Agentic Active Causal Discovery for Alzheimer's Disease Reversal: Closing the Genomic Experimental Loop

Designing robust interventions for complex genomic phenotypes like Alzheimer's Disease (AD) requires moving beyond static association to agentic systems capable of causal reasoning and experimental planning. We introduce an Agentic Active Causal Discovery (AACD) framework that “closes the loop” between in-silico hypothesis generation and wet-lab experimental design. By integrating a Multiscale Encoder with a Causal Graph of Thoughts (C-GoT) reasoning engine, our agent autonomously navigates the vast search space of metabolic pathways, iteratively refining its causal graph through “Specification-Gating”—a mechanism that rejects low-confidence interventions before they reach the validation stage. This approach directly addresses the challenge of translating high-dimensional genomic data into actionable, prioritized experimental protocols. To validate this agentic framework, we deploy it on the task of reversing AD pathology, specifically targeting the dysregulated kynurenine pathway. The agent successfully rediscovers critical therapeutic nodes, identifying IDO1 and TDO2 inhibition as high-value interventions for restoring metabolic homeostasis and reversing cognitive decline. We demonstrate that our Specification-Gated policy outperforms standard associative baselines, generating “Decision Cards” that serve as interpretable, biologically grounded blueprints for future in-vivo validation. These results establish AACD as a powerful paradigm for automating the discovery of mechanistic interventions in genomics, bridging the gap between computational reasoning and biological experimentation.