Genomic indicators in the blood predict drug-induced liver injury

Abstract

Genomic biomarkers for the detection of drug-induced liver injury (DILI) from blood are urgently needed for monitoring drug safety. We used a unique data set as part of the Food and Drug Administration led MicroArray Quality Control Phase-II (MAQC-II) project consisting of gene expression data from the two tissues (blood and liver) to test cross-tissue predictability of genomic indicators to a form of chemically-induced liver injury. We then use the genomic indicators from the blood as biomarkers for prediction of acetaminophen-induced liver injury and show that the cross tissue predictability of a response to the pharmaceutical agent (accuracy as high as 92.1%) is better than, or at least comparable to, that of non-therapeutic compounds. We provide a database of gene expression for the highly informative predictors which brings biological context to the possible mechanisms involved in DILI. Pathway-based predictors were associated with inflammation, angiogenesis, Toll-like receptor signaling, apoptosis and mitochondrial damage. The results demonstrate for the first time and support the hypothesis that genomic indicators in the blood can serve as potential diagnostic biomarkers predictive of DILI.

INTRODUCTION

Drug-induced hepatotoxicity is the most frequent cause for a drug to be withdrawn from the market, to have its use restricted or to have a warning on the label associated with it. Currently, preclinical models are not always predictive of an adverse response in humans since sensitivity to certain drugs can be influenced by human-specific genetic variability and other variables accounting for idiosyncratic drug reactions. Discovery of diagnostic indicators of liver injury from a minimally invasive bio-available source is of interest to monitor for adverse effects of a drug.

A few recent studies have shown that genomic markers obtained from blood gene expression data are predictive of adverse effects of a drug or chemical compounds. Bushel et al. ¹ demonstrated that gene expression profiles from rat blood samples could accurately predict exposure levels of acetaminophen to the rat liver better than traditional clinical panels. Lobenhofer et al. ² used gene expression data from rats exposed to a compendium of hepatotoxicants to show that blood gene expression patterns could be used to provide an indication of the severity level of liver injury. Wang et al. ³ recently demonstrated the potential of circulating microRNA molecules (small regulatory non-coding RNAs) as biomarkers of drug-induced liver injury (DILI) in an acetaminophen-overdosed mouse model system. Although these efforts are certainly pioneering they do not address the question of whether or not genomic indicators in the blood are truly predictive of DILI.

In this paper we used the Lobenhofer et al. ² gene expression data set that was contributed to the Food and Drug Administration (FDA) led MicroArray Quality Control Phase-II (MAQC-II) effort as a training data set and for internal validation to identify genes and biological processes in the blood that are predictive of liver necrosis (a particular form of DILI). This data set was chosen for analysis for several key reasons: 1) it was the only data set publicly available at the time which contained gene expression measurements from the two tissues of interest, histopathology, clinical chemistry and other ancillary biological data from exposure to a compendium of compounds, 2) the experimental design and data acquisition were performed in a rigorous, standardized fashion (i.e. a common array platform, experimental procedure, data acquisition and analysis methods) in order to reduce the amount of systematic variation in the data, and 3) it permitted a global view of the landscape of the transcriptome of the rat as a model system to explore the possibility of using expression profiles from a non-invasive tissue as potential biomarkers predictive of DILI. We show that the classifiers derived from the gene expression data are highly predictive across tissues (blood and liver) and microarray platforms (Agilent and Affymetrix). We then demonstrate that gene expression profile sets from the blood predict acetaminophen-induced liver injury (samples classified as the subjects having either some or no observable form of liver necrosis as an end-point) in an independent (validation) data set better than (at an accuracy as high as 92.1%), or at least comparable to, that of non-therapeutic compounds used in this study. Cumulatively, these data support the hypothesis that genomic indicators in blood can serve as biomarkers that are highly predictive of a form of DILI and as a model for the acquisition of gene expression signatures that potentially can be used in a clinical setting for monitoring drug treatments and diagnosing adverse drug effects in humans.

MATERIALS and METHODS

Compendium (standardized) gene expression data

The gene expression data set was derived from the studies of the exposure of rats to one of eight compounds (1,2-dichlorobenzene, 1,4-dichlorobenzene, bromobenzene, monocrotaline, N-nitrosomorpholine, thioacetamide, galactosamine and diquat dibromide). The data are publicly available in the Gene Expression Omnibus (GEO) database under the MAQC-II reference series accession GSE16716 and in the Chemical Effects in Biological Systems (CEBS) database ⁴ under accession number 001-00001-0020-000-4. All eight compounds were studied using standardized procedures, i.e. a common array platform, experimental procedures and data retrieving and analysis processes. For details of the experimental design see Lobenhofer et al.². Briefly, for each compound, four to six male, 12 week old Fischer F344/N rats were exposed to a low dose, mid dose(s) and a high dose of a compound and sacrificed at 6, 24 and 48 hrs later. For each time point, control animal groups were treated with vehicle alone. At necropsy, liver and blood were harvested for RNA extraction, histopathology, clinical chemistry and hematology assessments.

Both Agilent (Agilent Technologies, Inc., Santa Clara, CA) and Affymetrix (Affymetrix, Inc., Santa Clara, CA) platforms were used for the gene expression profiling. The cross tissue predictions used the Agilent data only whereas the cross platform predictions used both Agilent and Affymetrix data. For the Agilent platform, RNA isolated from the liver from each of the treated rats was labeled and hybridized against the time- and compound-matched control pool to Rat #G4130A oligonucleotide (22,075 probes) arrays. Fluorescent pixel intensities measurements were acquired using an Agilent DNA Microarray Scanner and processed with the Agilent Feature Extraction software. The averaged dye-swap ratio of the pixel intensity values (background subtracted and channel normalized red and green processed signals) were used to represent the gene expression profiles for 318 samples. The same approach was carried out for the rat blood samples. For the Affymetrix platform, the gene expression data were generated only for the rat liver. Specifically, the RNA samples from individual animals were profiled, one hybridization per animal on Rat Genome 230 2.0 (31,099 probe-sets) arrays for a total of 418 liver hybridizations. The data were processed using the MAS5 algorithm ⁵. The signals were background subtracted, averaged (across probes within a probe-set) using a mean Tukey biweight function and then scaled to account for differences between chips. The intensity data from the compound-treated samples were used to represent the gene expression profiles for 318 samples. The classification of the samples is described in the “Histopathology and sample classification” subsection.

Independent (non-standardized) validation gene expression data

The gene expression data is from the Agilent platform only and is comprised from liver samples of rats exposed to one of three hepatotoxicants (acetaminophen, carbon tetrachloride, and allyl alcohol) with a time-matched vehicle control pool made for each compound and tissue by pooling equal amounts of RNA from the control animals. Acetaminophen: Groups of four male Fischer F344/N rats each received 0 (vehicle), 50, 150, 1,500 or 2,000 mg/kg body weight of acetaminophen at two different times: between 12 (noon)-1PM (“light” subjects) or between 12 (midnight)-1AM (“night” subjects). The animals were sacrificed after 6, 18, 24 or 48 hrs. RNA samples from eight subjects were either not available or did not produce high quality RNA for hybridization leaving a total of 152 samples. Carbon Tetrachloride: Groups of six male Fischer F344/N rats each received 15, 750 or 2,000 mg/kg body weight of carbon tetrachloride. The animals were sacrificed after 3, 6, 24 or 72 hrs. There are a total of 72 samples. Allyl Alcohol: Groups of six male Fischer F344/N rats each received 10, 20, 40 or 50 mg/kg body weight of allyl alcohol. The animals were sacrificed after 6, 24, 48 or 72 hrs. The RNA from one subject was not available or did not produce high quality RNA for hybridization leaving a total of 95 samples.

Each treated animal was hybridized against a time matched control pool to Agilent Rat #G4130A oligonucleotide arrays with a dye-swap technical replicate. The data were acquired and extracted in a similar fashion as the Compendium gene expression data. The averaged dye-swap ratio of the pixel intensity values (background subtracted and channel normalized red and green processed signals) were used to represent the gene expression profiles for the 319 validation samples. For more details see Huang et al. ⁶ and Bushel et al. ¹. The acetaminophen data is publicly available in the CEBS database ⁴ under accession number 002-00001-0011-000-5. The allyl alcohol and carbon tetrachloride data are stored in the NIEHS MicroArray Project System (MAPS) database ⁷ under project ID 221 and 236 and is available upon request. The classification of the samples is described in the “Histopathology and sample classification” subsection. ArrayTrack ⁸ was used to manage the microarray data, histopathology observations and clinical chemistry measurements for this study.

For comparison of the prediction of the blood and liver samples across array platforms, the gene expression data (ratio values for Agilent [blood] and intensity measurements for Affymetrix [liver]) were batch corrected as follows: After log transformation, the mean of the gene expression for each array feature across all the samples within each batch (array platform type or Compendium and validation data sets) is set to zero. This approach is also referred to as Mean Shift, Mean-Centering, or One-Way ANOVA Adjustment.

Histopathology and sample classification

Two sections were taken from the left liver lobes and fixed in 10% formalin. After dehydration with ethanol, the liver sections were embedded in paraffin and H&E stained slides were made. The slides were evaluated by independent pathologists and any disagreements were resolved by a pathology working group review ⁹. The severity of necrosis was graded into five levels by the pathologists, i.e., 0, 1, 2, 3, and 4 representing none, minimal, mild, moderate and marked levels of necrosis respectively. The necrosis severity levels were used as a class label for the samples. Specifically, histopathological severity scores (1–4) of any one of four areas (centrilobular hepatocyte necrosis, centrilobular mid-zonal hepatocyte necrosis, mid-zonal hepatocyte necrosis or focal hepatocyte necrosis) were used to classify samples with at least some observable sign of necrosis (class 1; n = 154 Compendium data set samples and n = 127 validation data set samples). All other liver injuries and no injury observed samples were classified as having no observable sign of necrosis (class 0; n = 164 Compendium data set and n = 192 validation data set samples). See the Supplementary files A and B for the specific classification assignments of the Compendium and validation data sets samples respectively.

Clinical chemistry

At sacrifice, blood was collected into serum separation tubes (BD Microtainer® Tubes, Becton-Dickinson, Franklin Lakes, NJ) and serum was separated. Clinical chemistry analyses were performed on all rats in the Compendium data set at study termination. Serum levels of the established liver injury marker alanine aminotransferase (ALT) are used routinely to assess hepatocyte injury in both animals and humans. Data from the other analytes were not used in this study but are publicly available ^{2, 4}.

Classifier building and prediction

Classifiers were built and used for prediction according to the MAQC-II common practices for developing and validating microarray-based predictive models. For gene-based classifiers a sequential forward array feature selection approach with Welch t-test (fold change [FC] > 1.5 or 2 and P <0.05 criteria) comparisons of two groups (class 1: some observable form of necrosis vs class 0: samples with no observable form of necrosis) was used. The data are considered to be independent and assumed to be normally distributed. The measure of variability is the standard deviation. Optimization of the features selected as predictors was performed by a five-fold internal cross-validation strategy with the Compendium data set samples split into training (n=175) and testing (n=143) subjects. Support vector machines (SVM), k-nearest neighbors (KNN) and nearest centroid (NC) classifiers or a random forest (RF) classifier with 100 trees were used to predict the class of the samples (subjects having either some or no observable form of liver necrosis). The prediction using GeneGo’s canonical pathway maps (CPMs) was performed using the randomForest package in R. GeneGo’s canonical pathway maps were derived from an ontology of experimentally-confirmed signaling and metabolic multi-step pathways in human, mouse and rat ¹⁰. The pathways were manually inferred and curated from primary scientific literature. To date there are over 1100 pathway maps in total for normal and disease states. The genes of each of the CPMs were mapped to the array probes and considered as features for prediction using the RF classifier ¹¹. As a result, there are about 350 classifiers built. An internal out-of-bag (OOB) process (a variant of cross validation) was used to estimate prediction performance and to rank the CPMs classifiers by cross validation accuracy. Briefly, 2000 bootstrap (random) samples were taken where at each iteration about one-third of the blood samples from the Compendium data set are left out of the construction of the k^th tree [k=100] and then used for prediction. OOB error is estimated as the proportion of times that the predicted class of the samples is not equal to the true class averaged over all predicted cases. The best pathways (those with the highest cross validation prediction accuracy) were chosen for final prediction on the liver test data. On average, the probability of the informative genes in one of the best pathways to be highly predictive of drug-induced liver injury (DILI [characterized as the subjects having either some or no observable form of liver necrosis]) by chance given the blood and liver Compendium data set is 1×10⁻⁴. This means that the probability of any random set of genes with the same size as the one of the best pathway to be as predictive, or better than, the informative genes for the considered pathway is substantially small.

Coherent co-expression (cc) - biclustering

In the first step of cc-Biclustering ¹² of the Compendium (Agilent) microarray data set, we use a pairwise approach to obtain subsets of the liver gene expression samples and the genes as initial coherent biclusters. Then we apply the Extracting Patterns and Identifying co-expressed Genes (EPIG) method ¹³ to the initial coherent biclusters in order to further subset the genes into final biclusters that contain coherent and highly co-expressed genes. EPIG uses a filtering process to extract gene expression patterns and then categorizes each gene to one of the patterns for which it has the highest correlation with the gene profile. Briefly, the relationship between two gene expression vectors a_ik and a_jk (i^th and j^th genes) for the samples within the k^th group is assessed with a binary coherent matrix H(h_(ij),k) according to an inclusion\exclusion criterion function

$$h_{(ij),k}=\{\begin{array}{cc}1& ifCM(a_{i,k},a_{j,k})<p_t\\ 0& otherwise\end{array},$$ (1)

where i and j are from 1 to N number of genes and k is from 1 to K number of groups (a given compound used for exposure). The k^th group contains the samples exposed to the compound at the treatment doses and time points. CM represents a coherent measure between these two vectors. CM is the p-value of the Pearson correlation (r-value) between a_ik and a_jk. p_t is a user-defined threshold for the p-value and is set to 0.001.

Biological processes over-representation and pathway analysis

The Expression Analysis Systematic Explorer (EASE) ¹⁴ was used to identify biological processes over-represented by sets of genes identified as predictors or contained within biclusters. The over-represented processes were confirmed by the Gene Ontology (GO) Enrichment Analysis Software Toolkit (GOEAST) ¹⁵ using the Adrian Alexa's improved weighted scoring algorithm to account for the hierarchical structure of GO ¹⁶ and the Benjamini-Yekutieli procedure to control the false discovery rate under the assumption of gene-to-gene dependency ¹⁷.

RESULTS

Hepatocellular injury classified by severity of necrosis

As detailed in Table 1, exposure to any one of the eight compounds in the Compendium data set resulted in necrosis of the liver which was scored using a five-point scale (0, 1, 2, 3, and 4 representing none, minimal, mild, moderate and marked levels of necrosis respectively). The majority of the rats showed no histopathological (observable) evidence of necrosis of the liver. Few samples were found to have a necrosis score that was concordant with levels of alanine aminotransferase (ALT) and the variation of the ALT measure among samples sharing the same necrosis severity score is wide (Figure 1a). When the samples were classified based on the presence or absence of necrosis and a t-test was performed on the blood gene expression data, six genes (Table 2) were found to partition the liver samples fairly well (Figure 1b).

Table 1.

Severity scores and classification of necrosis

			Serverity scores
Compound	No necrosis	Some sign of necrosis	0	1	2	3	4	Total
1,2-Dichlorobenzene	17	17	17	8	5	2	2	34
1,4-Dichlorobenzene	31	5	31	4	1	0	0	36
Bromobenzene	16	20	16	7	5	0	8	36
Diquat dibromide	50	22	50	10	6	4	2	72
Galactosamine	18	18	18	7	8	2	1	36
Monocrotaline	16	16	16	11	1	0	4	32
N-nitrosomorpholine	12	24	12	17	2	1	4	36
Thioacetamide	4	32	4	18	1	5	8	36
Total	164	154	164	82	29	14	29	318

No necrosis means that the liver samples from the rats exposed to a given compound received a histopathology severity score of 0 (no sign of necrosis seen). Any sign of necrosis means a sample received a histopathology score of 1 to 4 (at least some sign of necrosis seen).

Figure 1.

The Compendium data set samples partitioned by necrosis of the liver. a) Distribution of the samples by ALT (x-axis) and necrosis score (y-axis). Necrosis score: 0 (164 samples), 1(82 samples), 2 (29 samples), 3 (14 samples), and 4 (29 samples). The total number of samples is 318. b) Principal component analysis (PCA) of liver samples labeled according to an indication of necrosis. PCA was performed on the liver expression data from the six genes selected from the blood signature using a Welch t-test with P <0.05 and a fold change (FC) >2.0 filtering criteria in ArrayTrack to compare the two classes of samples. Triangles = class 1 (154 samples showing some form of liver necrosis), circles = class 0 (164 samples showing no sign of necrosis)]. The percent of variation captured by the first three principal components (PCs): PC1=76.3% (x-axis), PC2=15.9% (y-axis), and PC3=2.8% (z-axis).

Table 2.

Genes that separated the samples and were found most frequent in the gene-based classifiers that predicted liver necrosis as a form of DILI.

Separates samples by necrosis*	Probe ID	Symbol	Description
✓	A_42_P820657	Il1r2	interleukin 1 receeptor, type II
✓	A_42_P695401	Ccl2, MCP1	chemokine (C-C motif) ligand 2
✓	A_42_P597580	Cxcl10	chemokine (C-X-C motif) ligand 10
✓	A_43_P12944	S100a8	S100 calcium binding protein A8 (calgranulin A)
✓	A_42_P457572	Serpinb1a	serine (or cysteine) proteinase inhibitor, clade B, member 1a
	A_43_P12170	Mmp8	matrix metallopeptidase 8 (neutrophil collagenase)
	A_42_P733209	Dgat2	diacylglycerol O-acyltransferase 2
✓	A_43_P17175	Serpinb1a	serine (or cysteine) proteinase inhibitor, clade B, member 1a
✓	A_42_P620915	S100a9	S100 calcium binding protein A9 (calgranulin B)
	A_43_PP11474	Hp	haptoglobin

^*Check mark denotes the six genes (seven probes) which separated the samples by necrosis in the PCA.

Cross tissue predictability

Several classifier building strategies were used to predict the Compendium data set liver tissue samples (classified as subjects having either some or no observable form of liver necrosis) using gene expression data from the blood and the reciprocal prediction (liver to blood) (Figure 2a). Figure 2b and Table 3 present the accuracies of the predictions using the gene-based models. When considering all cases of the models, the genes selected using the fold change (FC) > 2.0 criteria coupled with a p-value < 0.05 performed better than using a 1.5 cut-off with the same p-value. The nearest centroid (NC) classifier yielded the highest accuracy (~90%) of all predictions except when the blood training set was used for training the model and then for prediction of the blood test set (Line 1, accuracy = 77.7%). Building the classifiers on the samples from the entire blood data set to predict the same samples but profiled in the liver (Line 0, accuracy = 86.2%-88.9%), using the classifiers built on the blood training data set samples to predict the liver training data set samples (Line 2, accuracy = 84.0%-89.7%) and to predict the liver test data set samples (Line 3, accuracy = 82.0%-87.9%) performed better than directly predicting the blood test set samples (Line 1, accuracy = 77.7%-80.4%).

Figure 2.

Prediction across tissues. a) Strategies for building classifiers and making predictions. The line numbers represent the strategy taken. Line 0: Building the classifiers on the entire blood data set to predict the same data set profiled in the liver. Line 1: blood training set was used for training the model and for prediction of the blood test set. Line 2: using the classifiers built on the blood training data to predict the liver training data. Line 3: using the classifiers built on the blood training data to predict the liver test data. Line 0' and 4–6 are the reciprocal predictions from liver to blood. b) Gene-based classifier predictions from the blood to the liver. The x-axis represents the strategies taken to build classifiers and make predictions. The line numbers are as denoted in Figure 2a. FC means the fold change used to select the predictor genes (P <0.05). The y-axis represents the accuracy of prediction (from the average of 100 trials). RF-random forest (# of trees = 100), SVM – support vector machines (RBF kernel), KNN – k-nearest neighbors (k=15), NC-nearest centroids. SVM, KNN and NC were individually combined with a forward array feature selection method (Welch t-tests), evaluated with a five-fold internal cross validation to select best genes in the model construction.

Table 3.

Prediction accuracies of the gene-based classifiers

	8 hepatotoxicants- blood classifier predict the liver
Classifier	Line 0	Line 1	Line 2	Line 3
RF	0.862	0.804	0.840	0.867
SVM	0.874	0.785	0.867	0.844
KNN	0.859	0.763	0.845	0.820
NC	0.889	0.777	0.897	0.879
	8 hepatotoxicants- liver classifier predict the blood
Classifier	Line 4	Line 5	Line 6	Line 0'
RF	0.895	0.589	0.503	0.556
SVM	0.867	0.629	0.517	0.560
KNN	0.884	0.587	0.491	0.520
NC	0.839	0.620	0.489	0.529

Lines numbers denoted as in Figure 2a.

The frequency of the genes selected for prediction revealed that nine genes occurred most often in the classifiers (Figure 3 and Table 2). The genes for interleukin 1 receptor-type II (Il1r2), chemokine (c-c motif) ligand 2 (Ccl2 also known as monocyte chemoattractant protein-1 [MCP-1]) and chemokine (c-x-c motif) ligand 10 (Cxcl10) top the list. Six of these predictor genes were found to have blood gene expression profiles that partitioned the liver samples well based on the presence or absence of necrosis (Figure 1b).

Figure 3.

Frequency of the predictor genes from the overlap between the gene-based classifiers. The x-axis denotes the gene that the Agilent array probe represents and the y-axis denotes the count. See the legend to Figure 2b for identification of the classifier.

As revealed in Table 3, when considering the reciprocal prediction (i.e., from liver to blood) the training of the classifiers on the liver training data set samples and using them to predict the liver test data set samples (Line 4, accuracy = 83.9%-89.5%) performed much better than a) predicting the blood data samples (Lines 5 and 6, accuracy = 58.7%-62.9%) and b) the classifiers that were built on the entire liver data samples directly in order to predict the blood samples (Line 0’, accuracy = 52.0%-56.0%). Moreover, the liver-to-liver prediction (Line 4, accuracy = 83.9%-89.5%) performed better than the blood-to-blood prediction (Line 1, accuracy= 77.7%-80.4%). In all cases, using the classifiers built with the blood data set samples to predict the liver samples performed much better than the converse (Lines 0, 2, 3, versus Lines 0’, 5, 6) (Figure 2b and Table 3).

Cross-tissue predictability was further evaluated at the pathway level based on Line 0 (Figure 2a). In this case, all the genes on the Agilent Array were mapped to the ontology of about 350 canonical pathway maps (CPMs). Classifiers for each pathway were constructed using the genes that were annotated as being present in the pathway. The highest ranked pathways identified in the blood based on the internal cross validation (column 2 of Table 4) are related to PIP3 signaling in B lymphocytes, the Toll-like receptor (TLR) signaling pathway leading to a cell proinflammatory response, and regulation of apoptosis by mitochondrial proteins. These and other top predictive pathways are related to an inflammatory response, apoptosis, mitochondrial damage and angiogenesis (see the VEGF-, TPO- and angiotensin- signaling/activation processes). When pathway-based classifiers from the blood were used to predict the liver samples (column 3 of Table 4), two of the top three pathways identified in the blood as highly predictive, ranked high in the cross-tissue prediction along with the anti-apoptotic TNFs, NF-kB, Bcl-2 pathway, which conferred high degrees of predictability of necrosis between the blood and liver tissues (accuracies ranging between 83.6 % and 89.3%). Interestingly, the cumulative impact of the gene expression signal in the regulation of apoptosis by mitochondrial protein pathway was found to be higher in liver than in the blood. This is evident by a larger number of pro-apoptotic genes up-regulated in case of necrosis in liver compare to blood (Supplementary materials Figure 1).

Table 4.

Prediction accuracies of the pathway-based classifiers

Pathway	blood.acc*	liver.acc**
PIP3 signaling in B lymphocytes	81.1	78.9
Toll-like receptor (TLR) ligands and common TLR signalling pathway leading to cell proinflammatory response ***	78.6	83.6
Regulation of Apoptosis by Mitochondrial Proteins ***	78.3	88.1
Cytoplasm/mitochondrial transport of proapoptotic proteins Bid, Bmf and Bim	78.3	72.3
Role of IAP-proteins in apoptosis	78.3	82.4
VEGF-family signaling	78.0	64.2
Leukocyte chemotaxis	78.0	78.6
Angiotensin signaling via STATs	77.7	77.4
Receptor-mediated axon growth repulsion	77.7	79.6
Immune response BCR pathway	77.7	80.8
Membrane trafficking and signal transduction of G-alpha (i) heterotrimeric G-protein	77.4	66.7
Anti-apoptotic TNFs/NF-kB/Bcl-2 pathway ***	76.7	89.3
G-Protein alpha-12 signaling pathway	76.1	70.4
Brca1 as transcription regulator	76.1	73.6
CXCR4 signaling via second messenger	76.1	78.9
G-Protein alpha-i signaling cascades	76.1	70.4
Caspases cascade	75.8	76.4
Apoptotic TNF-family pathways	75.8	80.2
CCR3 signaling in eosinophils	75.8	71.7
FAS signaling cascades	75.5	77.4
Activation of PKC via G-Protein coupled receptor	75.5	74.2
Galactose metabolism	75.5	60.1
Heme metabolism	75.5	81.4
G-Proteins mediated regulation p38 and JNK signaling	75.2	75.5
ChREBP regulation pathway	75.2	69.8
TPO signaling via JAK-STAT pathway	75.2	75.8
Angiotensin signaling via beta-Arrestin	75.2	74.5
PDGF signaling via STATs and NF-kB	75.2	79.6
Angiotensin activation of ERK	74.8	74.8
Angiotensin signaling via PYK2	74.8	76.7

^*Blood.acc is the internal cross validation accuracy based on the entire blood data set (see Methods);

^**Liver.acc is the accuracy of the blood classifiers to predict the liver based on Line 0 in Figure 2a;

^***Top three pathways using the blood to predict the liver are highlighted

Expression signatures transferable across tissues

The investigation of the blood gene signatures transferred to the liver was carried out in order to determine whether or not the high accuracy of predictability across tissues is sustainable. Classifiers for the blood using random forest (RF), support vector machine (SVM), k-nearest neighbor (KNN) and NC were constructed and the resulting expression signatures were used to develop classifiers based on the gene expression data from the liver samples (Supplementary materials Figure 2). The transferability of expression signatures results summarized in Supplementary materials Table 1 (top section) used a gene expression filter selection criteria of FC > 2 or FC >1.5 and P<0.05. Similar to cross tissue predictability, the prediction using the blood training classifiers on the blood test data set samples (Supplementary materials Table 1 top section: 75.9%-78.6%) was slightly worse when compared to the prediction of the liver training classifiers on the liver test data set samples (Supplementary materials Table 1 bottom section: 83.4%-88.8%). However, using the transferred expression signatures from the blood classifiers to the liver test data set samples gave much better prediction results (Supplementary materials Table 1 top section 81.7%- 88.8%). When the transferability was evaluated in the reverse order, the gene expression signatures transferred from the liver (Supplementary materials Table 1 bottom section: 58.6%-72.0%) performed worse than the gene expression signatures transferred from the blood.

Cross tissue predictability extendable across platforms

The two-way cross platform predictability of genomic indicators was also assessed from the blood (Agilent platform) to predict liver samples (Affymetrix platform) and vice versa by correcting the data for batch (array platform differences) and building classifiers using SVM, KNN and a diagonal linear discriminant analysis (DLDA). As shown in Supplementary materials Table 2, the accuracy of prediction from blood to liver was much higher with the batch corrected data (Lines 2 vs 2’ and 3 vs 3’, i.e., before vs after the batch correction) but about the same as the uncorrected data when predicting liver to blood (Line 5 vs 5’ and 6 vs 6’). In addition, the accuracy of the within-tissue prediction (blood to blood and liver to liver) generally produced higher accuracies than predicting cross tissue (Lines 1 vs 2’ and 3’ and Line 4 vs 5’ and 6’), with the only exception of using the DLDA classifier for prediction from blood to liver (Lines 1 vs. 2’ and 3’). In all cases of the classifiers, higher accuracies were observed when the blood from the Agilent platform was used for training to predict the liver on the Affymetrix platform than the converse (Lines 3’ vs 6’). For all three classifiers, the within-tissue (and platform) training and testing using the liver data is higher than that of the blood (Lines 4 vs 1). The highest (cross tissue) prediction accuracy obtained was 81.0% when the DLDA classifier with sequence mapping and the blood Agilent data set samples were used for training and then applied to predict the liver samples profiled by the Affymetrix platform.

External validation of the predictors across tissues

To validate the ability of the gene-based and pathway-based classifiers constructed from the blood to predict DILI, we leveraged an independent gene expression data set derived from rat liver samples exposed to a different set of hepatotoxicants, one of which is a pharmaceutical agent (acetaminophen) and the other two are non-therapeutic compounds (carbon tetrachloride and allyl alcohol). The accuracy of prediction was determined to be the proportion of samples predicted correctly according to their class label (samples classified as subjects displaying either some or no observable form of liver necrosis as an end-point [see the Materials and Methods section for the number of samples binned in each class and the exposure conditions]). As shown in Table 5, the accuracies of the blind predictions using the four gene-based classifiers (RF, KNN, SVM and NC) and three pathway-based classifiers (corresponding to the ones with high accuracy in cross-tissue prediction as shown in Table 4) are typically higher for acetaminophen and carbon tetrachloride than for allyl alcohol. The gene-based classifiers performed slightly better than the pathway-based classifiers. The NC gene-based classifier performed the best across all the independent validation data set samples and achieved a 92.1% accuracy of prediction on the acetaminophen data. The RF pathway-based classifiers consisted of genes in the a) regulation of apoptosis by mitochondrial proteins pathway, b) anti-apoptotic TNFs/NF-kB/Bcl-2 pathway or c) Toll-like receptor (TLR) ligands and common TLR signaling pathway and performed poorly on the allyl alcohol data set (average accuracy = 66.3%). The RF-b pathway-based classifier, which exhibited the best cross-tissue predictability (Table 4), predicted the acetaminophen and carbon tetrachloride samples slightly better than the other two pathway-based classifiers.

Table 5.

Prediction accuracies of gene- and pathway-based classifiers for independent (validation) data set samples.

	Classifier	Independent data sets			Mean Acc
		Acetaminophen	Carbon Tetrachloride	Allyl Alcohol
Pathway- based classifiers*	RF-a	0.855	0.875	0.642	0.791
	RF-b	0.888	0.903	0.642	0.811
	RF-c	0.882	0.889	0.705	0.825
Gene-based classifiers**	RF	0.816	0.889	0.684	0.796
	KNN	0.836	0.889	0.684	0.803
	SVM	0.888	0.931	0.747	0.855
	NC	0.921	0.917	0.747	0.862
	Mean Acc	0.869	0.899	0.693	0.820

^*Pathway-based random forest (RF) classifiers (Table 2) consist of genes in the a) regulation of apoptosis by mitochondrial proteins pathway, b) Anti-apoptotic TNFs/NF-kB/Bcl-2 pathway, or c) Toll-like receptor (TLR) ligands and common TLR signaling pathway.

^**Gene-based pathways are developed using the entire blood data set.

DISCUSSION

As part of the Food and Drug Administration led MicroArray Quality Control Phase-II (MAQC-II) effort to develop and validate predictive signatures, we used gene expression data acquired from the blood of rats chemically stressed to identify gene- and pathway-based indicators of liver necrosis. Although others have used blood to either predict the exposure of a single drug ¹ or to survey a compendium of hepatotoxicants ^{2, 6}, we took a more formal and comprehensive approach to evaluate the genomic indicators in blood for prediction of liver necrosis across a variety of chemical compounds that target the liver. Our work is the first demonstration of the usefulness of blood as a surrogate tissue to extract genomic indicators for predicting the manifestation of necrosis in the liver based on hepatocellular stress from a drug, therapeutic or across a wide variety of hepatotoxicants. Importantly, the findings are verified by an independent data set comprised of gene expression data from samples stressed by compounds with different characteristics. Acetaminophen is a therapeutic agent while carbon tetrachloride and allyl alcohol are compounds with no pharmacologic benefit. Furthermore, while as hepatotoxicants acetaminophen and carbon tetrachloride require more P450 isoenzyme for bioactivation, allyl alcohol differs in that it requires higher oxygen levels for oxygen-dependent bioactivation ¹⁸. Despite these salient differences, from our analysis, the results demonstrate that using the genomic indicators in blood to predict liver necrosis is somewhat of a general phenomenon and is presumably independent of the choice of hepatotoxicant, the extent of chemical stress or the use as a therapeutic.

Our findings are consistent with the role of organ-to-organ communication that has been previously reported for acetaminophen-induced toxicity ¹⁹ and the role of transmigration of leukocytes into the liver vasculature by inflammatory mediators at the onset of hepatotoxicity contributing to acute liver injury ^{20, 21}. In contrast to the blood-to-liver prediction, the gene and pathway signatures from the liver to predict the blood were not as highly predictive as those acquired from the blood to predict the liver (Figure 2b and Table 3). A possible reason for this phenomenon may be the fact that the dynamic range and overall changes in gene expression that are statistically significant in the liver are quite different and much greater than what is detected in the blood (Supplementary materials Figure 3). Another possibility could be that many of the animals in this study had lesions other than just necrosis or that the phenotypic response that the classifier captured was for a general necrotic lesion whereas the end-point for the validation data set samples was for a specific form of necrosis. An area for further investigation is the determination of a more complex classification based on the histopathology data to predict a composite representation of liver injury which encompasses many end-points.

The genes and pathways acquired from the blood expression data that comprised the classifiers for prediction of necrosis of the liver represent biological mechanisms related to a severe immune response, induction of apoptosis, targeting of the mitochondria and angiogenesis. These mechanisms agree with the current literature on drug-related hepatotoxicity ^22–24 but the latter may be related to the formation of new blood vessels during the regeneration of the liver to compensate for the loss of hepatocytes. Interestingly, one of our top ranking pathway-based classifiers for predicting necrosis in the liver points to the Toll-like receptor (TLR) signaling pathway leading to a cell proinflammatory response. TLRs are a class of single membrane-spanning non-catalytic receptors that recognize structurally conserved molecules derived from microbes and activate immune cell responses. Recently, Yohe et al. ²⁵ reported the role of TLR4 in acetaminophen-mediated hepatotoxicity in endotoxin-responsive mice.

We found that the genes for interleukin 1 receptor-type II (Il1r2), chemokine (c-c motif) ligand 2 (Ccl2) and chemokine (c-x-c motif) ligand 10 (Cxcl10) were most frequently selected for prediction among all the classifiers built and six of the nine most frequent genes have blood gene expression profiles that separated the liver samples fairly well based on the presence or absence of necrosis (Figure 1b). Two pathways with high predictability, the regulation of apoptosis by mitochondrial proteins and the anti-apoptotic TNFs, NF-kB, Bcl-2, have three genes that overlap: B-cell CLL/Lymphoma 2 (Bcl2), TNF receptor superfamily member 1a (Tnfrsf1a) and Bcl2-related protein A1 (Bcl2a1). The latter encodes a member of the Bcl2 protein family. The proteins of this family form hetero- or homodimers and act as anti- and pro-apoptotic regulators. Coincidently, the biological processes that these predictor genes represent match several of the enriched Gene Ontology (GO) categories and KEGG pathways from the biclusters of up- and down-regulated (co-expressed) genes from the Compendium data set liver samples (Supplementary materials Figure 4, Supplementary materials Table 3 and Supplementary file C).

In order to assess the possible mechanisms that the predictor genes contribute to the liver injury phenotype, we built a direct interaction (DI) network using signature genes as seed nodes and the MetaCore collection of over 300,000 curated protein interactions as the source of edges and connected genes (Figure 4). The network revealed that nine signature genes are commonly regulated by 10 transcription factors (TFs) with Ccl2 regulated by seven TFs and S100A9 by six. The downstream targets of the signature genes belong to many of the biological processes ranked highly significant in enrichment and are involved in liver injury: inflammation, extracellular matrix remodeling and apoptosis.

Figure 4.

Network analysis of the upstream and downstream regulation. The nine genes (marked with solid circles) are direct targets of 10 transcription factors. The downstream genes belong to three processes implicated in liver injury.

Although blood serum levels of alanine aminotransferase (ALT) have been historically used as a gold standard clinical chemistry marker of liver injury, the enzyme measurements do not always correlate well with histopathologic data ²⁶ (i.e., the true nature and extent of the liver damage is not always proportional to the elevation in the serum enzyme activity ²⁷). Recently, a study was performed that measured the level of gene expression of haptoglobin (Hp) in blood and compared it to serum ALT as a marker of liver damage ²⁸. The group found that Hp gene expression was more sensitive as an indicator of liver damage. Other genes in our predictor list play a role in inflammation. For instance, the chemokine Cxcl10 is a marker of inflammation found in many models of inflammatory liver diseases ^{29, 30} and is thought to be mainly expressed by hepatocytes but also by macrophages and stellate cells ³¹. S100A8 and S100A9 make up a complex found in leukocytes that appears to be an anti-inflammatory protein ³². Finally, matrix metallopeptidase 8 (Mmp8), a neutrophil collagenease, is involved in the control of the polymorphonuclear cell feed-forward mechanism in an inflammatory process ³³ and others have correlated peripheral blood expression of Mmp8 as a marker of idiopathic pulmonary fibrosis ³⁴.

Genome-wide expression profiling using microarray technologies provides a practical way of surveying the global transcriptional response of a stressor on biological systems³⁵. Using this system to assay peripheral blood for the identification of novel biomarkers of drug-induced liver injury (DILI) is intriguing and may be a useful diagnostic test in the near future ³⁶. Other assay systems have been proposed or used as a model for identifying serum biomarkers as candidates for liver injury ^{3, 37–42}. Our results strongly support the claim that genomic indicators in the blood can serve as biomarkers of necrosis as a form of a chemically-stressed adverse effect on the rat liver and give credence to the acquisition of gene expression signatures from minimal invasive biomaterial sources potentially for diagnostic testing of DILI in humans.