Moremi Bio Agent: Using Neisseria meningitidis Reference Data For The Double Blinded Validation of A General Purpose Biology-Trained Reasoning Model for Pathogen and Antigen Discovery

Abstract

Antibodies serve as vital diagnostic and therapeutic agents due to their exceptional specificity toward antigenic targets. Mapping antibody-antigen interactions is essential for understanding immune responses and developing vaccines or biologics. Traditional antigen identification relies on labor-intensive wet-lab techniques such as phage display, peptide microarrays, and ELISA, while computational methods employ sequence alignment, epitope mapping, and structure prediction. Despite progress, to our knowledge, no existing AI framework has demonstrated the ability to blindly inference--predicting an antibodys antigen target solely from its amino acid sequence without prior biological context. This research employed Moremi Bio Nano, a general agentic reasoning large language model (LLM), to infer the antigen and pathogen targets of an anonymized monoclonal antibody sequence from Imperial College London. The model received only the VH and VL chain sequences and autonomously hypothesized, ranked, and validated probable targets. Of ten independent inference tests, four were completed successfully, with three correctly identifying the experimentally validated antigen and pathogen; with SARS-CoV-2 Spike RBD, Neisseria meningitidis fHbp v1.1, and SARS-CoV Spike emerging as the top 3-ranked candidates across both reranking strategies. Validation of the models predictions with experimental wet-lab data confirmed its capacity for correct antigen inference, marking Moremi Bio Nano as a first-of-its-kind AI system demonstrating reasoning-driven antigen discover; complementing experimental immunology and advancing automated biological inference.