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

Area of Science:

• Computational biology
• Systems biology
• Artificial intelligence in medicine

Background:

• Constraint-based metabolic modeling offers mechanistic insights into cellular states and diseases.
• Effective metabolic modeling requires significant computational expertise and multi-step analysis coordination.
• Existing methods present a high barrier for researchers without specialized computational skills.

Purpose of the Study:

• To develop MechAInistic, an AI-powered system to lower the barrier for complex biological question-answering using metabolic models.
• To enable researchers to query metabolic models using natural language.
• To automate the generation of executable workflows and structured reports from natural language queries.

Main Methods:

• MechAInistic is a multi-agent system utilizing large language models (LLMs) with an Architect-Reviewer pattern.
• It converts natural language biological questions into executable, model-grounded workflows.
• The system supports pathway comparison, perturbation analysis, drug-target exploration, and literature interpretation for paired healthy and disease states.

Main Results:

• MechAInistic successfully generated therapeutic hypotheses in two immune-cell use-cases.
• For rheumatoid arthritis (Naive B cells), it identified mitochondrial metabolic rewiring and proposed Devimistat/CPI-613 targeting OGDH.
• For multiple sclerosis (CD4+ Th17 cells), it identified NADP-dependent isocitrate dehydrogenase as a target and suggested Ivosidenib for drug repurposing.

Conclusions:

• MechAInistic effectively democratizes constraint-based metabolic modeling by enabling natural language querying.
• The system demonstrates potential for therapeutic hypothesis generation in complex diseases.
• MechAInistic facilitates the exploration of cellular metabolism for drug discovery and repurposing.