In all, 9- to 11-week-old B6C3F1 male mice, weighing about 30 g (Charles River Breeding Laboratories, Raleigh, NC) were acclimatized, randomized into treatment groups, and housed individually in each study. Room environmental conditions had design specifications as follows: minimum of 12 air changes per hour with air filtered through 90–95% efficiency filters and then through high efficiency particulate air filters, relative humidity of 50 ± 10%, temperature of 70°F ± 5°F, and a 12-h light/dark cycle. Certified Rodent Diet 5002 (PMI Feeds, Inc., St. Louis, MO) and municipal drinking water, further purified by reverse osmosis, were provided ad libitum. The animal care and experimental procedures of these studies were conducted in compliance with the U.S. Animal Welfare Act and were performed in accordance with the standards of the Institute of Laboratory Animal Resources Guide (1996). The Association for Assessment and Accreditation of Laboratory Animal Care International has accredited the Pfizer facility in which the studies were conducted.

Mice were administered either 900 mg/kg 2-BE (Sigma-Aldrich, St Louis, MO) in deionized water or deionized water alone by oral gavage daily for 7 days (10 animals per treatment group). Each dose group was further divided into two arms, one (n = 5) for Hypoxyprobe (NPI, Burlington, MA) immunohistochemical (IHC) staining and the other (n = 5) for RNA isolation and transcriptional profiling. Hypoxyprobe was injected i.p. 2.5 h prior to sacrifice. Additional treatment groups (four animals per group) received only a single dose of 2-BE (900 mg/kg) or deionized water by oral gavage and were sacrificed 4 h after dosing for gene expression analysis. Red blood cell (RBC) counts and hemoglobin (Hgb) were determined using an Advia 120 hematology analyzer (Siemens, Tarrytown, NY) to verify hemolytic effects. All animals were anesthetized with ketamine and sacrificed by cervical dislocation. BM (left femur), liver, and spleen were collected in 10% neutral buffered formalin (NBF) and processed using routine methods for microscopic evaluation. The dose chosen for the 2-BE studies (900 mg/kg) was based on Siesky et al. (2002), where treatment of mice with 900 mg/kg 2-BE for 7 days produced a significant decrease in hematocrit, and on Park et al. (2002b), where 900 mg/kg 2-BE for 7 days also produced oxidative stress in the liver.

Mice were subjected to either whole-body hypoxia or normoxia. Unrestrained mice were treated in a plethysmography chamber (PLY 3211; Buxco Research Systems, Wilmington, NC) connected to a Bias Flow Regulator (PLY1040; Buxco Research Systems), which provided normobaric airflow of 1.0–1.5 ml/min. Mice (n = 4) were exposed to hypoxia that followed a course of 8% O₂ pretreatment for 30 or 60 min followed by a 60- to 90-min 6% O₂ treatment. Control mice (n = 4) were exposed to normal air in the plethysmography chambers for an equal duration. Mice were sacrificed by cervical dislocation immediately following hypoxia treatment. Tissues were evaluated for gene expression or Hypoxyprobe immunohistochemistry.

Hypoxyprobe (pimonidazole HCl) uses an IHC method for the detection of local tissue hypoxia (http://www.hypoxyprobe.com). In a low oxygen environment, the pimonidazole is reductively activated and binds to cysteine residues and thiols forming protein adducts in hypoxic cells that are detected using standard IHC methods. The reagent has been used to detect local tissue hypoxia in liver (Arteel et al., 1996). Liver, spleen, and BM (from femurs) from mice injected with Hypoxyprobe were extracted and immediately placed in ice-cold 10% NBF for 24 h. BM samples were subsequently placed in ImmunoCAL (Decal Chemical, Tallman, NY) for an additional 24 h. All tissues were processed and embedded in paraffin blocks according to standard procedures. Tissue sections were cut at 5 μm and placed on glass slides. Citrate antigen retrieval was done using high heat and pressure (Chemicon, Temecula, CA) followed by peroxidase block (Sigma-Aldrich) and protein block (Dako, Carpinteria, CA). IHC staining was performed at room temperature on an automated Dako Autostainer (Dako). The Hypoxyprobe reagent was run at a dilution of 1/500 for 60 min, detected using Dako Envision+ Rabbit HRP Polymer for 30 min at room temperature, and then visualized with Dako Liquid DAB+. Slides were counterstained with Mayer's hematoxylin and cover slipped. Negative controls for the study included a rabbit isotype control antibody (Dako) incubated for 60 min. In addition, tissue samples from animals not treated with pimonidazole were stained with the same protocol to show the absence of nonspecific anti-Hypoxyprobe antibody staining. All images were captured at ×20 using a Hammamatsu NanoZoomer (Olympus, Center Valley, PA) slide scanner and stored in a Bacus image format. Definiens (Munich, Germany) image analysis software was used to calculate the stain ratio (stain area/tissue area). See Supplementary Materials for more details on image quantification.

A portion of the left lateral liver lobe and a half cross-section of spleen were rapidly dissected from mice and flash-frozen in liquid nitrogen. The BM cells were collected via a BM flush utilizing 700 μl Paxgene BM buffer (Qiagen, Inc., Valencia, CA) from the right humerus and frozen at −80°C. Total RNA was extracted from a minimum of four mice per treatment group, along with corresponding vehicle controls. The frozen liver and spleen were crushed, and 50–100 mg of the crushed tissue was used to prepare total RNA by extracting in Trizol (Invitrogen, Carlsbad, CA) followed by purification on a Qiagen RNeasy Mini column as recommended by the manufacturer (Qiagen, Inc.). Total RNA was isolated from the BM in Paxgene buffer as recommended by the manufacturer (Qiagen, Inc.). The quantity and purity of the RNA were determined by absorbance at 260 and 260 nm/280 nm absorbance ratio, respectively. Each of the total RNA preparations was individually assessed for RNA quality based on the 28S/18S ratio measured on an Agilent 2100 Bioanalyzer system using the RNA 6000 Nano LabChip Kit. complementary DNA (cDNA) was synthesized using 2 μg of high-quality total RNA. SuperScript II (Invitrogen) was used to reverse transcribe the mRNA in the presence of a T7-containing oligo-dT24 primer followed by second-strand synthesis as recommended by Affymetrix (Santa Clara, CA) cDNA was purified using RNAClean magnetic beads (Agencourt Bioscience Corp., Beverly, MA) and used as a template for in vitro transcription using the GeneChip One-Cycle Labeling Kit from Affymetrix. Complmentary RNA (cRNA) was purified using magnetic beads from Agencourt Bioscience Corp., and the quantity and purity of the cRNA were determined by absorbance at 260 and 260 nm/280 nm absorbance ratio, respectively. The quality of the cRNA was evaluated by assessing the size distribution of the cRNA using a 1.25% 3-(N-morpholino)propanesulfonic acid gel.

The labeled cRNA was fragmented as recommended by Affymetrix and 10 μg was added to a hybridization cocktail prior to loading onto individual Mouse 430 2.0 GeneChip microarrays. The microarrays were hybridized at 45°C for 16–24 h and washed and stained on an Affymetrix FS450 fluidics station according to the manufacturer recommendations. The microarrays were scanned on a GeneChip Scanner 3000.

Array data quality was evaluated using a proprietary high-throughput application that assesses the data against multiple objective standards including 5′/3′ glyceraldehyde 3-phosphate dehydrogenase ratio, signal/noise ratio, and background as well as other additional metrics (e.g., outlier, vertical variance) and a visual inspection of the chip image for surface defects. All samples passed these data quality assessments.

RNA expression data were analyzed using the “affy” and “limma” packages of the Bioconductor suite of microarray analysis tools available for the R statistical environment (Gentleman, 2005; Gentleman et al., 2004; Irizarry et al., 2003; R Development Core Team, 2007). Robust microarray analysis background correction and quantile normalization were used to generate microarray expression values. Prior to analysis of changing genes, data were visualized using principal components analysis and hierarchical clustering to identify outlier samples. Three samples (one control each in 2-BE 7-day liver, 2-BE 4-h liver, and hypoxia BM) were judged to be sufficiently different from others in their treatment cohorts that they were not used for further analysis. An overall linear model was fit to the data for all sample groups, and specific contrasts of interest were evaluated to generate raw p values for each probe set on the expression array (Smyth, 2004). The Benjamini-Hochberg false discovery rate method was then used to correct for multiple testing effects. Probe sets were considered to have changed qualitatively in a specific comparison if an adjusted p value of 0.05 was obtained, their average expression intensity was above 150 in either treatment group, and they had an absolute fold change greater than 1.3. Genes represented by multiple probe sets were considered to have changed if at least one probe set was observed to change. Gene expression changes that met these criteria were called statistically significant RNA expression changes and have the directional qualities of “up” or “down” (i.e., they can be upregulated or downregulated in response to treatment).

The substrate for analysis of statistically significant RNA gene expression changes in response to a given treatment is the mouse knowledge assembly Model, which is derived from the global Genstruct Knowledgebase. This knowledge base is a collection of biological concepts and entities and their causal relationships that is derived from peer-reviewed scientific literature as well as public and proprietary databases. The Mouse Knowledge Assembly Model is the set of mouse-specific causal assertions that has been augmented with orthologous causal assertions derived from either rat or human sources and is competent for causal modeling. An example causal assertion would be increased transcriptional activity of nuclear factor kappa-light-chain-enhancer of activated B cells (NF-κB) complex causing an increase in the gene expression of the insulin receptor substrate 1 (Irs1) (Ruan et al., 2002). Each causal assertion has a specific scientific citation, and the assembled collection of these causal assertions is referred to as the Mouse Knowledge Assembly Model in this article.

In the case of RNA expression data, CNM interrogates the Mouse Knowledge Assembly Model to identify upstream controllers for the statistically significant gene expression changes observed in the experiment. These upstream controllers are called hypotheses as they are statistically significant potential explanations of the gene expression changes. Hypothesis generation is performed automatically by a computer program that utilizes the Mouse Knowledge Assembly Model to identify hypotheses that explain the input gene expression changes, prioritized by multiple statistical criteria. Each hypothesis is scored according to two probabilistic scoring metrics, richness and concordance, that examine distinct aspects of the probability of a hypothetical cause explaining a given number of statistically significant gene expression changes. Richness is the probability that the number of observed gene expression changes connected to a given hypothesis could have occurred by chance alone. Concordance is the probability that the number of observed gene expression changes that match the directionality of the hypothesis (e.g., increased or decreased kinase activity for a kinase, increased or decreased transcriptional activity for a transcription factor, etc.) could have occurred by chance alone. For each hypothesis “Score” is the number of gene expression changes that support the hypothesis in its statistically significant direction, and “Contra” is the number of gene expression changes that support an opposite direction for the hypothesis. A hypothesis was considered to be statistically (although not necessarily biologically) significant if it met richness and concordance probability cutoffs of 0.05 and marginally significant if it met richness and concordance probability cutoffs of 0.1. These hypotheses are further investigated and prioritized by evaluation of their biological relevance to the experimental context, whether they are causally linked to phenotypes and processes relevant to hemangiosarcoma formation in the literature and if they are causally downstream of the experimental treatments. The models for each organ in response to 2-BE were merged to create the overall mechanistic model for hemangiosarcoma formation.

To build a model of 2-BE–induced hemangiosarcoma, CNM was used to identify statistically significant explanations or hypotheses for the observed gene expression changes induced by 4 h and 7 days 2-BE treatment in the BM, spleen, and liver of B6C3F1 mice (Table 2); 7-day treatment was used to identify molecular mechanisms that can lead to hemangiosarcoma formation, while the 4-h treatment served to identify potential initiating events for these mechanisms. The model was then compared to the transcriptomic fingerprint generated in mice treated directly with hypoxia (2 h, 6–8% O₂). Table 2 lists the number of gene expression changes and hypotheses that explain these changes for all treatments and tissue combinations.

Analysis of mice treated with 2-BE for 4 h identified increased transcriptional responses to inflammation and oxidative stress in spleen and liver (Supplementary Figs. 1A–D). At the 7-day time point, CNM also predicted an increase in inflammation in the liver (Fig. 1), which, like the results at 4 h, is consistent with publications from Klaunig and colleagues (Corthals et al., 2006; Klaunig and Kamendulis, 2005). An increase in inflammation was supported by increased catalytic activity of Tlr4, evidenced by eight gene expression changes (Supplementary Table 2). Increases in the levels of the proinflammatory cytokines Il1b and Il6 in the liver were also supported by eight and six gene expression changes, respectively (Fig. 1). Increases in cytokines are consistent with macrophage activation. As illustrated in Figure 1, Tlr4 promotes inflammation by inducing the transcriptional activity of NF-κB, which leads to proinflammatory cytokine production (O'Neill and Bowie, 2007) that in turn activates AP-1 signaling, which has been shown to stimulate cell proliferation (Ballermann et al., 1998).

In addition to confirming previously reported increases in inflammation and oxidative stress, CNM also identified hypoxia as a molecular response in spleen and liver in mice treated with 2-BE for 4 h and in spleen and BM in mice treated with 2-BE for 7 days. This response was then compared to the transcriptomic fingerprint generated in mice treated directly with reduced oxygen as a positive control for hypoxia. In addition, Hypoxyprobe immunohistochemistry was used to quantitate hypoxia in the same organs following these treatments. The sections that follow describe the support for hypoxia as an initiating event. See Supplementary Materials for a more detailed overview of the CNM and IHC results following treatment with hypoxia. Briefly, CNM provided evidence for oxidative stress and a transcriptional response to hypoxia in liver, spleen, and BM of mice following exposure to 6–8% O₂ for 1.5–2.5 h.

In addition to the confirmation of previous findings supporting inflammation and macrophage activation, and the complementary evidence of hypoxia, 2-BE was also found to induce several other processes not previously identified but relevant for events that contribute to hemangiosarcomas.

2-BE–induced hypoxia in the liver, spleen, and BM was demonstrated by transcriptomics, CNM, and phenotypically via the use of Hypoxyprobe. We propose that hypoxia may mediate previously identified mechanisms of 2-BE–induced hemangiosarcoma. For example, hypoxia can serve as an initiating event to iron-induced oxidative stress previously observed in response to 2-BE treatment (Klaunig and Kamendulis, 2005; Nyska et al., 2004). Local tissue hypoxia generates an acidic environment that stimulates iron release from hemosiderin leading to the production of hydroxyl radicals in vitro (Ozaki et al., 1988), and without an acidic environment, iron is not as efficiently released from hemosiderin (Ozaki et al., 1988). Furthermore, Hif1a, the hypoxia-responsive transcription factor, has been shown to induce ROS production, likely via the Fenton reaction (Hervouet et al., 2008), which has also been implicated in the oxidative stress effects of 2-BE (Klaunig and Kamendulis, 2005). Therefore, it is possible that hypoxia may function upstream of and contribute to iron-induced oxidative stress in response to 2-BE.

In addition to its role in oxidative stress, hypoxia and activation of Hif1a have been shown to induce processes that can increase the supply of oxygen. This is achieved through the transcription of genes that function in processes such as angiogenesis and EC proliferation (Otrock et al., 2009; Yamakawa et al., 2003). Among these genes is Vegf, a well-known growth factor that has been shown to induce angiogenesis (Forsythe et al., 1996; Levy et al., 1995). Since hemangiosarcomas are EC-derived tumors (Mendenhall et al., 2006), the discovery of high expression levels of Hif1a, Hif2a, as well as the Hif target genes, Vegf and Epo, in a renal angiosarcoma (Rathmell et al., 2004) is consistent with the role of these factors in EC proliferation.

As described above, the proposed mechanistic model for induction of hemangiosarcoma identifies continuous Epo signaling in the BM as a key event leading to the differentiation and proliferation of EPCs and ECs in hematopoietic organs (BM and spleen). The RBC hemolysis induced by 2-BE likely serves as the initial stimulus for Epo signaling (Leng et al., 1999). The induction of Epo serves to ameliorate the anemia resulting from RBC hemolysis, by signaling for production of erythrocytes. Interestingly, the present study found transcriptomic data consistent with a decrease in Bnip3l (also called Nix) protein levels in the BM of 2-BE–treated mice compared to controls. Bnip3l is a Bcl2-interacting protein family member (Schweers et al., 2007) that, when knocked down, has been shown to hinder the process of mitochondria removal and apoptosis in erythrocytes during their maturation (Diwan et al., 2007; Schweers et al., 2007; Sandoval et al., 2008). Loss of Bnip3l has also been characterized using Nix−/− mice, which demonstrate anemia, reduced erythrocyte lifespan in vivo, reticulocytosis, and splenomegaly (Diwan et al., 2007). Decreased Bnip3l protein levels in response to 2-BE, therefore, results in an increase in mitochondria-containing erythrocytes, which are targeted for removal by the spleen. Increased elimination of mature erythrocytes results in a sustained decrease in the number of mature erythrocytes despite robust reticulocytosis. This decrease in mature erythrocyes, in turn, leads to continuous Epo signaling in an attempt to replenish erythrocyte levels to normal values (Noguchi et al., 2008; Weidemann and Johnson, 2009). Reduced apoptosis of spleen and BM erythrocyte populations results in the aberrant survival of committed erythroblasts contributing to erythroblastosis (Diwan et al., 2007). Additionally, Bnip3l has been shown to be regulated by hypoxia, and deregulation has been associated with tumor growth (Zhang and Ney, 2009). Experimentally, downregulation of Bnip3l increased the growth of osteosarcoma cells in a tumor transplant model, suggesting that Bnip3l is needed to inhibit tumor growth under hypoxic conditions.

Epo has been shown in previous studies to stimulate EC and EPC proliferation (Bahlmann et al., 2003). EPCs derived from hematopoietic organs that circulate in the blood have the ability to differentiate into ECs (Shi et al., 1998). Sustained Epo signaling due to decreased Bnip3l levels and sustained effects on erythropoiesis would result therefore in continuous proliferation of ECs and EPCs, contributing to hemangiosarcoma formation. Supporting the notion of EPCs as central to hemangiosarcomas, EPC-like hematopoietic precursors have been shown to be the main source of ECs in canine hemangiosarcomas (Lamerato-Kozicki et al., 2006). Moreover, preliminary results have demonstrated that after 7-day treatment with 2-BE, there are increases in proliferation of ECs/EPCs in both spleen and BM (Criswell, K. A., Sokolowski, S., and Aubrecht, J.), consistent with the proposed mechanistic model.

The liver is the main site of hemangiosarcoma formation in mice in response to 2-BE. Therefore, processes identified in the liver define mechanisms that may preferentially lead to hemangiosarcoma formation in this organ. CNM of the liver gene expression data revealed activation of inflammatory processes after acute and chronic 2-BE treatment (Fig. 1). This response is consistent with previous 2-BE studies where activation of Kupffer cells and IL-6 induction were found (Corthals et al., 2006; NTP, 2000). These inflammatory processes can contribute to local hemangiosarcoma formation by recruiting circulating EPCs/ECs to the site of inflammation (Fan et al., 2008). Therefore, while EPC/ECs proliferate and acquire mutations secondary to increased cell proliferation in the BM and spleen, hemangiosarcomas are formed at the sites of recruitment of these EPC/ECs. Indeed, EPC/EC recruiting signals are commonly associated with sites of hemangiosarcoma formation. This has been observed for compounds such as p-nitroaniline and Elmiron, which induce hemangiosarcomas in the liver where iron accumulation and Kupffer cell activation is observed (Abdo et al., 2003; NTP, 1993, 2004). For other compound classes that induce mouse-specific hemangiosarcoma, such as peroxisome proliferator-activated receptor gamma agonists, pro-angiogenic growth factors have been observed (Sotiropoulos et al., 2006), and these may serve as EPC/EC recruiting signals (Chen et al., 2008). Taken together, these studies indicate a key role for EPC/EC recruiting signals in hemangiosarcoma formation, something that was also suggested in a recent review on hemangiosarcoma in rodents (Cohen et al., 2009).

Hemangiosarcomas are rare in humans but are more commonly found in mice. Of the processes identified in this study, increased hypoxia exhibits species differences that may help explain the different rates of hemangiosarcoma occurrence between mice, rats, and humans. Because 2-BE causes hemolysis only in mice and rats, humans should not be subject to the subsequent hypoxia (Udden, 2002); therefore, the stimulus for sustained Epo signaling that results from 2-BE treatment does not occur in humans. Mice may also be more vulnerable to oxidative stress stemming from hypoxic conditions since they have been shown to have lower hepatic levels of the antioxidant, vitamin E, compared to rats and humans (Siesky et al., 2002). Thus, hypoxia as an initiating event to 2-BE–induced hemangiosarcoma is consistent with previously identified species-specific differences that predispose mice to hemangiosarcoma formation. Additionally, there are no reports of increased hemangiosarcomas in humans with chronic hemolytic disease or in people living at high altitudes, so hypoxia alone is likely not an operational component of hemangiosarcoma formation in humans. Interestingly, Mori-Chavez et al. (1970) report that exposure of mice to high altitude does increase the incidence of spontaneous ovarian angiomas.

In summary, this is the first study to use transcriptomic data linked to pathways analysis to develop a mechanistic, temporal, and multiorgan model of hemangiosarcoma. The resulting model confirms that 2-BE induces macrophage activation and inflammation in the liver as previously proposed (Corthals et al., 2006; Klaunig and Kamendulis, 2005). In addition, the model specifically supports local tissue hypoxia, increased Epo signaling and erythropoiesis in the spleen and BM, and suppression of mechanisms that contribute to genomic stability as contributing factors to hemangiosarcoma in mice (Fig. 6). The key processes emerging from this model are sustained Epo signaling in the BM leading to proliferation of EPCs/ECs and recruitment signals for EPCs/ECs in the liver, the main site of hemangiosarcoma formation. Finally, these mechanisms, and in particular hypoxia as an initiating event, are consistent with a generalized mode of action framework that was developed with multiple other mouse-specific hemangiosarcoma-causing agents and that was published in a recent review (Cohen et al., 2009).