MechAInistic: A Reviewer-Supervised Multi-Agent LLM System for Auditable Mechanistic Drug-Hypothesis Generation

LLM agents are increasingly used for scientific reasoning, but their fluent-sounding outputs can diverge from verifiable computational evidence, limiting their reliability for biomedical hypothesis generation. We developed MechAInistic, a multi-agent system in which an independently configured Reviewer agent supervises a planning Architect agent at each stage of the workflow, with all reasoning grounded in executable mechanistic-model analyses rather than language-model text alone. The Reviewer scores plans and intermediate results against pre-specified rubrics and triggers re-planning or re-execution when scores fall below threshold, producing an auditable chain from a natural-language question to model-derived evidence and cited literature. We instantiate the system over paired constraint-based metabolic models using COBRApy, supporting pathway comparison, perturbation analysis, drug-target exploration, and literature interpretation across healthy and disease states. We evaluated MechAInistic on two immune-cell therapeutic hypothesis-generation tasks. For rheumatoid arthritis versus healthy naive B-cell models, it identified mitochondrial metabolic rewiring and nominated Devimistat/CPI-613 as an investigational OGDH-centered hypothesis. For multiple sclerosis CD4+ Th17 versus healthy models, it identified NADP-dependent isocitrate dehydrogenase as a candidate target and proposed ivosidenib, with vorasidenib as a mechanistically complementary alternative. Comparator analyses against general-purpose LLM systems showed that plausible biological narratives can lack auditable model grounding, whereas MechAInistic preserves the computational reasoning path from prompt to result.