Gene co-regulation and co-expression in the aryl hydrocarbon receptor-mediated transcriptional regulatory network in the mouse liver
Four decades after its discovery, the aryl hydrocarbon receptor (AHR), a ligand-inducible transcription factor (TF) activated by the persistent environmental contaminant 2,3,7,8-Tetrachlorodibenzo-p-dioxin (TCDD), remains an enigmatic molecule with a controversial endogenous role. Here, we have assembled a global map of the AHR gene regulatory network in female C57BL/6 mice orally gavaged with 30 µg/kg of TCDD from a combination of previously published gene expression and genome-wide TF-binding data sets. Using Kohonen self-organizing maps and subspace clustering, we show that genes co-regulated by common upstream TFs in the AHR network exhibit a pattern of co-expression. Directly bound, indirectly bound, and non-genomic AHR target genes exhibit distinct expression patterns, with the directly bound targets associated with highest median expression. Interestingly, among the directly bound AHR target genes, the expression level increases with the number of AHR-binding sites in the proximal promoter regions. Finally, we show that co-regulated genes in the AHR network activate distinct groups of downstream biological processes. Although the specific findings described here are restricted to hepatic effects under short-term TCDD exposure, this work describes a generalizable approach to the reconstruction and analysis of transcriptional regulatory cascades underlying cellular stress response, revealing network hierarchy and the nature of information flow from the initial signaling events to phenotypic outcomes. Such reconstructed networks can form the basis of a new generation of quantitative adverse outcome pathways.
KeywordsLigand-activated transcription factors DNA binding Dioxin response element Signaling Co-regulation Co-expression Phenotypic outcomes
Aryl hydrocarbon receptor
Aryl hydrocarbon nuclear translocator
ChIP-X enrichment analysis
Dioxin response element
Nuclear factor erythroid 2-related factor
Peroxisome proliferator-activated receptor alpha
Signal transducers and activators of transcription
Transcription start site
Intracellular signaling pathways, when sufficiently perturbed by exogenous chemicals, can lead to an adverse outcome at the cellular level, and potentially at the level of tissues and the whole organism. These perturbed pathways have been described as “toxicity pathways” (NRC 2007; Whelan and Andersen 2013). Signaling, transcriptional, and post-transcriptional regulatory networks underlie toxicity pathways and their dynamic behavior in response to endogenous and exogenous perturbation. It is crucial to understand the organization, structure, and dynamics of these networks through mapping and modeling approaches for a quantitative assessment of the risks of chemical exposure to biological systems. Tissue-specific network models of chemical-induced perturbation can improve our understanding of the intracellular events leading to adverse effects and eventual injury from chemical exposure.
The major cellular response pathways are governed both transcriptionally and post-translationally. A core set of master regulatory transcription factors (TFs) are central actors in most molecular pathways leading to altered expression of suites of genes in response to exposure to a variety of chemical compounds (Jennings et al. 2013). These TFs, including the nuclear receptors, p53, nuclear factor erythroid 2-related factor (NRF2), nuclear factor-κB (NF-κB), the STAT (signal transducers and activators of transcription) family, and the aryl hydrocarbon receptor (AHR), typically coordinate a broad range of physiological processes like metabolism, oxidative stress response, differentiation, tumor suppression, reproduction, development, and homeostasis (Audet-Walsh and Giguére 2015; Evans and Mangelsdorf 2014; Ma 2013; Tyagi et al. 2011; Wright et al. 2017). They thus act as sentinels of normal biological activity, but their inappropriate activation or inhibition can lead to adverse outcomes at the cellular or tissue level (Andersen et al. 2013).
Here, we describe a network model of the AHR pathway in the mouse liver, assembled from previously published genomic data sets, and newly analyzed using various computational methods. The AHR is a ligand-activated TF that belongs to the basic helix–loop–helix (bHLH)–PER–ARNT–SIM (PAS) family of proteins, which serve as sensors of developmental and environmental signals (Gu et al. 2000). The prototypical AHR ligand is TCDD (Poland et al. 1976), a persistent environmental toxicant that produces a variety of adverse effects in laboratory animals, including immune suppression, reproductive and endocrine effects, neurochemical alterations, developmental toxicity, chloracne, and tumor promotion (Birnbaum 1994; Pohjanvirta and Tuomisto 1994). These effects are mediated by the transcriptional activity of the AHR, as shown by their absence or amelioration in AHR-null mice and mice with low-affinity AHR alleles (Gonzalez and Fernandez-Salguero 1998; Okey et al. 1989; Peters et al. 1999), as well as in mice with mutations in the DNA-binding domain or nuclear localization sequence of the AHR (Bunger et al. 2003, 2008). Ligand binding causes the AHR in the cytosol to undergo a conformational change, resulting in dissociation from its chaperone protein complex and translocation to the nucleus, where it forms a heterodimer with the related nuclear protein aryl hydrocarbon nuclear translocator (ARNT) (Hoffman et al. 1991; Whitelaw et al. 1993). The AHR–ARNT complex then binds to specific DNA sequences on target genes called dioxin response elements (DRE) containing the core sequence 5′-GCGTG-3′ (Denison et al. 1988), leading to the regulation of a diverse battery of genes (Hankinson 1995; Poland and Knutson 1982). While the 5′-GCGTG-3′ nucleotide core is substitution-intolerant, the flanking 5′ and 3′ nucleotides adjacent to the core sequence also contribute to a functional AHR-binding site (Denison et al. 1988; Gillesby et al. 1997; Lusska et al. 1993; Shen and Whitlock Jr 1992). DRE-independent mechanisms of AHR binding have also been reported (Dere et al. 2011b; Huang and Elferink 2012).
While the density of AHR-bound regions in the genome of hepatic tissue from TCDD-treated mice is greatest in proximal promoter regions close to the transcription start site (TSS) of annotated genes, AHR also binds to sites distal from a TSS, e.g., in intergenic regions and 3′ UTRs (Dere et al. 2011b). Moreover, only a third of the differentially expressed genes identified by microarray analysis showed AHR binding at a DRE in their proximal promoter regions, suggesting additional mechanisms of gene regulation by AHR beyond the canonical model described above (Dere et al. 2011b). These mechanisms may include target gene regulation from distal AHR-bound regions through DNA looping, or indirect regulation by AHR through tethering with a secondary TF (Farnham 2009). Such an indirect mechanism has been demonstrated in the regulation of the rat CYP1A2 gene by AHR (Sogawa et al. 2004).
Here, we have mapped the TCDD-induced AHR regulatory network from a combination of previously published gene expression and ChIP-on-chip data from the liver of female C57BL/6 mice orally gavaged with 30 µg/kg of TCDD (Dere et al. 2011b), which provides us a system-wide view of AHR-mediated gene regulation under short-term TCDD exposure. Specifically, statistical and visualization tools were used to establish a relationship between gene co-regulation by multiple TFs and gene co-expression, and link groups of co-regulated genes to distinct downstream functional outcomes. Such reconstructed networks can form the basis of a new generation of quantitative adverse outcome pathways (Conolly et al. 2017; Perkins et al. 2019). Our focus here is on the early stages of hepatic response to TCDD exposure—longer term exposure may lead to a different suite of adaptive responses at the cellular and tissue level.
Materials and methods
Our network analysis was based on results from a previous study of gene expression profiling using whole-genome oligonucleotide arrays (Agilent Technologies, Santa Clara, CA) of hepatic tissues from female C57BL/6 mice orally gavaged with 30 µg/kg of TCDD (Boverhof et al. 2005; Dere et al. 2011b). The gene expression analysis was performed in hepatic tissue from mice exposed to TCDD for 2, 4, 8, 12, 18, 24, 72, and 168 h. Differentially responsive genes were identified using previously described cutoffs for fold change and statistical significance (|fold change| ≥ 1.5 and posterior probabilities P1 (t) ≥ 0.999) (Dere et al. 2011b; Eckel et al. 2004).
Genome-wide AHR location data were taken from the previously described ChIP-on-chip experiments (Dere et al. 2011b), where ChIP assays were performed with hepatic tissue from female C57BL/6 mice exposed to TCDD for 2 and 24 h. Genes were associated with AHR-enriched regions if the position of maximum fold enrichment was within 10 kb upstream of a transcriptional start site (TSS) through to the end of the 3′ UTR. For the present analysis, the ChIP data for 2 and 24 h were combined to obtain a unique list of ChIP-enriched regions associated with annotated genes (Supplementary Methods; Supplementary Code 1). The choice of data sets for our analysis was constrained by the requirement of matched mouse liver gene expression and ChIP data sets under similar conditions.
DRE analysis in ChIP-enriched regions
The ChIP-enriched regions for the differentially expressed (DE) genes were computationally searched for the presence of 5′-GCGTG-3′ DRE core sequences to infer the nature of AHR binding to the target genes. The putative DRE search algorithm, written in R (R Core Team 2016) (Supplementary Methods; Supplementary Code 2), was based on a previously described approach (Sun et al. 2004). Briefly, the genomic sequences of the enriched regions were obtained from UCSC Genome Browser (https://genome.ucsc.edu) and scanned for exact matches to the DRE core sequences on both positive and negative strands. For each matched region, the 5-bp core sequence was extended 7 bp upstream and downstream of the core. The matrix similarity (MS) scores (Quandt et al. 1995) for the 19-bp DRE sequences were calculated and compared to an MS score threshold of 0.8473 based on the lowest MS score of 13 bona fide AHR-binding sequences (Dere et al. 2011a) (i.e., sites from the literature confirmed to bind AHR). The DRE sequences with high MS scores (MS score ≥ 0.8473) were defined as putative DREs capable of binding AHR. The DE genes that were AHR-enriched and had a putative DRE in the enriched region were described as “directly bound” by AHR, while AHR-enriched genes without a putative DRE were described as “indirectly bound”. The remaining DE genes that were not AHR-enriched were regarded as “unbound”/“non-genomic” targets.
Construction and visualization of the AHR transcriptional regulatory network
The DE genes from the Agilent oligonucleotide array data were searched against online databases to obtain a list of TFs that regulate these genes. The ChIP-X Enrichment Analysis (ChEA2) database (Kou et al. 2013) was used to obtain the list of regulatory TFs. To obtain the mouse-liver specific list of transcription factors, the mouse-specific TFs from ChEA2 were screened for expression in the liver using the TRANSFAC® database (Matys et al. 2003). The ensemble of DE genes including the directly and indirectly AHR-bound genes, together with their inferred transcriptional regulators, form a comprehensive network for TF–gene interactions under AHR-mediated TCDD induction. The landscape of this regulatory network was rendered using the open-source network visualization tool Cytoscape (Shannon et al. 2003). The gene expression values at each time point of TCDD exposure were superposed on this network to visualize the temporal changes associated with each gene. A |fold ratio| threshold of 1.20 was used to identify the key target genes that are themselves TFs regulating other genes in the data set (a less stringent fold change threshold was used for TFs than other genes as TFs tend to be more tightly regulated). To generate and annotate the network in Cytoscape, three input files describing the network topology and gene expression values were used: an AHR–gene interaction file and a TF–gene interaction file (“network files”), and a gene expression file (“attributes file”). Log2 scaling of the fold ratios was used for visualizing gene expression. The network files were merged together to form the complete layout.
Gene expression analysis based on transcriptional groupings
A binary TF–gene interaction matrix with 43 TFs in addition to AHR was created indicating which TFs interact with which target genes. If a gene is regulated by a particular TF, then the corresponding interaction is represented as ‘1’; otherwise, it is represented as ‘0’. We used this TF–gene interaction matrix to classify target genes into co-regulated groups in a transcriptional cascade, to examine any possible relation between co-regulation and co-expression. To generate this grouping, AHR and other key TFs that were also target genes were considered in all possible combinations to identify the expression trends for target genes in each group. The total number of genes in each co-regulated group was counted by referring to the TF–gene interaction matrix, and all groups with at least five genes were considered for examination of the expression patterns. A graphical analysis was performed in R to identify the expression patterns of target genes for each combination of regulatory TFs (Supplementary Methods; Supplementary Code 3).
Kohonen self-organizing maps to visualize gene co-expression
To further examine the relationship between the transcriptional groups and target gene expression patterns, a self-organizing map (SOM) for the AHR network was generated using the Kohonen SOM package in R (Wehrens and Buydens 2007). The same TF–gene interaction matrix described above was used as input for this analysis. The SOM algorithm follows a clustering technique to group the target genes according to their TF-binding patterns. Target genes with similar TF-binding patterns are grouped into the same cluster or adjacent clusters, referred to as ‘units’ (Supplementary Methods; Supplementary Code 4).
The ORCLUS subspace clustering algorithm (Aggarwal and Yu 2000) and corresponding R package (Szepannek 2013) were used to cluster the differentially expressed genes into 16 non-overlapping groups. The number of clusters k = 16 and the dimensionality of each cluster l = 4 were chosen so as to minimize the cluster sparsity coefficient (Aggarwal and Yu 2000) (Supplementary Code 5).
Functional categorization of genes in each cluster
Gene ontology (GO) functional analysis was performed for the DE genes present in each ORCLUS cluster. Enriched GO “process” categories were identified for genes in each cluster using the GOrilla tool (Eden et al. 2009) with a p-value threshold of 10–3 and the list of all DE genes as background. REViGO (Supek et al. 2011) was used to arrange the enriched processes into a “treemap”, which was then rendered as an image using the downloadable R script generated by the program (Supplementary Code 6, Supplementary Code 7).
Differential gene expression
The raw array data set (Dere et al. 2011b) consisted of 41,267 records with annotated genes, fold ratio and significance [P1 (t) values] at 2, 4, 8, 12, 18, 24, 72, and 168 h post-TCDD exposure. For genes with multiple occurrences in the dataset, the fold ratios and P1(t) values were averaged, resulting in a total of 21,307 unique gene records. After applying the statistical cutoff values for fold change and P1(t) at each expression time point, the resulting number of unique differentially expressed (DE) genes was 1407. All 1,407 DE genes were used to generate the AHR regulatory network map.
Analysis of AHR-enriched genomic regions associated with DE genes
Regions with one or more 5-bp DRE cores centrally located, such that a 7-bp upstream and downstream extension was possible for MS score calculations.
Regions with DRE cores present only at the edge of the region, so that the 7-bp extension in both directions was not possible.
Regions with no DRE core.
A total of 144 genes were associated with AHR-enriched regions where MS score calculations were possible, and that had putative DREs, i.e., 19-bp DRE sequences with an MS score ≥ 0.8473 (see “Materials and methods”). These genes were considered to be “directly bound” by AHR. For the AHR-enriched regions with (1) non-putative DRE core (i.e., MS score < 0.8473), (2) DRE core located at edges, or (3) DRE core not present in the enriched region, the associated genes were considered to be “indirectly bound” by AHR. In total, among the 1407 differentially expressed genes, 632 were bound by AHR with 144 genes directly bound and 488 indirectly bound, and the remaining 775 genes unbound by AHR.
Other transcriptional regulators of the DE genes
The AHR regulatory network
All interactions of the DE genes with AHR and the other 43 identified TFs together form the mouse liver AHR regulatory network (Fig. 1a), which consists of 44 “source” nodes interacting with 1,241 “target” nodes.
AHR and the other seven hub TFs act as both source and target nodes (AHR regulates itself). Two of these hub TFs are regulated by AHR: NRF2 is a direct target and Fli1 an indirect target (network schematic in Fig. 1a). The expression levels for up- and down-regulated genes were superposed on this network layout for each of the eight time points in the gene array study (Supplementary Fig. 1a–h), illustrating that the gene expression levels were not monotonic in time.
Count of genes in each transcriptional grouping
Co-regulation and co-expression in the AHR network
Localized clustering of co-regulated genes
We further attempted to cluster the 1,191 target genes considered in the SOM analysis above based on regulation by the 44 TFs. Fundamentally, the clustering problem may be stated as: “Given a set of data points, partition them into a set of groups which are as similar as possible” (Aggarwal 2014). If we consider the binary TF–gene connectivity matrix, with genes in rows (observations), TFs in columns (features), and each matrix element equaling 1 or 0 depending on whether a TF binds a gene, we have a high-dimensional clustering problem with feature localization, i.e., different groups of genes are regulated by different subsets of TFs. Global clustering methods like k-means or dimensionality reduction approaches like principal components analysis do not perform well in this situation, which motivated the development of high-dimensional subspace clustering methods (Aggarwal 2014). These methods include “projected clustering” or “subspace clustering” approaches like PROCLUS (Aggarwal et al. 1999), CLIQUE (Agrawal et al. 2005), and ORCLUS (Aggarwal and Yu 2000), where feature selection or transformation is performed specific to different localities of the data (Aggarwal 2014). ORCLUS in particular is suited for data sets like ours where relevant subspaces may be arbitrarily oriented due to inter-feature correlations (Aggarwal and Yu 2000), i.e., many TFs are correlated in term of which genes they regulate.
Distinct gene clusters activate distinct biological processes
We carried out gene ontology (GO) analysis on the six major clusters of genes labeled in Fig. 3a. The genes in the six clusters enrich for different groups of biological processes (Fig. 3b). In particular, the genomic cluster (Cluster 6) is enriched for genes associated with metabolic processes and ribosome biogenesis, whereas the major GO categories associated with the non-genomic cluster (Cluster 2) are immune regulatory processes. Interestingly, Cluster 5 is enriched for cell migration and activation of cellular defense mechanisms. Presumably, this reflects immune cell infiltration into the mouse liver under exposure to TCDD (Fader et al. 2015). Cluster 16 is also enriched for immune system response. Thus, co-regulated genes in the AHR network in the mouse liver show patterns of co-expression and lead to differential downstream activation of biological processes.
AHR binding and gene expression
These differences between direct, indirect, and unbound AHR target genes are also highlighted in overlaid box and violin plots (Fig. 5d–g; Supplementary Fig. 3), showing the respective distributions of expression level of the three groups of genes at multiple time points. At each time point shown, the middle 50% (first to third quartile) of the directly bound genes are all up-regulated, while the indirectly bound group is symmetrically distributed with about half of the genes up-regulated. In the unbound group, most genes are down-regulated at earlier time points, but at 168 h, the distribution is considerably right-skewed with many genes up-regulated. Overall, the directly regulated group has the highest median expression (except at 168 h), and also has the most outliers on the high expression end, the furthest outlier being the CYP1A1 gene.
Conclusions and discussion
Ligand-activated transcription factors underlie most major cellular response pathways. These TF-governed molecular pathways tend to have a similar organizational structure with key functional components that act as signal sensors (co-binding proteins) and transducers (protein kinases) to complement the central role of the TF (Simmons et al. 2009). The inactivated TF is typically sequestered in the cytoplasm or nucleus. Upon activation by its ligand (endogenous or exogenous molecule), the TF is able to bind specific response elements in the promoter regions of target genes and activate or inhibit expression of suites of genes in a coordinated manner. Beyond these “direct target” genes, there are additional genes that bind the master regulatory TF indirectly through tethering interactions with secondary TFs (George et al. 2011; McMullen et al. 2014; Shen et al. 2011). In fact, combinatorial control of gene expression by TFs is a common feature of cellular pathways, since binding sites are often clustered in the genome, allowing multiple TFs to act in a coordinated fashion to induce or suppress groups of genes in specific cell types under particular conditions (George et al. 2011). In addition, a surprisingly large number of genes are activated or inhibited in a “non-genomic” manner, showing no evidence of binding by the master regulatory TF of the stimulated pathway in their promoter regions (Dere et al. 2011b; McMullen et al. 2014; Shen et al. 2011; van der Meer et al. 2010). These observations collectively suggest that combining gene expression data from transcriptome profiling with high-throughput genome-wide analysis of TF binding can provide an integrated, systems-level view of the structure and function of transcription factor-governed molecular pathways (Blais and Dynlacht 2005; Dere et al. 2011b; Limonciel et al. 2015; Walhout 2006).
Accordingly, we have integrated TCDD-induced gene expression and multiple genome-wide TF-binding data sets for a global view of the AHR regulatory pathway in the mouse liver. Using a combination of self-organizing maps and subspace clustering, we show that there is a pattern of co-regulated genes in the AHR pathway being co-expressed, as previously observed in Saccharomyces cerevisiae (Allocco et al. 2004; Yu et al. 2003). In particular, directly bound, indirectly bound, and unbound AHR target genes have distinct patterns of gene expression, with the directly bound group showing higher median expression. Furthermore, among the direct AHR target genes, the expression level increases with the number of AHR-binding DRE sites in the proximal promoter regions. Finally, we found that co-regulated gene clusters activated distinct groups of downstream biological processes, with the AHR-bound genomic cluster enriched for metabolic processes and the AHR-unbound non-genomic cluster primarily activating immune processes. This work, together with the other recent studies of the peroxisome proliferator-activated receptor alpha (PPARα) and estrogen receptor pathways (McMullen et al. 2014; Pendse et al. 2016), illustrates the application of bioinformatic and statistical tools for reconstruction and analysis of the transcriptional regulatory cascades underlying cellular stress response. While these network reconstructions are species, tissue, and condition-specific, we anticipate that next-generation models that use machine learning to predict network structure and dynamics from genomic sequence and epigenomic features will soon be available. Such models will reduce our reliance on expensive assays for gene expression and genome-wide protein binding for different animal models, tissues, and exposures.
The work presented here describes a detailed map of the AHR transcriptional regulatory network activated by TCDD exposure in the mouse liver. This map can be used to derive a predictive model of TCDD dose-dependent genomic response. The overall map (Supplementary Fig. 1) contains a large number of network edges that may be difficult to model quantitatively. However, our findings regarding gene co-regulation and co-expression imply that simplified representations of the network with groups of co-expressed genes represented as individual entities (as in Fig. 1a) may be sufficient to model the overall network response. Such a simplified representation would also be more easily quantifiable in terms of parameterizing the regulatory interactions. We foresee such quantitative dose–response models as a crucial step in the development of more rigorous mechanistic risk assessment protocols.
The authors would like to thank Agnes Karmaus, Arindam Banerjee, Rory Conolly, and Qiang Zhang for helpful discussions. This work was supported by the US EPA STAR Program (EPA Grant number: R835000), the USDA National Institute of Food and Agriculture and Michigan AgBioResearch, and the Superfund Research Program of the National Institute of Environmental Health Sciences (Grant number: P42ES04911). No funding body played any role in the design of the study, collection, analysis, and interpretation of data, or in writing the manuscript.
The study was conceived and coordinated by SB. NJ and SB analyzed and interpreted the data. The manuscript was written by NJ and SB with input from MEA, NEK, ED, and TRZ. All authors read and approved the final manuscript.
Compliance with ethical standards
Conflict of interest
The authors declare that they have no conflict of interest.
- Aggarwal CC (2014) An introduction to cluster analysis. In: Aggarwal CC, Reddy CK (eds) Data clustering: algorithms and applications. Chapman and Hall Press, Boca Raton.Google Scholar
- Bunger MK, Glover E, Moran SM et al (2008) Abnormal liver development and resistance to 2,3,7,8-tetrachlorodibenzo-p-dioxin toxicity in mice carrying a mutation in the DNA-binding domain of the aryl hydrocarbon receptor. Toxicol Sci 106(1):83–92. https://doi.org/10.1093/toxsci/kfn149 CrossRefPubMedPubMedCentralGoogle Scholar
- Bunger MK, Moran SM, Glover E et al (2003) Resistance to 2,3,7,8-tetrachlorodibenzo-p-dioxin toxicity and abnormal liver development in mice carrying a mutation in the nuclear localization sequence of the aryl hydrocarbon receptor. J Biol Chem 278(20):17767–17774. https://doi.org/10.1074/jbc.M209594200 CrossRefPubMedGoogle Scholar
- Gillesby BE, Stanostefano M, Porter W, Safe S, Wu ZF, Zacharewski TR (1997) Identification of a motif within the 5' regulatory region of pS2 which is responsible for AP-1 binding and TCDD-mediated suppression. Biochemistry 36(20):6080–6089. https://doi.org/10.1021/bi962131b CrossRefPubMedGoogle Scholar
- Kou Y, Chen EY, Clark NR, Duan Q, Tan CM, Ma‘ayan A (2013) ChEA2: gene-set libraries from ChIP-X experiments to decode the transcription regulome. In: Cuzzocrea A, Kittl C, Simos DE, Weippl E, Xu L (eds) Availability, reliability, and security in information systems and HCI: IFIP WG 84, 89, TC 5 international cross-domain conference, CD-ARES 2013, Regensburg, Germany, September 2–6, 2013 proceedings. Springer, Berlin, pp 416–430CrossRefGoogle Scholar
- NRC (2007) Toxicity testing in the 21st century: a vision and a strategy. The National Academies Press, Washington, DCGoogle Scholar
- R Core Team (2016) R: a language and environment for statistical computing. R Foundation for Statistical Computing, ViennaGoogle Scholar
- Szepannek G (2013) orclus: ORCLUS subspace clustering. R package version 0.2-5Google Scholar
- Whelan M, Andersen ME (2013) Toxicity pathways—from concepts to application in chemical safety assessment. Publications Office of the European Union, JRC, LuxembourgGoogle Scholar
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.