We chose 11 model compounds, namely methylmercury chloride (MeHg), CdCl₂ (Cd), PbCl2 (Pb), As₂O₃ (As), Aroclor 1254 (PCB), acrylamide (AA), tert-butylhydroquinone (tBHQ), 4-chloroaniline (4CA), 1,1-bis-(4-chlorophenyl)2,2,2-trichloroethane (DDT), 2,3,7,8-tetrachlorodibenzo-p-dioxin (TCDD) and valproic acid (VA). These are compounds known for their environmental toxicity 24 and VA is a teratogen and an anti-epileptic drug 25. VA is known to inhibit histone deacetylases and Wnt signaling in mammals, thus adding an additional mode of toxic action 26.

We first established exposure protocols with which one can trigger toxicogenomic alterations with high likelihood and at the same time cause only a small amount of cell death or embryo mortality. We limited the exposure time to 20-24 hours in the expectation of focusing predominantly on primary responses rather than indirect, secondary effects. Finally, we decided to carry out these assays in embryos before they begin to feed, that is, before 120 hours post-fertilization (hpf). We tested a range of toxicant concentrations to determine the one that caused a morphologically visible toxic/teratological effect in the treated embryos after exposure at 96-120 hpf (Figure 1, Table 1, Additional data file 1). We were not able to discriminate unequivocally between toxicant-specific morphological effects (see Figure 1). Frequently the tails were bent, and the animals had difficulty swimming correctly; in some instances they developed pericardial edema (see Figure 1a). Vehicle-treated embryos did not show alterations (see Figure 1l-n) or did so only at very low frequency. Cell death as monitored by acridine orange staining was not, or only rarely, obvious immediately after treatment when animals were sacrificed for microarray analysis.

Between 96 and 120 hpf organogenesis has proceeded so far that the animals feed for the first time 8, marking the end of the embryonic stage. At this stage, gut, liver, pancreas, nervous system, musculature and the cardiovascular system are assumed to reflect adult physiology in many respects, including the response to toxicants. Younger embryonic stages are likely to have different responses to the toxicants. We therefore included two more stages in our initial experiments. The 4-24 hpf treatment covers late blastula, gastrula and segmentation stages, during which the overall body plan is laid down 8. The treatment phase between 24 and 48 hpf coincides with the onset of organogenesis 8. Early embryonic stages appear more sensitive to toxicant exposure than the older embryos (compare the 24-48 hpf and 96-120 hpf treatment groups in Table 1). The concentrations of the toxicants were adjusted accordingly (see Table 1).

To assess possible stage-specific differences, we analyzed and compared the toxicogenomic response to six compounds - MeHg, Cd, 4CA, DDT, TCDD, and VA - at the three different stages. We treated several hundred embryos with each of these compounds at each of the three stages (see Materials and methods and Additional data files 1). Principal component analysis (PCA) revealed distinct toxicogenomic responses to exposure with the six toxicants in the 24-48 and 96-120 hpf treatment groups (Figure 2a). Principal components were derived by singular value decomposition (SVD). SVD is based on the decomposition of the gene-expression matrix, whose entries are the log-transformed fold changes (M values) of gene expression, into unique orthonormal superpositions of genes and treatments. Expression changes of at least twofold and a p_adj < 0.025 were taken into account. The p_adj-value was adjusted for multiplicity testing by controlling the false discovery rate 27.

The differences between the transcriptional profiles induced by the six toxicants were less prominent in the datasets from the 4-24 hpf treatment groups (see Figure 2a). This may be due to the fact that different toxicants caused similar gene effects at 24 hpf. For example, the expression of the gene for fast muscle troponin T (BE693169) was downregulated by Cd, MeHg, TCDD, and VA in embryos treated between 4 and 24 hpf but not at later stages (data not shown). Furthermore, many genes that are involved in organ physiology may not yet be responsive by 24 hpf, as organ development has not proceeded far enough. In agreement with this, the expression levels of only 57 genes were significantly altered by the 4-24 hpf treatment. In contrast, the expression levels of 476 and 311 genes were significantly affected by the 24-48 hpf and 96-120 hpf treatment regimens, respectively (see Figure 2b). Moreover, very few genes in the 4-24 hpf treatment set overlapped with the 24-48 and 96-120 hpf treatment groups (15 and 10 genes, respectively). The latter groups (24-48 h and 96-120 hpf) shared more gene responses (74 genes) but 393 and 233 gene responses were stage specific (see Figure 2b). The smaller number of affected genes in the 4-24 hpf regimen may also have been caused by the lower concentrations of toxicants that we had to apply to ensure sufficient survival at these younger stages. Irrespective of this, these data indicate a high stage specificity of the toxicogenomic effects in the three treatment windows.

We focused further analysis on the 96-120 hpf stage and used the full set of 11 toxicants by including treatments with AA, PCB, As, tBHQ and Pb. Replicate hybridizations with mRNA from at least three independent toxicant treatments were performed (see Table 1). Toxicant effects were clustered based on their Euclidean distance to each other and the similarity of gene responses was determined by a Pearson correlation proximity measure. The expression profiles summarize clustering results for a subset of 199 genes across all 11 toxicant responses (Figure 3). The gene-selection criteria applied take into account the extent and significance of changes in gene expression (at least twofold, p_adj < 0.025) as well as differences and similarities in expression changes between toxicants (see Materials and methods). Distinct patterns of gene expression were noted for each of the 11 compounds. However, similarities in gene responses were also detected. One group of chemicals with related gene responses comprises Pb, As, Cd, tBHQ, MeHg and VA (see Figure 3, lanes 6 to 11). Another subgroup of related responses was induced by TCDD, 4CA, DDT and AA (see Figure 3, lanes 2-5), whereas the PCB triggered a more distinct expression profile (see Figure 3, lane 1).

As verification, we carried out blind tests to identify the chemicals by their induced gene-expression profile. Fourteen out of the 15 chemicals were unambiguously identified (Table 2). In the case of 4CA, close matches were scored to the 4CA, the DDT and the AA response profiles (see Table 2). Thus, we identified the correct group of chemicals (see Figure 3, lanes 2-4). Taken together, the results from these blind trials underscore the reliability of the toxicogenomic profiles and furthermore suggest that it is possible to derive signatures of toxicogenomic responses predictive for specific chemicals or chemical groups from whole animal exposure experiments.

The induced genes fell into different gene ontology groups such as genes involved in combating oxidative stress (Table 3) and genes encoding chaperones (Table 4). Another major class of genes that was significantly regulated by a number of toxicants comprised solute carriers (Table 5). We also carried out a computational analysis of the affected genes using the GoTreeMachine algorithm to identify more complex pathways and processes (Additional data files 2-9). An inflammatory response was induced by several compounds (As, 4CA, Cd, MeHg, Pb, PCB and tBHQ), whereas inductions characteristic of an immune response were evoked by MeHg and tBHQ. The latter compound also triggered genes involved in G-protein-coupled signaling and phototransduction. Induction of genes with a function in base-excision repair was noted in the case of exposure to As and PCB, suggesting that these compounds cause DNA damage in the embryo.

We next verified the observed gene responses by methods other than microarray hybridization. First, the changes in gene expression were confirmed by re-evaluating a subset of gene responses by semi-quantitative reverse transcription PCR (RT-PCR). Out of 14 gene responses analyzed, all showed the up- or downregulation expected from the array data (Figure 4). This suggests that the changes in transcript levels measured by the microarray hybridizations reflect genuine responses to the toxicants.

We used in situ hybridization with selected probes to toxicant-treated and control embryos to assess the tissue-specific expression patterns of the response genes and whether these are altered in response to toxicant. Cytochrome P4501A1 mRNA (AF057713) was induced by 500 ng/l TCDD in endothelial cells (15/15 embryos, Figure 5a,b). The levels of glutathione peroxidase 1 (AW232474) mRNA in stomach and gut were repressed by 60 μg/l MeHg (11/15 embryos, Figure 5c,d), in agreement with microarray and RT-PCR data (see Figure 4).

The neuromasts of the zebrafish lateral line are very sensitive to a number of compounds including CdCl₂ 282930. the mRNA for oncomodulin A (also called parvalbumin3a), which is expressed in the hair cells and supporting cells of neuromasts in untreated embryos, is barely detectable in the neuromasts of embryos treated with 500 μg/l CdCl₂ (13/15 embryos, Figure 5e,f), in concordance with the Cd-induced, 4.4-fold decrease of oncomodulin A mRNA measured by microarray hybridization (see Figure 4). In contrast, thioredoxin-like mRNA (BI864190) is upregulated in hair cells (12/13 embryos, Figure 5g,h) in response to Cd. This suggests that Cd does not cause a complete loss of hair cells, even though staining with the dye DASPEI suggests that hair cells are strongly reduced (data not shown). The thioredoxin-like mRNA is also expressed in selected areas of the brain. These regions show also increased levels of expression in response to Cd (data not shown). In summary, these in situ expression studies show that the microarray procedure used permits detection of organ- and cell-specific gene responses with very high sensitivity. Moreover, these results also suggest that the gene responses occur in almost all of the embryos exposed to the toxicants.

The concentrations of the toxicants were adjusted in the initial experiments so that they caused morphologically visible defects in exposed animals. We asked next whether one could measure changes in the expression profiles at lower concentrations that do not have apparent morphological effects. TCDD, DDT, Cd, 4CA, MeHg, and VA were used as a set of test compounds. We could detect significant changes in gene expression (at least twofold and p_adj < 0.025) in response to 0.5 mg/l Cd, 6 μg/l MeHg, 5 mg/l VA, 25 mg/l 4CA, 15 mg/l DDT, and 50 ng/l TCDD (Figure 6a-c, Table 6, and data not shown). With the exception of 6 μg/l MeHg and 25 mg/l 4CA, these low concentrations did not cause obvious morphological or behavioral defects (data not shown), suggesting that this assay can detect responses to toxicant concentrations that do not cause acute morphological effects. It is clear, however, that the number of genes with a significant response to the toxicants decreases (see Table 6). Cytochrome P4501a1 was fivefold upregulated by 50 ng/l TCDD, oncomodulin A was reduced 4.5-fold by 0.5 mg/l Cd and peroxiredoxin was still 3.5-fold induced by 6 μg/l MeHg. Thus, even though fewer genes respond to these lower concentrations, the measured changes in transcript levels are robust.

In the environment we are normally confronted with compound mixtures rather than pure substances. The components of these mixtures could act synergistically, thereby potentiating the toxic effect 31. We therefore investigated whether synergistic effects of compound mixtures can be observed in toxicogenomic profiles. To this end, 96-hpf embryos were exposed to a mixture of low concentrations of Cd (50 μg/l), Pb (280 μg/l), MeHg (6 μg/l) and As (7.9 mg/l). About twice as many genes (158 genes) were significantly up- or downregulated (absolute change at least twofold, p_adj < 0.025) than the sum of the genes regulated by exposure to the individual toxicants (81 genes: Cd 48 genes; As 12 genes; MeHg 5 genes; Pb 16 genes). Complex expression profiles composed of both additive and synergistic as well as novel patterns of gene responses (at least twofold change, p_adj < 0.025; Figure 6d) were scored for the mixture. In the case of the genes with similarity to peroxiredoxin (BI980610) or the solute carrier family members 6 and 9 (BE016639, AW421040), the response to the mixture appears to be purely additive (see Figure 6d, dots; Table 6, and Additional data file 13). In other instances, for example the Hsp70-related genes (AB062116, AF210640, AF006007) or the sequestosome1 gene (AW343560), the mixture induced a strong increase in transcript levels, whereas significant gene responses (more than twofold, p_adj < 0.025) were not induced by administration of the individual compounds (see Figure 6d, arrowheads, Table 6, Additional data file 13). These genes can, however, be induced by higher concentrations of the individual compounds (see Figures 3, 6a,b and Additional data files 10, 11), suggesting that the observed synergy is the result of a lowered response threshold. Curiously, we also noted loss of gene responses on exposure to the compound mixture (see Figure 6d), suggesting suppressive effects of the combination. For example, the transcript levels of glutathione-S-transferase omega 1 (AW019036) are significantly altered by exposure to PbCl₂, but not by the mixture (see Table 3 and Additional data file 13). In a few instances we observed opposing effects, such as in the case of suppressor of cytokine signalling 3 (BI878700), which was 4.9-fold downregulated by As and 2.6-fold upregulated by the mixture (see Additional data file 13). Taken together, these results show potentiated, additive, and nonadditive effects of the mixture in comparison to the individual compounds.

AA (acrylamide; CH₂ = CHCONH₂), PCB (Aroclor 1254), As (arsenic (III) oxide; As₂O₃), tBHQ (tert-butylhydroquinone; (CH₃)₃CC₆H₃-1,4-(OH)₂), Cd (cadmium chloride; CdCl₂2H₂O), 4CA (4-chloroaniline; ClC₆H₄NH₂), DDT (1,1-bis-(4-chlorphenyl)-2,2,2-trichlorethane; ClC₆H₄)₂CHCCl₃), Pb (lead (II) chloride; PbCl₂), MeHg (methylmercury chloride; CH₃ClHg), TCDD (2,3,7,8-tetrachlorodibenzo-p-dioxin) and VA (valproic acid; CH₃(CH₂)₄CO₂H) were purchased from Sigma-Aldrich (St Louis, MO).