DDIAgents Predicts Drug Interactions with Mechanism-Conditioned Context.

Researchers have developed DDIAgents, a novel multi-agent framework designed to enhance the accuracy and interpretability of drug-drug interaction (DDI) prediction. Predicting DDIs is crucial for medication safety but requires complex reasoning over diverse biomedical evidence, whose relevance varies depending on the interaction mechanism. DDIAgents addresses this by employing a mechanism-conditioned context flow. The framework operates with a planner agent that, given a drug pair, instantiates specialized expert agents. It then intelligently routes only mechanism-relevant knowledge sources to each expert agent, thereby reducing noise from irrelevant information. A conclusion agent aggregates the analyses from these experts. This dynamic knowledge orchestration supports complementary reasoning and generates interpretable, agent-level rationales for predictions. Extensive experiments on realistic DDI benchmarks show that DDIAgents consistently outperforms various existing methods, including feature-based, graph-based, LLM-based, and other agent-based baselines, demonstrating its effectiveness in organizing heterogeneous scientific knowledge for adaptive AI4Science reasoning.