Knowledge Graph‑Enabled Adverse Outcome Pathways for Environmental Risk Assessment of Chemicals: A Critical Review

Over 350,000 synthetic chemicals are in global use, yet adverse outcome pathway (AOP) frameworks remain largely qualitative, static, and disconnected from real-world exposure scenarios. The widespread adoption of knowledge graphs (KGs) and artificial intelligence (AI) has raised expectations that more data will automatically resolve these limitations. This review challenges that assumption, arguing that uncritical data accumulation, without explicit causal alignment (directionality, temporality, intervention, and provenance), exacerbates the three classic AOP bottlenecks, including data fragmentation, quantitative gaps, and exposure–mechanism disconnect. Through a causal lens, we demonstrate that high predictive accuracy in correlation tasks often masks poor causal directionality, and that AI-generated mechanistic hallucinations are systemic rather than occasional. To address these issues, we propose the EE‑ID‑BR (External Exposure–Internal Dose–Biological Response) framework, which directly embeds physiologically based pharmacokinetic (PBPK) models and dose–response functions into KG edges, transforming static AOP diagrams into quantitatively predictive networks. We further articulate a dual-engine architecture coupling exposure-to-dose bridging with topology-enhanced causal reasoning, and provide a tiered weight-of-evidence framework for regulatory integration. Future directions include causal benchmarking datasets, minimal causal provenance standards, and AOP-specific hallucination detection. This review offers a roadmap for transitioning AOP-based environmental risk assessment from data-driven correlation toward causality-driven, regulation-ready decision support.