The Persistent Agent: Orchestrating AI, Knowledge Graphs, and Predictive Toxicology for a Safer World

The Persistent Agent framework unifies large language models, predictive tools, and interoperable knowledge graphs to automate toxicological reasoning. Built around ToxIndex, it integrates retrieval-augmented prediction with the BioBricks-OKG data registry, enabling agents to collect evidence, run models, and generate structured hazard assessments. This system links chemical structure, mechanistic pathways, and literature evidence across sources, turning fragmented data into coherent, interpretable predictions. Benchmarks show strong performance across major toxicity datasets (AUC > 0.9) and substantial gains from retrieval-augmented prediction for endpoints such as LC₅₀ inhalation toxicity (R² = 0.74). Case studies, including PFAS hepatotoxicity and neurotoxicity analysis, demonstrate how persistent, agentic workflows can support hazard surveillance and safer material design. Together, these components illustrate a scalable approach to predictive toxicology: continuously updated, interpretable, and aligned with the goal of a safer world.