MedDiscovery: A Multi-Agent Scientific Discovery System for Cross-Domain Therapeutic Hypothesis Generation

Background: Current AI-assisted drug discovery systems predominantly operate within single domains, limiting their ability to identify novel therapeutic mechanisms through cross-disciplinary insights. We present MedDiscovery, a multi-agent scientific discovery system that generates testable therapeutic hypotheses by synthesizing knowledge across disparate scientific fields. Methods: MedDiscovery employs a tri-core architecture combining BioBERT (knowledge processing), TinyLlama (reasoning/consciousness), and DeepSeek (articulation) with autonomous agent generation, episodic memory, and cross-domain analogical reasoning. We validated the system across six diverse medical conditions: Alzheimer's disease, chronic inflammatory demyelinating polyneuropathy (CIDP), glioblastoma multiforme (GBM), Type 1 diabetes (T1D), Tourette syndrome, and idiopathic pulmonary fibrosis (IPF). Results: Across all cases, MedDiscovery generated novel therapeutic hypotheses with mean novelty score of 0.73 (range: 0.45-0.78), synthesizing evidence from 682 sources per case on average (total: 4,122 sources). The system achieved 4 Tier-1 (high novelty, strong evidence) and 2 Tier-2 (moderate novelty or evidence) classifications. Each hypothesis incorporated cross-domain reasoning from 3-4 scientific fields and included SPIRIT-compliant experimental protocols with explicit rejection criteria. Conclusions: MedDiscovery demonstrates that autonomous multi-agent systems can generate scientifically rigorous, experimentally testable therapeutic hypotheses by leveraging cross-domain analogical reasoning. The system's ability to identify non-obvious mechanistic connections across fields suggests a new paradigm for computational drug discovery that complements rather than replaces human scientific intuition.