Ethics Statement

The here analyzed data is publicly available from the TG-GATEs database (http://toxico.nibio.go.jp/english/), which has been established during the Japanese Toxicogenomics Project (TGP).

Microarray Dataset

The Affymetrix dataset used for classifier training and evaluation was compiled on the basis of the recently published TG-GATEs database [13]. We selected 2 GC, 9 NGC and 11 NC compounds which could be unambiguously assigned to one of these three compound classes. Experimental evidence for class membership was obtained from genotoxicity tests and long-term animal studies (Table 1). The three compounds Methapyrilene hydrochloride (MP), Pirinixic acid (WY) and Monocrotaline (MCT) were considered as undefined, as these showed characteristics of both genotoxic and nongenotoxic compounds.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Selected compounds from TG-GATEs database.

https://doi.org/10.1371/journal.pone.0097678.t001

For equipotent dose selection, we considered the doses which were tested in long-term animal studies and finally resulted in the development of first hepatocellular carcinomas. In parallel, we investigated the TD₅₀ rates, i.e., the compound dose that leads to tumor formation in half of the tested animals, published in the CPDB database (http://potency.berkeley.edu/). Furthermore, we compared histopathology and clinical chemistry data between different compound administrations within the TG-GATEs database. Altogether, from the three dose levels available for each compound in the database, we selected for each compound individually an appropriate dose which corresponds to 5–30-fold of the TD₅₀ rate. Since dose range finding studies up to 14 days are routinely performed for drug candidates in the pharmaceutical industry, this time point was of major practical interest.

Download and Preprocessing of the Microarray Data

The Affymetrix raw data (CEL files) and the corresponding metadata were downloaded from the TG-GATEs FTP site (ftp://ftp.dbcls.jp/archive/open-tggates/). After importing the Affymetrix raw data into the R software environment, the data was normalized by using the robust multi-chip average (RMA) method implemented in the affy package. In a subsequent annotation step gene symbols and Entrez Gene IDs were assigned to each Affymetrix probeset, based on the corresponding metadata package rat2302.db provided by Bioconductor [17], [18].

Signature Extraction by using an Ensemble of Feature Selection Techniques

In this study we conceived and implemented an approach for the extraction of predictive and generalizable mRNA signatures for the detection of nongenotoxic hepatocarcinogenesis in rat. For the ranking of candidate signature genes we used weights of Support Vector Machines (SVM) with linear kernels, Prediction Analysis for Microarrays (PAM), Recursive Feature Elimination (RFE) and Golub’s Signal-to-noise Ratio (Golub-Ratio) [15], [19]–[21]. However, in principle, any other algorithm producing a complete ranking of the genes depending on their potential to discriminate NGCs from NCs could have been used.

Optimization of the Number of Informative Genes

Each method was applied on 25 random subsamplings of the data (bootstrap samples), each containing 90% of the training data. The remaining 10% of the training data (out-of-bag samples) were used for evaluating the classification performance based on the area under the ROC curve. For this purpose, 11 different signature sizes between 2 and 100 genes were defined. Then ROC scores were estimated based on the predictions of a fast KNN classifier trained on each signature obtained for a certain combination of a method, bootstrap and signature size. By averaging ROC scores across methods, bootstraps and cross-validation folds, a performance could be assigned to each signature size. The optimal number of informative genes was then determined by fitting a spline function which describes the relation between signature size and classification performance and computing its analytical maximum.

Assessment of the Stability

In order to assess the robustness of the signatures depending on the selection method and the number of signature genes, a stability index proposed by Kuncheva was used [22]. Given two feature subsets A and B of size k, which have r features in common and which were selected from a set of n features, the Kuncheva index (KI) serves as a measure for the consistency between A and B and is formally defined as [22].

Obviously, KI(A,B) monotonically increases with r, the number of features shared between A and B. The term k²/n was included to correct for the common selection of features by chance and transforms the KI to a value range between −1 and 1. A value of 0 is expected for independently drawn feature subsets which overlap purely by chance. While negative values of the KI suggest that the selected subsets are more complementary than expected by chance, positive values indicate higher overlaps than expected by chance. The maximum is obtained for two equal subsets and the minimum is reached for two disjoint subsets with |A| = |B| = n/2. The KI can also be computed for m>2 feature subsets by averaging across all pairs of subsets:

Merging of the Signatures

In order to obtain a robust consensus signature which is less dependent on the employed selection method and the compounds used for training, the sets of informative genes selected by diverse methods on different bootstraps were merged. For this purpose, the gene ranks obtained from individual bootstraps and methods were summed up. Then the genes were sorted in ascending order based on their rank sums to obtain a consensus ranking. Given the ranking and the previously determined optimal number of informative genes, the final multi-gene signature was generated.

Specificity Ratio for Signature Extraction

As an alternative to the Ensemble Feature Selection (EFS) approach, which makes extensive use of diverse machine learning algorithms, we conceived a second, purely statistical approach. Our Specificity Ratio (SR) relies on a similar concept as Golub’s ranking method, which selects informative genes, based on high expression differences between classes and small deviations within classes [15]. Golub et al. proposed to calculate for each gene the signal-to-noise ratio , where µ_1g and µ_2g denote the fold-changes of g averaged across all samples in class c₁ and c₂, respectively. σ_1g and σ_2g denote the corresponding standard deviations [15].

Please note that high positive or negative values of r_g may indicate significant expression differences between the two classes. However, r_g does not ensure that the selected genes have class-specific expression profiles in the sense that high expression changes are only present in the class of primary interest. As we aimed at selecting genes, which are up- or downregulated in the primary class c₁ (here: NGC) while being not differentially expressed in the secondary class c₂ (here: NC), we replaced and by the average absolute fold-changes and with in the denominator of Golub’s ratio. In this equation denotes the log₂(fold-change) of a gene g in the i-th compound of class c_j which contains n_j compounds. The total number of compounds is .

In the next step, the gene-specific standard deviations σ_g were shrunken towards the median variance σ₀ averaged across all genes, in order to avoid the overvaluation of genes with marginal differential expression, due to an underestimation of the sample variance. For this purpose, we replaced the gene-specific standard deviations σ_g by the moderated standard deviations used by Opgen-Rhein and Strimmer in their shrink-t statistic [16]. The authors proposed to adjust the variances by using a weighed sum of the gene-specific variances and the overall variances , where the weight of is given by a pooling parameter λ with .

The optimal value for λ is determined based on the variances of the gene-specific variances as defined in the following equations:

Intuitively, the main idea is that if the variance of the gene-specific variances is high, these values could not be reliably obtained from the data and thus, should be adjusted by shrinkage towards the target [16]. Otherwise, if all could be reliably estimated with small variance , the adjustment should be smaller. The complete mathematical derivation of the formulas used for the estimation of the moderated standard deviations was provided earlier by Opgen-Rhein and Strimmer [16]. Finally, on the basis of Golub’s ratio , we defined the specificity ratio .

Post-Filtering of Candidate Signature Genes

Since our ratio s_g is based on the absolute values of the fold-changes, the direction of a gene’s deregulation may vary across the compounds in the primary class c₁. In order to filter genes which are either consistently up- or downregulated, we applied a ROC-based filter to the obtained gene lists. This filter computes a ROC score for each gene by comparing the fold-changes observed for the different compounds to their corresponding compound class. The ROC score reaches its maximum () for a particular gene g, if the fold-changes measured for the compounds of the primary class (here: NGC) are all greater than the fold-changes observed for the secondary class (here: NC). For an equally informative gene which is consistently more strongly downregulated upon treatment with primary class compounds, the ROC score would be 0. If there is no correlation between gene expression and the compound class, we would expect a ROC score of 0.5. As we aimed at filtering both consistently up- and downregulated genes, we computed for each gene the maximum ROC score obtained for regular and inverted class labels, respectively. Then we used a cutoff of to filter relevant signature genes.

Assessment of the Classification Performance

Ultimately, the selected informative genes were incorporated into different prediction models which were evaluated on independent sets of test compounds. To this end, we trained 6 different classifiers based on the learning algorithms SVM, KNN, PAM, Random Forest, Weighted Voting and Naive Bayes and evaluated the performance of these classifiers based on a nested 3×3-fold cross-validation [15], [19], [21], [23]. Within each fold of the outer cross-validation, an inner 3-fold cross-validation was performed on the corresponding training data to determine the optimal model parameters.

Scaling of Classifier Inputs and Outputs

The standardized fold-changes (i.e., z-scores) of the informative genes were used as features for model construction. Specifically, for each gene the mean μ and the standard deviation σ of the fold-changes were calculated and then each fold-change was transformed into a z-score . For the sake of simplicity and better comparability, the prediction scores returned by the classifiers were mapped to a value range between 0 and 1. As SVM outputs were contained in the interval , where n varies between models depending on the used training data and hyperparameter C, the sigmoidal function with was used for scaling. The confidences obtained from Weighted Voting classifiers, which originally ranged between −1 and 1, were transformed using the linear function with . For KNN we computed continuous outputs, based on the function with where k denotes the number of nearest neighbors and x ≤ k is the number of positively labeled ones among these.

Pathway Analysis

Enrichment analysis against KEGG pathways was performed to investigate putative mechanisms of action [24]. Overrepresentation of genes from specific pathways among the signature genes was determined by a hypergeometric test. Given a set of n signature genes of which m are contained in a certain pathway and provided that the union of all pathways contains N genes of which M are in the considered pathway, the p-values were computed according to the following formula:

The resulting p-values were corrected for testing multiple pathways using the FDR method by Benjamini and Hochberg [25]. As criterion for significant enrichment we used a p-value cutoff of 0.05.

Implementation

All plots shown in the results section were generated using the R programming language for statistical computing and packages from the Bioconductor library [17]. For SVM training and classification the libsvm implementation from the SHOGUN toolbox was used [19]. RFE was implemented in R on top of the linear kernel SVM from SHOGUN [19]. Golub-Ratio and Weighted Voting were implemented from scratch based on the description of the algorithm from Golub et al. [15]. For KNN, PAM, Naive Bayes and Random Forest the R packages knnflex, pamr, klaR and randomForest were used [21], [23]. In order to assess the prediction accuracy of the classifiers ROC curves were generated using the ROCR package [26].

Classification Performance Achieved with Novel EFS and SR Method

The standard gene selection methods Golub-Ratio, SVM, PAM and RFE as well as the statistical inference methods t-test, Wilcoxon rank-sum test, and permutation test were employed to infer predictive signatures for the reliable discrimination of NGCs from NCs, based on Affymetrix data from Sprague-Dawley rats treated for 14 days in a repeated dosing study. Next, a consensus signature was compiled by integrating the individual gene rankings obtained from diverse ensembles of gene selection methods across varying subsamplings of the training data (bootstraps). The idea of this approach is firstly, that the integrated signature contains only genes which were deemed informative by multiple independent selection methods. Secondly, it is characterized by increased robustness against variations in the set of training compounds, as the respective genes were highly ranked for multiple random subsamplings of the data. We inferred predictive signatures based on three different variants of our EFS methodology, which involve the use of standard gene selection methods, the statistical inference methods t-test, Wilcoxon rank-sum test, and permutation test, or both types of ranking algorithms. Depending on the employed ensemble of methods the resulting signature size ranges between 36–54 probesets referring to 24–45 annotated genes (Table S1).

A second signature was compiled by selecting the 100 genes with the strongest NGC-specific expression pattern according to our Specificity Ratio. As described in more detail in the methods section we refined the signature by using a ROC-based filter, which excludes genes showing inconsistent regulation states upon NGC treatment. The final SR signature contains 29 probesets which are attributed to 24 genes (Table S1).

We evaluated the classification performance by incorporating each of our signatures into prediction models, which were constructed using established supervised learning algorithms. For each classifier an averaged ROC curve was generated based on the results from a 3×3 nested cross-validation (Figure 2). Apparently, the performance of the EFS method heavily depends on the employed ensemble of feature selection techniques. While depending on the adopted classifier good or excellent classification accuracy was found for standard gene selection methods, the performance achieved with t-test, Wilcoxon rank-sum test, and permutation test was considerably weaker (Figure 2A, B). By combining both standard methods for gene ranking and the three evaluated statistical methods it was also possible to construct fairly well generalizable prediction models (Figure 2C). However, despite the increased complexity, the models did not improve in terms of prediction accuracy. Therefore, we did not pursue this approach any further. The steep initial slope of the ROC curves shown in Figure 2D indicates that the SR method also permits the specific detection of NGCs at a low false positive rate. However, when compared to the EFS signature, the SR signature is clearly limited in terms of sensitivity, which may in part be owed to the stringent selection of genes with NGC-specific expression patterns. PAM, Weighted Voting and Random Forest were found to be the most reliable classification methods for both the EFS and the SR signature. When compared to these classifiers, KNN and Naive Bayes tend to result in worse prediction accuracies. The SVM classifier performs well in conjunction with the EFS signature, but is not recommended for use with the SR signature.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 2. Evaluation of EFS-based and SR-based signatures on TG-GATEs data.

The ROC curves obtained from different cross-validation folds were averaged based on the thresholds for class discrimination and drawn separately for each of the six classification methods. The classifiers evaluated here were trained on features selected using our (A–C) EFS methodology in conjunction with (A) the standard gene selection methods Golub-Ratio, PAM, SVM and RFE, (B) the statistical inference methods t-test, Wilcoxon rank-sum test and permutation test or (C) all previously stated methods. (D) The prediction accuracy was also determined for the SR signature-based models and the corresponding ROC curves were generated as described previously.

https://doi.org/10.1371/journal.pone.0097678.g002

Comparison of Feature Selection Methods in Terms of Accuracy and Robustness

Besides assessing the predictive power of EFS-based signatures depending on the applied ensemble of methods, we also evaluated the robustness of each individual method against variations of the training data. For this purpose, we used the stability index KI proposed by Kuncheva (see methods section) [22]. Concurrently, the average performance of the individual signatures inferred from varying subsets of the training data was measured in terms of the ROC scores on the 10% out-of-bag samples, which were excluded from each bootstrap. For both scores the optimum is 1, but the values assigned to the worst cases are different. While a ROC score of 0.5 corresponds to the performance achieved by a random guesser, a KI of 0 would be expected for randomly drawn feature subsets.

The accuracy and stability of the signatures selected by different methods is depicted depending on the number of informative genes in Figure 3. According to the curves in Figure 3A the highest average classification accuracy could be achieved with the methods t-test, Wilcoxon rank-sum test and permutation test, which select informative genes based on statistical inference. For these statistical tests the maximum accuracy was attained with at most 30 informative genes. However, as illustrated in Figure 3B a relatively low consistency was observed between the gene sets selected on different bootstraps. Similarly, a relatively low signature stability was found for Golub’s ranking method, which also relies on a statistical comparison of the classes. Golub’s method reached its maximum accuracy in conjunction with a signature size of 40 (average ROC score: 0.79). The decrease of performance which is observable for increased signature sizes (>40 genes) may be partly due to overfitting, caused by the selection of genes with minor relevance for the problem of distinguishing NGCs from NCs. A marginally weaker performance than for the Golub-Ratio was found for the PAM method, which also reaches its maximum accuracy when only the 40 top ranked genes are used for model construction. However, the informative genes selected by PAM are clearly less dependent on the subsampled part of the training data, which becomes apparent from Figure 3B. Strikingly, a considerably higher robustness against sampling variation was found for the two SVM-based methods. However, the classification accuracy achieved with signatures from SVM and RFE was significantly lower when compared to Golub-Ratio and PAM, especially for smaller signature sizes. A comparable, moderate performance, but significantly increased robustness can only be observed for larger signature sizes, for instance, when 70 informative genes are selected.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 3. Accuracy and stability of signatures depending on gene selection method and signature size.

(A) The line plots illustrates the mean classification performance achieved by a KNN classifier based on diverse signatures selected by the methods Golub-Ratio, SVM, PAM, and RFE. The performance was measured in terms of area under the ROC curve for 11 linearly spaced numbers of informative genes ranging from 2 to 100. Points and whiskers correspond to the means and the standard deviations of the ROC scores assessed on the out-of-bag samples of 25 bootstraps during 3-fold cross-validation. The adjacent bar plots depict the corresponding ROC scores averaged across signature sizes and cross-validation folds. (B) The Kuncheva index (KI) was used to score the correspondence of signatures selected on different bootstraps depending on the selection method and the number of selected genes. The average KI obtained for each method is shown in the bar plot on the right.

https://doi.org/10.1371/journal.pone.0097678.g003

Comparison Against Existing Signatures for NGC Prediction

We then compared our novel signatures to several already available ones which had been derived using widely used supervised classification methods [6], [10], [11], [27]–[30]. Furthermore, we determined which genes have been selected most frequently. Besides emphasizing the most relevant informative genes, we also point out shared mechanistic characteristics, which were identified based on pathway enrichment analysis.

The classification performance, that has been achieved by 6 representative machine learning methods, was assessed for all signatures using cross-validation and measured based on the area under the ROC curve. The distributions of the resulting ROC scores are illustrated as box plots in Figure 4A. Another illustration, which depicts the predictions made for the individual compounds in more detail, is shown in Figure 4B.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 4. Performance comparison against signatures known from the literature.

(A) Box plots show the distribution of ROC scores resulting from 3-fold cross-validation of 6 different classifiers incorporating either one of 7 published signatures or one of our 2 novel signatures. Each box corresponds to a certain multi-gene signature. The lower, center, and upper line of each box indicates the lower quartile, median, and upper quartile of the ROC scores obtained for a certain signature. The whiskers extend to the most extreme ROC score which is within 1.5 times the interquartile range from the box. Outliers are depicted as individual points. (B) The heatmap shows the binary classification outcomes of 6 predictors used for NGC vs. NC discrimination, depending on the incorporated signature. The colors indicate the number of methods which classified a certain compound as NGC and NC, respectively. Cells are drawn in blue if a compound was classified as NGC and in green if it was classified as NC by the majority of predictors. Grey color indicates a tie between the two classes. In order to discretize the continuous prediction scores, the class discrimination cutoff was chosen individually for each predictor as an optimal tradeoff between sensitivity and specificity.

https://doi.org/10.1371/journal.pone.0097678.g004

The boxes shown in Figure 4A clearly indicate that the highest average classification performance is achieved by classifiers incorporating our novel 54-probeset EFS signature (mean ROC score: 0.95). In this performance comparison the second best signature was the one proposed by Uehara in 2011 followed by our SR signature (Table 2). The signatures by Ellinger-Ziegelbauer and Nakayama also show a trend towards higher accuracies when compared to the rest. Consistent with our expectations the Uehara (2008) signature was outperformed by the more recent Uehara (2011) signature, as it primarily captures changes in gene expression present after 24 hours upon one-time NGC exposure, which are of minor relevance for the 14-day administration setting evaluated here. Similarly, the low performance found for the Auerbach signature, which was designed for longer treatment durations up to 90 days, must be seen in the context of its limited applicability to the 14-day setting.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Published and novel signatures for NGC prediction in rat liver.

https://doi.org/10.1371/journal.pone.0097678.t002

In order to enable a conclusive and fair comparison to each of the existing signatures, we also applied our approaches to the original datasets from which these signatures were identified. Consistent with our evaluation on the TG-GATEs data, the models constructed from the EFS-based signatures achieved the highest prediction accuracy on 5 of 6 published toxicogenomics datasets, when compared to classifiers trained on the original and SR-based signatures, respectively (Figure 5). The corresponding ROC curves can be found in Figure S1-S6. Only on the dataset compiled by Auerbach et al., the SR-based signature performed better. Considering the average performance across all datasets and signatures, the best classifier was Weighted Voting (mean ROC score: 0.88), followed by PAM and Random Forest.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 5. Evaluation of EFS-based and SR-based signatures on datasets from previous studies.

The bar plots depict the area under the ROC curve achieved by specific prediction models built and evaluated on datasets which have been used in related toxicogenomics studies. The prediction models differ in the adopted classifiers and in the incorporated signatures. On each dataset, two signatures were extracted using the EFS and SR method, respectively, and compared to the signature from the original study in terms of classification performance. Each dataset corresponds to a certain panel (see panel headers), each signature is represented by a group of bars, and the classifiers are indicated by different colors (see legend).

https://doi.org/10.1371/journal.pone.0097678.g005

Besides comparing our novel signatures to published ones in terms of classification accuracy, we determined which informative genes were commonly selected (Figure 6A–C). Among these genes is, for instance, App, which is a key protein in the pathomechanism of Alzheimer’s disease, but was also reported to be upregulated in different cancer types, and shown to play a crucial role in the growth control of pancreatic and colon cancer [31]. Another highly informative gene is Cdkn1a, which is a p53-dependent key regulator of cell fate, as it triggers cell cycle arrest in the G1 phase under various stress conditions, such as DNA damage [32]. Mechanistically similar growth-inhibitory effects were reported for Ccng1, which is also known to be transcriptionally regulated by p53, and which was shown to induce pRb-dependent G1 phase arrest when being highly expressed [33]. Phlda3, which was proposed as a potential biomarker in diverse toxicogeniomics studies [6], [10], [29], is a known target gene of p53, too. Phlda3 may act as a tumor suppressor by inhibiting the translocation and activation of Akt, thereby preventing negative regulation of p53-dependent apoptosis via Akt [34]. The Akr7a3 transcript, which is part of our proposed signature and which was mostly found upregulated in response to carcinogen treatment, was previously considered as informative by Ellinger et al., Nakayama et al. and Uehara et al. [6], [10], [29]. Consistent with these results, in a recent explorative study aiming at the identification of novel candidate biomarkers for liver cancer, an upregulation of Akr7a3 on the protein level was observed in rat hepatomas [35]. Another frequently used informative gene is Abcb4 alias Mdr3. This multidrug-resistance gene is typically expressed in diverse tumors, rendering them less sensitive to treatment with anti-cancer drugs (Figure 6C) [36].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 6. Comparison of predicted and known signature genes.

(A) The Venn diagram depicts the overlap between the informative genes from the EFS and SR signature and the genes contained in 7 previously reported signatures. The symbols of known marker genes independently confirmed by our approach are listed in the intersection of the two sets. (B) This diagram illustrates the overlaps between the two novel signatures predicted using the Ensemble Feature Selection (EFS) method and the Specificity Ratio (SR), respectively. (C) The histogram shows the absolute selection frequencies of informative genes which are part of the 2 novel signatures and the 7 previously published ones. The genes which are included in 2 or more signatures are named on top of the corresponding bars. (D) For each signature a pathway enrichment analysis against KEGG was performed and then a histogram was generated that shows which pathways were most frequently detected as enriched among the informative genes. Some of the KEGG pathways which were detected to be significantly enriched in two or more signatures are listed on top of the respective bars.

https://doi.org/10.1371/journal.pone.0097678.g006

In addition to determining the most commonly selected informative genes, we also compared the signatures with respect to their associated pathways (Table S2). It can be concluded from Figure 6D that signatures for early NGC detection frequently include genes involved in energy metabolism, anabolic metabolism, drug metabolism (e.g., by CYP enzymes), and DNA-damage response (e.g., via p53 pathway). These changes clearly reflect a response to carcinogens including starting cellular hyperplasia upon treatment. However, it should be noted that the early molecular events captured by the proposed signature genes are not specifically related to nongenotoxic mechanisms, of which a wide variety have been proposed for rat liver [3]. As representing the complete spectrum of NGC mechanisms in one gene list is hardly feasible, most of the current signatures for NGC detection are also sensitive to genotoxic carcinogens (Figure S7).

Expression Profile Analysis of Undefined Compounds

For three of the analyzed compounds, namely MP, WY and MCT a definite assignment to either GCs or NGCs was not possible based on published animal studies and genotoxicity assays [37]–[39]. As only two unambiguously classified GCs (AAF and DEN) are available from TG-GATEs, we take the view that it is not feasible to construct generalizable toxicogenomics models for reassigning the undefined compounds to either GCs or NGCs. Nevertheless, we compared the expression profiles of these undefined substances to the ones observed for unambiguously classified compounds, and evaluated whether a detection of their carcinogenic potential was possible based on our multi-gene signatures.

In order to generate a graphical representation of the compounds’ expression profiles, the dimensionality of our predicted signatures was reduced to a two-dimensional space using PCA (Figure 7A–B). Obviously, a clear separation between NGC- and NC-treated sample groups is possible, based on both the EFS and the SR signature after 14 days of repeated dosing. While for most published signatures considerable overlaps could be observed between the NGC and the NC clusters (Figure S8), only a small overlap, caused by the compound CMA, was found for the EFS and SR signature, respectively. Consistent with this observation, the expression changes in the top 10 genes of our EFS signature (Figure 8) and SR signature (Figure S9) are less pronounced for CMA when compared to the other NGCs. It also becomes apparent from Figure 8 that the expression differences between NGCs and NCs are much more striking than the differences between NGCs and GCs. This finding is consistent with the fact, that the EFS signature is also sensitive to GCs. In general, it can be concluded that owing to the circumstance that the GC class was not considered during informative gene selection, the inferred signatures are not specifically related to nongenotoxic mechanisms, but rather to hepatocarcinogenesis in general. Nevertheless, individual informative genes (e.g., Me1) exist, that are clearly differentially expressed between GC and NGC (Figure 8), which may in part explain the isolated positions of the genotoxic compounds DEN and AAF in PCA space (Figure 7A).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 7. Separation and classification of compounds based on EFS and SR signature.

(A) The dots correspond to different treatment groups and are colored according to the classes of the compounds used for treatment. Each treatment group was originally represented by a vector composed of the fold-changes of the 54 signature genes measured after 14 days of repeated dosing. In order to inspect the compound-specific expression profiles in a lower-dimensional space, these vectors were transformed to the first and second principal component resulting from PCA. In order to highlight clusters of NGCs and NCs, convex hulls were drawn around the respective compounds. The compounds WY, MP and MCT were considered as undefined, due to ambiguous outcomes of published studies. (B) PCA plot similar to (A), but generated on the basis of the SR signature. (C) The heatmaps depict the confidence of the predictions made by diverse classifiers for assessing the carcinogenic potential of GCs (AAF, DEN) and undefined compounds (MP, WY, MCT). Columns represent compounds and rows correspond to classifiers. The compound classes are indicated by the colorbar on top. The discrimination between carcinogens (blue) and non-carcinogens (green) was done based on the EFS signature. (D) Toxicogenomics-based assessment of the carcinogenic potential of GCs and undefined compounds using diverse classifiers which incorporate the SR signature genes as predictive features.

https://doi.org/10.1371/journal.pone.0097678.g007

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 8. Expression profiles of EFS signature genes.

Shown is a heatmap depicting the fold-changes of the top 10 informative genes from the predicted EFS signature. Rows represent genes and columns represent treatment groups. Cell colors indicate the strength and direction of differential expression relative to the corresponding control groups (red: upregulation, green: downregulation). Treatment groups which belong to different compound classes are separated by solid vertical lines. The respective classes are indicated by the color bar on top of the heatmap.

https://doi.org/10.1371/journal.pone.0097678.g008

In the PCA plots the three undefined compounds MP, WY and MCT were found to be very distinctly positioned (Figure 7A–B), which indicates compound-specific differences in the expression profiles of the selected informative genes. While the profile of MP resembles the one of the genotoxic compound AAF, WY is located near the nongenotoxic substances CFB and FFB (Figure 7A–B). MCT also tends to show a rather NGC-like profile similar to the ones observed for CCL4, PHE, and EE. Whereas a certain similarity between the profile of WY and those of other NGCs was also apparent from PCA representations obtained from most published signatures (Figure S8), this was not the case for MCT which mostly clustered in the vicinity of the genotoxic compound DEN.