An AI agent just got unusually good at finding rare virus-fighting antibodies in human blood.
Researchers built ImmuneAgent, a closed-loop AI system that combines multimodal reasoning, continual meta-learning, and wet-lab feedback to screen natural antibody repertoires from vaccinated or infected patients. Out of 110 cloned candidates, it found 60 that neutralized viruses - a 55% hit rate - and 12 broadly neutralizing antibodies, an 11% yield. That beat sequence-based prediction tools and cofolding models tested with the same cloning budget. Five of the discovered antibodies gave mice complete protection against a lethal flu challenge, matching results from MEDI8852, a clinical-stage antibody therapy.
The bigger finding is what the system learned about where these antibodies come from. It flagged FCRL5+CD27+ atypical memory B cells as a recurring source of broadly neutralizing antibodies and hydrophobic interface enrichment as a shared structural trait across viruses. That pattern held up on viruses the system wasn't trained to target, turning up cross-neutralizing antibodies against human metapneumovirus and HPV-neutralizing antibodies without needing antigen-specific sorting first. If that generalization holds, it points toward antibody discovery that moves faster than the traditional one-virus-at-a-time screening pipeline.
Worth noting: this is a 110-candidate study with mouse data, not human trials, and a generalizable framework claim still rests on one research team's cloning budget. The comparison to MEDI8852 is encouraging, not confirmation this ships as a drug.