ToxSci Advance Access originally published online on February 14, 2008
Toxicological Sciences 2008 103(1):28-34; doi:10.1093/toxsci/kfn022
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Interlaboratory Evaluation of Genomic Signatures for Predicting Carcinogenicity in the Rat
a Roche Palo Alto, Palo Alto, California b Johnson & Johnson Pharmaceutical Research & Development, Raritan, New Jersey c GlaxoSmithKline, Research Triangle Park, North Carolina d Abbott Laboratories, Abbott Park, Illinois e Amgen, Thousand Oaks, California f AstraZeneca, Boston, Massachusetts g AstraZeneca, Södertälje, Sweden h Boehringer-Ingelheim, Ridgefield, Connecticut i Bristol-Myers Squibb, Princeton, New Jersey j Bristol-Myers Squibb, Wallingford, Connecticut k Critical Path Institute, Tucson, Arizona l Federal Institute for Drugs and Medical Devices, Bonn, Germany m Food and Drug Administration, Washington, DC n Iconix Biosciences, Mountain View, California o Lilly Research Laboratories, Eli Lilly and Company, Greenfield, Indiana p Merck Research Laboratories, West Point, Pennsylvania q Novartis Pharma AG, Basel, Switzerland r Pfizer, Groton, Connecticut s Sanofi-Aventis, Porcheville, France t Schering-Plough Research Institute, Summit, New Jersey u Wyeth Research, Chazy, New York
1 To whom correspondence should be addressed at Roche Palo Alto LLC, 3431 Hillview Avenue, Palo Alto, CA 94304. Fax: (650) 855-558. Tel: (650) 855-5136. E-mail: mark.fielden{at}roche.com.
Received November 20, 2007; accepted January 30, 2008
 |
ABSTRACT |
|---|
 TOP ABSTRACT INTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONSUPPLEMENTARY DATAREFERENCES |
|---|
The Critical Path Institute recently established the Predictive Safety Testing Consortium, a collaboration between several companies and the U.S. Food and Drug Administration, aimed at evaluating and qualifying biomarkers for a variety of toxicological endpoints. The Carcinogenicity Working Group of the Predictive Safety Testing Consortium has concentrated on sharing data to test the predictivity of two published hepatic gene expression signatures, including the signature by Fielden et al. (2007, Toxicol. Sci. 99, 90–100) for predicting nongenotoxic hepatocarcinogens, and the signature by Nie et al. (2006, Mol. Carcinog. 45, 914–933) for predicting nongenotoxic carcinogens. Although not a rigorous prospective validation exercise, the consortium approach created an opportunity to perform a meta-analysis to evaluate microarray data from short-term rat studies on over 150 compounds. Despite significant differences in study designs and microarray platforms between laboratories, the signatures proved to be relatively robust and more accurate than expected by chance. The accuracy of the Fielden et al. signature was between 63 and 69%, whereas the accuracy of the Nie et al. signature was between 55 and 64%. As expected, the predictivity was reduced relative to internal validation estimates reported under identical test conditions. Although the signatures were not deemed suitable for use in regulatory decision making, they were deemed worthwhile in the early assessment of drugs to aid decision making in drug development. These results have prompted additional efforts to rederive and evaluate a QPCR-based signature using these samples. When combined with a standardized test procedure and prospective interlaboratory validation, the accuracy and potential utility in preclinical applications can be ascertained.
Key Words: biomarkers; nongenotoxic; toxicogenomics.
|
INTRODUCTION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONSUPPLEMENTARY DATAREFERENCES |
|---|
The standard genotoxicity testing battery consisting of in vitro and in vivo assays has been established for evaluating the potential of chemicals or their metabolites to damage DNA (Müller et al., 1999
). As pharmaceutical companies rarely if ever advance compounds with known genotoxic activity for indications other than oncology, there is little justification for developing additional in vivo biomarkers of genotoxic carcinogenicity. Although there is a mechanistic connection between genotoxicity and cancer development, a substantial number of carcinogenicity findings in the 2-year rodent bioassay are caused by nongenotoxic mechanisms (Snyder and Green, 2001). In the case of nongenotoxic carcinogens, it is commonly accepted that a no-effect level exposure can be determined and an exposure margin to projected human exposure that possesses an acceptable level of risk can be established. This is reflected in current guidelines for dose selection for carcinogenicity studies (International Conference on Harmonization, 2005). In addition, several examples exist for which nongenotoxic carcinogenicity in rodents has been conclusively shown to be not relevant to human risk (Holsapple et al., 2006; McDonald, 2004). As a result, an understanding of the mode of action of compounds inducing tumors in a chronic rodent bioassay can be used to make the case for safe human use and thus support product approval and labeling. The time and resources involved in determining the mode of action and relevance to humans of a nongenotoxic carcinogen is considerable. Therefore, biomarkers providing an early prediction of nongenotoxic carcinogenicity would permit compound selection decisions to be made at an earlier stage in development, and would facilitate the earlier initiation of the risk assessment process, thus reducing delays to the introduction of important medicines to the clinic.
Many genomic-based biomarkers have been published for a variety of toxicities in nonclinical species, primarily the rat. However these putative biomarkers are often not comprehensively evaluated using large independent test sets, and are rarely validated across laboratories (Minami et al., 2005; Osada et al., 2006; Zidek et al., 2007), thus hampering wider adoption of these new predictive tools with the potential to improve safety assessment. Transcriptional multigene biomarkers, or signatures, predictive of nongenotoxic carcinogenicity have been published. The signature described by Fielden et al. (2007) has been shown to have the potential to predict nongenotoxic hepatocarcinogens. The signature described by Nie et al. (2006) has been shown to have the potential to predict nongenotoxic carcinogens inducing tumors in hepatic and multiple nonhepatic tissues. The goal of the collaborative effort of the Carcinogenicity Working Group within the Critical Path Institute Predictive Safety Testing Consortium was to further evaluate the utility of these published signatures as an early predictor of nongenotoxic carcinogenesis in rat using large databases of microarray data collected independently in three laboratories. These signatures were chosen primarily due to the extensive size of their training sets relative to others and the extent of existing, published validation data suggesting they had relatively good accuracy. Because these predictive tools are not intended to replace the chronic rodent bioassay for carcinogenicity, but rather to guide internal decision-making and/or mechanistic studies prior to initiation of chronic rodent studies, a rigorous validation of the biomarker as a replacement of chronic rodent studies is not an objective. Instead, this interlaboratory evaluation was an opportunistic meta-analysis of existing hepatic gene expression data to evaluate retrospectively the predictivity of classifiers for nongenotoxic carcinogenicity. Due to differences in microarray technologies and treatment regimens used across laboratories, this study provides a real-life example of biomarker evaluation and learning for guiding additional experiments necessary to standardize an assay platform that could be implemented more broadly across testing laboratories, and more importantly to guide definitive validation studies for judging acceptability for use in preclinical safety assessments.
|
MATERIALS AND METHODS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONSUPPLEMENTARY DATAREFERENCES |
|---|
Description of the carcinogenicity signatures evaluated.
The literature was reviewed to determine the putative biomarkers of nongenotoxic carcinogenicity that the consortium could evaluate for their ability to classify correctly independent gene expression data for nongenotoxic carcinogenicity. To maximize the probability of success, two signatures were identified from the literature as being derived from a relatively large training set and also having undergone additional internal validation (Fielden et al., 2007
; Nie et al., 2006). Based on these criteria, the signatures were further evaluated with hepatic gene expression data that were available from three member companies.
The first gene expression-based signature to predict nongenotoxic hepatocarcinogens in rats was described in detail in Fielden et al. (2007). Briefly, Codelink RU1 gene expression data on 100 compounds (25 compounds positive and 75 compounds negative for nongenotoxic hepatocarcinogenicity) from the livers of 7- to 8-week-old male Sprague–Dawley rats treated at a maximum tolerated dose (MTD) for 5 days were used to derive a signature (Table 1). The Iconix signature consisted of 37 genes and corresponding weights (or coefficients), whereas a version of the signature for use with Affymetrix (Santa Clara, CA) RG230 v2.0 and RAE230A data contained 42 and 39 genes and weights, respectively (Supplemental Table 1). The Codelink version of the Iconix signature was validated by Iconix Biosciences (Mountain View, CA) on 47 independent compounds not used in building the signature and found to have a sensitivity and specificity of 86% and 81%, respectively (Fielden et al., 2007).
|
A second gene expression-based signature to predict nongenotoxic carcinogens across tissues was described in detail using 52 compounds (24 compounds positive and 28 compounds negative for nongenotoxic carcinogenicity in any tissue of rodents) (Nie et al., 2006
). Briefly, groups of 7- to 8-week-old rats were dosed once and sacrificed 24 h later. Liver RNA was then prepared and complimentary DNA (cDNA) microarray gene expression data generated to derive a signature (Table 1). Two classes of compounds were excluded from this set of training samples: macrophage activators such as Zymosan A and Coumarin; and peroxisome proliferators such as WY-14643 and Clofibrate. It was reasoned that both types of compounds induce many pronounced gene expression changes and inclusion of these compounds may bias the training set toward these classes and away from other classes that cause more subtle effects. The published J&J signature consists of 6 genes and corresponding weights (Supplemental Table 1). Cross-validation estimated the signature has an accuracy (percentage of all test cases correctly predicted) of 88.5%, and 84% when tested on the Codelink Rat Whole Genome platform (Nie et al., 2006).
Genotoxic carcinogens were excluded from both training sets in these signatures, and as such, the signature is unlikely to accurately predict genotoxic carcinogens. This has been confirmed for a number of genotoxic carcinogens in Fielden et al. (2007). The consortium focused on classifying nongenotoxic carcinogens from nongenotoxic noncarcinogens, as it was felt that current assays already exist to accurately detect genotoxic compounds.
Interlaboratory evaluation.
The data used as training sets for the Iconix and J&J signatures described above were also useful as independent test sets to evaluate each others’ signature. In addition, GSK had a large collection of gene expression data in rat liver that was useful for evaluation of the two signatures. Therefore, to evaluate the sensitivity and specificity of the individual signatures (hereafter referred to as the "Iconix" or "J&J" signatures) to classify nongenotoxic compounds as carcinogenic or noncarcinogenic, the genes and corresponding weights of the individual signatures were shared among consortium members and evaluated against independent test sets from Iconix, J&J, and GSK. The overlap in the compounds used in the three test sets is given in Figure 1 and fully detailed in Supplemental Table 2. To avoid bias in the evaluation, only compounds that were not used to derive the signature were considered for testing.
|
The general form of the linear-discriminant based classifiers used by both Iconix and J&J is defined by n variables, r1, r2, ... rn, and n associated constants or weights, w1, w2, ... wn, where ri is the log ratio of gene i, wi is the weight of gene i in the signature, and b is the bias term, all derived mathematically by their respective algorithms:
 |
Evaluation for S (signature score) for a test experiment across the n genes in the signature determines which side of the hyperplane in multidimensional space the experiment lies, and thus the classification of the test experiment. Test experiments with a signature score greater than 0 are predicted as nongenotoxic carcinogens, whereas experiments with a signature score less than 0 are predicted as not being nongenotoxic carcinogens.
Evaluation of the Iconix signature.
To evaluate the Iconix signature, test sets from J&J and GSK were used. A total of 32 compounds from J&J were used as a test set for the evaluation of the Iconix signature (Table 2). This included eight nongenotoxic hepatocarcinogens and 24 noncarcinogens (Supplemental Table 2). Groups of three adult male Sprague–Dawley rats, aged 7–8 weeks, were administered a single MTD of test compound, or matched vehicle control. Liver samples were obtained 24 h after a single dose of test compound or vehicle for extraction of total RNA. The gene expression intensities were normalized and log2 ratios calculated as previously described (Nie et al., 2006). Because the Iconix signature was based on log10 ratios, it was necessary to adjust the log transform. Additionally, differences in dynamic range between platforms also necessitated adjusting the bias term, that is, the classification threshold, to improve prediction accuracy. The signature score was calculated by summing the products of the gene weights and their corresponding log10 ratio for each individual hybridization, and then subtracting the bias term (Supplemental Table 1). A compound treatment was predicted to be a nongenotoxic carcinogen when the median signature score of the replicate hybridizations was positive.
|
A total of 74 compounds from GSK were used as a test set for the evaluation of the Iconix signature (Table 2). This included 22 nongenotoxic hepatocarcinogens and 52 noncarcinogens (Supplemental Table 2). Groups of five adult male Sprague–Dawley rats, aged 7–8 weeks, received daily administration of test compound or vehicle control by oral gavage for 4 days at an MTD. Liver samples were obtained 24 h after the last dose of test compound or vehicle for extraction of total RNA and profiled on the Affymetrix Rat Genome 230A GeneChip (RAE230A) according to the manufacturer's instructions. Probe sets for each hybridization were summarized and normalized using the Affymetrix MAS 5 algorithm. Log10 ratios were 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. Each of the 37 genes from the Iconix signature, based on the Codelink RU1 platform, was mapped in silico to the same gene on the RAE230A platform based on sequence identity. In the cases where there were multiple Affymetrix probe sets per gene, the RAE230A probe set with the highest correlation to the Codelink probe when measured across a number of tissue standards was used (data not shown). The signature score was then calculated by summing the products of the gene weights and their corresponding log10 ratio and then subtracting the bias term (Supplemental Table 1). A compound treatment was predicted to be a nongenotoxic carcinogen when the average signature score was greater than zero in any dose group.
Evaluation of the J&J signature.
To evaluate the J&J signature, test sets from Iconix and GSK were used. A total of 61 compounds from Iconix were used as a test set for the evaluation of the J&J signature (Table 2). This included 30 nongenotoxic carcinogens, primarily inducers of liver tumors, and 31 noncarcinogens (Supplemental Table 2). Groups of three adult male Sprague–Dawley rats 7–8 weeks of age were dosed daily at an MTD with test compound or vehicle as previously described (Fielden et al., 2007) and as above. In order to evaluate the J&J signature, Affymetrix expression data from Iconix were used because data on the Codelink RU1 microarray does not contain two of the six signature genes. Livers from compound and vehicle-treated rats were obtained 24 h after a single dose. Signature results were calculated using day 1 gene expression data only because the J&J signature was designed to predict carcinogenicity 24 h after a single dose. Total RNA from the liver samples were then profiled on the Affymetrix RG230 v2.0 GeneChip according to the manufacturers instructions. Probe sets were summarized and normalized using MAS 5 software. Log2 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. In the cases where there were multiple probe sets per gene, the average log2 ratio across the probe sets was used. The signature score was calculated by summing the products of the gene weights and their corresponding log2 ratio, averaged across the three treated rats and their controls, and then subtracting the bias term (Supplemental Table 1). A compound treatment was predicted to be a nongenotoxic carcinogen when the average signature score was greater than zero.
A total of 78 compounds from GSK were used as a test set for the evaluation of the J&J signatures (Table 2). This included 25 nongenotoxic carcinogens and 53 noncarcinogens (Supplemental Table 2). Groups of five adult male Sprague–Dawley rats, aged 7–8 weeks, received daily administration of test compound or vehicle control, by oral gavage for 4 days at an MTD. Liver samples were obtained 24 h after the last dose of test compound or vehicle for extraction of total RNA was profiled on the Affymetrix RAE230A GeneChip according to the manufacturer's instructions. Probe sets for each hybridization were summarized and normalized using the Affymetrix MAS 5 algorithm. Log2 ratios were 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. Each of the six genes from the J&J signature was then mapped to the same gene on the RAE230A platform. In the cases where there were multiple probe sets per gene, the probe set with the highest overall average signal across the GSK experimental data was selected as the representative probe set for that gene. The signature score was calculated by summing the products of the gene weights and their corresponding log2 ratio and then subtracting the bias term (Supplemental Table 1). A compound treatment was predicted to be a nongenotoxic carcinogen when the average signature score was greater than zero in any dose group.
|
RESULTS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONSUPPLEMENTARY DATAREFERENCES |
|---|
Evaluation of the Iconix Signature
The Iconix signature for nongenotoxic hepatocarcinogenicity was tested against 32 compounds in the J&J test set and 74 compounds in the GSK test set (Supplemental Table 2). The J&J test set consisted of eight nongenotoxic hepatocarcinogens and 24 noncarcinogens. The overall prediction accuracy was 71.9% (p-value = 0.002), with a sensitivity and specificity of 100% and 62.5%, respectively (Fig. 2A; Supplemental Table 2). This level of test sensitivity is highly similar to that reported by Iconix (86%); however, the specificity is approximately 20% lower.
|
The GSK test set consisted of 22 nongenotoxic hepatocarcinogens and 52 noncarcinogens. The overall prediction accuracy was 63.5% (p-value = 0.011), with a sensitivity and specificity of 72.7% and 59.6%, respectively (Fig. 2B; Supplemental Table 2). Some of false negatives include the hepatotoxicants acetaminophen, isoniazid, and carbon tetrachloride. A notable false positive was mestranol, which was positive at all tested doses. This is likely reflective of the fact that steroidal hormones are in the positive class of the Iconix training set, and the estrogenic activity of mestranol is leading to a positive prediction. This is also the case for dexamethasone, which is a potent Cyp3a inducer and liver tumor promoter. This suggests that the signature is highly sensitive to mechanisms of action that are linked to tumorigenicity. Additionally, noncarcinogenic hepatotoxicants are also incorrectly predicted as positive, including 1-naphthylisothiocyanate, 1,2-dichlorobenzene, dimethylformamide, phalloidin, and perhexiline. The reason for these false positives is unclear. The performance of the Iconix signature against the GSK test data is lower than reported by Iconix. The sensitivity and specificity was about 15 and 20% lower, respectively; the latter of which is consistent with the specificity obtained with the J&J test set.
Evaluation of the J&J Signature
The J&J signature for nongenotoxic carcinogenicity was tested against 61 compounds in the Iconix test set and 78 compounds in the GSK test set (Supplemental Table 2). The Iconix test set consisted of 30 nongenotoxic carcinogens and 31 noncarcinogens. The overall prediction accuracy was 63.9% (p-value = 0.026), with a sensitivity and specificity of 53.3% and 74.2%, respectively (Fig. 3A; Supplemental Table 2). The false negative rate of the J&J signature is likely to be an overestimate because the signature was purposely not trained on PPAR-alpha agonists. As a result, the four PPAR-alpha agonists in the Iconix test set were predicted negative. Removing these cases from consideration increases the sensitivity to 66.7%. The two carcinogens in the test set that induce tumors outside the liver (catechol and fluvastatin) were correctly predicted positive by the signature, suggesting that some mechanisms of carcinogenicity that occur in other tissues may be represented in the liver. Further evaluation of the signatures ability to predict carcinogenicity in other tissues outside the liver is warranted. A number of hepatotoxicants and steroidal compounds were not correctly predicted, including the Ah receptor agonist beta-naphthoflavone, chloroform, and carbon tetrachloride, and the steroidal compounds progesterone and ethynylestradiol. However, the steroidal compound megestrol was incorrectly predicted as positive, as was the Cyp3a inducer clotrimazole. The tested accuracy is approximately 20% lower than originally reported by J&J, despite similarity in the time points.
|
The GSK test set consisted of 25 nongenotoxic carcinogens and 53 noncarcinogens. The overall prediction accuracy was 55.1% (p-value = 0.002), with a sensitivity and specificity of 96.0 and 35.9%, respectively (Fig. 3B; Supplemental Table 2). The J&J signature correctly predicted amiodarone, a thyroid tumor inducing compound, and phenytoin, a lymphoma inducing compound, as positive. Surprisingly the one false negative in the GSK test set was tamoxifen, which could be argued as being a genotoxic and/or epigenetic hepatocarcinogen (Tryndyak et al., 2006
). This mechanism would not be expected to be positively identified because the training set did not include genotoxic carcinogens, nor did the Iconix training set. As one might expect, the very high sensitivity against this test set came at the cost of low specificity, as 34 of the 53 (64%) noncarcinogens were incorrectly predicted as carcinogens. The discrepancy in specificity between the Iconix and GSK test sets is not clear. Nonetheless, the accuracy of the J&J signature, in addition to the Iconix signature, is statistically greater than expected by chance and may be improved by standardizing the treatment regime and microarray platform. |
DISCUSSION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONSUPPLEMENTARY DATAREFERENCES |
|---|
The interlaboratory evaluation of the two genomic biomarkers presented herein represents a realistic example of the effective performance of published genomic biomarkers across independent laboratories. Although it was recognized a priori that the experimental designs and conditions of evaluation in other laboratories would not match those under which the original signatures were derived, it was nonetheless of great interest to understand how such differences in experimental conditions would affect classification accuracy and whether additional investment in deriving and validating such predictors is warranted. Although the importance of interlaboratory validation is obvious, the lack of substantial amounts of genomic data in multiple laboratories has hindered cross-site validation efforts within industry. Furthermore, genomic biomarkers derived within industry are often proprietary or only commercially available, thus limiting wider validation efforts. Only by sharing data and combining efforts across companies can such genomic biomarkers be sufficiently evaluated for wider use across industry.
The results reveal that the signatures do maintain significant classification accuracy despite differences in test conditions, as indicated by the significant (p < 0.05) Fisher's Exact test, however the decreased level of test sensitivity and specificity are not likely sufficient to justify routine use of the signatures as tested outside their laboratory of origin. However, in interpreting such signatures it is important to note that the value of a signature should be interpreted in the context of sensitivity and specificity for its intended use, which may be diverse depending on whether it is used in drug discovery or development, or for hazard identification of chemicals or environmental pollutants. At this point in the evaluation phase, we were looking to evaluate general trends in classification accuracy to justify more extensive and definitive validation studies, rather than defining absolute cutoffs for sensitivity and specificity as determinants of feasibility.
The decreased test accuracy in comparison to the reported accuracy as measured in the same laboratory argue that differences in both the animal treatment protocol and the microarray platform, in addition to stochastic error, are likely to contribute to the discrepancies observed. Because the test chemicals were distinct from those used to train the model, differences in chemotype or mode of action may also affect the relative accuracy estimates between laboratories, however, this scenario is more realistic to the anticipated application where unique structures are likely to be evaluated for carcinogenicity. The classification accuracy of the Iconix signature when validated by Iconix was approximately 85% (n = 47 compounds), whereas the test accuracy in other laboratories ranged from 62.5% (n = 32) to 71% (n = 74). Likewise, the J&J signature had a reported accuracy of 88.5%, compared with a test accuracy of 55.1% (n = 78) to 63.9% (n = 61) when tested in other laboratories. A loss of 15–30% accuracy can be substantial, however, it is understood that a loss of performance is anticipated given the differences in experimental conditions, treatment protocols and microarray platforms between laboratories.
An additional source of test error is the annotation of the test compounds. Although the classification of the test compounds reflects results in 2-year rodent bioassays or known mechanisms of action related to tumorigenicity, we cannot discount the possibility that differences in dose, route of administration, rat strain, age, feeding conditions or other experimental variables may have influence on the expected outcomes after 2 years of dosing. Unfortunately this cannot be rectified without repeating 2 years of dosing in rats. Another source of test error is the annotation of test compounds between laboratories. For example, the J&J group accepted some positive findings in mouse, such as cyproterone acetate, as positive across rodents. As a result, several compounds classified as noncarcinogens in the present study, such as coumarin, pyrilamine, and sulpiride, would have been classified as nongenotoxic carcinogens by the J&J signature. Other compounds could be argued as being genotoxic, such as 2-ethylhexanol (liver tumors in female mice), allyl alcohol (its metabolite acrolein is a carcinogen), the aneugen/polyploidy inducer thiabendazole, the topoisomerase inhibitor etoposide, and perhaps chloroquine, which has been reported as mutagenic (Obaseiki-Ebor and Obasi, 1986). Although we accept some uncertainty in our knowledge of the mechanisms of such compounds, the overall interpretation of our results would not change if we were to remove these compounds from consideration given the majority of the compounds are well understood toxicologically. Ultimately, the utility of these nongenotoxic carcinogen signatures should also be derived from the identification of common mechanisms of action involved in nongenotoxic carcinogenesis to guide human risk assessment. This appears promising with these published gene sets as detailed in the original publications (Fielden et al., 2007; Nie et al., 2006).
Recent publications confirm that the underlying biology of compound treatments can be consistently revealed across sites and measurement platforms (Guo et al., 2006), however, the expression of that biology is likely to differ temporally between samples exposed to a single dose or multiple repeated doses of the test compound. Additionally, differences in relative exposure to the compounds may also contribute to the differences in sensitivity brought on by the distinct treatment regimens. Assuming the Iconix and GSK dosing paradigm was meant to induce some level of toxicity after 4 or 5 doses, the magnitude of drug exposure and resulting expression changes may be significantly higher than that obtained in a single dose paradigm. This may explain the increased sensitivity of the J&J signature when evaluated against the 4-day GSK test data (96%) relative to the day 1 Iconix test data (53.3%). However, the increased sensitivity comes at the cost of low specificity (36%). Considering that the signatures are likely to represent both primary and secondary modes of action that contribute to carcinogenicity, it is likely that a combination of single and repeat doses are necessary to realize gains in assay sensitivity.
Another source of discrepancy in experimental design that may contribute to differences in classification accuracy is the signature itself, and by extension the platform upon which it was derived. Classification models are very sensitive to the magnitude of the expression changes for highly weighted genes. As a result, small differences in the probe sequences between microarray platforms for the same gene can greatly affect the measured response and thus the overall signature score outcome. This is particularly true for cDNA microarrays relative to oligonucleotide-based microarrays where the sequence being measured can differ dramatically. This points to the importance of deriving, validating and using a signature that is based on the same measurement platform in order to minimize differences due to probe set sequences.
Overall, these results indicate that short-term gene expression profiles are likely to have sufficient classification ability to warrant future investments for refining methodologies and designing definitive validation studies for predicting carcinogenicity in the rat. Arguably more important than the identification of potential carcinogenicity of a compound is the identification of the mechanism of action and its relevance to man. As a result, these observations, collective experience and theoretical arguments have prompted efforts by the consortium to rederive a carcinogenicity signature on a single and readily accessible cost-effective platform, such as quantitative PCR (QPCR). This would necessitate the identification of the key genes with appreciable dynamic range that are useful for classification and measure their expression in the original training and test samples by QPCR to rederive a QPCR-based signature. This would likely facilitate wider interlaboratory analytical and biological validation of the signature on existing or new test compounds to judge the strength and limitations of the biomarker in preclinical screening paradigms.
|
SUPPLEMENTARY DATA |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONSUPPLEMENTARY DATAREFERENCES |
|---|
Supplementary data are available online at http://toxsci.oxfordjournals.org/.
|
REFERENCES |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONSUPPLEMENTARY DATAREFERENCES |
|---|
Fielden MR, Brennan R, Gollub J. A gene expression biomarker provides early prediction and mechanistic assessment of hepatic tumor induction by nongenotoxic chemicals. Toxicol. Sci. (2007) 99:90–100.
Guo L, Lobenhofer EK, Wang C, Shippy R, Harris SC, Zhang L, Mei N, Chen T, Herman D, Goodsaid FM, et al. Rat toxicogenomic study reveals analytical consistency across microarray platforms. Nat. Biotechnol. (2006) 24:1162–1169.[CrossRef][Web of Science][Medline]
Holsapple MP, Pitot HC, Cohen SM, Boobis AR, Klaunig JE, Pastoor T, Dellarco VL, Dragan YP. Mode of action in relevance of rodent liver tumors to human cancer risk. Toxicol. Sci. (2006) 89:51–56.
International Conference on Harmonization. S1C(R1): Dose selection for carcinogenicity studies of pharmaceuticals & limit dose (2005) Available at: http://www.ich.org/LOB/media/MEDIA491.pdf. Accessed January 25, 2008.
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.
Minami K, Saito T, Narahara M, Tomita H, Kato H, Sugiyama H, Katoh M, Nakajima M, Yokoi T. Relationship between hepatic gene expression profiles and hepatotoxicity in five typical hepatotoxicant-administered rats. Toxicol. Sci. (2005) 87:296–305.
Müller L, Kikuchi Y, Probst G, Schechtman L, Shimada H, Sofuni T, Tweats D. ICH-harmonised guidances on genotoxicity testing of pharmaceuticals: Evolution, reasoning and impact. Mutat. Res. (1999) 436:195–225.[CrossRef][Web of Science][Medline]
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]
Obaseiki-Ebor EE, Obasi EE. Aspects of chloroquine mutagenicity. Mutat. Res. (1986) 175:51–59.[CrossRef][Web of Science][Medline]
Osada S, Naganawa A, Misonou M, Tsuchiya S, Tamba S, Okuno Y, Nishikawa J, Satoh K, Imagawa M, Tsujimoto G, et al. Altered gene expression of transcriptional regulatory factors in tumor marker-positive cells during chemically induced hepatocarcinogenesis. Toxicol. Lett. (2006) 167:106–113.[CrossRef][Web of Science][Medline]
Snyder RD, Green JW. A review of the genotoxicity of marketed pharmaceuticals. Mutat. Res. (2001) 488:151–169.[CrossRef][Web of Science][Medline]
Tryndyak VP, Muskhelishvili L, Kovalchuk O, Rodriguez-Juarez R, Montgomery B, Churchwell MI, Ross SA, Beland FA, Pogribny IP. Effect of long-term tamoxifen exposure on genotoxic and epigenetic changes in rat liver: Implications for tamoxifen-induced hepatocarcinogenesis. Carcinogenesis (2006) 27:1713–1720.
Zidek N, Hellmann J, Kramer PJ, Hewitt PG. Acute hepatotoxicity: A predictive model based on focused illumina microarrays. Toxicol. Sci. (2007) 99:289–302.
CiteULike 
Connotea 
Del.icio.us What's this?
This article has been cited by other articles:
|
|
 |
|
  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]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||