Elucidating the mechanism of toxicity is crucial in drug safety evaluations. We focus on toxic processes and developed a toxic process ontology, designated TXPO. Here, we outline the TXPO, which systematizes toxic processes within the liver in a consistent manner. The TXPO makes processes explicit across granularity using a functional decomposition tree. Concerning the course of toxic processes, we present a framework of causal relationships between processes from latent to toxicity manifestation. In applied work, we introduce a prototype of TOXPILOT, a toxic process interpretable knowledge system. TOXPILOT provides visualization maps of the toxic course, which facilitates capturing the comprehensive picture for understanding toxicity mechanisms. Our ontological approach will help develop new knowledge regarding drug safety evaluations.
ontology; process; drug-induced liver toxicity
Drug-induced liver injury is a major cause of drug withdrawal
from the market and discontinuation of drug development [
From textbooks [
Table 1 PubMed search terms for hepatotoxic course We used the ontology editing tool Protégé 5.2.0 [7] to develop the TXPO in the Web Ontology Language (OWL) and HermiT reasoner [8] as a Protégé Plug-in.
Figure 1 shows examples of the TXPO development process. First, 1) each toxic course was defined, and related information was annotated using the Annotation Properties. Next, 2) the processes constituting each toxic course were described using a 'has part' relation as Object Property. Then, 3) each process was generalized using an is-a hierarchy: processes common to multiple toxic courses, biological processes, and biomedical-independent processes. Furthermore, 4) each process was decomposed into subprocesses (has part relation), and 5) the biological structure in which the process takes place was described (occurs in). In addition, 6) molecules/drugs and their roles in the process were also defined. Finally, 7) causal relationships between process were defined by using a 'has result' relation. (6) (7) (1) (5) In generalizing the is-a tree construction, we reused existing ontologies. Domain- independent general entities were based on BFO [9], and biomedical entities were imported manually from existing ontologies in NCBO BioPortal [10]. These biomedical ontologies include UBERON [11], Cell Ontology [12], NCBI Taxon [13], ChEBI [14], Gene Ontology [15], PATO [16], INOH [17], and Ontology of Genes and Genomes (OGG) [18].
TOXPILOT consists of an ontology library, a Resource Description Framework (RDF) database, and a Web application (2) (3) system. The TXPO file is stored in the ontology library, and the file is converted to RDF format represented by a triple Subject, Predicate, and Object by Protégé. The RDF data are then stored in an RDF triple store using Apache Fuseki [19] to construct the SPARQL endpoint. Regarding the web application system for TOXPILOT, necessary information is dynamically acquired via SPARQL queries. Moreover, TOXPILOT generates graphs using D3.js [20] of the JavaScript library. (4)
Figure 2 shows an overview of the TXPO, which is a three-layer model organized in an is-a hierarchy of general terms to specialized toxicologic terms. The top layer is domain-independent (domain-neutral) and provides general terms. Most of the entities in the top layer refer to the upper ontology, Basic Formal Ontology (BFO). Upper ontologies support generic categories and relations based on a philosophical orientation. Accordingly, we could construct our ontology with inheritance of the intrinsic nature in a consistent manner. All entities of the TXPO are classified into the basic categories continuant or occurrent. Continuant refers to an entity that persists, endures, or continues to exist through time while maintaining its identity and includes objects, roles, and qualities.
An object is an independent continuant, such as a thing. Roles and qualities are dependent continuants that can only exist depending on something else. Occurrent includes entities that unfold over time, such as processes.
The intermediate layer is biomedical domain dependent and Figure 2. Overview of the ToXic Process Ontology (TXPO). The TXPO contains an is-a hierarchy that is organized into three layers: the top layer contains general terms, mostly derived from the Basic Formal Ontology. The intermediate layer contains biomedical terms, and the lower layer contains more granular, toxicology terms. consists of entities commonly used in biomedicine. As lower entities of continuant, biological structures such as molecules, compounds, organelles, cells, organs, and species are defined.
The open community OBO Foundry [21] seeks to share knowledge and standardize terms among the biological community, and OBO ontologies utilize BFO as an upper ontology.
Accordingly, the TXPO imports existing terms and reuses them from biomedical ontologies of OBO foundries. These terms include anatomic structures from UBERON, cells from Cell Ontology, organisms from NCBI Taxon, compounds from ChEBI, biological processes and cellular components from Gene Ontology, qualities from PATO, some molecule families from INOH, and genes from OGG.
The lower layer encompasses entities specific to toxicology (i.e., entities that are toxicological domain dependent).
Process is a central category in the TXPO. In order to elucidate
a toxicity mechanism adequately, we provide two
subcategories, namely, primitive process and process sequence. The
former is defined as a single unit of process, whereas the latter is
defined as a series of processes, which includes pathways and
toxic courses.
(1) Functioning Process
Many biological defense processes function to protect
organisms from toxicity-associated injury. Therefore, we focused on
functioning processes in the present study. Functioning
processes in organisms are diversified in granularity from the molecular
level to the organelle, cell, tissue, and organ level. In order to
define functioning processes in a consistent fashion, we
systematize the functioning tree based on functional ontology [
The intermediate layer is biomedical domain dependent. For instance, a lower level of transmitting includes biological (a) Examples of functions in functional ontology (c) An example of functioning decomposition of protein quality control during ER stress Figure 3 functioning process (b) The TXPO functioning is-a hierarchy transport processes such as nuclear transport and Golgi vesicle transport. Decomposing includes proteolysis and lipid degradation; splitting includes cell division; and detaching includes complex dissociation. These processes are generally consistent with the GO Biological Process, as some of the GO biological processes can be interpreted as functional processes common to biomedicine.
The lower layer is a toxicology domain dependent. In this study, we define a "toxic process" as a process that constitutes a specific toxic course. For example, by specializing the biomedical process "apoptotic process (GO:0006915)", we define an "apoptotic process [ER stress]" that constitutes a course of ER stress, and an "apoptotic process [Phospholipidosis]" that constitutes a course of Phospholipidosis, and so on.
One of the difficulties of capturing a toxic process is that
some toxic effects are protective responses to xenobiotic
substances (drugs) [
Developing a toxicity-dependent process subtree is based on the low-hanging fruit policy. From toxicology-related textbooks and published articles, terms were extracted and manually annotated.
Here, as a function-related process, in addition to the functionexecution process, the TXPO defines meta-functioning processes. Meta-functioning processes are functioning processes specific to other functions and include controlling, for example. Subtypes of controlling include the regulation of apoptosis and cell cycle control. (2) Decomposition of Functioning
The TXPO specifies a functioning process based on a
function decomposition framework. As an ontological engineering
approach, a device (system) consists of sub-devices
(subsystems). In a function decomposition tree, the whole function of
a system is achieved by a sequence of sub-functions of the
subsystems. As biological functions can be considered
specializations of systemic functions [
In the present study, we developed a framework called the
“toxic course map” to represent toxic courses uniformly. The map
represents a toxic course as causal relationships between
processes (Fig. 4). With regard to development of toxicity, we applied
the imbalance theory [
1) a functioning process (supply) for biological defense and maintaining homeostasis; 2) a functional demand process (demand) as toxic activity; 3) balance/imbalance between toxic activity and defense processes; and
4) outcome from organelles, cells, or tissues to the organ exhibiting toxicity manifestations The degree of functioning performance can change according to changes in demand; however, if demand exceeds the performance of functioning, an imbalance occurs and results in an outcome that is no longer latent and manifests toxicity. Table 2 shows examples describing the imbalance framework in ER stress. In general, a molecule plays multiple roles in the body. Therefore, in the present research, we tried to explicate the roles of molecules participating in specific processes in the toxic course. For example, in ER stress, GRP78 participates in the protein refolding process and can play the role of a "chaperone" that assists protein refolding (Fig. 5). GRP78 also plays the role of "autophagy inducer" in the positive regulation of autophagy process during ER stress. As viewed relative to the role of a molecule, the TXPO contributes to identifying biomarkers that participate in the turning points of processes that cause cell injury during the course of toxicity manifestation. As for drugs, TXPO makes explicit the role of drugs in a specific toxic process. For example, tunicamycin plays a 'protein glycosylation inhibitor' role and participates in the negative regulation of glycosylation process. Tunicamycin also plays an 'apoptosis inducer' role in the positive regulation of apoptosis in the liver (Fig. 5). As of February 1, 2018, the TXPO defined approximately 6000 entities, and Figure 5 shows the major relationships between terms defined in the TXPO.
We developed a prototype toxic process interpretable support knowledge system, known as TOXPILOT. The TOXPILOT provides varied useful information based on the TXPO (Fig. 6). The TOXPILOT visualizes toxic course maps (Fig. 6 (a)), as described in the previous section. Since our map can visualize molecules that participate in toxic processes, we can apply the map to facilitate explanation of biomarkers for toxicity prediction by machine learning. Our preliminary data show that by using maps, in vivo and in vitro data of predicted marker genes of liver toxicity can be comparatively analyzed. As a result, we can identify genes predicted to participate in common processes in ER stress based on rat in vivo and human in vitro analyses. Thus, this toxic course map facilitates evaluation and extrapolation to humans for translational research.
The TXPO also provides process maps (Fig. 6 (b)), in which sub-processes can be displayed according to the whole-part relationship of systemic functioning across granularities. These maps also enable visualization of pathologic findings associated with a process.
The TXPO also provides a general course map that visualizes general toxic courses common to multiple specific toxic courses (Fig. 6 (c)). In safety evaluation, toxicologists sometimes want to know whether one phenomenon that occurs in a particular toxic course could occur in other toxic courses. For instance, in the course of lipidosis, "lipid accumulation" can cause "increasing hepatocyte volume." The TXPO system extracts information from the RDF database by SPARQL and automatically generates a general course map. In the general course map, common processes are represented as large nodes. As a result, users can obtain information indicating that "increasing hepatocyte volume" is common to other toxic courses, such as cholestasis. Moreover, users can obtain information regarding different causes associated with other courses. Each toxic course is colored, so users can see easily that, for example, "bile acid accumulation" occurs specifically in the course of cholestasis.
Our system also provides a function for searching routes from specific processes (Fig. 6 (d)). When users want to conduct retrospective analyses, TOXPILOT provides an illustration of ‘upstream’ of the focused process in the toxic course, which can help identify critical causes during the early stages of toxicity. In the same way, if users wish to know how a process unfolds with the progress of the toxicity development, our system provides a ‘downstream’ illustration that supports severe manifestation risk management. o c
o u f s nd leew tsech isonw to is a v tf e iz a h isu ssee .seT tsohw sohw l s a v c s ap rop sc l
h d e ig in m n rpo ihhg reww se eew icx nd lo r
o a e cou teb t
n ,y h ic sp i (a ti vo ieachh .rsceouT .PTXO ox ih ed r )T son lv r sse icx een cou sg ic .rs ize tox loo l c a u i s f c t
n i i v c r
e ap sep iff e d m lep in se ti d ruo lu ey c m la l p ra to is
d een no is G m e ()c com tehm e a s e
r te ts u n )(doR rdoow .ssce ) T O L I T s y s e g d e l w O n k e L b a t e r p r e t n I s s e c o r P c i X O T f o w e i v r e v o n A 6 e r u g i
There are many biomedical pathway databases, including
KEGG [
Understanding toxicity mechanisms is a hard task. Among the many issues involved, one aspect is the complexity of various interactions in the toxic course. We demonstrated that our imbalance model can make the context clearer and distinguish toxic actions from body defense functions in each granularity, thus facilitating interpretations of toxic mechanisms. Interestingly, we found that sometimes one functioning process plays both a biological defense role and toxic role. For example, as shown in Table 2, during the course of ER stress, apoptosis plays a defensive role in removing abnormal cells accumulating unfolded proteins, whereas increasing apoptosis has a toxic effect at the organ level that can lead to liver failure. Furthermore, our imbalance model is possible to explain that an imbalance also occurs when defensive functioning becomes excessive. For instance, when the cell proliferation function becomes excessive, liver carcinogenesis can develop at the organ level. We are currently trying to introduce the imbalance model for other toxic courses and clarify the relationships between functioning demand and the defense function.
The identification of biomarkers for toxicity prediction using machine learning techniques is a frequent objective of computational toxicology research. However, such machine learning approaches often lack accountability. By annotating markers based on the ontology of TXPO, associating markers with the toxicity process as a progression of toxicity development, and by visualizing them, it is possible to provide accountability for marker genes. Therefore, the TXPO and TOXPILOT will contribute to the enhancement of safety evaluations. Moreover, the general course map in TOXPILOT provides an indication of causal relationships across various mechanisms of toxicity. Therefore, it could be used to discover previously unknown relationships and contribute to the identification of new risks.
Using TOXPILOT, researchers can obtain an overall picture of the mechanism of toxicity in the liver and explore the systemic effects of biological functions. Moreover, from a fragmented knowledge perspective, our maps facilitate the discovery of new knowledge through commonality. Also, our system supports both retrospective and forward analyses. In this way, TOXPILOT enables the generation of knowledge cycles based on the TXPO (Fig. 6).
In the present work, we developed a TXPO to organize toxic process knowledge. As an application, we developed the TOXPILOT as a prototype system for supporting the interpretation of toxicity mechanisms. We are currently annotating more toxic courses and enhancing the level of sophistication of the terms in the TXPO. In the future, we plan to cover toxic courses in other organs, such as the kidney. We are also planning to reuse various ontologies, such as the Disease Ontology (http://disease-ontology.org/), and the Monarch Disease Ontology (Mondo, https://github.com/cmungall/tbd-disease-ontology.) Bridging domains on toxicity knowledge from basic to clinical medicine could help elucidate multiple mechanisms of toxicity. In furthering the applications of the TOXPILOT, we are striving to enhance its functions. The first version of the TXPO is available via the NCBO BioPortal, and the prototype TOXPILOT is open at the following site: https://toxpilot.nibiohn.go.jp. New term requests and reporting of issues can be made via a GitHub tracker (https://github.com/txpo-ontology/TXPO/issues.) We plan to submit TXPO to the OBO foundry for collaboration and knowledge sharing among not only toxicologists but also other biomedical communities.
This research was supported by AMED under grant number 19nk0101103h0005. The authors would like to thank Dr. K. Horimoto, Dr. K. Fukui, Prof. Y. Uesawa, and S. Ueda. The authors also thank Prof. R. Mizoguchi for useful discussions related to the ontological approach.