Understanding and predicting the potential hazardous effects of chemicals is a major goal in current efforts to protect the health of human and ecological receptors. As a result, numerous strategies have been and continue to be developed to measure and assess chemical hazards. Recent efforts have focused on moving away from traditional animal testing towards the use of in vitro assays, biological pathways, and mechanisms of action with adverse outcomes to assess chemical hazards (National Research Council of the National Academies, 2007; Cote et al., 2016). The development and application of inexpensive high throughput and high content assays providing broad coverage of biological effects has also been identified as a critical need in assessing the potential impacts of chemicals (National Academies 2017). The Adverse Outcome Pathway (AOP) framework is one approach that is gaining acceptance in linking biological pathways, represented by in vitro assays and other nontraditional toxicological data, to adverse outcomes (Ankley et al., 2010). The AOP framework provides a useful construct in which to document biological pathways and develop predictive models for chemical hazards (Perkins et al., 2019).

The AOP framework is attractive as it creates linear biological pathways composed of essential components, or Key Events (KE), leading to an Adverse Outcome. As a result, AOPs are simplifications of complex biological pathways where KE are measurable biological events. Conversely, systems biology uses detailed descriptions of biological networks (Edwards and Preston 2008; Cote et al., 2016). Systems biology provides the ideal context in which to interpret omics technologies, but its complexity is problematic in a decision-making context (van Ravenzwaay et al., 2017). While the AOP framework and systems biology are complementary, they are generally not combined, limiting the application of ‘omics in a decision-making context.

Since KE are often the result of interactions of several genes, proteins, and/or metabolites, KE provide an opportunity to map the biological networks of systems biology to AOPs in a relatively discrete manner using causal biological networks. AOPs themselves are considered to be linear causal networks where the occurrence of one KE causes the next downstream KE to occur and ultimately cause an Adverse Outcome to occur e.g. a Molecular Initiating Event causes KE1 which causes KE2 which causes an Adverse Outcome A causal subnetwork can also be developed that represents the cascading biological events or nodes required to cause a KE. In other words, the subnetwork represents a network that is contained underneath, or within, a KE. Using a KE causal subnetwork with causal network theory allows one to use ‘omics data to infer the biological state or activity of the overall KE network.

Causal networks can be used to understand the state of individual KE using ideas of conditional dependence/independence from causal network theory (Burgoon, et al., 2017). In causal network theory, specifically the Markov property, the state of a node is only dependent upon the node(s) immediately preceding it, e.g. a predecessor or parent with the causal effect either positive or negative. Building on that is the idea of d-separation, which states that a node is conditionally independent of another node if it is blocked by a third node. As a result, the adverse outcome above would be conditionally independent of the state of KE1 since it is blocked by KE2. Therefore, the adverse outcome will not occur if only KE1 occurs, KE2 must also occur. Using this relationship, if KE2 occurs, then the adverse outcome will occur and can be sufficient to infer the occurrence of an outcome in AOPs. If this can be applied to causal networks representing the molecular events essential for a KE to occur, then this concept may be useful in applying omics to assess if individual KE are occurring in toxicological studies. When combined with dose response analysis, as proposed by Burgoon et al. (2017) causal networks may be useful in a regulatory context to understand at what dose a KE might occur. When applied to multiple KE, this approach might be useful in assessing entire AOPs.

Here, we explore how transcriptomics and systems biology can be merged with AOPs to understand chemicals effect liver biology through the use of causal subnetworks that depict KE in AOPs. We examined the hypothesis that if chemicals are known to cause an adverse effect, regenerative proliferation, and that adverse effect is caused by upstream events (gene/protein activity), then activation of upstream events adjacent to the adverse effect should occur when animals are exposed to chemicals that cause the adverse effect. To do this, we first developed a causal subnetwork for the KE for regenerative proliferation in human/rodent liver using literature-based experimental evidence. We then examined the activation of nodes (genes) in the network in relation to exposure to chemical known to cause regenerative proliferation in rats using five case study chemicals including three carcinogens (carbon tetrachloride, aflatoxin B₁, and thioacetamide) and two non-carcinogens (diazepam and simvastatin) and data from the toxicogenomics database Open TG-GATEs (Igarashi et al., 2015; https://toxico.nibiohn.go.jp/english/). Our results suggest that the combination of ‘omics and causal subnetworks could be a useful approach to examine the activation of individual KE, and eventually collections of KEs, to better understand the hazards of chemical exposure using the AOP framework.

Here we propose the development of a simple systems biology, causal subnetwork to represent critical events within a KE in order to define and understand the state of a KE using omics and systems biology approaches. It should be noted that we do not use the causal network to predict if a chemical causes liver injury/regenerative cell generation. We use the network to measure the state of the tissue after exposure. By developing the causal network, we were able to identify essential events leading from tissue damage to dysregulation of cell division and genes/proteins whose activity is sufficient to indicate activation of the KE Regenerative proliferation.

We have intentionally focused only on events within an individual KE rather interactions between KE to maintain the hierarchy where an AOP organizes KE and KE organize subevents/subnetworks. To illustrate how such a subnetwork could be developed and applied, we developed a KE subnetwork that describes the development of regenerative cell proliferation due to repeated liver injuries from several mechanisms including oncotic necrosis, chronic inflammation and oxidative stress caused by electrophilic metabolites from cytochrome p450 metabolism of xenobiotics (Smith et al., 2016). The subnetwork represents a common KE in cancer AOPs such as AOP220 Cyp2E1 activation leading to liver cancer (https://aopwiki.org/aops/220; Figure 1). The regenerative proliferation subnetwork (Figure 1) is composed of 28 events (genes, proteins or processes). In this simplified network model, events are considered either as on/active or off/inactive, connections are causal (activate/promote, inactivate/inhibit), dose-dependence and time dependence are inherent characteristics of the causal linkage. However, dose-response features could be incorporated into an event using point of departure analysis of gene expression or assay activity as described in Burgoon et al. (2017).

Several assumptions were made in applying the subnetwork. Parent events have been shown to cause daughter event, therefore if a parent event occurs the daughter event will also occur. Since the effect of time and dose dependence is not explicitly accounted for in network interactions but the linkage is causal, activated/inactivated events were assumed to cause their nearest downstream event given sufficient time and stimulus. To demonstrate how omics, or transcriptomics, can be used to monitor activity in the subnetwork, we used gene expression as a surrogate to assess if a protein or gene changed levels/activity, realizing that this does not necessarily reflect protein activity temporally or in abundance and may not capture certain protein interactions represented in the subnetwork.

The subnetwork was used to identify events (genes or proteins) whose activity could be used to infer the occurrence of the outcome (the KE regenerative proliferation) using causal network theory. Five proteins were identified as sufficient to infer regenerative proliferation events based on their proximity to cell proliferation: Cyclin D1 (CCND1), RB transcriptional corepressor 1 (pRb), proteasome 26S subunit, non-ATPase 10 (PSMD10) and Cyclin-dependent kinase 4 and 6 (CDK4/6). These were confirmed in the scientific literature as sufficient to infer cell proliferation as removal of action (activation or inhibition) of these proteins prevents the occurrence of cell proliferation (Rivadeneira et al., 2010; Jing et al., 2014; Zhu et al., 2015; Marampon et al., 2016; Ding et al., 2020; Chicas et al., 2010; Doan et al., 2021). Activity of these genes/proteins were used to infer whether or not regenerative cell proliferation was occurring in liver tissue.

We examined the behavior of the subnetwork in conditions that cause liver injury and promote regenerative cell generation by examining livers of rats exposed to chemicals with well characterized effects on liver injury and regenerative proliferation (carbon tetrachloride, aflatoxin B₁, thioacetamide) and chemicals with no known effect on liver injury and regenerative proliferation (diazepam and simvastatin). The behavior of genes in the network were consistent with liver pathology data for tissue level effects related to hepatocellular carcinoma development and known effects of the chemicals on liver. Genes sufficient to infer cell proliferation were affected only where animals were exposed to chemicals known to cause liver injury and cell proliferation. Interestingly, gene expression sufficient to infer that cell proliferation was occurring was detectable much earlier than effects on the tissue levels. This is expected given that changes at the gene expression levels frequently occur before changes at the tissue level.

These analyses demonstrate that the combination of transcriptomics and causal subnetworks could be used to describe or infer the sate of a KE and provides a promising approach to merge omics, systems biology and AOPs. Measuring sufficient sets of genes in causal networks for Key Events in AOPs or AOP networks can provide more efficient and informative measurements related the toxicity or Adverse Outcome of a chemical. Applying causal theory reduces the uncertainty associated with measuring an event distal from the adverse outcome and interfering communication from other pathways. If combined with targeted sequencing, transcriptomics-based point of departure and high throughput assays this approach could greatly facilitate our understanding of adverse effects caused by chemicals. Future research into the utility and applicability of this approach should test to see if evidence from in vitro liver cell systems (and the causal network presented here) are accurate predictors of damage/injury in whole organs/animals. Additional research is needed to determine if combining the activity of sufficient genes in reference chemical exposures with classification modeling be useful in identifying chemicals that activate this KE, if these causal subnetworks are sensitive and/or specific enough to be used in screening for chemical effects and whether or not the subnetworks serve to organize and explain existing data rather than make prospective predictions of in vivo effects using in vitro data.

Table 1. Liver pathology data for carbon tetrachloride exposures. Low = 30 mg/kg, medium = 100 mg/kg and high = 300 mg/kg. Data for individual rats are designated by glass image numbers and highlighted by yellow or white rows.Table 2. Liver pathology data for aflatoxin B₁ exposures. Low = 1 mg/kg, medium = 3 mg/kg and high = 10 mg/kg. Data for individual rats are designated by glass image numbers and highlighted by yellow or white rows.Table 3. Liver pathology data for thioacetamide exposures. Low = 4.5 mg/kg, medium = 15 mg/kg and high = 45 mg/kg. Data for individual rats are designated by glass image numbers and highlighted by yellow or white rows.Table 4. Liver pathology data for diazapam exposures. Low = 25 mg/kg, medium = 75 mg/kg and high = 250 mg/kg. Data for individual rats are designated by glass image numbers and highlighted by yellow or white rows.Table 5. Liver pathology data for simvastatin exposures. Low = 40 mg/kg, medium = 120 mg/kg and high = 400 mg/kg. Data for individual rats are designated by glass image numbers and highlighted by yellow or white rows.