The Environmental Health Language Collaborative (EHLC) Adverse Outcome Pathway (AOP) Standards Workshop Report

1.1 Background

The international Adverse Outcome Pathway (AOP) communities, consisting of government, academic, and industry scientists, stakeholders and partners, have recently proposed several similar initiatives aimed at the improvement of AOP data standards and the coordination of AOP data processing and tool pipelines. The FAIR AOP Cluster Working Group was initiated to identify and evaluate existing approaches to improve AOP data standards for AOP communities as well as review and coordinate AOP tool developers in the concerted processing of AOP data and metadata, including biomedical entity mapping, from the AOP-Wiki to specifically improve AOP data content. The FAIR AOP Cluster Working Group developed and published a FAIR AOP Roadmap (Mortensen et al. 2025), which highlights existing AOP tools and efforts to standardize and coordinate processing of AOP data to include biomedical information. While the FAIR AOP Roadmap identified existing AOP efforts, the workgroup was unable to independently establish a pipeline/workflow for the processing of biomedical information for inclusion in the AOP-Wiki or other repository without community agreement.

The NIH-led Environmental Health Language Collaborative (EHLC), established in 2021, is a community of practice dedicated to advancing environmental health research by promoting standardized language to achieve harmonized datasets and interoperable tools. The EHLC Executive Council organized a 2-day event focused on AOP standards, which was motivated by the FAIR AOP Cluster to engage other aligned groups (e.g. Biomedical Knowledge Graph, New Approach Methodologies (NAMs), AI/NLP biomedical, the Exposome, and Data Sharing and Standards communities) in order to discuss current methods that improve data content, interoperability and machine-actionability that could be applied to AOPs, and facilitate NAMs, thereby improving the mechanistic underpinnings of AOPs to support risk and regulatory applications.

1.2 Community effort to Identify and evaluate approaches to improved mapping of biomedical entities

Through the EHLC AOP Standards event, we have sought out feedback from the aligned groups, many of which have had more consistent US funding, on how to improve AOP data standards and trustability. Invited presenters across disciplines were selected based on their specific expertise and unique contributions to data standardization, applications, tools or methods directly relevant to human health, disease, or toxicological understanding that could be used to improve AOP development or expedite the application of AOPs for NAMs for regulatory application. With many emerging activities across these domains, EHLC is fostering collaboration and knowledge exchange among them.

1.3 Identify community needs and decision contexts for regulatory application

Understanding agency requirements directly affects the types of information required in different decision contexts, whereby agency coordination has the ability to provide a unified framework for how NAMs can be practically applied (Stucki et al. 2022; Ouedraogo et al. 2025). Though the regulatory processes for US and EU regulatory agencies are different and adhere to differing policy constraints in terms of what type of mechanistic content is needed and required, the AOP community as a whole works to meet regulatory requirements and needs across agencies. Therefore, any approach to the implementation of methodology or tools for the coordinated alignment of biomedical or other relevant information to AOPs must be generalizable across regulatory processes. The AOP community, with these stakeholder needs in mind, is then obliged to respond in the creation of a coordinated strategy for improved data standards that meets the needs of all users. This need is formalized by FAIR AOP mechanistic output that can support NAMs and accelerate next-generation risk assessment (NGRA) through harmonized templates and community agreed upon reporting strategies to improve regulatory assessment processes.

2.1 Conference presentations

The conference presentations were kicked off by Dr. Maria Shatz (NIH/NIEHS; EHLC EC), who welcomed all attendees and shared the workshop agenda. Dr. Shatz briefly defined the charge of the EHLC and described the workshop motivation as an opportunity to bring the AOP community together with other related communities. Example communities relevant to the AOP Community and included in the 2-day event were the Biomedical Knowledge Graph Community, NAMs community, AI/NLP Biomedical Communities, Exposome Community, and Data Sharing and Standards Community.

2.1.1 Day 1

Charles Schmitt, NIH/NIEHS; EHLC EC – Introduction

Dr. Charles Schmitt led the charge by first introducing the EHLC community of practice and providing an introduction to the workshop. EHLC began in 2021 with the mission of advancing environmental health research by promoting harmonizable language to achieve harmonized datasets and interoperable tools. Dr. Schmitt stated that EHLC works through several mechanisms, including hosting webinars, workshops and events to communicate progress and discuss relevant issues, and facilitating working groups to advance solutions. Dr. Schmitt noted that there are many new and ongoing community activities and that EHLC serves to promote and foster the sharing of knowledge across communities. He stated that the charge for the workshop is to raise awareness, increase alignment with other communities, and seek advice to inform future efforts, including how to improve the generation of knowledge and data and how to improve the import of data into AOPs in automated and semi-automated ways.

Holly Mortensen, US EPA – The FAIR AOP Roadmap: Community Efforts Related to AOP Standards

Dr. Holly Mortensen began her presentation with the timeline for AOP tool development. Dr. Mortensen discussed how AOPs have evolved from the inception of AOP-Wiki and with the mapping of gene and protein information in the EPA AOP database (AOP-DB), where early efforts focused on mapping key events (KEs) to protein ontologies (Ives et al., 2017; Pittman et al., 2018). Mortensen et al. (2021) introduced the integration of AOP and biomedical (meta) data (e.g. pathway, disease, population and individual genetic variability). In 2022, AOP data started to be reused, with the AOP-Wiki and AOP-DB RDFs (Martens, Evelo et al. 2022; Mortensen, Martens et al. 2022). AOP initiatives to implement the use of artificial intelligence (AI) are quickly progressing, underlining the importance of coordinated approaches. AOP tools depend on the structure of the data within AOP-Wiki. Dr. Mortensen elaborated with discussion of the FAIR AOP Cluster Workgroup efforts related to AOP standards, discussed in the recently published FAIR AOP Roadmap (Mortensen, Gromelski, et al. 2025). The FAIR AOP Roadmap discusses how AOP tools (e.g., AOP-Wiki, AOP-DB, AOP-DB-RDF, AOP-Wiki RDF, ComptoxAI, Vizit, and AOPWiki-Explorer) are unique in their processing and mapping of AOPs to biomedical data, and though data content of all tools is overlapping, no tool is currently coordinated or interoperable with other AOP tools. Each tool developer maps AOP-Wiki data independently, so a future goal is to streamline this process and make it less onerous so that each tool does not have to repeat this critical step. Dr. Mortensen discussed three current OECD proposals addressing aspects related to improved inclusion of mechanistic AOP data for use in the regulatory process that are relevant to NAMs, specifically the OECD proposal for Improved AOP Reporting Standards for Biomedical Data, which focuses on the coordination of mapping of biomedical entities to AOPs. Dr. Mortensen closed her presentation with a call for papers to the open collection in F1000, entitled, “AOP Mechanistic Data-Oriented Coordination” centered on human health relevant AOPs and biomedical data standards that she hoped would serve as a platform for submission for workshop presenters (Mortensen et al. 2020).

Ginnie Hench, Open BioData Modeling – AOPs for EHLC and the Biomedical Research Community

The AOP framework was introduced in 2010 to support chemical risk assessment decision-making and has been informed by a systems biology perspective since its inception. Identifying ways to enhance AOP computability by leveraging biomedical resources and tools could reduce the time and labor burden associated with AOP development and help accelerate translation of biomechanistic knowledge in service of regulatory decision making. The AOP-Wiki and its underlying database, the AOP-Knowledge Base, serve as a central repository for AOPs, with the AOP-Wiki being the sole AOP authoring tool for the AOP-KB. The AOP-Wiki has an important role to play in enhancing the computability of AOPs, based on the organization of its data model and through the information input and output interfaces that it offers. As such, AOP-Wiki evidence model (EMOD) prototyping work was initiated in 2022 and later extended to align with efforts underway in the Methods-2-AOP collaboration. To date, the EMOD work has led to defining essential properties relevant to structuring evidence, methods, and references. Attention has been given to both human accessibility and machine readability in the design and implementation of EMOD prototypes.

Penny Nymark, Karolinska Institute – Why Adverse Outcome Pathways Need to Be FAIR

FAIR data and AOPs are useful in predictive toxicology research, as they enable explainability and trust in toxicological modeling. FAIR principles can be used to guide solutions and tools for machine-actionable data and metadata. AOPs must be sufficiently complex to be useful for modeling data, and by being FAIR, AOPs could retain high utility while being more streamlined with the use of machine learning. There is a difference between social and technical FAIR implementation for the AOP-Wiki, which includes a balance between meeting community needs and meeting necessary data standards. Regulatory risk assessment is dependent on trust, validation, and efficiency to achieve legal certainty and pre-market SSbD (safe and sustainable by design) depends on trust to achieve scientific confidence. As animal testing is phased out of chemical safety assessments, FAIR AOPs will be key in supporting both regulatory risk assessment and pre-market SSbD by increasing process speed and trust in data models.

Scott Lynn, US EPA – AOP Applications in the US EPA Endocrine Disruptor Screening Program

Currently, the EPA requires pesticide chemicals to be screened for potential estrogenic effects in humans. The Endocrine Disruptor Screening Program (EDSP) uses a tiered testing approach to screen chemicals for adverse effects, including assays that are intended to reflect estrogenic effects in humans. AOPs are useful for this type of chemical screening as the pathways of interest and endpoints of interest are already known. The US EPA Toxicity Forecaster (ToxCast) Estrogen Receptor (ER) Pathway Model integrates data from multiple assays and provides a score (area under the curve) for the chemical’s estrogenic potential. The ToxCast ER Pathway Model has been validated for regulatory use by EPA. AOPs for EDSP can be used to validate NAMs for estrogenic and androgenic effects, and to prioritize chemicals for further screening. Certain KEs and adverse outcomes are in need of further development to select appropriate assays and NAMs for regulatory assessment.

Helena Hogberg, NIEHS – Methods2AOP: A Collaboration to Strengthen the Integration of Test Methods into the Adverse Outcome Pathway Framework

Method2AOP is an international collaboration working to integrate test methods and protocol information into the AOP Framework and to promote acceptance of NAMs data in regulatory applications. Methods2AOP has developed a proposal for including structured method information (e.g., test system, assay readout) in the AOP-Wiki. This proposal organizes and standardizes the information to be included, such as study type (e.g., in vitro, in vivo), documentation or standard operating procedures (SOPs), species, cell or culture type, culture conditions, and limitations. Methods2AOP seeks input from the broader community, as requiring the inclusion of additional documentation may increase the users’ burden. By promoting test methods and finding underserved KEs with no available data or methods, this effort would demonstrate the data gaps for method developers to focus on. Harmonizing documentation (e.g., templates) and refining information is ongoing.

Madison Feshuk, US EPA – Connecting ToxCast to the AOP Framework to Support Assay Interpretation and Data Use

The US EPA ToxCast program is a world leader in providing an accessible bioactivity data resource for toxicology. This data can be used to support chemical evaluation, providing alternatives to traditional animal testing for hazard characterization. ToxCast assays employ heterogeneous assay technologies, looking at diverse event types and biological targets. Data are released on an annual basis via the ToxCast database, invitrodb, after new data processing and assay curation. Users can access the public data from CompTox Chemicals Dashboard (CCD), Computational Toxicology and Exposure (CTX) Application Programming Interfaces (APIs), and the ToxCast downloadable data webpage. Annotation standardization practices are necessary due to the number and diversity of assays across the database. Useful annotations include the tissue, cell type, and intended target, and some annotations are hierarchical (e.g., general target family) or map to larger processes. A table of manually curated mappings of AOPs and KEs for individual ToxCast assay endpoints will be included in the invitrodb v4.3 release (anticipated Fall 2025). Accessing this ToxCast-AOP data can be currently done through the CCD, such as chemical-specific ToxCast bioactivity summary grid or the “AOP Information” section under the Executive Summary. A future goal is to integrate AOP and KE searching into the CCD to help users more easily identify the relevant ToxCast assays (Feshuk, 2025).

Anne Thessen, UNC Chapel Hill – Ontology Support for the Human Exposome Project

The Human Exposome Project launched in May 2025 with the release of the Washington Declaration, which specifically identified ontologies as a technology needed to advance exposome research. Advances in genomics and phenomics made possible through the application of ontologies and knowledge graphs (KGs) in biomedicine, have yet to extend to the effects of environmental determinants of health (including social, biological, and physical factors), the so-called “exposome”. Here we examine some existing biomedical knowledge resources: the Monarch Knowledge Graph, the Environments, Conditions, and Treatments Ontology (ECTO), and the Mondo disease ontology, to identify gaps in the semantic infrastructure and propose approaches to understand how the environment impacts disease. We propose development of ECTO as a unifying ontology for the exposome, that can link together data and existing exposure ontologies across disciplines to fully realize the Global Human Exposome Project.

David Hines, RTI – Leveraging Standards to Link Mechanistic Frameworks Across the Source-to-Outcome Continuum

Data and metadata standards are interrelated; metadata standards contextualize and describe data, while data standards ensure consistency and interoperability across a field. Environmental human health research requires multiple data streams, which must be integrated. Environmental health sciences (EHS) data present a challenge as they are often siloed, lack appropriate metadata, or are pulled from diverse domains with varying standards. The Source to Outcome (S2O) continuum links factors such as contaminant sources and demographics to adverse outcomes. Work is currently ongoing to develop a standards and terminology model and data infrastructure to enhance interoperability, through a mechanistic frameworks use case. FAIR adherence and metadata availability can enhance the use of machine learning and AI tools for EHS, especially when pushing to accelerate AOP discovery and validation.

Sierra Moxon, Berkeley Lab – Modeling the Path from Source to Outcome with Biolink Model

Biolink is a high-level, open-source data model designed to standardize types and relationships in biological KGs, covering entities like genes, diseases, chemical substances, organisms, genomics, phenotypes, and more. The goal of the Biolink Model is to provide a consistent framework for representing biological knowledge across various databases and formats. We are using Biolink in the Source to Outcome project to support the development of standards and terminology models for environmental health science data. Biolink is highly annotated with knowledge from various domains and can map to external ontologies, which helps with interoperability. In this way, Biolink can be useful to the AOP community because AOPs function to integrate and organize information and evidence from across research domains to explain the causal biomechanistic pathways linking interactions with chemicals and other hazards to adverse effects. Biolink associations can help link exposures to downstream effects on individuals or populations. Biolink also offers ways to leverage curation efforts behind ontologies, like the Monarch Disease Ontology (MonDO). The culmination of these steps is a Biolink Model that describes data inputs, outputs, and context from diverse domains.

Ginnie Hench, Open BioData Modeling – AOP-Wiki Data Model Vision to Enhance Computability of AOPs

AOP Event Components (ECs) were introduced to the AOP-Wiki to define KEs in computable terms – using an action, object, and process term – that map to ontologies and controlled vocabularies. The EPA’s Endocrine Disruptor Screening Program explored the possibility of using the AOP EC concept to annotate thyroid research abstracts, alongside a larger systematic review effort. The manual curation effort led to expansion of annotated entities beyond the 3 core EC entities to include causal agents, methods of measurement, biological context, and domain of applicability concepts (sex, life stage, and taxa). Though it is not scalable for humans to manually complete this type of text annotation work, it served as a useful proof of concept for how we can envision computable approaches to aggregating contextual information, which is necessary for AOPs to provide context for NAMs. Importantly, exploration of AOP-based information extraction workflows has informed the EMOD data model prototyping work and plans for improving the data model of the production AOP-Wiki.

2.1.2 Day 2

Karamarie Fecho, Copperline Professional Solutions, LLC; UNC Chapel Hill – Ontologies, Tools, and the ROBOKOP Knowledge Graph System in Support of AOPs

KGs have become a common approach for knowledge representation in many fields, including biomedicine. The basic unit of “knowledge” in a KG is the subject-predicate-object “triple”. The publicly accessible Reasoning Over Biomedical Objects linked in Knowledge Oriented Pathways (ROBOKOP) KG contains >10 million nodes and >140 million edges, drawn from dozens of integrated and harmonized biomedical knowledge sources. ROBOKOP achieves semantic harmonization by leveraging several key resources. Specifically, ROBOKOP invokes Biolink Model as an upper-level ontology and graph-based data model that provides the semantic glue to create a harmonized data ecosystem. Tools such as Node Normalizer and Name Resolver provide name resolution across disparate identifier systems. For example, Node Normalizer takes a Compact Uniform Resource Identifier (CURIE) and returns bio-equivalent CURIEs, as prescribed by Biolink Model. ROBOKOP recently was applied to develop an AOP between clofibrate, an environmental hazard and activator of peroxisome proliferator-activated receptor alpha (PPARA), and hepatic fibrosis. ROBOKOP was successful in building the AOP. For instance, ROBOKOP identified numerous edges or associations between clofibrate and PPARA (e.g., increases expression of ). ROBOKOP additionally identified downstream genes and biological processes or activities that link PPARA activation to hepatic fibrosis, replicating what is currently known. However, not all aspects of the AOP were identified, primarily contextual steps involving cell types and anatomical entities. Knowledge sources such as Cell-Cell Interaction Database (CCIDB) are being explored to fill in gaps in the ROBOKOP KG, which should increase the granularity of AOPs identified by ROBOKOP.

Nyssa Tucker & Alex Tropsha, UNC Chapel Hill – Automating AOP Elucidation Using Knowledge Graph Mining

The construction of AOPs can be done through the use of KGs. KGs can be used to determine novel functional relationships and rules by extrapolating from known interactions, such as in the metal-gene interaction case study that used ROBOKOP knowledge graph built at UNC with funding provided by NIEHS, to generate hypothetical AOPs. Cobalt was found to have a potential contribution to congestive heart failure via chronic upregulation of HIF1A, a connection that was not previously described in literature. ROBOKOP mining can identify hypothetical candidate genes both within and outside of the AOP-Wiki. RADisH (ROBOKOP Assisted Discovery of in silico Hypotheticals) is an in-development project aimed at automated development of AOPs with the use of ROBOKOP knowledge graph, which is under continuous development. Future studies are ongoing to address critical elements of AOP, such as temporality and directionality, currently missing in ROBOKOP.

James Balhoff, RENCI – Representing AOPs as GO Causal Activity Models

Gene Ontology (GO) organizes and describes concepts related to the function of gene products, using an annotation knowledgebase underpinned by a rigorous semantic structure of interrelated terms. A new approach, GO causal activity models, or GO-CAMs, can place function annotations into an interconnected context of gene activities, biological processes, cellular locations, and causal relationships across activities. GO-CAM can incorporate the structures of external ontologies including Uberon, Cell Ontology, ChEBI, and others, and is modeled within the same logical framework the bio-ontologies themselves. After development of a pipeline to provide GO-CAMs as a NCATS Biomedical Data Translator component, we developed a transformation of AOPs to CAMs. The pathway structure of AOP naturally lends itself to modeling in the CAM framework, allowing incorporation of AOPs into the Biomedical Data Translator via the CAM ingest pipeline. This process and related products are being validated and introduced to the community for use.

Jennifer Fostel, NIEHS – Standard Terminology: Supporting FAIR Data and Extending to AOPs

The Division of Translational Toxicology (DTT) Data Dictionary (DDD) is aligned with OBO Foundry ontologies, and its use will increase the value, findability, and interpretability of data. OBO Foundry covers broad and non-overlapping domains of knowledge and contains an upper-level ontology, based on the Basic Formal Ontology (BFO) standard, to promote interoperability. In the Ontology for Biomedical Investigations (OBI), an assay has a specified input and output and is a planned process with a parent. Cell viability, cell migration, and cell differentiation assays have been added to OBI. Other developmental neurotoxicity (DNT) screening assays are currently being created with OBI, such as a neurite outgrowth assay. An individual can request the addition of an assay to OBI. Steps include developing a good term label and definition, identifying the parent assay, and adding the target, input, measured analyte, and method. KEs are designed to be measurable and could be analogous to an assay. However, in AOP-Wiki many KEs are lacking method details and redundant terms (e.g., decrease vs. decline).

Harry Caufield, Berkeley Lab – Enhancing data harmonization for AOPs with agentic information extraction

Turning pathways into new AOPs can be done by enriching them with detail and structuring their provenance and evidence. Large language models (LLMs) are good at summarizing multiple knowledge sources and can be useful for transforming loosely structured text into structured AOPs. LLMs follow examples in structure and terms and work well with other computational tools and human curators; however, because they are grounded in language and not fact, LLMs do not produce outputs based on knowledge unless explicitly instructed in how to do so. Using LLMs to provide corresponding ontology identifiers for specific terms (e.g., Gene Ontology identifiers) will result in incorrect matches, but pairing LLMs with other approaches can be successful. The OntoGPT and CurateGPT tools are built to complement the strengths of LLMs with curated knowledge, including terms from OBO Foundry and other vocabularies. OntoGPT works well for normalizing terms within text with ontologies common to AOPs and minimizes the chance of imaginary or incorrect IDs. It will improve as LLM models get more powerful. OntoGPT requires use of an extraction schema, which may not be ideal and will require some adaptation. It also relies on lexical similarity or known synonyms to assign IDs so it may not be aware of some similarities. CurateGPT serves as more of a retrieval and augmentation tool. For CurateGPT, lexical similarity is not necessary. However, it is limited by the available examples and does not tell the user which agent to use. The next steps are to work towards more agentic approaches and frameworks (e.g., Aurelian).

Julija Filipovska, Independent Researcher, ERT, Ohrid, N. Macedonia – AI4AOP: The Crowd Dedicated to Harnessing the Power of AI Tools for Enhanced Development and Use of Adverse Outcome Pathways

The AOP development process requires identification and structuring of data and information relevant to building evidence and knowledge to support regulatory application in the area of chemical safety. It is based on five fundamental principles: (i) modularity of AOP building blocks (such as KEs), (ii) chemical, and more generally stressor, agnosticism, (iii) pragmatism aiming for simplicity of linear representation of biological pathways while allowing (iv) interconnectivity and (v) responsiveness to emergence of new data, information and knowledge. All of these, but particularly the principle of modularity and interconnectivity, require standardized language for the identification and description of the AOP building blocks, presenting a number of interesting challenges.

AOP-Wiki, the primary user interface for AOP development, maintenance, dissemination and usage, relies on crowdsourcing contributions that undergo review within the OECD and the wider scientific community (https://aopwiki.org/). At present the majority of the evidence and information within the AOP-Wiki is represented as a free text narrative (i.e. human generated language) format. Thus, the chance that AI tools can potentially be used to develop new AOPs and facilitate the modularity and interconnectivity of existing building blocks into functional networks, increases with more formal language standardisation for machine readability and interoperability. Certain processes are more likely to benefit from AI implementation, such as metadata management, systematic review and evidence mapping, detecting AOP networks, and quality control. For example, Key Event Components (KECs), grounded in standardized language of relevant existing ontologies, can be used as standardised KE identifiers. However, the possibility to tag KEs with appropriate KECs has been underused by developers of AOPs to date. In fact, a large number of KEs have not been tagged with KECs at all, and various higher biological level KEs have been tagged by more than one lower level KEC. Appropriate KE identifiers (i.e. appropriate KEC tagging) are necessary to support AOP modularity (and KE re-use in networks), but it is labor-intensive and difficult for humans to identify, select, add and use the appropriate KEC terms. This is partially due to the fact that in the AOP-Wiki KECs are not easily searchable, and ontology sources, structures, and hierarchies are not visible to developers and users. AI tools have not yet been successful in assigning KECs to KEs, but ongoing work is continuing to search for use cases and solutions.

Allan Peter Davis, NCSU – Using the Comparative Toxicogenomics Database to Provide Environmental Chemical Content and Inform Adverse Outcome Pathways from the AOP-Wiki

The mission of the Comparative Toxicogenomics Database (CTD; https://ctdbase.org/) is to address chronic diseases by understanding health effects and exposure toxicity responses from environmental chemicals. CTD manually curates, harmonizes, and centralizes literature-based exposome and toxicogenomic information across diverse species, allowing users to survey the environmental health data landscape from molecular mechanisms to population exposures (Davis et al., 2025). Additionally, CTD integrates data sets to find computational solutions that fill knowledge gaps and build chemical-induced disease pathways. The user-friendly online tool CTD Tetramers quickly computes four-unit blocks of information (“tetramers”) organized as a stepwise pathway linking an initiating chemical with a gene to affect an intermediate phenotype (biological process) linked to a disease outcome (Davis et al., 2023). This tetramer modular framework is analogous to the AOPs represented in the AOP-Wiki, allowing CTD chemical data to be intersected with each event in an AOP to look for potential environmental influences, discover putative new KEs to be added to established AOPs, and interconnect different AOPs to form disease networks. Importantly, this method can be used for any chemical, gene, phenotype, or disease in CTD or mapped from the AOP-Wiki, empowering users to construct testable frameworks for diverse studies that span the entire environmental health continuum from exposome source (AEP) to outcome (AOP).

Julia M. Malinowska, European Commission – Joint Research Centre (Ispra, Italy) – Omics2AOPs: Relating Genes, Proteins and Metabolites to Adverse Outcome Pathways through Established Ontologies

A key challenge in using omics data (e.g., transcriptomics, proteomics, and metabolomics) in regulatory toxicology is the lack of formal standardization of various elements across an omics-based workflow, including interpretation of omics data. Improved embedding of omics data within the framework of AOPs could provide a new approach for interpreting omics data to accelerate its use in regulatory toxicology. The European Commission – Joint Research Centre (EC-JRC) and Tampere University (the group of Professor Dario Greco) proposed a new project to the Working Party on Hazard Assessment (WPHA) of the Organisation for Economic Co-operation and Development (OECD) to address this. Omics2AOPs aims to relate sets of genes, proteins and metabolites from established knowledgebases and ontologies to KEs and AOPs. The project was inspired by work from the group of Professor Dario Greco that employed Natural Language Processing and manual curation to link gene sets to KEs (DOI: 10.1038/s41597-023-02321-w). The objectives of Omics2AOPs are 1) establishing a methodology for mapping KEs and AOPs to sets of genes, proteins, and metabolites, using high-quality resources (i.e., knowledgebases and ontologies), 2) developing a user-friendly tool to implement the aforementioned methodology, and 3) evaluating and demonstrating the utility and usability of such tool through appropriate use cases. In October 2025, following the acceptance of Omics2AOPs onto the workplan of WPHA, an in-person workshop and kick-off meeting took place at the EC-JRC (Ispra, Italy). The meeting convened international experts from disciplines pertinent to the project and provided a forum to review and discuss sector-leading activities that are important to Omics2AOPs, laying the groundwork for addressing its objectives.

Vikas Kumar, TecnATox (IISPV – URV, Spain) and German Federal Institute for Risk Assessment (BfR, Germany) – Evidence Extraction and Harmonization for AOP Development

The PARC (Partnership for the Assessment of Risks from Chemicals) is a broad collaborative initiative with a key objective: to facilitate the exchange, reuse, and interoperability of data across toxicology and chemical risk assessment. Among its primary activities is the integration of AOPs into regulatory toxicology through various data modalities, including omics data, image-based data, and associated workflows, as well as the FAIRification of AOPs. To support these goals, the partnership is developing a suite of tools and ontologies centered on AOPs. One major effort is the development of AOP-BOT (Kumar et al., 2023), a modular system for AI-assisted text mining. AOP-BOT is built upon three foundational components: Biological concept recognition, Biological concept normalization, and Visualization. This framework includes the S2CIE architecture (Syntactic, Semantic, and Contextual Information Extraction), enabling robust and context-aware extraction of information from unstructured text (Kumar et al., 2025). Another innovative tool under development is the AOP-SKETCH PAD, which provides a schema-agnostic representation of AOPs and KEs within a decentralized environment. The goal is to reduce the extensive curation effort typically required for harmonizing evidence and capturing and storing data according to FAIR standards. Complementing these efforts, AOPWIKI-Explorer is an interactive graph-based query engine that leverages large language models to enhance access to the AOP-Wiki archive (Kumar et al., 2024). By converting the data into a Labelled Property Graph (LPG) schema and offering a visual, multi-query interface, the tool enables users—including non-technical stakeholders, to construct flexible queries using either natural language or structured formats. This significantly lowers the technical barrier for exploring AOP data and facilitates more effective retrieval and analysis of graph-based knowledge. A workshop on AOP Ontology is proposed for the end of 2025, aiming to gather experts and stakeholders to advance these efforts further.

Marvin Martens, Maastricht University – Integration and Reuse of AOP Knowledge with RDF, SPARQL, and Ontologies

The AOP-Wiki serves as a central resource for capturing mechanistic toxicological knowledge. However, challenges remain in enabling dynamic organization, machine-readability, and integration into computational workflows. To address these limitations, a semantic, ontology-based representation of the AOP-Wiki has been developed using the Resource Description Framework (RDF), transforming the content into a FAIR knowledge graph.

The AOP-Wiki RDF includes all core components of the AOP-Wiki along with additional elements, structured using standard metadata vocabularies and domain-specific ontologies (Martens, Evelo et al. 2022). Data access is provided through three complementary services: a SPARQL endpoint for advanced querying, a SNORQL interface for interactive navigation, and a RESTful API for programmatic access. These services support a wide range of use cases, including querying chemicals linked to specific KEs and direct integration with external resources such as the AOP-DB RDF to extract ToxCast assay data.

The RDF is updated weekly and is already in active use by initiatives such as PARC and VHP4Safety. Ongoing development is focused on improving user interfaces and embedding the services into AOP development workflows. In addition, the concept of molecular AOPs was introduced to enable the integration of omics data with AOPs. Taken together, these efforts aim to support transparent, reproducible, and computationally accessible AOP research.

2.2 Conference breakout sessions

Following speaker presentations, workshop attendees were arbitrarily split into moderator-led breakout session rooms. Each moderator was provided with written questions developed by the workshop organizers, as well as schematics of current and proposed AOP data models with that could accommodate biomedical data inclusions. Outcome areas to be addressed included: 1. Proposed AOP-Wiki features and data model changes; 2. Knowledgebases, repositories and other tools that support AOP development; 3. Reporting templates relevant to AOP inclusion in regulatory decision making; and 4. AI/LLM to assist in generating and reviewing AOPs.

2.2.1 AOP-Wiki features and data model changes

Preliminary assessment of locations within the existing AOP-Wiki data model, and Evidence Model Prototypes (https://emod.aopwiki.org/prototypes), where biomedical annotations could be included was presented (Hench et al., 2024). Needs were assessed according to breakout group discussions on how the existing models could be improved. Multiple, new AOP-Wiki components that could provide utility and improve interoperability for the AOP and biomedical communities were discussed. Specifically, how strategic use of Event Components (ECs), as added according to Ives et al., 2017, could be modified to improve biomedical annotation within the AOP-Wiki to support interoperability between the AOP-Wiki and other knowledge resources continues to evolve (Hench, 2025). It is not clear whether everything can fit into the proposed model, which underlines the breadth of the community. How standardized output format for biomedical entities is incorporated into existing data models and joined with tools and methods in the future, will determine how selections can be incorporated into OECD Harmonized Templates (OHT) framework for ease of access by the regulatory communities across the globe.

Further, data structure improvements in the form of integration with existing Linked Open Data (e.g., GO-CAMs, Biolink Model, LinkML) and biomedical KG models and methods (ROBOKOP/RADisH; CTD) were seen to have the potential to improve semantic clarity and interoperability as well as increase the level of granularity of events, helping to represent complex disease more accurately. However, no one method of integration was identified, agreed upon or preferred by the group. Supported automation and AI-assisted knowledge extraction (e.g., CurateGPT, AutoGPT) was agreed to be an area of future improvement for AOP author submission assistance as well as automated AOP extraction. Both were viewed as potentially improving the quality, accuracy, and reproducibility of AOP data capture. The scope of biomedical data and tool integration with AOP data was not defined in these proceedings, however, it was overwhelmingly agreed that these changes would enable better mapping and evidence tracing from data sources, and improve modeling of mechanistic relationships, supporting regulatory confidence in AOPs.

Establishing causality was raised as a main challenge for regulatory use of AOPs, where AOP reporting standards need to describe more than a general association. Reporting templates were seen as one possible way to address more discrete mechanistic reporting. Some criticism of the AOP Handbook as being descriptive rather than quantitative underlined the need for more structured data for certain user groups. Prospective studies that could test the causality framework of individual AOPs were suggested, again underlining the need for leveraging quantitative data to establish biological causality.

2.2.2 Knowledgebases, repositories and other data resources that support AOP development

Many biomedical knowledge bases and data repositories were discussed in the breakout sessions and found to be central to AOP development and mechanistic modeling . Resources were in general use-specific, where participants’ preferences indicated desired integration with species-specific, mechanistic pathways, omics, phenotype contents, among others and, in relation to AOP development and mechanistic modeling. Cross-species extrapolation was described to rely currently on molecular-level information obtained from repositories in order to annotate and interpret the results of omics analyses, in lieu of resources specific to a particular species of interest. Integration with ontological information is one area that could encompass many of the data specific needs of the AOP community, where A Simple Standard for Sharing Ontology Mappings (SSSOM) was seen to be an excellent solution for future iterations of AOP-Wiki. As an example, the PINK EU project (https://pink-project.eu/) was described, where biomedical knowledge bases and data repositories are central to AOP development and mechanistic modeling. The PINK project integrates toxicogenomics data, biokinetic modeling, and computational tools to refine and support AOP construction, especially for DNT and metabolic disorders. Specific uses and resources include ToxCast for chemical testing data, AOP-Wiki, Gene Ontology (GO), ChEBI, Cell Ontology, and Uberon for standardized biological info, and biokinetic modeling platforms- like PKSim and IndiMeSH. With the PINK project example, as well as individual data sources, a clear need was observed in the prerequisite need for standardized, persistent identifiers as a minimum requirement for any inclusion of new data to the AOP-Wiki. Description of structured identifiers for KEs and stressors using standard ontologies was also indicated as a continued need.

2.2.3 Reporting templates relevant to AOP inclusion in regulatory decision making

In support of FAIR principles for AOPs, and the desire of communities to support and improve the regulatory process with regard to human health and ecological outcomes with existing and publicly available data, there is a need to align current data workflows and pipelines with AOPs,and report this information in standardized reporting templates. Structured formats exist for some parts of this overall workflow; however, translation is still needed to connect many aspects. In order to satisfy regulatory requirements, the AOPs should include information expected in regulatory templates. It is critical that proposed reporting templates be evaluated and merged into a single evidence structure that is representative across countries and regulatory communities. Templates from EFSA (food safety), ECHA, and FDA could be helpful because they also support the use of mechanistic data important for AOPs and decision-making. Some specific templates discussed were EU eco-guidance documents (chemical safety), and US Office of Pesticides Program (US OPP) guidelines for literature-based data for risk assessment, and the Criteria for Reporting and Evaluating Ecotoxicity Data (CRED) project, for example. It is important to ensure that regulatory guidance is timely and up-to-date, and keeping pace with current research and scientific understanding of biology and exposure science.

2.2.4 AI/LLM to assist in generating and reviewing AOPs

What AI approaches are currently used to support AOP development, what concerns exist and what QC can be applied?

Most participants are currently exploring potential use of LLMs in their AOP research. Since LLMs are probabilistic, there was concern that these models would tend to find correlations rather than the evidence of causation needed for regulation. Suggestions for the future were in the writing of AOP descriptions with generative AI, extracting mechanistic information from literature, suggesting ontology terms for KEs using CurateGPT and MechSpy, and expanding AOPs currently under development into a larger AOP network. Another suggested use area was implementing AI was in searching published literature to extract data associated with identified NAMs or specific chemicals of interest with relation to KE, adverse outcome, or disease. Auto-populating ontological information was another use of AI methods within the AOP-Wiki environment. Lastly, highlighting inconsistencies or missing data linkages to improve AOP-Wiki entries in terms of completion and consistency was viewed as a viable use of AI for AOPs. This process would not be seen to replace the need for expert human review.

Though many concerns were expressed by attendees regarding the use of AI, there was support for using AI, in general, to assist in improved accuracy of human curation of AOP information as well as lessen the number of curators needed and time-burden on individual AOP curators. Concerns for an established process of accuracy, validation and transparency for assisted curation of some types of information was underlined by discussants. QC approaches need to be developed with specific suggestions, such as cross-referencing results from several LLMs; clear citation or source tracking for AI-suggested content; use of evidence codes or confidence scores; expert review before publication and quality confirmation; version control; and transparency flags for human-curated vs. AI-generated information.