Original Article

The Pharmacogenomics Journal (2010) 10, 267–277; doi:10.1038/tpj.2010.33

Genomic indicators in the blood predict drug-induced liver injury

J Huang1,8,12, W Shi2,12, J Zhang3,12, J W Chou4,13, R S Paules5, K Gerrish5, J Li4,14, J Luo3, R D Wolfinger6, W Bao6, T-M Chu6, Y Nikolsky2, T Nikolskaya2,11, D Dosymbekov7, M O Tsyganova7, L Shi8, X Fan1,8, J C Corton9, M Chen8, Y Cheng1, W Tong8, H Fang10 and P R Bushel4

  1. 1Pharmaceutical Informatics Institute, College of Pharmaceutical Sciences, Zhejiang University, Hangzhou, China
  2. 2GeneGO, St Joseph, MI, USA
  3. 3Department of Bioinformatics, Systems Analytics, Waltham, MA, USA
  4. 4Biostatistics Branch, National Institute of Environmental Health Sciences, Research Triangle Park, NC, USA
  5. 5Microarray Group, National Institute of Environmental Health Sciences, Research Triangle Park, NC, USA
  6. 6Genomics Division, SAS, Cary, NC, USA
  7. 7Vavilov Institute for General Genetics, Russian Academy of Sciences, Moscow, Russia
  8. 8Center for Toxicoinformatics, National Center for Toxicological Research, Food and Drug Administration, Jefferson, AR, USA
  9. 9National Health and Environmental Effects Research Laboratory, US Environmental Protection Agency, Research Triangle Park, NC, USA
  10. 10Division of Bioinformatics, Z-Tech Corporation, ICF International Company at NCTR/Food and Drug Administration, Jefferson, AR, USA
  11. 11Systems Biology Laboratory, Institute for General Genetics, Moscow, Russia

Correspondence: Dr PR Bushel, Biostatistics Branch, National Institute of Environmental Health Sciences, P.O. Box 12233, Research Triangle Park, NC 27709, USA. E-mail: bushel@niehs.nih.gov

12These authors contributed equally to this work.

13Current address: Department of Biostatistical Sciences, Wake Forest University School of Medicine, Medical Center Boulevard, Winston-Salem, NC 27157, USA.

14Current address: Lineberger Comprehensive Cancer Center, University of North Carolina, Chapel Hill, 130 Mason Farm Road, Chapel Hill, NC 27599, USA.

Received 22 November 2009; Revised 28 February 2010; Accepted 25 April 2010.



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 show for the first time and support the hypothesis that genomic indicators in the blood can serve as potential diagnostic biomarkers predictive of DILI.


prediction; acetaminophen; blood; cross-tissue; liver injury; microarray gene expression



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 as 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 bioavailable 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 showed 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 showed 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 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 manner (that is a common array platform, experimental procedure, data acquisition, and analysis methods) 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 (Agilent Technologies, Santa Clara, CA, USA) and Affymetrix (Affymetrix, Santa Clara, CA, USA). We then show 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 database under the MAQC-II reference series accession GSE16716 and in the Chemical Effects in Biological Systems database4 under accession number 001-00001-0020-000-4. All eight compounds were studied using standardized procedures, that is 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 killed at 6, 24, and 48h 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 and Affymetrix 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 (22075 probes) arrays. Fluorescent pixel intensity 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 (31099 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 are 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.


Groups of four male Fischer F344/N rats each received 0 (vehicle), 50, 150, 1500, or 2000mgkg−1 body weight of acetaminophen at two different times: between 1200 (noon) and 1300 hours (‘light’ subjects) or between 2400 (midnight) and 0100 hours (‘night’ subjects). The animals were killed after 6, 18, 24, or 48h. 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 2000mgkg−1 body weight of carbon tetrachloride. The animals were killed after 3, 6, 24, or 72h. There are a total of 72 samples.

Allyl alcohol

Groups of six male Fischer F344/N rats each received 10, 20, 40, or 50mgkg−1 body weight of allyl alcohol. The animals were killed after 6, 24, 48, or 72h. 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 manner 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 Chemical Effects in Biological Systems database4 under accession number 002-00001-0011-000-5. The allyl alcohol and carbon tetrachloride data are stored in the NIEHS MicroArray Project System database7 under project ID 221 and 236 and is available on request. The classification of the samples is described in the ‘Histopathology and sample classification’ subsection. ArrayTrack8 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 analysis of variance 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, that is, 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 the time of killing, blood was collected into serum separation tubes (BD Microtainer Tubes, Becton-Dickinson, Franklin Lakes, NJ, USA) 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 versus class 0: samples with no observable form of necrosis) was used. The data is considered to be independent and assumed to be normally distributed. The measure of variability is the s.d. Optimization of the features selected as predictors was performed by a fivefold internal cross-validation strategy, with the compendium data set samples split into training (n=175) and testing (n=143) subjects. Support vector machines (SVMs), k-nearest neighbors (KNNs) 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 the GeneGo's canonical pathway maps (CPMs) was performed using the randomForest package in R. GeneGo's CPMs were derived from an ontology of experimentally confirmed signaling and metabolic multistep 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 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 kth tree (k=100) and then used for prediction. Out-of-bag 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 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-biclustering

In the first step of co-expression-biclustering12 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 method13 to the initial coherent biclusters to further subset the genes into final biclusters that contain coherent and highly co-expressed genes. Extracting patterns and identifying co-expressed genes 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 aik and ajk (ith and jth genes) for the samples within the kth group is assessed with a binary coherent matrix H(h(ij),k) according to an inclusionexclusion criterion function

Unfortunately we are unable to provide accessible alternative text for this. If you require assistance to access this image, please contact help@nature.com or the author

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 kth 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 aik and ajk. pt is a user-defined threshold for the P-value and is set to 0.001.

Biological processes overrepresentation and pathway analysis

The expression analysis systematic explorer14 was used to identify biological processes overrepresented by sets of genes identified as predictors or contained within biclusters. The overrepresented processes were confirmed by the Gene Ontology Enrichment Analysis Software Toolkit15 using the Adrian Alexa's improved weighted scoring algorithm to account for the hierarchical structure of Gene Ontology16 and the Benjamini–Yekutieli procedure to control the false discovery rate under the assumption of gene-to-gene dependency.17



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 5-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 the levels of 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).

Figure 1.
Figure 1 - Unfortunately we are unable to provide accessible alternative text for this. If you require assistance to access this image, please contact help@nature.com or the author

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).

Full figure and legend (128K)Download PowerPoint slide (1,568 KB)

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 FC>2.0 criteria coupled with a P-value <0.05 performed better than using a 1.5 cutoff with the same P-value. The 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.
Figure 2 - Unfortunately we are unable to provide accessible alternative text for this. If you require assistance to access this image, please contact help@nature.com or the author

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. Lines 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 machine (RBF kernel); KNN, k-nearest neighbor (k=15); NC, nearest centroid. SVM, KNN, and NC were individually combined with a forward array feature selection method (Welch t-tests), evaluated with a fivefold internal cross-validation to select the best genes in the model construction.

Full figure and legend (91K)Download PowerPoint slide (953 KB)

The frequency of the genes selected for prediction revealed that nine genes occurred most often in the classifiers (Figure 3; 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.
Figure 3 - Unfortunately we are unable to provide accessible alternative text for this. If you require assistance to access this image, please contact help@nature.com or the author

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 of Figure 2b for identification of the classifier.

Full figure and legend (66K)Download PowerPoint slide (542 KB)

As revealed in Table 3, when considering the reciprocal prediction (that is 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 (1) predicting the blood data samples (Lines 5 and 6, accuracy=58.7–62.9%) and (2) the classifiers that were built on the entire liver data samples directly 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; 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 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 phosphatidylinositol-3,4,5-trisphosphate 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 vascular endothelial growth factor, thrombopoietin, 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-κB, 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 upregulated in case of necrosis in liver compared with blood (Supplementary Figure 1).

Expression signatures transferable across tissues

The investigation of the blood gene signatures transferred to the liver was carried out to determine whether or not the high accuracy of predictability across tissues is sustainable. Classifiers for the blood using RF, SVM, 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 Figure 2). The transferability of expression signatures results summarized in Supplementary 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 Table 1 top section: 75.9–78.6%) was slightly worse when compared with the prediction of the liver training classifiers on the liver test data set samples (Supplementary 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 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 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. As shown in Supplementary Table 2, the accuracy of prediction from blood to liver was much higher with the batch corrected data (Lines 2 versus 2′ and 3 versus 3′, that is, before versus after the batch correction) but about the same as the uncorrected data when predicting liver to blood (Line 5 versus 5′ and Line 6 versus 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 versus 2′ and 3′ and Line 4 versus 5′ and 6′), with the only exception of using the diagonal linear discriminant analysis classifier for prediction from blood to liver (Lines 1 versus 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 (Line 3′ versus 6′). For all three classifiers, the within-tissue (and platform) training and testing using the liver data is higher than that of the blood (Line 4 versus 1). The highest (cross-tissue) prediction accuracy obtained was 81.0% when the diagonal linear discriminant analysis 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 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 (1) regulation of apoptosis by mitochondrial proteins pathway, (2) anti-apoptotic TNFs/NF-κB/Bcl-2 pathway, or (3) TLR ligands and common TLR signaling pathway and performed poorly on the allyl alcohol data set (average accuracy=66.3%). The RF-(2) 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.



As part of the Food and Drug Administration led 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 drug1 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 whereas carbon tetrachloride and allyl alcohol are compounds with no pharmacologic benefit. Furthermore, although 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 show 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 toxicity19 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; 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 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, 23, 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 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-κB, 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 categories and KEGG pathways from the biclusters of up- and downregulated (co-expressed) genes from the compendium data set liver samples (Supplementary Figure 4; Supplementary Table 3; Supplementary File C).

To assess the possible mechanisms that the predictor genes contribute to the liver injury phenotype, we built a direct interaction network using signature genes as seed nodes and the MetaCore collection of over 300000 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 with Ccl2 regulated by seven transcription factors 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.
Figure 4 - Unfortunately we are unable to provide accessible alternative text for this. If you require assistance to access this image, please contact help@nature.com or the author

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.

Full figure and legend (157K)Download PowerPoint slide (1,295 KB)

Although blood serum levels of 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 data26 (that is the true nature and extent of the liver damage is not always proportional to the elevation in the serum enzyme activity27). Recently, a study was performed that measured the level of gene expression of haptoglobin (Hp) in blood and compared it with 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 have a function in inflammation. For instance, the chemokine Cxcl10 is a marker of inflammation found in many models of inflammatory liver diseases29, 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 seems 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 process33 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 systems.35 Using this system to assay peripheral blood for the identification of novel biomarkers of 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, 38, 39, 40, 41, 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.


Conflict of interest

The authors declare no conflict of interest.



  1. Bushel PR, Heinloth AN, Li J, Huang L, Chou JW, Boorman GA et al. Blood gene expression signatures predict exposure levels. Proc Natl Acad Sci USA 2007; 104: 18211–18216. | Article | PubMed
  2. Lobenhofer EK, Auman JT, Blackshear PE, Boorman GA, Bushel PR, Cunningham ML et al. Gene expression response in target organ and whole blood varies as a function of target organ injury phenotype. Genome Biol 2008; 9: R100. | Article | PubMed | ChemPort |
  3. Wang K, Zhang S, Marzolf B, Troisch P, Brightman A, Hu Z et al. Circulating microRNAs, potential biomarkers for drug-induced liver injury. Proc Natl Acad Sci USA 2009; 106: 4402–4407. | Article | PubMed
  4. Waters M, Stasiewicz S, Merrick BA, Tomer K, Bushel P, Paules R et al. CEBS—chemical effects in biological systems: a public data repository integrating study design and toxicity data with microarray and proteomics data. Nucleic Acids Res 2008; 36(Database issue): D892–D900. | Article | PubMed | ChemPort |
  5. Hubbell E, Liu WM, Mei R. Robust estimators for expression analysis. Bioinformatics (Oxford, England) 2002; 18: 1585–1592. | Article | PubMed | ChemPort |
  6. Huang L, Heinloth AN, Zeng ZB, Paules RS, Bushel PR. Genes related to apoptosis predict necrosis of the liver as a phenotype observed in rats exposed to a compendium of hepatotoxicants. BMC Genomics 2008; 9: 288. | Article | PubMed | ChemPort |
  7. Bushel PR, Hamadeh H, Bennett L, Sieber S, Martin K, Nuwaysir EF et al. MAPS: a microarray project system for gene expression experiment information and data validation. Bioinformatics (Oxford, England) 2001; 17: 564–565. | Article | PubMed | ChemPort |
  8. Tong W, Harris S, Cao X, Fang H, Shi L, Sun H et al. Development of public toxicogenomics software for microarray data management and analysis. Mutat Res 2004; 549: 241–253. | PubMed | ChemPort |
  9. Boorman GA, Haseman JK, Waters MD, Hardisty JF, Sills RC. Quality review procedures necessary for rodent pathology databases and toxicogenomic studies: the National Toxicology Program experience. Toxicol Pathol 2002; 30: 88–92. | Article | PubMed
  10. Nikolsky Y, Kirillov E, Zuev R, Rakhmatulin E, Nikolskaya T. Functional analysis of OMICs data and small molecule compounds in an integrated ‘knowledge-based’ platform. Methods Mol Biol (Clifton, NJ) 2009; 563: 177–196. | ChemPort |
  11. Liaw A, Wiener M. Classification and regression by randomForest. R News 2002; 2: 18–22.
  12. Chou JW, Bushel PR. Discernment of possible mechanisms of hepatotoxicity via biological processes over-represented by co-expressed genes. BMC Genomics 2009; 10: 272. | Article | PubMed | ChemPort |
  13. Chou JW, Zhou T, Kaufmann WK, Paules RS, Bushel PR. Extracting gene expression patterns and identifying co-expressed genes from microarray data reveals biologically responsive processes. BMC Bioinformatics 2007; 8: 427. | Article | PubMed | ChemPort |
  14. Hosack DA, Dennis Jr G, Sherman BT, Lane HC, Lempicki RA. Identifying biological themes within lists of genes with EASE. Genome Biol 2003; 4: R70. | Article | PubMed
  15. Zheng Q, Wang XJ. GOEAST: a web-based software toolkit for Gene Ontology enrichment analysis. Nucleic Acids Res 2008; 36(Web Server issue): W358–W363. | Article | PubMed | ChemPort |
  16. Alexa A, Rahnenfuhrer J, Lengauer T. Improved scoring of functional groups from gene expression data by decorrelating GO graph structure. Bioinformatics (Oxford, England) 2006; 22: 1600–1607. | Article | PubMed | ChemPort |
  17. Benjamini Y, Yekutieli D. The control of the false discovery rate in multiple testing under dependency. Ann Stat 2001; 29: 1165–1188. | Article
  18. Casarett LJ, Doull J, Klaassen CD. Casarett and Doull's Toxicology: The Basic Science of Poisons, 6th edn. McGraw-Hill Medical Pub. Division: New York, 2001, xix, 1236 pp.
  19. Neff SB, Neff TA, Kunkel SL, Hogaboam CM. Alterations in cytokine/chemokine expression during organ-to-organ communication established via acetaminophen-induced toxicity. Exp Mol Pathol 2003; 75: 187–193. | Article | PubMed | ChemPort |
  20. Jaeschke H, Hasegawa T. Role of neutrophils in acute inflammatory liver injury. Liver Int 2006; 26: 912–919. | Article | PubMed | ChemPort |
  21. Ramaiah SK, Jaeschke H. Role of neutrophils in the pathogenesis of acute inflammatory liver injury. Toxicol Pathol 2007; 35: 757–766. | Article | PubMed | ChemPort |
  22. Jaeschke H, Gores GJ, Cederbaum AI, Hinson JA, Pessayre D, Lemasters JJ. Mechanisms of hepatotoxicity. Toxicol Sci 2002; 65: 166–176. | Article | PubMed | ISI | ChemPort |
  23. Kaplowitz N. Biochemical and cellular mechanisms of toxic liver injury. Semin Liver Dis 2002; 22: 137–144. | Article | PubMed | ISI | ChemPort |
  24. Lee WM. Drug-induced hepatotoxicity. N Engl J Med 2003; 349: 474–485. | Article | PubMed | ISI | ChemPort |
  25. Yohe HC, O’Hara KA, Hunt JA, Kitzmiller TJ, Wood SG, Bement JL et al. Involvement of Toll-like receptor 4 in acetaminophen hepatotoxicity. Am J Physiol Gastrointest Liver Physiol 2006; 290: G1269–G1279. | Article | PubMed | ChemPort |
  26. Ozer J, Ratner M, Shaw M, Bailey W, Schomaker S. The current state of serum biomarkers of hepatotoxicity. Toxicology 2008; 245: 194–205. | Article | PubMed | ChemPort |
  27. Kaplowitz N, DeLeve LD. Drug-Induced Liver Disease, 2nd edn. Informa Healthcare: New York, 2007, xv, 808 pp.
  28. Kudo Y, Ochi T, Shimada H, Ogawa S, Shinjo K. Utility of plasma circulating mRNA as a marker to detect hepatic injury. J Vet Med Sci 2008; 70: 993–995. | Article | PubMed
  29. Nakken KE, Nygard S, Haaland TK, Berge KE, Odegaard A, Labori KJ et al. Gene expression profiles reflect sclerosing cholangitis activity in abcb4 (−/−) mice. Scand J Gastroenterol 2009; 44: 211–218. | Article | PubMed | ChemPort |
  30. Zhai Y, Shen XD, Gao F, Zhao A, Freitas MC, Lassman C et al. CXCL10 regulates liver innate immune response against ischemia and reperfusion injury. Hepatology 2008; 47: 207–214. | Article | PubMed | ChemPort |
  31. Yoneyama H, Kai Y, Koyama J, Suzuki K, Kawachi H, Narumi S et al. Neutralization of CXCL10 accelerates liver regeneration in carbon tetrachloride-induced acute liver injury. Med Mol Morphol 2007; 40: 191–197. | Article | PubMed | ChemPort |
  32. Ikemoto M, Murayama H, Itoh H, Totani M, Fujita M. Intrinsic function of S100A8/A9 complex as an anti-inflammatory protein in liver injury induced by lipopolysaccharide in rats. Clin Chim Acta 2007; 376: 197–204. | Article | PubMed | ChemPort |
  33. Tester AM, Cox JH, Connor AR, Starr AE, Dean RA, Puente XS et al. LPS responsiveness and neutrophil chemotaxis in vivo require PMN MMP-8 activity. PLoS ONE 2007; 2: e312. | Article | PubMed | ChemPort |
  34. Rosas IO, Richards TJ, Konishi K, Zhang Y, Gibson K, Lokshin AE et al. MMP1 and MMP7 as potential peripheral blood biomarkers in idiopathic pulmonary fibrosis. PLoS Med 2008; 5: e93. | Article | PubMed | ChemPort |
  35. Minami K, Saito T, Narahara M, Tomita H, Kato H, Sugiyama H et al. Relationship between hepatic gene expression profiles and hepatotoxicity in five typical hepatotoxicant-administered rats. Toxicol Sci 2005; 87: 296–305. | Article | PubMed | ChemPort |
  36. Burczynski ME, Dorner AJ. Transcriptional profiling of peripheral blood cells in clinical pharmacogenomic studies. Pharmacogenomics 2006; 7: 187–202. | Article | PubMed | ChemPort |
  37. Gangadharan B, Antrobus R, Dwek RA, Zitzmann N. Novel serum biomarker candidates for liver fibrosis in hepatitis C patients. Clin Chem 2007; 53: 1792–1799. | Article | PubMed | ChemPort |
  38. Miller T, Knapton A, Adeyemo O, Noory L, Weaver J, Hnig J. Cytochrome c: a non-invasive biomarker of drug-induced liver injury. J Appl Toxicol 2008; 28: 815–828. | Article | PubMed | ChemPort |
  39. Xu JJ, Henstock PV, Dunn MC, Smith AR, Chabot JR, de Graaf D. Cellular imaging predictions of clinical drug-induced liver injury. Toxicol Sci 2008; 105: 97–105. | Article | PubMed | ChemPort |
  40. Merrick BA, Bruno ME, Madenspacher JH, Wetmore BA, Foley J, Pieper R et al. Alterations in the rat serum proteome during liver injury from acetaminophen exposure. J Pharmacol Exp Ther 2006; 318: 792–802. | Article | PubMed | ChemPort |
  41. Amacher DE, Adler R, Herath A, Townsend RR. Use of proteomic methods to identify serum biomarkers associated with rat liver toxicity or hypertrophy. Clin Chem 2005; 51: 1796–1803. | Article | PubMed | ChemPort |
  42. Merrick BA. The plasma proteome, adductome and idiosyncratic toxicity in toxicoproteomics research. Brief Funct Genomics Proteomic 2008; 7: 35–49. | Article | ChemPort |


We thank the National Center for Toxicogenomics at the National Institute of Environmental Health Sciences (NIEHS) for the hepatotoxicant compendium data. In addition, we thank the participants of MAQC-II for comments, feedback, and discussions on the topic of this paper during teleconferences and face-to-face project meetings. We also thank K Shockley, A Merrick, S Hester, B Ward, and D Mendrick for their critical review of the manuscript. JH acknowledge the support of the Oak Ridge Institute for Science and Education (ORISE) for the Post-graduate Research Program at the National Center for Toxicological Research (NCTR), US Food and Drug Administration (FDA). JH also acknowledges the support of the China State-funded Study Abroad Program that is organized by the China Scholarship Council (CSC). JH and XF both acknowledge the Chinese Key Technologies R&D Program (No. 2005CB23402) and the National Science Foundation of China (No. 30801556) for support to participate in the MAQC-II project at the NCTR/FDA. This research was supported, in part by, the Intramural Research Program of the NIH and NIEHS (Z01 ES102345-03). This document has been reviewed in accordance with the US FDA and Environmental Protection Agency (EPA) policies and is approved for publication. Approval does not signify that the contents necessarily reflect the position or opinions of the FDA or EPA nor does mention of trade names or commercial products constitute endorsement or recommendation for use. The findings, views, and conclusions in this report are those of the authors and do not necessarily represent or reflect the views of the FDA or EPA.

Supplementary Information accompanies the paper on the The Pharmacogenomics Journal website