Dixa-fp7.eu

Carcinogenesis vol.35 no.1 pp.201–207, 2014
Advance Access publication August 12, 2013 Assessing compound carcinogenicity in vitro using connectivity mapping
Florian Caiment, Maria Tsamou, Danyel Jennen and
a classifying in vitro model for the human situation in vivo remains a challenge, due to the high heterogeneity of human cancers and the difficulty to build accurate cancer biomarkers.
Department of Toxicogenomics, Maastricht University, 6200 Maastricht, The Recently, a systematic approach referred to as ‘connectivity map- ping' (Cmap) was proposed to establish putative connections between a *To whom correspondence should be addressed. Department of signature profile characteristic of any biological states and a core data- Toxicogenomics, Maastricht University, Universiteitsingel 50, 6200 base of gene e). For instance, Cmap allows the Maastricht, The Netherlands. Tel: +31 43 3881089; Fax: +31 43 3884146; prediction of the molecular response to a new chemical entity by linking its observed effect on the gene expression profile with similar (in case One of the main challenges of toxicology is the accurate predic-
of positive connection) or antagonist (negative connection) compounds. tion of compound carcinogenicity. The default test model for
Applied to a disease state signature, Cmap may also predict whether assessing chemical carcinogenicity, the 2 year rodent cancer
small bioactive molecules are capable of causing or preventing this par- bioassay, is currently criticized because of its limited specific-
ticular disease. The authors demonstrated the viability of the method ity. With increased societal attention and new legislation against
with respect to several biological processes, including complex disease animal testing, toxicologists urgently need an alternative to the
states such as Alzheimer disease or diet-induced obesity, be it with vari- current rodent bioassays for chemical cancer risk assessment.
able success. One of the main advantages of this method, based on the Toxicogenomics approaches propose to use global high-through-
non-parametric Kolmogorov–Smirnov test, is the possibility to use any put technologies (transcriptomics, proteomics and metabolomics)
platform technology combination. Cmap has consequently been used to study the toxic effect of compounds on a biological system.
for a variety of purposes including finding disease alternative treat- Here, we demonstrate the improvement of transcriptomics assay
ments, elucidation of mode of action of drugs, drug repurposing and consisting of primary human hepatocytes to predict the putative
systems biology approaches (revie liver carcinogenicity of several compounds by applying the con-
In this study, we propose to exploit hepatocellular carcinoma (HCC) nectivity map methodology. Our analyses underline that connec-
as a model to investigate the power of the Cmap method for predicting tivity mapping is useful for predicting compound carcinogenicity
and classifying the putative liver carcinogenicity of a compound in by connecting in vivo expression profiles from human cancer tis-
vitro, the liver being the main target organ in the 2 year rodent cancer sue samples with in vitro toxicogenomics data sets. Furthermore,
assay. HCC is the most common form of liver cancer in humans, and the importance of time and dose effect on carcinogenicity predic-
the third cause of cancer mortality. Liver carcinogenesis is a complex tion is demonstrated, showing best prediction for low dose and
multifactor mechanism with several possible etiologies among which 24 h exposure of potential carcinogens.
several viruses, such as Hepatitis B virus and Hepatitis C virus, are known to play a role, along with environmental and chemical expo- sures such as ethanol or aflin vitro liver systems have been well studied and multiple relevant data sets are publicly available, thus allowing the collection of sufficient data to perform a genome-wide Cmap analysis.
For over 40 years, carcinogenic properties of both natural and syn- thetic compounds have been estimated using the 2 year rodent bioas- Materials and methods
say. This laborious, time consuming and expensive test kills a large number of animals. However, only 60% of its predictions are relevant According to the Cmap method design ( controversy on the model's between liver carcinogenicity and any given compound first necessitates the ). Moreover, chemical manufacturing companies face an building of a signature query describing the disease (in our case HCC) and of a increasingly restrictive legislation against animal testing (,). Taken widely ranging in vitro gene profiling ‘reference' data set ().
together, toxicology urgently needs alternative non-animal testing HCC signature geneset methods to improve compound carcinogenicity predictions. Several To establish our HCC signature geneset, we took advantage of publicly avail- alternative methods already exist, such as the quantitative alterna- able microarray e tive structure relationship method, the quantum mechanics/molecular ) and in several steps selected the final list of genes. First, we used mechanics or the threshold of toxicological concern, but all feature the Expression Project for Oncology (expO) study (E-GEOD-2109 on array- major flaws (for reviews, see refs Express), which performed gene expression analyses on a clinically annotated Global gene expression profiling (i.e. transcriptomics), applied to in set of a large panel of different tumor samples using a total of 2158 arrays. Among the 42 samples labeled as ‘liver sample' in expO (each belonging to vitro systems representative of target organs in vivo, is also considered a unique donor), we selected the 10 arrays strictly identified as HCC. These a putative alternative to animal testing. Some promising results, nota- 10 arrays, all hybridized on Affymetrix GeneChip Human Genome U133 Plus bly using the cancer liver cell line HepG2, have been obtained show- 2.0 Arrays, were subjected to a quality control using the ArrayAnalysis pipe- ing that toxicogenomic-based approaches are capable of significantly discriminating carcinogenic subclasses (,The overall accuracy of Averaging method (for Multichip Average) and reannotated using the Custom these toxicogenomics approaches have been estimated to be around CDF version 14 with Entrez Gene identifiers.
80% for predicting in vivo). A recent study, Second, healthy liver sample gene expression profiles generated by the based on genesets selected after stratification of chemicals combined same Affymetrix genechip were used as control. Three studies corresponding with results from the classical Ames mutagenicity assay, reached a to these characteristics were found: E-GEOD-11045 (vestigating three normal liv) analyzing two normal liver sam- compound prediction accuracy of 89%, with a sensitivity of 91% and ples and E-GEOD-15238 (ving studied three liver samples from donors a specificity of 87% (). However, demonstrating the relevance of of different age (1.5, 42 and 81  years). The prenatal liver sample from the E-GEOD-15238 study was not used. All normal liver arrays were subjected to quality control and subsequently normalized and annotated similar to the Abbreviations: CHL, carcinogenic in human liver; Cmap, connectivity map-
cancer samples.
ping; CNL, non-liver carcinogens; CRL, carcinogenic in rodent liver; expO, Reference data set Expression Project for Oncology; HCC, hepatocellular carcinoma; IARC, International Agency for Research on Cancer; NC, non-carcinogens; PHH, To build the reference data set, all human liver in vitro microarray data files primary human hepatocytes; UC, unknown carcinogenicity.
from the Open TG_Gates database ) were The Author 2013. Published by Oxford University Press. All rights reserved. For Permissions, please email: [email protected] Fig. 1. Analysis flowchart. Global flowchart of the sscMAP analysis, with the in vivo HCC signature building and the reference data set composition.
downloaded. This in vitro toxicogenomics data set presents a total of 158 com- connectivity score with each reference file from the TG_Gates data set, pounds applied in duplicate to primary human hepatocytes (PHH) at up to three by summing the product of each individual gene in the signature with its doses (low, middle and high doses) and 3 time points (2, 8 and 24 h) per com- signed ranked in the reference file. Thus, a gene in the query described pound. The high dose was determined for each compound as the maximally as upregulated in HCC (+1) will increase the connectivity score if this tolerated dose. Middle and low doses were defined by taking a concentration of gene is upregulated in the reference (positive rank) or decrease the score 1/5 and 1/25 of the highest, respectively. Each treated compound condition is if downregulated in the reference (negative rank). A P value was gener- associated with a control, using PHH from the same donor. After curation, 10 ated for each connectivity score using random permutation of the geneset. compounds with inconsistent array data (missing 24 h time point condition or The analysis was carried out using the default parameter recommended duplicates; available at Carcinogenesis Online) were by sscMAP: a number of random permutations of 10  000 (leading to a removed, as well as a further nine compounds for which effects on other hepatic minimum P value of 10−4), a random seed of 0 and an expected number of endpoints of toxicity than cancer were av, false connections to tolerate of 1. The threshold of significance was set at available at Carcinogenesis Online). This curation thus resulted in a reference 1/N, with N the number of treatment instances in the reference database data set of 139 different compounds, corresponding to 2320 microarrays, all in the given analysis.
similarly generated by Affymetrix GeneChip Human Genome U133 Plus 2.0 In order to parse the results, a scoring system was used together with plot- Arrays. Each array was processed in the same way as described above for the ting algorithms. The connection between the query signature and a reference arrays used for building the HCC signature geneset. Each treatment condition instance were assessed by computing significance score (S score) for the CHL was converted into an instance reference file (as defined by Lamb et al.), and NC compounds: S score pCHL i.e. a treatment associated to its control pair) according to the sscMAP soft- With pCHL and pNC representing the percentage of instances significantly w). Briefly, each gene was associated with a score correspond- and positively connected for CHL and NC, respectively.
ing to its signed rank in the microarray global expression panel.
Ideally, we expect the entire CHL group compounds, our testing set, to be positively connected and above the threshold of significance after correction for multiple testing. All NC group compounds (non-liver carcinogens) are Based on publicly available carcinogenicity data collected by different inter- expected to be at least below this same threshold, or even negatively connected national or national agencies specialized in cancer research [International to the query. Our scoring system will give an S score of 100 for this perfect Agency for Research on Cancer (IARC), Carcinogenic Potency Database, condition, and this S score will decrease for each CHL or NC compound mis- National Toxicology Program and Environmental Protection Agency], classified. A  score of 0 would indicate a random distribution of the various together with additional research on literature and personnel knowledge, all groups and a score of −100 would correspond to the extreme opposite scenario the compounds used in the open TG-Gates project were classified according (all NC compound are significantly and positively connected to the HCC sig- to their known carcinogenicity (, available nature, and all CHL are not).
at Carcinogenesis Online). Taking into account the two main end points for carcinogenicity of the compounds in view of the aims of this study, namely target test system (liver) and target species (human–rodents), a new classifica- tion list for hepatocarcinogenicity was created. The list consists of five groups, where two groups contain compounds with liver carcinogenic activities clearly HCC signature geneset composition reported on humans [carcinogenic in human liver (CHL)] or only available on rodents [carcinogenic in rodent liver (CRL)]. The group comprising of non- The HCC signature geneset was built by taking all the commonly liver carcinogens present the known human and/or rodent carcinogenic com- up and downregulated genes in the 10 liver cancer samples from the pounds in non-liver target organs (species being distinguished by the IARC expO project compared with the normal liver samples, regardless of classification). The non-carcinogens (NC) group consists of established non- any arbitrary cutoff threshold. The HCC final signature contained carcinogens in both species. Finally, the unknown carcinogenicity (UC) group 7520 upregulated and 1414 downregulated genes that were assigned comprises compounds for which no clear information is available. The final with a score of +1 for the upregulated genes and −1 for the downregu- , available at Carcinogenesis Online).
sscMAP analysis We then set out to check whether the gene expressions com- Connectivity between the geneset with all individual instances compos- monly associated with HCC carcinogenesis in literature were actu- ing our reference data set was computed using the sscMAP JAVA soft- ally present in our selected HCC signature geneset. We established w, our HCC geneset signature was used to compute a a list of the 36 most commonly reported genes ( Compound carcinogenicity using connectivity mapping
Table I.  Reference compound classification
Chlorpheniramine maleate Carbon tetrachloride Interleukin 1 beta, human Tumor Necrosis Factor α Interleukin 6, human Transforming Growth Factor β1 Iproniazid phosphate Phenylanthranilic acid Enalapril maleate Methylene dianiline Rosiglitazone maleate Buthionine sulfoximine Butylated hydroxyanisole Fluoxetine hydrochloride Naphthyl isothiocyanate Hepatocyte growth factor Open TG_Gates dataset contains 158 different compounds applied on PHH from which we selected 139. Each compound is classified based on its predicted carcinogenicity [CHL, CRL, CNL (non-liv, available at Carcinogenesis Online, for a complete data set annotation.
, available at Carcinogenesis Online) from literature reviews sscMAP analysis strictly restricted to human cases combined with the EHCO II The sscMAP analysis of our HCC signature geneset against the TG_ database (gulated Gates data set produce 829 connectivity scores, one for each indi- (among which COL5A2, GPC3, MDK, TP53BP2, XPO1) and vidual compound dose and time point. All the S scores are presented 12 downregulated (including HGFAC, IGFALS, LCAT, MT2A, SLC22A1). Surprisingly, only 40% (15/36) of those genes were Plotted all together, no particularly important pattern appears found in common with our HCC signature geneset. Notably, GPC3 (wever, where the complete set fails to classify the (glypican 3), a gene known to be overexpressed in most HCC and carcinogenic compounds correctly, dividing the arrays according to earlier proposed as a HCC biomarker(as not conserved in experimental factors greatly improv our signature. Although GPC3 was indeed highly expressed in Hence, experiments using a low concentration of the compounds five of the expO HCC cancer samples, in contrast, two samples with an exposure duration of 24 h demonstrate the best classification, presented a downregulation compared with the normal liver con- trol pool, causing this particular gene to be excluded from our signature geneset.
Table II. sscMap S score
Cmap reference data set The reference data set was built from the in vitro human microarray data in the Open TG_Gates ) (), all derived from the same cell model (PHH) and microarray platform (Affymetrix hg-u133 plus 2.0). Data from 139 compounds were con- verted to 829 individual instances. We found clear liver carcinogenic reports in human for three compounds, i.e. ethanol, aflatoxin B1 and S Score computed for our HCC signature geneset for each experimental factor condition on the TG_Gates reference geneset. A score of 0 would azathioprine, which thus defined our positive control testing group indicate a random distribution of the various groups, a score of 100 (CHL group). The full reference details of all compounds are available correspond to the perfect classification (all CHL compound are significantly in and vailable at Carcinogenesis and positively connected to the HCC signature and all NC are not) and a score of −100 would represent the opposite scenario.
Fig. 2. sscMap plot results. sscMAP plot using our HCC signature geneset against the TG_Gates reference data set. Each dot represents a compound instance
(unique dose and time condition compared with the control). The color code indicates the compound classification, as indicated in the legend. For the CRL group,
the IARC classification information is added (i1 for group 1, i2 for group 2 and i3 for group 3 or no IARC classification). The green horizontal line present the
threshold corrected for multiple testing. The S score is indicated on the bottom left of each graph (A) Complete analysis, displaying all doses and time points
available in the reference data set. (B) Same plot displaying only arrays made at low concentration and 24 h compound exposure. (C) Same plot displaying only
arrays made at high concentration and 8 h compound exposure. (D) Same plot displaying only arrays made at high concentration and 24 h compound exposure.
yielding a S score of 80 (o concentration and exposure time a positive connection for 66.7% (10 CHL compounds (ethanol was not tested at low concentration in our of 15) of the NC compounds is displayed. This tendency toward mis- reference dataset) are correctly positively connected (at the low- classification decreases with time but remains at the level of 40% (6 est P value) but also all NC compounds, except for cimetidine, are out of 15) for the NC compounds after 24 h of exposure ( significantly negatively connected (thus a total of 85.7% compound The same observation but with less amplitude is made for the medium correctly classified). The rodent hepatocarcinogenic compounds (CRL), for which no data on human carcinogenicity are available, are represented here characterized by their available IARC (Agency for Research on Cancer) classification. The highest connectivity score In order to validate predictions of hepatocarcinogenic properties in comes from azathioprine, a CHL compound, followed by omeprazole vitro based on our in vivo HCC signature geneset, we used in vitro tox- (UC) and by cyclosporine A that represents the only CRL compound icogenomic data from two compounds, tested in-house also in PHH: classified as IARC group 1 (a group of 111 agents reported to be aflatoxin B1, a CHL compound also present in the TG_Gates data set, carcinogenic to humans, independently of tissue specificity) in this and Benzo[a]pyrene, an IARC group 1 compound known for its hepa- setting. Several other CRL compounds with lower (or no) classifica- tocarcinogenicity in mice. This validation data set has been generated tion in IARC also yield a significantly high positive connection score, similarly to the study protocol of the TG_Gates assays (three doses, like methapyrilene, thioacetamide and ethinylestradiol, tending to Affymetrix platform, PHH) but has been derived only after one expo- prove a possible carcinogenic role in human liver. However, three CRL sure period (24 h). Those data were added to our reference data set, compounds (phenytoin, phenobarbital and gemfibrozil) with demon- and a new sscMAP analysis was performed (). Interestingly, strated carcinogenicity in rodent are negatively connected to the in the resulting data set generated a significant positive connection at vivo human signature and thus could be questioned regarding their all experimental condition for aflatoxin B1, similarly as obtained for carcinogenicity in humans. Globally, we observed that the classifica- aflatoxin using the TG_Gates data set, thus confirming the high poten- tion improves when the dose decreases at 24 h e).
tial of aflatoxin B1 for inducing liver carcinogenicity. However, the The lowest S score (−26,6) is observed at high-dose concentra- scores from our data set are always higher. Benzo[a]pyrene displays tion and 8 h everal a significant positive connection when using data from both low- and NC compounds are positively connected with HCC. Hence, at this high-incubation concentrations. Data obtained upon BaP exposure at Compound carcinogenicity using connectivity mapping
Table III. sscMap results for low concentration and 24 h only
Transforming Growth Factor β1 Tumor necrosis factor α Carbon tetrachloride Naphthyl isothiocyanate Interleukin 1 β Buthionine sulfoximine Butylated hydroxyanisole Rosiglitazone maleate Methylene dianiline Hepatocyte growth factor Fluoxetine hydrochloride Table III. Continued
Each row corresponds to a microarray using a given compound, at low concentration and at 24 h exposure. The connectivity score (setscore) and the corresponding P value are given. A significance mark of 1 indicates significance above the threshold corrected for multiple testing.
usually linked to HCC in literature and databases failed to establish any significant connection with the CHL compounds (data not shown). We believe this could be explained by the complexity of the carcino- genic mechanisms that cannot be resumed within a small list of bio- marker genes. This observation supports the debate on the difficulty to define effective biomarkers for cancer (The ability of Cmap methodology to virtually use an unlimited amount of genes to define a signature seems, however, to be a powerful asset in carcinogenicity Importantly, our reference set (and validating set) used only PHH, a non-cancer cell model, and thus our positive connection using in vivo cancer signature, cannot be biased by deviations in normal expression patterns resulting from the common criteria with the can- cer phenotype, as the immortal state. Indeed, we tried to mix PHH assays with HepG2 cells assays to build our reference data set and observed that HepG2 assays were globally more connected to query- defined HCC states than PHH assays, independently of the com- pound carcinogenicity classification (data not shown). To avoid this putative bias, and because our ultimate goal is to find an alternative in Fig. 3. SscMap plot results on the validation set. Each dot represents a
vitro method to predict carcinogenicity at the earliest possible stage, compound instance (three different concentrations per compound, always at we decided to remove all liver cancer cell models and work only with 24 h exposure time). The color code indicate the compound concentration (green for low dose, blue for middle dose and red for high dose), and form Moreover, Cmap seems to be able to differentiate compound carci- represent the compound (a dot for the original TG_Gates aflatoxin, a triangle nogenicity based on their incubation concentration in vitro. Thus, our for our validation aflatoxin and a diamond for Benzo[a]pyrene validation set). signature geneset better classifies the known human hepatocarcino- The green horizontal line present the threshold corrected for multiple testing.
gens (CHL) at low concentration in vitro than at high concentration. More importantly, a higher number of false positives appear at high medium concentration, even if still positive, do not pass the threshold dose for NC compounds. These results underline the importance of for multiple testing.
applying low exposures in in vitro toxicological assays for predicting genotoxicity and carcinogenicity, in order to avoid possible cytotoxic dose responses. Unfortunately, the low-concentration condition is missing for several compounds in the TG_Gates database. It would be Here, we studied the reliability of the Cmap method for classify- interesting to add more compounds, both known human liver carcino- ing and predicting a compound's hepatocarcinogenicity. To reduce gens and non-carcinogens, to our reference set to see how this query all possible variables, the study was performed on a single cancer would behave.
type (HCC) and used a unique liver cell model (PHH) exposed to Since our best setting to predict carcinogenicity is at low compound a wide range of different compounds, of which some were known concentration and after 24 h of exposure, evaluating connectivity to human liver carcinogens (CHL) and some known non-liver carcino- the HCC signature of compounds with unknown carcinogenic proper- gens. Using the sscMap software, our Cmap analysis establishes sig- ties (UC group) at this setting may be quite informative. Of the 36 nificantly positive connections with all hepatocarcinogens available in compounds in the TG_Gates data set within the UC group and tested the TG_Gates data set at low concentration and at 24 h of exposure, at low concentration, 10 are significantly positively connected with and a negative connection with all but one of the NC compounds at our query: omeprazole, diclofenac, glibenclamide, carbamazepine, this same concentration.
triazolam, N-nitrosomorpholine, naphthyl isothiocyanate, proprano- Most of the Cmap studies used the Cmap data set available at the lol, methyltestosterone, amphotericin B, buthionine sulfoximine and ) as a reference set. diethyl maleate (listed here in order of decreasing connectivity score). Cmap reference data set contains more than 7000 expression profiles However, as those compounds are mainly drugs used for diverse representing 1309 compounds used on five different cultured human applications, some of them could have been taken by the patient cancer cell lines (MCF7, ssMCF7, HL60, PC3 and SKMEL5). We associated with the liver cancer samples. For instance, omeprazole, decided not to use this reference data set as none of the five cell a proton pump inhibitor, is one of the most widely prescribed drugs lines are related to liver and thus the Cmap reference data set is not internationally. Diclofenac is often used to treat chronic pain associ- optimized for liver carcinogenicity prediction.
ated with cancer, and thus may have induced a change of expression Interestingly, our signature geneset, based on an in vivo liver cancer in the patient's liver, correlated with the gene expression induced by expression profile, contains many more genes than usual biomarker this compound in the in vitro TG_Gates and then inducing a connec- lists representing a biological signature. However, with a geneset of tion between the two profiles. The complete donor drug prescription 8934 genes (7520 upregulated and 1414 downregulated) selected would be necessary to sort out the real carcinogenic compound from without any expression threshold, the advantage is that we were able the false positive.
to keep all small genes expression variations observed ‘in liver' can- Another possible way to increase prediction accuracy would be to cer samples in vivo. Internal testing with many different lists of genes use other types of ‘omics' information in the reference dataset such Compound carcinogenicity using connectivity mapping
as proteomics, thus increasing the amount of possible connections. 9. Ellinger-Ziegelbauer,H. et  al. (2009) Application of toxicogenomics to Indeed, in view of the fact that the Cmap method essentially is a non- study mechanisms of genotoxicity and carcinogenicity. Toxicol. Lett., 186,
parametric method, complementing transcriptomics results with data on proteomics and/or metabolomics experiments derived from the 10. Hu,T. et al. (2004) Identification of a gene expression profile that discrimi- same samples, may bring added value. All those connections could nates indirect-acting genotoxins from direct-acting genotoxins. Mutat.
Res., 549, 5–27.
then be integrated in a putative model predicting cancer pathways and, 11. Le Fevre,A.C. et  al. (2007) Characterization of DNA reactive and non- consequently, may allow a better carcinogenicity prediction. In order DNA reactive anticancer drugs by gene expression profiling. Mutat. Res., to achieve such cross-omics analyses, the Toxicogenomics commu- nity would benefit from a central database to regroup all the available 12. Li,H.H. et al. (2007) Toxicogenomics: overview and potential applications experiments. Several international projects, including the European for the study of non-covalent DNA interacting chemicals. Mutat. Res., 623,
), work in this direction.
To conclude, we believe that the Cmap methodology offers an 13. Mathijs,K. et  al. (2010) Gene expression profiling in primary mouse interesting perspective of how to predict a compound's carcinogenic- hepatocytes discriminates true from false-positive genotoxic compounds. ity based on data from representative in vitro models by adding rel- Mutagenesis, 25, 561–568.
14. Tsujimura,K. et al. (2006) Prediction of carcinogenic potential by a toxi- evance for human disease end points to the prediction model and may cogenomic approach using rat hepatoma cells. Cancer Sci., 97, 1002–1010.
be considered as a classification method for new compounds 15. Magkoufopoulou,C. et al. (2012) A transcriptomics-based in vitro assay for predicting chemical genotoxicity in vivo. Carcinogenesis, 33, 1421–1429.
16. Lamb,J. et al. (2006) The Connectivity Map: using gene-expression sig- natures to connect small molecules, genes, and disease. Science, 313,
can be found at 17. Qu,X.A. et al. (2012) Applications of Connectivity Map in drug discovery and development. Drug Discov. Today, 17, 1289–1298.
18. Farazi,P.A. et  al. (2006) Hepatocellular carcinoma pathogenesis: from genes to environment. Nat. Rev. Cancer, 6, 674–687.
19. Marioni,J.C. et al. (2008) RNA-seq: an assessment of technical reproduc- The European Union Seventh Framework project diXa (283775); ibility and comparison with gene expression arrays. Genome Res., 18,
The Netherlands Organisation for Health Research and Development project Data Integration and Mining Towards Risk Assessment 20. Yu,Y. et al. (2010) A comparative analysis of liver transcriptome suggests divergent liver function among human, mouse and rat. Genomics, 96,
281–289.
21. Tzur,G. et al. (2009) Comprehensive gene and microRNA expression pro- Conflict of Interest Statement: None declared.
filing reveals a role for microRNAs in human liver development. PLoS
One, 4, e7511.
22. Uehara,T. et al. (2011) Prediction model of potential hepatocarcinogenic- ity of rat hepatocarcinogens using a large-scale toxicogenomics database. 1. Hartung,T. (2009) Toxicology for the twenty-first century. Nature, 460,
Toxicol. Appl. Pharmacol., 255, 297–306.
23. Zhang,S.D. et al. (2009) sscMap: an extensible Java application for con- 2. Ward,J.M. (2008) Value of rodent carcinogenesis bioassays. Toxicol. Appl. necting small-molecule drugs using gene-expression signatures. BMC Pharmacol., 226, 212.
Bioinformatics, 10, 236.
3. Knight,A. et  al. (2005) Which drugs cause cancer? For. BMJ, 331,
24. Hsu,C.N. et  al. (2007) Detection of the inferred interaction network in hepatocellular carcinoma from EHCO (Encyclopedia of Hepatocellular 4. Knight,A. et al. (2006) Animal carcinogenicity studies: implications for the Carcinoma genes Online). BMC Bioinformatics, 8, 66.
REACH system. Altern. Lab. Anim., 34 (suppl. 1), 139–147.
25. Capurro,M. et  al. (2003) Glypican-3: a novel serum and histochemical 5. Braga,R.C. et al. (2012) QSAR and QM/MM approaches applied to drug marker for hepatocellular carcinoma. Gastroenterology, 125, 89–97.
metabolism prediction. Mini Rev. Med. Chem., 12, 573–582.
26. Acevedo,L.G. et al. (2008) Analysis of the mechanisms mediating tumor- 6. Hennes,E.C. (2012) An overview of values for the threshold of toxicologi- specific changes in gene expression in human liver tumors. Cancer Res., cal concern. Toxicol. Lett., 211, 296–303.
7. van Delft,J.H. et  al. (2004) Discrimination of genotoxic from non-gen- 27. Diamandis,E.P. (2010) Cancer biomarkers: can we turn recent failures into otoxic carcinogens by gene expression profiling. Carcinogenesis, 25,
success? J. Natl Cancer Inst., 102, 1462–1467.
28. Ransohoff,D.F. (2005) Bias as a threat to the validity of cancer molecular- 8. van Delft,J.H. et al. (2005) Comparison of supervised clustering methods marker research. Nat. Rev. Cancer, 5, 142–149.
to discriminate genotoxic from non-genotoxic carcinogens by gene expres-
sion profiling. Mutat. Res., 575, 17–33.
Received May 28, 2013; revised May 28, 2013; accepted August 4, 2013

Source: http://www.dixa-fp7.eu/IManager/Download/544/45990/10190/576438/EN/10190_576438_Y02K_florian_Cmap.pdf