Providing Adverse Outcome Pathways from the AOP-Wiki in a Semantic Web Format to Increase Usability and Accessibility of the Content

Since its establishment in 2010, the adverse outcome pathway (AOP) concept has become a prominent tool for the risk assessment community.^1,2 AOPs are a chain of biological processes, called Key Events (KEs), starting from a molecular perturbation with a stressor toward an Adverse Outcome (AO), connected by Key Event Relationships (KERs). AOPs exist to capture all mechanistic toxicological knowledge from literature and data, to direct future studies to fill gaps of existing knowledge, and to drive Integrated Approaches to Testing and Assessment (IATA) development.^1,3 This was demonstrated with the AOP-based IATA for skin sensitization, resulting in various IATA with combinations of in vitro and in silico assays outperforming animal tests.⁴

The majority of the AOPs are developed and stored in the AOP-Wiki (https://aopwiki.org/), which is part of the AOP knowledge base, released in 2014 as a result of the AOP development program initiated by the Organisation for Economic and Collaborative Development.⁵ This wiki is designed to facilitate collaborative development of qualitative AOP descriptions, and thereby promote their incorporation into risk assessments and stimulate effective reuse of mechanistic toxicological knowledge.^6,7

The resulting AOPs describe much of the biological context surrounding toxicological processes; most of the information on genes, chemicals, biological pathways, phenotypes, among other things, is already captured in specialized databases or ontologies outside of AOP-Wiki.⁸ However, the AOP-Wiki has limited possibilities for linking external information and data, mostly consisting of free-text descriptions and links to the U.S. CompTox Chemistry Dashboard⁹ and to NCBI for taxonomic applicability.¹⁰ An initiative to make the reporting more consistent was the introduction of KEs Components¹¹ for the annotation of Biological Processes, Biological Objects, and Biological Actions for KEs, and annotations of cell types and organs in which KEs can occur.

Since the AOP-Wiki is the central repository for AOPs and therefore a key player in the shift toward animal-free testing strategies, it is essential that its contents can be queried and utilized effectively to answer biological questions and to reuse existing knowledge. However, accessing the data computationally or linking with other resources is hardly possible when only downloadable eXtensible Markup Language (XML) data dumps are provided that consist mostly of free text. Because of these aspects, parsing and querying the continuously growing amount of information in the AOP-Wiki is a complex, time-consuming task. This is a problem because it prevents the integration of AOP knowledge with other data and resources.

This could be resolved by applying Linked Open Data solutions, such as structuring the data in a resource description framework (RDF) model,¹² introducing persistent identifiers and semantic annotations, and implementing Application Programming Interfaces (APIs) for accessing the data. RDF represents knowledge as semantic triples, in which a subject, predicate, and object together define a statement and assist in the meaningful representation of knowledge in a machine-readable manner.

These concepts are generally in line with the FAIR principles¹³ for data and knowledge management, developed to enhance the Findability, Accessibility, Interoperability, and Reusability of data and allow computational support of data usage; for example, the solutions applied by the Swiss Institute of Bioinformatics with the development of neXtProt Linked Data by implementing RDF annotations for easier exploration and retrieval of data through web services.^14,15

Also, the use of ontologies and vocabularies for semantic annotations allows for the integration of data between resources, such as the direct linking of chemical or protein databases with WikiPathways.^16,17

In this article, we show how using RDF makes the AOP-Wiki content more usable for automated exploration in combination with other existing semantic web-based information sources. We describe our implementation of Linked Open Data solutions for the AOP-Wiki to introduce new, effective ways of accessing and using the data. These solutions will enhance the usefulness of the AOP-Wiki to risk assessors, developers, and modelers, and facilitate answering complex research questions, also across databases or as part of automated workflows.

We hypothesize that with the implementation of RDF, with the use of standard ontologies for semantic modeling of information captured in AOPs, the data can be better exploited.¹⁸ Furthermore, the use of persistent, unique, and resolvable identifiers allows interoperability with other related data sources. When combined with computational tools that can access experimental data, these approaches can make AOP information a core element for predictive modeling.¹⁹

The main result of this project is an RDF schema and scripts that lead to the production of RDF content for all AOP-Wiki content with additional semantic annotations, persistent identifiers, and extended identifiers for genes and chemicals, and consists of 122,576 unique triples consisting of 15,132 unique subjects, 158 unique predicates, and 53,087 unique objects (Fig. 1). The semantic annotation was done using 8 standard metadata vocabularies and 17 domain-specific ontologies and vocabularies. We here detail these results.

The work described in this article has led to the creation of AOP-Wiki RDF based on the existing AOP-Wiki XML, combined with a variety of ontologies and enriched with persistent identifiers. Besides, the data are extended with additional identifiers for chemicals, proteins, and genes. The RDF has been validated, loaded on a SPARQL endpoint, and tested using a Jupyter notebook, and is indexed in YummyData for external validation of the SPARQL endpoint. These developments made AOP-Wiki content ready for use in risk assessment workflows, through coding environments, or in federated SPARQL queries.

Because the AOPO is developed for consistent reporting in the domain of AOPs and allowing the integration of data and tools,²⁸ it was an obvious choice to implement the AOPO for semantic annotations in the AOP-Wiki RDF. The ontology includes a variety of AOP-specific definitions for properties and classes, which are directly applicable to AOP-Wiki content in the RDF. However, these do not fully cover all types of entities and relationships that exist in the AOP-Wiki. For example, while it has terms to describe the connections between AOPs, KEs, and KERs, there is no annotation for the link with stressors. Similarly, terms are lacking for sex and taxonomic applicability, KE components, KER-specific information, and AOP assessment sections such as KE essentiality and quantitative considerations.

For the terms missing in the AOPO, we selected definitions from a wide range of other ontologies and vocabularies for the semantification of predicates and subjects, including NCI Thesaurus,³¹ NCBI of Organismal Classification, Gene Ontology, and Measurement Method Ontology. Whereas the majority of the AOP-Wiki contents are generic and can be described with well-established metadata ontologies, some properties of AOPs, KEs, and KERs could not be found, leading to the selection of more general terms, lacking detail that would be preferred. With the conversion to RDF, most of the necessary terms have been uncovered and documented, and these will be added to the AOPO.

Because the realm of AOPs includes many types of data, knowledge, repositories, and services, the development and implementation of a central, community-wide vocabulary would facilitate their integration. Since the AOPO has been developed to fill that purpose, it could be extended to include descriptions of classes and properties for all ontology terms used in the AOP-Wiki RDF. Having a central, field-wide ontology for AOPs helps maintain a high-quality vocabulary through continuous development and involvement of the community. Such an ontology would facilitate the annotation and integration of data, resources, and tools, as is done with the eNanoMapper ontology in the nanotoxicology community.⁷⁰

With the increased importance of consistent use of identifiers to integrate knowledge and data, our implementation of persistent identifiers for AOPs, KEs, KERs, stressors, chemicals, proteins, and genes will benefit the integration of the AOP-Wiki with other resources, data, and tools.⁷¹ These persistent identifiers stored in the MIRIAM registry are stable, unique, resolvable, documented, and directly linked to the corresponding entries in the databases.^20,72

Furthermore, our efforts have introduced additional content into the AOP-Wiki RDF through ID mappings and text-mapping for chemicals and genes, providing more ways of extrapolating the data and linking with other resources and data. While the addition of molecular descriptors of genes and proteins in KEs and KERs can provide more insights into the actual processes described, it is merely a small step toward complete machine-readable biological processes.

Although we have added molecular identifiers to increase the number of link-outs and improve the usefulness of the database, our addition of genes through textual ID mapping does introduce errors to the AOP-Wiki RDF. The automated process on free-text content assumes good practice in writing gene symbols and names according to the HGNC guidelines.⁷³ Although HGNC strives for stable gene symbols and makes justified changes for problematic ones,⁷⁴ some gene symbols still overlap with free-text abbreviations in the AOP-Wiki and are therefore falsely recognized.

Opportunities exist to improve the AOP-Wiki machine readability by having more structured text and annotations for molecular entities, pathways, organs, species, and other biological concepts that are relevant for AOPs and not yet covered by the KE components. Text-mining tools, such as ProMiner,⁷⁵ ContentMine,⁷⁶ and PolySearch2,⁷⁷ could be implemented for extracting biological concepts and understanding associations to add more structured information to the AOP-Wiki, and facilitate the integration with other databases and tools. Once such concepts are recognized and extracted, the RDF could be extended and increase the interoperability of the AOP-Wiki with external databases, such as pathway databases.

The AOP-Wiki RDF allows for new and efficient ways of accessing the data and using it to answer questions. By loading the RDF in a SPARQL endpoint, SPARQL queries can be used to access the data and extract all necessary information. It allows complex queries across the complete AOP-Wiki database, optional filtering for any variable, and requesting an output suitable for answering the research question. Furthermore, these SPARQL queries can be executed from most coding environments as part of larger workflows or data pipelines.

It also facilitates direct linkage of databases through federated SPARQL queries, which returns information across databases with a single query. Any database with a SPARQL endpoint can be used for such questions across databases, such as WikiPathways,^16,17 Wikidata,^59,60 neXtProt,¹⁵ UniProt,⁵³ ChEMBL-RDF,⁷⁸ DisGeNET,^79,80 Rhea,⁸¹ and Pathway Commons.⁸² Examples of federated queries are stored in the SPARQL examples panel in the AOP-Wiki SNORQL User Interface https://aopwiki.rdf.bigcat-bioinformatics.org/ (Fig. 2).

Another way of using the RDF to extract AOP-Wiki content is through a web service such as the git repository linked data API constructor (grlc),⁸³ which can build a Web API on top of a SPARQL endpoint with predefined SPARQL queries. While more straightforward than SPARQL queries, the API is limited to the predefined SPARQL queries and variables implemented in these.

An advantage of creating RDF for the AOP-Wiki is the ability to link and expand AOP-Wiki content with information from other databases. For example, the Adverse Outcome Pathway Database combines knowledge from the AOP-Wiki with annotations of genes, chemicals, diseases, tissues, pathways, ontologies, and ToxCast data.^84,85 Future work should focus on integrating such efforts by developing RDF, and thereby allow full integration of their data and tools⁸⁶ with the AOP-Wiki RDF and other databases.

In terms of compliance with Linked Data standards according to the YummyData registry, the AOP-Wiki SPARQL endpoint consistently scores above average in availability, freshness, operation, usefulness, validity, and performance.⁶⁹ With a consistent A rank, the AOP-Wiki SPARQL endpoint is placed among the 10% best scoring of the 70+ SPARQL endpoints registered in YummyData. However, incidentally the Umaka score dropped due to slow server response, or when the service has been down for maintenance or data loading. Based on the feedback given by YummyData, a point for improvement would be supporting more response formats of the SPARQL endpoint to increase the usefulness score.

The implementation of compact identifiers and the development of a formal, machine-readable RDF schema make the content of the AOP-Wiki more findable and interoperable to other components of the database, and by allowing SPARQL queries and API to explore the data, the AOP-Wiki database was made accessible through new methods. Furthermore, the addition of link-outs to various chemical, gene, and protein databases, as well as the data storage in an RDF format and implementing Linked Open Data standards, has made the data more interoperable with other databases and tools.

Furthermore, the AOP-Wiki has recently introduced licenses on its content, and the code for the creation and validation of the AOP-Wiki RDF is available under the MIT license. These provide clear statements and terms of using, sharing, and modifying the content. Taken together with the addition of metadata and semantic information represented by ontology annotations, the content of the AOP-Wiki has been made more accessible and reusable. Therefore, the development of the AOP-Wiki RDF addresses all major FAIR principles.¹³

Overall, the AOP-Wiki RDF allows for new ways of exploring the data, using it in automated workflows, from coding environments, or directly through a SPARQL endpoint. With the implementation also comes the possibility to execute federated queries to combine data of multiple resources and answer more elaborate questions. For example, KEs can be linked to the results of ToxCast assays that measure the activity of the protein described, or molecular pathways can be explored for a more detailed description of mechanistic processes.