected given the composition of the training set.
|
Comparison of Biomarker Accuracy
A highly standardized procedure was used to generate the gene expression and ancillary clinical pathology and histopathology data in the DrugMatrix database. This allowed the accuracy of the signature for predicting chemical-induced hepatic tumors to be directly compared to other putative biomarkers in the same animals from which the gene expression profiles were generated. The early preclinical biomarkers evaluated include increased liver weight, hepatocellular hypertrophy, hepatic necrosis, and P450 induction (Allen et al., 2004
; Elcombe et al., 2002). Serum ALT was also evaluated since it is an alternative indicator of hepatic injury and because of its association with P450 induction (O'Brien et al., 2002). Although these markers may have increased predictive accuracy in longer-term subchronic studies, this comparison serves to establish their utility as subacute markers of hepatic tumorigenicity relative to the signature. The repression of Tsc-22 mRNA was also tested as a biomarker, since it has been shown to be repressed by a variety of nongenotoxic hepatocarcinogens in both mouse and rat (Kramer et al., 2004; Iida et al., 2005; Michel et al., 2005) and has been observed to be repressed in some human tumor samples (Rentsch et al., 2006; Shostak et al., 2003). Repression of alpha2-macroglobulin mRNA was also tested as this gene was found to best classify the 100 chemical training set based on being significantly (p < 0.05) repressed by nongenotoxic hepatic tumorigens (Supplementary Table 2). Paradoxically, an increase in alpha2-macroglobulin mRNA and protein has been associated with preneoplastic and neoplastic lesions in the rat (Sukata et al., 2004). This discrepancy may be due to the differences in the time points of expression measurement. To facilitate comparison of prediction accuracy between these biomarkers and the signature, sensitivity, and specificity were determined based on an evaluation of the 47 independent chemical test set.
A direct comparison of the putative biomarkers indicates that the transcript-based signature was more accurate than other clinical, histological, or genomic methods for predicting nongenotoxic hepatic tumorigenicity (Table 2). The sensitivity of increased liver weight, hepatocellular hypertrophy, and hepatic necrosis as predictors were less than 50%, which is similar to previous estimates using longer term study data (Allen et al., 2004; Elcombe et al., 2002), and approximately 35% lower than the signature (Table 2). The induction of cytochrome P450 genes was a sensitive biomarker (81%), however, the false positive rate (50%) was too high for practical use. Results from individual P450 genes were no more accurate (data not shown), although they may be useful to diagnose specific mechanisms of receptor activation. Likewise, Tsc-22 and alpha2-macroglobulin were not highly specific (< 77% TN) and were repressed by approximately 45% of the positive compounds. When considered in the context of the entire 147 chemical data set, however, alpha2-macroglobulin was 54% sensitive and 91% specific (Table 2). Cytotoxicity and subsequent regenerative proliferation have been proposed as a mode of hepatic tumor formation for a number of chemicals, yet only 16% of the positive test chemicals were diagnosed with histopathological evidence of necrosis. In contrast, serum ALT was the most accurate biomarker with a sensitivity of 57% and a specificity of 88% (Table 2). When comparing the concordance between the predictions based on pathological endpoints relative to the signature across all test chemicals, serum ALT predictions were most coincident with the signature prediction (Fig. 2). This is primarily based on the concordance between TN predictions, as TP signature predictions were most concordant with P450 induction. There was little correspondence between the signature predictions and repression of alpha2-macroglobulin or Tsc-22, suggesting that neither of these genes in isolation is a good classifier of nongenotoxic tumorigenicity. Indeed neither gene forms part of the signature. A number of the false positives predicted by the signature, including the benzodiazepines lorazepam and diazepam, were also incorrectly predicted positive by other biomarkers. While the signature was trained to identify rat hepatic tumorigens, diazepam is a tumor promoter in mice suggesting the signature may identify expression changes indicative of tumor formation that may not be fully manifested in the rat under the long-term treatment conditions used (Diwan et al., 1986). These results clearly establish the gene expression signature as having increased predictive accuracy relative to current short-term pathology observations or genomic biomarkers, and also superior to estimates based on longer term subchronic studies (Allen et al., 2004; Elcombe et al., 2002).
|
Identification of MOA
Predictive accuracy in the rat is important for identifying hazards, however, the assessment of cancer risk to humans must take into account the mechanism(s) of tumorigenicity. Investigating the function of the genes in the signature and their relationship to neoplasia is challenging due to a limited functional understanding of most genes in mammalian genomes. Restricting the classification algorithm to genes with known function considerably reduces the performance of the signature (data not shown), indicating that as yet uncharacterized genes have classification power and are necessary for optimal predictive performance. The poor accuracy of individual genes for classification (Supplementary Table 2) and the observation that signature genes are not necessarily consistently regulated across all positive compounds suggests that multiple genes are required for accurate classification due to the multiple MOAs that the model must account for. This being the case, the effects of chemicals with similar MOA are likely to be reflected in similar expression profiles across the signature genes. To test this hypothesis, the products of the log10 ratio and weight (impact scores) for the 37 probes in the signature were hierarchically clustered across the 25 nongenotoxic hepatic tumorigens in the training set using Pearson's correlation as the similarity metric. Positive impact scores are preferred to log10 ratio data since they reflect not just the direction of change contributing to a positive prediction, but also the relative importance of the gene toward classification. This has the benefit that the similarity between chemicals is driven by the expression changes most important for prediction, and is less dependent on the minor expression changes that contribute to only a fraction of the classification power. The 25 compounds clustered into at least four distinguishable clusters, and some subclusters, suggesting this signature is capturing at least four distinct modes of action of hepatic tumor formation, including activation of xenobiotic receptors (Constitutive Androstane Receptor (CAR), Pregnane X Receptor, Aryl Hydrocarbon Receptor (AhR)) and concurrent P450 induction, peroxisome proliferation, hepatotoxicity and regenerative proliferation, and steroidal hormone–mediated mechanisms (Fig. 3A). These results suggest that the gene expression signature can be used not only to predict the ability of a chemical to induce hepatic tumors in rats, but also to indicate a potential MOA of a test chemical by virtue of similarity of the signature profile to chemicals with known mechanism.
|
To test this approach, the signature impact profiles for a number of mechanistically well-defined test chemicals that were correctly predicted positive by the signature were chosen for evaluation, including beta-naphthoflavone, diethylhexylphthalate (DEHP), chloroform, and DES (Table 1). The signature impact profiles were hierarchically clustered with a set of well-defined reference compounds in the training set representing at least four discrete MOAs, including 2,3,7,8-tetrachlorodibenzo-p-dioxin (TCDD) and phenobarbital (xenobiotic receptor agonists), fenofibrate (peroxisome proliferator), methapyrilene (cytotoxic/regenerative proliferation), and beta-estradiol (steroid hormone). In all four cases, the test chemical clustered most closely to the reference chemical with a similar known MOA (Fig. 3B). For example, beta-napthoflavone clustered with phenobarbital and TCDD. Beta-napthoflavone is well-known to activate the AhR, but not CAR. The genes in the signature mediating tumor formation induced by xenobiotic receptors appear highly related, such that differentiation among xenobiotic receptors is likely not possible, at least with this gene set. Secondary evaluation of P450 genes is likely to discriminate these submechanisms. DEHP was most similar to fenofibrate, suggesting that DEHP causes hepatic tumors through activation of PPAR-
. This is supported by the evidence of peroxisome proliferation induced by DEHP (Willhite, 2001). Chloroform was most similar to methapyrilene, suggesting that this compound induces hepatic tumors through cytotoxicity and subsequent regenerative proliferation, which is consistent with hypothesized mechanisms (Steinmetz et al., 1988). Finally, DES, a potent estrogen receptor agonist, clustered with the endogenous estrogen beta-estradiol, further validating the approach (Fig. 3B). |
DISCUSSION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONSUPPLEMENTARY DATAREFERENCES |
|---|
The current study demonstrates the successful derivation of a gene expression–based signature to predict the ability of chemicals to induce hepatic tumors in the rat through nongenotoxic mechanisms using a short-term repeat dose in vivo study design. The assay design facilitates an efficient means for evaluating chemicals for prioritization or further study, and unlike existing short-term assays for carcinogenic activity, it can provide an early understanding of potential modes of action that can contribute to a human risk assessment. This is particularly advantageous in drug development since carcinogen bioassays are typically performed at very late stages of development, and early proactive approaches initiated before development begins may circumvent years of additional investigations and development time. While similar approaches to identify biomarkers of carcinogenic activity using in vivo and in vitro models have been published (Iida et al., 2005
; Kramer et al., 2004; Tsujimura et al., 2006), more extensive independent validation is needed to increase the confidence in their general utility. Likewise, other attempts using larger data sets (Nie et al., 2006) show promise and await further independent validation.
The independent validation performed in the current study on 47 chemicals demonstrates a sensitivity and specificity of at least 80%. For a novel biomarker to be of practical use, it must provide a more efficient or accurate means of predicting the phenotype of interest than current practices. It is highly unlikely that a predictive biomarker will achieve 100% accuracy, however, an increase in predictivity over existing methods would demonstrate that the biomarker has practical utility. When the signature was compared to alternative predictive endpoints or other putative genomic biomarkers, it was found to have a greater overall accuracy (Table 2). While some pathological endpoints had better specificity, they lacked adequate sensitivity to be reliable. The second most accurate endpoint was serum ALT elevation. This may be rationalized based on the link between induction of cytotoxicity and subsequent regenerative proliferation that often ensues following tissue injury. Such a mechanism has been used to argue against the relevancy of the 2-year rodent bioassay since dose levels are relatively high and sustained for long periods relative to human exposure (Gaylor, 2005). Similar approaches using other biochemical parameters, including ALT and P450, have yielded similar estimates of accuracy as observed here (Kitchin et al., 1991). Single genes such as Tsc-22 were not as sensitive as expected given the results previously published using smaller sized data sets (Iida et al., 2005; Kramer et al., 2004; Michel et al., 2005). This is not surprising given the heterogeneous nature of carcinogenicity and the requirement for a single gene to capture all the unique mechanisms that may initiate the carcinogenic process. The validation results for these genes, compared to published reports, also illustrate how biomarker discovery efforts using small data sets can lead to inaccurate estimates of predictive accuracy.
It must be noted that the estimated accuracy of prediction is only as good as the truth of the outcome that is being modeled. This makes claims of predictive accuracy for this and other methods difficult given the experimental variables that can influence the reproducibility, severity, and incidence of tumor development in long-term studies. It also makes comparison of signature accuracy to other established or putative biomarkers difficult. However, by simultaneously measuring a number of pathological and genomic biomarkers in the same treated animals for their ability to classify chemicals, a fair comparison of their relative merits is assessed. Future biomarker derivation studies may benefit from considering both genomic and pathological endpoints in combination since these endpoints are typically collected simultaneously, and risk assessments necessarily integrate multiple endpoints into the overall hazard characterization. It is likely that tumorigenicity in other target organs could also be more accurately predicted using transcript-based biomarkers in combination with other histological or clinical endpoints, particularly where short-term hazard identification assays do not exist.
Ideally, the analyte(s) comprising a mechanism-based biomarker should be functionally associated with the predicted endpoint. However, given that the majority of genes in mammalian models are unknown or poorly characterized, and that our understanding of tumorigenicity is incomplete, this requirement is not likely to be met for some time. This is likely true for other signature-based predictions or diagnoses of rodent and human tumors. Screening approaches employing biomarkers without an entirely understood mechanistic foundation have utility in various contexts of product development or clinical practice, where mechanistic information is not needed for decision making or diagnosis. For chemically induced hepatic tumors, however, it is desirable, if not necessary, to obtain MOA information in order to guide a human risk assessment (Holsapple, 2006).
As a first step to understand the biological basis of classification, the 37 genes in the signature were evaluated for their function and relevance to hepatic neoplasia. Since the majority of the genes in the rat genome are not annotated or only poorly characterized, it was expected that a proportional number of genes in the signature were also not well annotated and functionally characterized. However, of the 27 genes with a Gene Ontology annotation (http://www.geneontology.org), a number of them point to proliferation, antiapoptosis, and processes previously implicated in hepatic tumorigenicity, as expected (Supplementary Data). While the genes in the signature were selected by the algorithm to accurately classify nongenotoxic hepatic tumorigens from nonhepatocarcinogens, such a small set is unlikely to mechanistically explain all pathways leading to tumor formation. Many genes that are mechanistically associated with tumorigenesis may play an important role in the chemicals used in this study, but may not be sensitive or specific enough to adequately contribute to the classification model, or may be redundant with other genes that are preferentially selected due to lower variability. While a number of signature genes can be hypothesized to play an important role in tumorigenesis, our understanding of such a complex phenotype is unlikely to be resolved by evaluating classifiers. Nonetheless, mechanistically associating classifier genes to the phenotype being modeled adds to the level of certainty around this biomarker.
It is recognized that many chemically induced hepatic tumors in rodents are not expected to induce tumors in humans based on differences in species biochemistry or pathophysiology (Holsapple, 2006). As a means of assessing the cancer risk to humans after positive findings in rodents are observed, a thorough analysis of the chemical's MOA is often undertaken to assess the relevancy of the MOA to humans. While understanding the function of the genes in the signature is helpful to rationalize the biological basis of classification, it does not necessarily assist in the evaluation of MOA for a new test chemical. By comparing the signature profile of a test chemical to chemicals of known MOA, a rapid and early testable hypothesis for a potential MOA can be derived. While it is encouraging that hierarchical clustering of the signature profiles is able to accurately subclassify test compounds, it is likely that a multiclass linear-discriminant classifier may provide more robust subclassification of carcinogenic MOAs given a larger training set of reference compounds with known mechanism. A potential limitation of this approach that has yet to be tested is whether the signature genes are able to predict hepatic tumorigens that act through unique mechanisms not represented by positive chemicals in the training set. Nonetheless, the signature profile could be used in these cases to rule out potential hypotheses to avoid unnecessary experiments that are unlikely to yield supportive results.
Despite encouraging results from comprehensive evaluations of gene expression results across multiple measurement platforms, including microarrays and PCR-based methods (Shi et al., 2006), differences in probe set sequence, dynamic range, and sensitivity may adversely impact the ability this signature derived on the Codelink RU1 platform to perform as well on other platforms. While ongoing evaluations will address this concern (Fielden et al. unpublished data), these findings open the possibility to recalibrate the samples on more universally accessible and cost effective platforms, such as PCR-based methods.
In summary, this novel short-term predictive assay was found to be more accurate than existing pathological or genomic approaches for predicting chemically induced hepatic tumors, thus providing a cost effective and reliable in vivo hazard identification test that can be applied to screen and prioritize chemicals for subsequent development and evaluation. Given the challenge in predicting such a complex and latent phenotype of diverse etiology, a model with the best predictive accuracy is needed. Rather than rely on incomplete knowledge of gene function, mechanistic information can be obtained by comparison of the signature profile to reference samples with known characteristics. Unlike existing predictive assays, this approach simultaneously provides evidence for a potential MOA to facilitate a proactive assessment of human cancer risk. It is expected that signature profiles generated for other heterogeneous human or rodent tumors, chemically induced pathologies, or a variety of disease states can be utilized in a similar manner to investigate underlying mechanisms or subclasses not otherwise apparent using current preclinical or clinical practices, thus increasing the value and cost effectiveness of short-term in vivo rodent testing to protect human safety and for understanding human diseases.
|
SUPPLEMENTARY DATA |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONSUPPLEMENTARY DATAREFERENCES |
|---|
Supplementary data are available online at http://toxsci.oxfordjournals.org/.
|
ACKNOWLEDGMENTS |
|---|
The authors would like to acknowledge the C-Path Predictive Safety Testing Consortium Carcinogenicity Working Group and the Microarray Quality Control Consortium Toxicogenomics Working Group for the many helpful discussions and suggestions. The authors are also grateful to Drs Donald Halbert and Antoaneta Vladimirova for their constructive feedback on the analysis and manuscript preparation.
|
REFERENCES |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONSUPPLEMENTARY DATAREFERENCES |
|---|
Allen DG, Pearse G, Haseman JK, Maronpot RR. Prediction of rodent carcinogenesis: An evaluation of prechronic liver lesions as forecasters of liver tumors in NTP carcinogenicity studies. Toxicol. Pathol. (2004) 32:393–401.[CrossRef][Web of Science][Medline]
Cohen SM. Human carcinogenic risk evaluation: An alternative approach to the two-year rodent bioassay. Toxicol. Sci. (2004) 80:225–229.
Cohen SM, Robinson D, MacDonald JS. Alternative models for carcinogenicity testing. Toxicol. Sci. (2001) 64:14–19.
Davies TS, Monro A. Marketed human pharmaceuticals reported to be tumorigenic in rodents. J. Am. Coll. Toxicol. (1995) 14:90–107.[Web of Science]
Diwan BA, Rice JM, Ward JM. Tumor-promoting activity of benzodiazepine tranquilizers, diazepam and oxazepam, in mouse liver. Carcinogenesis (1986) 7:789–794.
Elcombe CR, Odum J, Foster JR, Stone S, Hasmall S, Soames AR, Kimber I, Ashby J. Prediction of rodent nongenotoxic carcinogenesis: Evaluation of biochemical and tissue changes in rodents following exposure to nine nongenotoxic NTP carcinogens. Environ. Health Perspect. (2002) 110:363–375.[Web of Science][Medline]
El Ghaoui L, Lanckriet GRG, Natsoulis G. Robust classifiers with interval data (2003) Berkeley, California: Tech. Rep. UCB/CSD-03-1279, University of California.
Ganter B, Tugendreich S, Pearson CI, Ayanoglu E, Baumhueter S, Bostian KA, Brady L, Browne LJ, Calvin JT, Day GJ, et al. Development of a large-scale chemogenomics database to improve drug candidate selection and to understand mechanisms of chemical toxicity and action. J. Biotechnol. (2005) 119:219–244.[CrossRef][Web of Science][Medline]
Gaylor DW. Are tumor incidence rates from chronic bioassays telling us what we need to know about carcinogens? Regul. Toxicol. Pharmacol. (2005) 41:128–133.[CrossRef][Web of Science][Medline]
Gold LS, Manley NB, Slone TH, Rohrbach L, Garfinkel GB. Supplement to the Carcinogenic Potency Database (CPDB): Results of animal bioassays published in the general literature through 1997 and by the National Toxicology Program in 1997–1998. Toxicol. Sci. (2005) 85:747–808.
Holsapple MP. Mode of action in relevance of rodent liver tumors to human cancer risk. Toxicol. Sci. (2006) 89:51–56.
Iida M, Anna CH, Holliday WM, Collins JB, Cunningham ML, Sills RC, Devereux TR. Unique patterns of gene expression changes in the liver after treatment of mice for 2 weeks with different known carcinogens and non-carcinogens. Carcinogenesis (2005) 26:689–699.
Ito N, Tamano S, Shirai T. A medium-term rat liver bioassay for rapid in vivo detection of carcinogenic potential of chemicals. Cancer Sci. (2003) 94:3–8.[CrossRef][Medline]
Jacobs A. Prediction of 2-year carcinogenicity study results for pharmaceutical products: How are we doing. Toxicol. Sci. (2005) 88:18–23.
Kitchin KT, Brown JL, Kulkarni AP. Predictive assay for rodent carcinogenicity using in vivo biochemical parameters: operational characteristics and complementarity. Mutat Res. (1992) 266:253–272.[Web of Science][Medline]
Kramer JA, Curtiss SW, Kolaja KL, Alden CL, Blomme EA, Curtiss WC, Davila JC, Jackson CJ, Bunch RT. Acute molecular markers of rodent hepatic carcinogenesis identified by transcription profiling. Chem. Res. Toxicol. (2004) 17:463–470.[CrossRef][Web of Science][Medline]
McDonald JS. Human carcinogenic risk evaluation, part IV: Assessment of human risk of cancer from chemical exposure using a global weight-of-evidence approach. Toxicol. Sci. (2004) 82:3–8.
Michel C, Roberts RA, Desdouets C, Isaacs KR, Boitier E. Characterization of an acute molecular marker of nongenotoxic rodent hepatocarcinogenesis by gene expression profiling in a long term clofibric acid study. Chem. Res. Toxicol. (2005) 18:611–618.[CrossRef][Web of Science][Medline]
Natsoulis G, El Ghaoui L, Lanckriet GR, Tolley AM, Leroy F, Dunlea S, Eynon BP, Pearson CI, Tugendreich S, Jarnagin K. Classification of a large microarray data set: Algorithm comparison and analysis of drug signatures. Genome Res. (2005) 15:724–736.
Nie AY, McMillian M, Parker JB, Leone A, Bryant S, Yieh L, Bittner A, Nelson J, Carmen A, Wan J, et al. Predictive toxicogenomics approaches reveal underlying molecular mechanisms of nongenotoxic carcinogenicity. Mol Carcinog. (2006) 45:914–933.[CrossRef][Web of Science][Medline]
O'Brien PJ, Slaughter MR, Polley SR, Kramer K. Advantages of glutamate dehydrogenase as a blood biomarker of acute hepatic injury in rats. Lab. Anim. (2002) 36:313–321.
Rentsch CA, Cecchini MG, Schwaninger R, Germann M, Markwalder R, Heller M, van der Pluijm G, Thalmann GN, Wetterwald A. Differential expression of TGFbeta-stimulated clone 22 in normal prostate and prostate cancer. Int J. Cancer (2006) 118:899–906.[CrossRef][Web of Science][Medline]
Shi L, Reid LH, Jones WD, Shippy R, Warrington JA, Baker SC, Collins PJ, de Longueville F, Kawasaki ES, Lee KY, et al. The Microarray Quality Control Consortium (MAQC) project shows inter- and intraplatform reproducibility of gene expression measurements. Nat. Biotechnol. (2006) 9:1151–1161.
Shostak KO, Dmitrenko VV, Garifulin OM, Rozumenko VD, Khomenko OV, Zozulya YA, Zehetner G, Kavsan VM. Downregulation of putative tumor suppressor gene TSC-22 in human brain tumors. J. Surg. Oncol. (2003) 82:57–64.[CrossRef][Web of Science][Medline]
Sukata T, Uwagawa S, Ozaki K, Sumida K, Kikuchi K, Kushida M, Saito K, Morimura K, Oeda K, Okuno Y, et al. Alpha(2)-Macroglobulin: A novel cytochemical marker characterizing preneoplastic and neoplastic rat liver lesions negative for hitherto established cytochemical markers. Am. J. Pathol. (2004) 165:1479–1488.
Steinmetz KL, Tyson CK, Meierhenry EF, Spalding JW, Mirsalis JC. Examination of genotoxicity, toxicity and morphologic alterations in hepatocytes following in vivo or in vitro exposure to methapyrilene. Carcinogenesis (1988) 9:959–963.
Tsujimura K, Asamoto M, Suzuki S, Hokaiwado N, Ogawa K, Shirai T. Prediction of carcinogenic potential by a toxicogenomic approach using rat hepatoma cells. Cancer Sci. (2006) 97:1002–1010.[CrossRef][Medline]
Willhite CC. Weight-of-evidence versus strength-of-evidence in toxicologic hazard identification: Di(2-ethylhexyl)phthalate (DEHP). Toxicology (2001) 160:219–226.[CrossRef][Web of Science][Medline]
CiteULike 
Connotea 
Del.icio.us What's this?
This article has been cited by other articles:
|
|
 |
|
  R. S. Thomas, W. Bao, T.-M. Chu, M. Bessarabova, T. Nikolskaya, Y. Nikolsky, M. E. Andersen, and R. D. Wolfinger Use of Short-term Transcriptional Profiles to Assess the Long-term Cancer-Related Safety of Environmental and Industrial Chemicals Toxicol. Sci., December 1, 2009; 112(2): 311 - 321. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 M. J. Jonker, O. Bruning, M. van Iterson, M. M. Schaap, T. V. van der Hoeven, H. Vrieling, R. B. Beems, A. de Vries, H. van Steeg, T. M. Breit, et al. Finding transcriptomics biomarkers for in vivo identification of (non-)genotoxic carcinogens using wild-type and Xpa/p53 mutant mouse models Carcinogenesis, October 1, 2009; 30(10): 1805 - 1812. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
J. M. Phillips, L. D. Burgoon, and J. I. Goodman The Constitutive Active/Androstane Receptor Facilitates Unique Phenobarbital-Induced Expression Changes of Genes Involved in Key Pathways in Precancerous Liver and Liver Tumors Toxicol. Sci., August 1, 2009; 110(2): 319 - 333. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
S. Nesnow, W. Ward, T. Moore, H. Ren, and S. D. Hester Discrimination of Tumorigenic Triazole Conazoles from Phenobarbital by Transcriptional Analyses of Mouse Liver Gene Expression Toxicol. Sci., July 1, 2009; 110(1): 68 - 83. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
J. M. Phillips, L. D. Burgoon, and J. I. Goodman Phenobarbital Elicits Unique, Early Changes in the Expression of Hepatic Genes that Affect Critical Pathways in Tumor-Prone B6C3F1 Mice Toxicol. Sci., June 1, 2009; 109(2): 193 - 205. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 Y. Li, P. Hao, S. Zheng, K. Tu, H. Fan, R. Zhu, G. Ding, C. Dong, C. Wang, X. Li, et al. Gene expression module-based chemical function similarity search Nucleic Acids Res., November 1, 2008; 36(20): e137 - e137. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 D. Jacobson-Kram Commentary: Regulatory Toxicology and the Critical Path: Predicting Long-term Outcomes from Short-term Studies Vet. Pathol., September 1, 2008; 45(5): 707 - 709. [Full Text] [PDF]
|
|
|
|
|
|
M. R. Fielden, A. Nie, M. McMillian, C. S. Elangbam, B. A. Trela, Y. Yang, R. T. Dunn II, Y. Dragan, R. Fransson-Stehen, M. Bogdanffy, et al. Interlaboratory Evaluation of Genomic Signatures for Predicting Carcinogenicity in the Rat Toxicol. Sci., May 1, 2008; 103(1): 28 - 34. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||