Hepatotoxicants particularly target the liver and cause a variety of liver injuries. One type of damage is necrosis, a degenerative process leading to cell death. We analyzed a compendium of gene expression data (see Lobenhofer et al., in press) acquired from rat livers exposed to hepatotoxicants in an attempt to define gene expression patterns as a signatures that are highly predictive of the level of necrosis. We used necrosis as a phenotypic anchor to identify genes which predict the level of necrosis of the rat liver with a high degree of accuracy.
A prevailing advantage of our study was identifying genes related to necrosis which may be directly related to cell death caused by exposure to the compounds while filtering out genes related to other mechanisms related to a single compound. This was accomplished by using the histopathologic class label of the biological samples for necrosis as defined by board certified pathologists. We then leveraged ANOVA pairwise contrasts of the severity groups and GoMiner analysis of the gene expression data to subset the samples into groups with similar phenotypic changes exhibited by necrosis of the liver and comparable overrepresentation of biological processes (Figure ). Using these groups as class labels for the level of necrosis in the training set and two gene selectionification approaches (Random Forest and GEMS-SVM), we were able to identify subsets of genes which yielded a low prediction error rate during cross validation of the classifiers. Preliminary analysis of the liver gene expression data within each dose\time group by Lobenhofer et al. (in press) revealed compound-specific separation of the samples. Although classification of the blood data was better in higher dose groups at the later time points as compared to the liver data, classifiers derived from it was not able to predict animals in some cases where the hepatotoxicant elicited a different phenotypic response with the animals of a particular dose\time group. Interestingly, concordance analysis of ALT and AST enzyme levels with the class label of the level of necrosis and our predicted class labels revealed that the latter is more consistent with the enzyme levels. One potential reason for this could be that the microarray samples are from the whole liver whereas the histopathology samples used for scoring the extent of necrosis are only from two independent slices of the liver specimen. In a limited study, Heinloth et al. [13
] also showed that gene expression analysis is more informative than histopathologic evaluations and offers unique advantages to liver biopsy evaluations. Another explanation could be that certain animals may develop the phenotype at a later time point after treatment than the time point when the samples were taken for analysis. Therefore, the histopathology samples may not completely represent the liver toxicity. However, our gene expression analysis, redefining of the class labels for the level of necrosis exhibited in the samples and selection of predictor genes for necrosis are geared towards capturing the biological processes and mechanistic pathways that may govern the manifestation of the phenotype from a low level of necrosis to its highest level.
Using independent gene expression data sets acquired from the exposure of rat liver samples to a different set of hepatotoxicants, we show that the prediction accuracies of either of the two classifiers were roughly 80% overall but approximately 90%, about 80% and around 60% for acetaminophen-, carbon tetrachloride- and allyl alcohol-exposed samples, respectively with a p-value < 0.0005 for the significance of the prediction using the Random Forest classifier (Table ). The dramatic difference in prediction accuracy could be related to bioactivation mechanisms involved in the manifestation of centrilobular necrosis in the case of acetaminophen- and carbon tetrachloride-toxicity as opposed to periportal necrosis in the case of allyl alcohol-toxicity. In the former, the abundance of cytochrome P450 plays a critical role whereas in the latter, higher oxygen levels are responsible [2
]. Another reason may be a site-specific batch effect since the allyl alcohol microarray data was generated at a different location than the acetaminophen and carbon tetrachloride microarray data. The training and testing samples are plotted together using PCA of the expression data from the signature of the 21 selected genes (See Additional file 6
). The testing samples show similar distributions as the training samples and also indicate that the necrosis level increases from right to left along PC1. The acetaminophen- and carbon tetrachloride – exposed samples show a similar data dispersion range as the training samples while the ally alcohol – exposed samples are more compressed along the first PC.
Genes from our predictive models (See Tables and ) have biological functions related to the regulation of apoptosis (Ripk3 and Bcl2a1) or are involved in a chemokine\inflammatory response (Ccl13 also known as CCL2/MCP-1), Cxcl16 and Lgals3. Pathway analysis of the predictor genes revealed a central regulating role of tumor necrosis factor (TNF), Jun and TP53 (Figure ). The majority of the predictor genes in the signature (17 out of 21) are regulated in their expression by these transcription factors. Therefore, our results are generally in agreement with the current hypothesis that TNF mediates liver injury and genes such as Jun and TP53 are closely involved in necrotic changes in response to exposure to some hepatotoxicants [14
]. Surprisingly, monocyte chemoattractant protein-1 (MCP-1), a serum factor gene and chemokine that is in our predictor gene list, was shown to have its protein product differentially expressed in acetaminophen-treated rats [22
] and is induced by TNF-α [23
]. This regulation might be a reflection of a repair process following liver injury by acetaminophen-toxic exposure or could be a contributor to the insult. Although the role of MCP-1 in liver injury is controversial [22
], new evidence using MCP-1 deficient mice suggests that interference of the gene's expression is sufficient for altering the processes that lead to severe carbon tetrachloride-induced liver injury [24
]. However, caution must be taken as a more complicated biological response to liver injury is likely since there are hepatotoxicants, such as monocrotaline – (MCT, a pyrrolizidine alkaloid plant toxin), where an inflammatory response ensues secondarily to injury of the liver and TNF-α appears to not be primarily responsible for the hepatotoxicity [25
]. In addition, transcription factors such as TNF-α and TP53 have both pro- and anti-apoptotic effects. TP53 keeps the cell from progressing through the cell cycle if there is damage to DNA but can also cause the cell to enter apoptosis if the damage cannot be repaired. Similarly, TNF-α can induce pro-apoptotic signaling mechanisms [26
] or induce resistance against apoptosis [27
] depending on the overall condition of the cell and its microenvironment.
The reconstructed Bayesian network from the toxic exposures of the hepatotoxicants (Figure ) revealed several gene interactions that are consistent with interactions in the pathway that was generated from curated scientific literature (Figure ) and points to apoptosis-related genes in necrosis-mediated toxicity. Bear in mind that the network is a consensus one, has only positive, one-way, acyclic interactions and was generated from microarray data alone using a limited number of genes. However, the confidence of each gene-to-gene edge (interaction) was calculated by performing 500 simulated annealing searches.