Abstract

   The last decade has seen the adverse outcome pathways (AOP) framework
   become one of the most powerful tools in chemical risk assessment, but
   the development of new AOPs remains a slow and manually intensive
   process. Here, we present a faster approach for AOP generation, based
   on manually curated causal toxicological networks. As a case study, we
   took a recently published zebrafish developmental neurotoxicity
   network, which contains causally connected molecular events leading to
   neuropathologies, and developed two new adverse outcome pathways:
   Inhibition of Fyna (Src family tyrosine kinase A) leading to increased
   mortality via decreased eye size (AOP 399 on AOP-Wiki) and GSK3beta
   (Glycogen synthase kinase 3 beta) inactivation leading to increased
   mortality via defects in developing inner ear (AOP 410). The approach
   consists of an automatic separation of the toxicological network into
   candidate AOPs, filtering the AOPs according to available evidence and
   length as well as manual development of new AOPs and weight-of-evidence
   evaluation. The semiautomatic approach described here provides a new
   opportunity for fast and straightforward AOP development based on large
   network resources.

   Keywords: systems toxicology, adverse outcome pathway, causal network,
   toxicological network, neurotoxicity

Introduction

   A decade ago, adverse outcome pathways (AOPs) ([34]Ankley et al., 2010)
   have been put forward as a tool for organizing toxicological knowledge
   across different levels of biological organization, from the initial
   interaction of chemicals with the biological system (MIE = molecular
   initiating event) ([35]Allen et al., 2014) to the individual and
   population level effects relevant for environmental risk assessment (AO
   = adverse outcome). The main idea of AOPs is collecting basic knowledge
   about biological systems and their chemical perturbations, and
   organizing it in easy to understand sequences of causally connected
   biological events (KE = key events; KER = key event relationship, i.e.,
   how one KE is connected to another) ([36]Villeneuve et al., 2014b)
   which allows risk assessors and toxicologists to identify chemicals
   likely to cause environmental harm. Additionally, through
   identification of knowledge gaps, AOPs inform future research and the
   development of novel biological assays that allow more specific in
   vitro chemical testing and reduction of animal testing ([37]Groh et
   al., 2015). The long-term goal of the AOP framework is to develop AOPs
   that cover the whole space of chemically-induced biological
   perturbations and a complete set of assays required for comprehensive
   chemical risk assessment ([38]Zupanic et al., 2018).

   From their conception, several tools have been developed for easier
   development and management of AOPs. The AOP-wiki and the AOP knowledge
   base ([39]https://aopwiki.org/; AOP-KB:
   [40]https://aopkb.oecd.org/index.html) form an online portal hosted by
   the OECD that serves as the central AOP hub ([41]Groh et al., 2015;
   [42]Fay et al., 2017). The accepted core principles ([43]Villeneuve et
   al., 2014a) and a handbook for AOP development ([44]OECD iLibrary,
   2021) serve as a standard that enables the development of high-quality
   and structurally similar AOPs, with comparable weight-of-evidence (WoE)
   evaluations. On top, several helper tools for AOP development and
   visualization have been made available to the community (e.g.,
   [45]http://datasciburgoon.github.io/aopxplorer). The AOPs in the
   AOP-wiki have been useful resources for quite diverse toxicological
   studies, e.g., to find chemicals likely to activate the AOPs ([46]Jeong
   et al., 2019), to evaluate the hazard associated with specific
   chemicals and chemical groups ([47]Carvaillo et al., 2019; [48]Negi et
   al., 2021), to develop assays for in vitro assessment of mixture
   toxicity ([49]Pistollato et al., 2020), to develop a new tiered testing
   approach for thyroid hormone disruptors ([50]Knapen et al., 2020), to
   find the mechanisms of nanomaterial toxicity ([51]Murugadoss et al.,
   2021) and to develop quantitative AOPs, mathematical models that can be
   used directly in chemical risk assessment ([52]Margiotta-Casaluci et
   al., 2016; [53]Doering et al., 2018; [54]Perkins et al., 2019;
   [55]Burgoon et al., 2020; [56]Lillicrap et al., 2020).

   As of August 2021, more than 400 AOPs and 5017 KEs have been developed
   [a large increase from 219 AOPs of April 2018 ([57]Pollesch et al.,
   2019)], covering a range of species from nematodes to humans. However,
   these still cover only a very small part of the biological
   perturbations caused by chemical exposure and also only a few taxonomic
   groups. AOP development remains relatively slow, because each AOP
   requires searching for scientific literature, its manual curation, the
   formatting of the acquired knowledge into a user-friendly AOP and
   performing a WoE assessment ([58]Vinken, 2013). To come closer to the
   final goal of a complete AOP space, the development of new AOPs needs
   to be accelerated.

   There have been some attempts to do this already. The United States
   Environmental Protection Agency (EPA) has recently developed the
   Adverse Outcome Pathway database (AOP-DB), which can help with the
   annotation of AOP pathways under development, by connecting the
   information present on the AOP-wiki with various public resources
   [e.g., the NCBI gene, STRING ([59]Szklarczyk et al., 2021) and
   Comparative Toxicogenomics Database (CTD) ([60]Pittman et al., 2018;
   [61]Davis et al., 2020)]. Other studies have tried to integrate
   publicly available resources [e.g., (ToxCast ([62]Dix et al., 2007),
   CTD, Reactome ([63]Jassal et al., 2020)] to develop AOP-like networks,
   which can serve as starting point for computationally predicted AOPs
   (cpAOPs). Several studies have used association rule mining to generate
   computationally predicted AOPs, mostly at the molecular level ([64]Bell
   et al., 2016; [65]Oki et al., 2016). Doktorova et al. have further
   developed a filtering approach for refining the molecular AOP-like
   networks using gene expression data from TG-Gates database
   ([66]Igarashi et al., 2015) and manual curation to arrive at a putative
   AOP-network ending in the AO non-genotoxic induced hepatocellular
   carcinoma ([67]Doktorova et al., 2020). The AOP-helpFinder tool uses
   text mining to find potential connections between key events
   ([68]Jornod et al., 2021), while the computational pipeline developed
   by Jin et al. uses chemical-specific toxicogenomic data, pathway
   enrichment analysis and biomarker selection to develop putative cpAOPs
   ([69]Jin et al., 2021).

   Here we present a new approach for development of AOPs, based on a thus
   far AOP-untapped toxicological resource—causal toxicological networks
   (CTN) ([70]Boué et al., 2015). CTNs are computational networks, which
   describe causally connected, mostly molecular, cellular and tissue
   events. We hypothesize that CTNs are especially appropriate as a
   starting resource for AOP development, as they are highly curated and
   all the connections in such networks are annotated by evidence.
   Compared to de novo AOP development, starting from such a network
   resource should decrease the time needed to search and curate
   scientific literature, but, since the CTNs are made from connected
   nodes and relationships between them, also make it easier to format the
   AOP into KE and KERs. The described approach is semiautomatic. It
   allows for automatic generation of a large number of candidate AOPs
   from a CTN, while the WoE evaluation and formatting of the evidence
   remains a manual process, as described by ([71]Becker et al., 2015).

   As a case study, we developed two developmental neurotoxicity AOPs. The
   choice of developmental neurotoxicity was made, because AOPs related to
   developmental neurotoxicity are still underrepresented in the AOP-wiki
   ([72]Bal-Price et al., 2015; [73]Knapen et al., 2015; [74]Pistollato et
   al., 2020) and because of the recent activity in incorporation of
   AOP-based developmental neurotoxicity in vitro assays into chemical
   risk assessment ([75]Sachana et al., 2021). This also allowed us to use
   our recently developed zebrafish causal developmental neurotoxicity
   network as starting point ([76]Li et al., 2021). The new AOPs can be
   found on the AOP-wiki under Ids 399 and 410
   ([77]https://aopwiki.org/aops/399 and
   [78]https://aopwiki.org/aops/410). Here, we present the details of the
   taken approach, the analysis of the differences between the CTNs and
   AOPs and the consequences for using the former as a source for the
   latter. We also provide tools and guidance for the development of AOPs
   from large network resources for the toxicological community.

Methods

From the Causal Toxicological Network to AOP Candidates

   The zebrafish causal developmental neurotoxicity network, which is the
   basis of our AOP development, has been developed in an earlier
   publication, therefore we here only summarize the most relevant
   properties of the network for this study, and refer the reader to
   original publication for details ([79]Li et al., 2021). CTN network
   development usually starts from a known adverse outcome of interest.
   For the network used as starting point for AOP development here, these
   were known zebrafish developmental nervous system-related pathologies
   (megalencephaly, microcephaly, microphthalmos, seizures, neurogenic
   inflammation, hydrocephalus). Then evidence for events leading to the
   pathologies are found in the scientific literature, specifically the
   evidence needs to be present in a peer reviewed publication. The
   scientific literature is queried using search engines such as Google
   Scholar, using keywords representing the specific molecular events. The
   evidence needs to show causality, i.e., a performed upstream
   perturbation leading to a measured downstream perturbation. With
   literature curation, upstream nodes directly affecting the pathology
   are added, then further upstream nodes directly connected to these
   nodes and so on, until no more evidence (in the form of causal
   experimentally validated relationships) is found in the literature.
   From the time of the first publication of this causal developmental
   neurotoxicity network, the network has been further extended and the
   version used as starting point in this paper (NTOX_BEL;
   [80]Supplementary Material) features 515 nodes (ranging from protein
   activations and gene ontology terms to pathologies), with 682 edges
   between them. It was built based on evidence gathered from 90 zebrafish
   specific scientific articles and each edge in the network is annotated
   by the evidence behind it in the form of a BEL statement ([81]Slater,
   2014) and the reference from which the evidence is taken.

   The original toxicological network (NTOX_BEL, [82]Supplementary
   Material) contained nodes (molecular events) connected by edges (causal
   relationships between the nodes) and was written in Biological
   Expression Language (BEL) formalism. From the AOP development point of
   view it contained several non-necessary nodes (e.g., mRNA expression or
   protein activation of the same gene). To make it easier to manipulate
   with standard network analysis tools, we first generated a simplified
   abstracted network (NTOX_ABSTRACTED, [83]Supplementary Material), i.e.,
   the BEL formalism was simplified by removing unnecessary details from
   the BEL node and edge definitions ([84]Figures 1A,[85]1B;
   [86]Supplementary Table S1). This abstracted network was then imported
   into Cytoscape [v3.8.0 ([87]Shannon et al., 2003)] and further reduced
   by removing all self-loops. Finally, all nodes not connected to
   pathologies were removed by first selecting all nodes that had no
   outgoing connections, and then deleting anything that was not a
   pathology ([88]Figures 1B,C). This step had to be repeated several
   times, until all the nodes in the network were connected to a
   pathology.

FIGURE 1.

   [89]FIGURE 1
   [90]Open in a new tab

   Schematic diagram of the network preparation procedure. First, the BEL
   network (A) is converted into its abstract form (B). This
   simplification results in the formation of self-loops (protein-D),
   which are removed together with all nodes that do not lead to a
   pathology node (C).

   In addition to original annotations that the abstracted network
   inherited from the BEL formalism (e.g., references, experimental