Adverse Outcome Pathways (AOPs) are central to next-generation chemical hazard assessment. They map how a molecular interaction triggered by a chemical can lead, step by step, to an adverse outcome in a cascading manner. AOPs provide a shared framework through which scientists and regulators identify key biological events, clarify evidence gaps, and determine what data are needed. They also support the prediction and prioritisation of data-poor chemicals, a crucial capability at a time when thousands of substances require assessment while animal testing must be reduced, refined, and replaced.

However, building an AOP is demanding. It requires rigorous curation of information fragmented across the literature and databases, followed by critical analysis by experts to evaluate the evidence and organise it into a structured map of key events within the AOP. This is where AI can be particularly helpful assisting in the curation and synthesis of information from literature sources and presenting it to researchers to support critical decision-making, effectively serving as a co-pilot in the design of AOPs.

From months of manual review to AI-supported information extraction

Traditionally, AOP development starts from an observed adverse outcome or a suspected key event. This triggers a systematic and often extensive literature review to identify and organise mechanistic evidence. Even with established guidelines, the process can take up to a team of experts up to a year, depending on the complexity of the topic and the volume of available literature.

Prioritising, reviewing, extracting, and structuring this information is time-consuming and may be prone to inconsistencies simply due to the scale of the task and the inherent biases of human judgement.

To address this challenge, PARC scientists have developed an AI-based online tool that supports AOP development through structured information extraction. The approach, recently published by Vikas and colleagues, demonstrates that AI-assisted, pattern-based information extraction can be up to 99% faster than conventional workflows used for AOP literature evidence curation, while maintaining transparency and precision above 95%.

Rather than replacing expert judgement, the tool assists researchers by rapidly identifying relevant publications and evidence sentences, extracting mechanistic relationships, and uncovering the "unknown from the known." For example, given the query "Identify chemicals leading to steatosis," the tool can efficiently identify all chemicals mentioned in association with steatosis in the literature using pattern-based extraction. This can further reveal the genes linked to steatosis, thereby enabling researchers to formulate hypotheses grounded in literature evidence.

Strengthening mechanism-based hazard assessment

AOPs are the backbone of many new approach methodologies (NAMs). By anchoring in vitro and in silico methods to clearly defined biological pathways, they help ensure that alternative methods are mechanistically meaningful and regulatory-relevant.

Currently, at least six PARC tasks are applying and benchmarking the AI tool against traditional AOP development approaches. The goal is to evaluate how AI-supported workflows can improve efficiency, transparency, and reproducibility in hazard assessment.

In areas where the literature is overwhelming or conversely, fragmented and difficult to navigate artificial intelligence offers powerful support to experts working toward faster, mechanism-based, and increasingly animal-free chemical assessment.

Researchers interested in exploring the AI-based AOP development platform or collaborating with the team are encouraged to get in touch to test and adapt the tool to suit their specific needs.