MedDiscovery: Toward Systematic Discovery of Cross-Domain Breakthroughs in Biomedicine

BackgroundThe accelerating growth of biomedical literature and data makes it increasingly difficult for human researchers to detect cross-domain therapeutic opportunities, particularly in oncology. Most current AI tools for drug discovery act as opaque predictors or simple API wrappers, offering limited transparency about underlying evidence and mechanistic rationale. MedDiscovery is a multi-agent, evidence-first module within the MedResearch AI platform, designed to generate traceable therapeutic hypotheses by combining large-scale literature mining, cross-domain knowledge integration, and simulation. Here, we present a technical case study in HER2-positive, oestrogen receptor–negative (HER2⁺/ER–) breast cancer. MethodsMedDiscovery orchestrates a 10-stage multi-agent pipeline (Ideation, Visioner, ConceptLearner, EvidenceMiner, SuperBioBERTAgent, ReasoningSystem, CrossDomainMapper, Synthesizer, SimulationAgent, EthicsValidator). The system ingests and structures evidence from PubMed, arXiv, KEGG, ChEMBL, and ClinicalTrials.gov, using a BioBERT-based model (SuperBioBERT) for entity extraction, semantic clustering, and causal graph construction. For the query “What is a novel neoadjuvant breast cancer therapy for HER2⁺ and ER–?”, the platform generated candidate strategies, ranked them on evidence strength and novelty, and explored translational feasibility using synthetic neoadjuvant response datasets. The underlying MedResearch-BioBERT model is iteratively fine-tuned after approximately every 100 eligible queries from the system’s evidence pool. FindingsFor this scenario, MedDiscovery proposed a hypothesis in which HER2-targeted piezoelectric nanoparticles, activated by focused ultrasound, induce mitochondrial calcium overload, trigger immunogenic cell death, and enhance the efficacy of anti-HER2 therapy in HER2⁺/ER– breast cancer. The system grounded this hypothesis in 265 literature sources (112 high-quality), identified 1045 entities, and mapped causal relations across tumour metabolism, mitochondrial signalling, immuno-oncology, nanotechnology, and materials science. Simulation modules explored parameter spaces in synthetic HER2⁺/ER– neoadjuvant datasets (including pathological complete response, objective response, and chemotherapy dose-reduction scenarios) and flagged key risks around delivery, calcium-mediated toxicity, and off-target nanoparticle accumulation. The overall translational feasibility for this concept was classified as “amber”, indicating promising but clearly exploratory potential. InterpretationThis case study illustrates how a multi-agent, evidence-traceable AI system can generate mechanistically coherent and cross-domain therapeutic hypotheses in oncology, while preserving full visibility of supporting literature, causal assumptions, and ethical concerns. The proposed HER2-targeted piezoelectric nanoparticle strategy is not presented as a clinical recommendation, but as a structured starting point for human expert review and preclinical testing. More broadly, MedDiscovery is intended to augment, not replace, researchers by narrowing the search space, revealing non-obvious connections, and systematising “hypothesis-first” exploration. Future work should include prospective expert evaluation of AI-generated hypotheses, benchmarking against human research teams, and rigorous assessment of real-world impact on discovery pipelines. FundingThis work received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors. Development of the MedResearch AI platform and the MedDiscovery module has been internally supported by InfoSphere Technologies.