| Abstract | Purpose: Artificial intelligence (Ai) and Machine Learning (ML) are being explored to improve systematic evidence gathering and to identify patterns across datasets. Their integration into the development of radiation Adverse Outcome Pathways (AOPs) offers an opportunity to accelerate data consolidation in radiation protection. AOPs provide a structured, transparent framework that links molecular-level perturbations to adverse outcomes relevant to risk assessment. Despite their value, AOP development is hindered by manual evidence mapping, the complexity of multi-level biological responses, and fragmented data across platforms, experimental models, and epidemiological studies. Herein, we explore the role of Ai/ML in overcoming these challenges by enabling extraction, annotation, and integration of heterogeneous data sources. Ai assist in identifying Key events (Kes), inferring Key event Relationships (KeRs), and suggesting putative AOP structures by mining scientific literature and experimental datasets. we propose an Ai-driven AOP development plan that includes: (1) establishing curated, open-access training datasets annotated with AOP-relevant biological and exposure entities; (2) applying domain-specific natural language processing techniques to extract mechanistic insights from unstructured literature; (3) deploying supervised and unsupervised ML methods to identify and prioritize Kes; (4) constructing transparent causal models using knowledge graphs and probabilistic inference to capture mechanistic directionality; (5) enabling automated narrative generation and evidence scoring; and (6) integrating iterative expert feedback and new data for continuous model refinement. This phased approach bridges data readiness, computational modeling, and domain expertise to advance the integration of Ai/ML into AOP development. Challenges such as model interpretability, data sparsity for low-dose radiation effects, ethical considerations, hallucination in large language models and validation of Ai-inferred pathways are discussed.
Conclusions: while fully Ai-assisted radiation AOPs remain conceptual, the review provides a methodological foundation for their future development. Ai/ML offers a means to accelerate radiation AOP development, facilitating systematic organization, integration, and prioritization of biological and experimental data. |
|---|
