From neuroscience to oncology: Multi-agent AI replicating neural reasoning architecture for evidence-based colon cancer clinical decisions.

17

Background: Colorectal cancer management requires integration of complex genomic data, clinical guidelines, and patient-specific factors. Current clinical decision support systems often lack transparency and fail to ensure guideline adherence. We developed GenomAI, a multi-agent AI reasoning system that replicates prefrontal cortex decision-making to provide verifiable, guideline-based treatment recommendations. The platform uses a hybrid Cache-Augmented Generation and Retrieval-Augmented Generation architecture. Methods: GenomAI employs specialized multi-modal AI agents for diagnosis, biomarker analysis, treatment planning, and clinical trial matching. The system integrates NCCN guidelines with patient genetic profiles, pathology reports, clinical data, and imaging through HL7/FHIR-compliant APIs or document upload. We validated GenomAI using 100 retrospective colorectal cancer cases, evaluating: (1) accuracy of NCCN guideline identification and application, (2) appropriateness of clarifying questions when clinical information was incomplete, (3) concordance with multidisciplinary tumor board decisions, and (4) transparency through complete audit trails with exact NCCN page references. Performance was compared against direct-answer AI models. Results: GenomAI achieved 100% accuracy in identifying relevant NCCN guideline sections and 99% concordance with tumor board decisions. The system appropriately requested follow-up data in 14 cases with insufficient information. Compared to Claude Opus 4.5 and GPT-5.2 (85%), Gemini 3 Pro (82%), and MedGemma 1.5 (62%), GenomAI demonstrated superior performance. The platform integrated comprehensive clinical data: laboratory values, molecular profiling (KRAS, NRAS, BRAF, MSI, TMB), pathology, imaging, and colonoscopy while incorporating patient-specific factors including toxicity risks, comorbidities, and age. All oncologist-validated recommendations included NCCN page references and confidence scores. GenomAI reduced evidence synthesis time by 40% versus manual review and successfully identified eligible clinical trials. Conclusions: GenomAI demonstrates high accuracy in providing NCCN guideline-compliant colorectal cancer treatment recommendations. The multi-agent architecture ensures transparency, appropriate clarification-seeking, and refusal when guidelines are not applicable. This validation establishes GenomAI as a promising clinical decision support tool enhancing physician workflow while maintaining guideline fidelity and patient safety. Future work will expand to additional cancer types and prospective evaluation.