TL;DR: A closed-loop AI called ImmuneAgent found broadly neutralizing antibodies across virus families from label-free human B cell repertoires.
What happened: On 2 Oct 2026, Hantao Lou, Jianqing Zheng, Can Yue and colleagues posted arXiv:2610.03160 describing ImmuneAgent, a closed-loop AI system that integrates multimodal reasoning with continual meta-learning and wet-lab feedback to screen natural BCR repertoires from vaccinated or infected cohorts.
The headline result reported in the abstract is a ~55% neutralization antibody discovery rate, or 60 of 110 cloned candidates, and a ~11% bnAb yield, or 12 of 110, substantially outperforming a state-of-the-art sequence-based neutralization predictor or cofolding models evaluated at the same cloning budget, according to Multimodal reasoning for broadly neutralizing antibody discovery from label-free human B cell repertoires across virus families.
The problem the team set out to solve
Why broadly neutralizing antibodies are hard to find
Discovering broadly neutralizing antibodies, abbreviated bnAbs, from human natural immune repertoires remains a fundamental challenge in immunology, as stated in the 2 Oct 2026 abstract.
The authors list three barriers that hinder that discovery task, as summarized on the arXiv abstract page.
The first barrier is the extreme rarity of bnAb in natural repertoires.
The second barrier is incomplete understanding of their cellular origins across pathogens.
The third barrier is the inability of existing computational tools to generalize across emerging viral threats.
That framing matters because the same scarcity theme appears in independent influenza antibody work published days earlier, according to Epistasis and pleiotropy constrain the evolution of cross-reactive breadth in a broadly neutralizing influenza antibody.
In that PLOS Biology study, published 28 Sept 2026 by Katrine E. Dailey, Yiquan Wang, Qi Wen Teo and colleagues at the University of Illinois Urbana-Champaign and University of Florida, the highly conserved influenza virus hemagglutinin HA stem domain is described as a major target for bnAbs.
The PLOS Biology authors report that the IGHV1-69-encoded CR9114 remains the only HA stem bnAb that cross-reacts with both influenza A and B viruses, despite being discovered more than a decade ago.
The PLOS Biology team performed four deep mutational scanning experiments to compare binding affinity landscapes of germline and somatic CR9114 against H1 HA, H3 HA and influenza B virus HA.
Many mutations that minimally affected or even improved H1 HA binding were detrimental for binding to H3 HA and influenza B virus HA in that dataset.
The PLOS Biology authors further report prevalence of epistasis in IGHV1-69 HA stem bnAbs and present the work as a mechanistic explanation for scarcity of HA stem bnAbs with cross-reactivity against both influenza A and B viruses.
Together, the 28 Sept 2026 PLOS Biology paper and the 2 Oct 2026 arXiv preprint describe the same practical bottleneck from different angles: potent cross-reactive antibodies exist, but they are rare and difficult to evolve, isolate and generalize.
What ImmuneAgent is
A closed-loop system linking computation and wet-lab feedback
The preprint introduces ImmuneAgent as a closed-loop AI system.
The system integrates multimodal reasoning with continual meta-learning and wet-lab feedback, in the wording of the abstract.
The listing places the work across Quantitative Methods, Artificial Intelligence, Computational Engineering Finance and Science, and Cell Behavior, with subject codes q-bio.QM, cs.AI, cs.CE and q-bio.CB.
The author list comprises Hantao Lou, Jianqing Zheng, Can Yue, Meihan Zhang, Yuanchao Bao, Yu Chen, Mengting Huang, Yupeng Yang, Qianyu Pan, Nana Fu, Yansong Shi, Hongli Li, Yangyang Chai, Ruyi Chen, Wansheng Li, Zhu Liang, Rongmei Yao, Yuanhan Mo, Lei Wang, Chunmei Wang, Yun Quan, Qiong Zhang, Xiangxi Wang and Xuetao Cao.
The submission history shows version v1 posted Fri, 2 Oct 2026 at 11:35:34 UTC, with arXiv identifier 2610.03160v1 and DOI link https://doi.org/10.48550/arXiv.2610.03160.
The system was applied to screen natural BCR repertoires from vaccinated or infected cohorts, rather than to libraries constructed around a single pre-defined antigen workflow.
The abstract emphasizes label-free human B cell repertoires in the title, indicating that the input repertoires were not pre-labeled for a single target in the discovery approach described.
The closed-loop description indicates that computational selection and wet-lab feedback operate together inside ImmuneAgent.
The continual meta-learning component is presented as part of how the system overcomes barriers to generalization across emerging viral threats.
The multimodal reasoning component is presented as part of how the system integrates different signals relevant to bnAb activity.
How well it worked at the reported cloning budget
60 of 110 neutralizers and 12 of 110 broadly neutralizing antibodies
The preprint reports outcomes at a fixed experimental scale of 110 cloned candidates.
Within that 110-candidate set, 60 candidates showed neutralization antibody activity, corresponding to the reported ~55% neutralization antibody discovery rate.
Within the same 110-candidate set, 12 candidates qualified as bnAbs, corresponding to the reported ~11% bnAb yield.
Those two fractions — 60 of 110 and 12 of 110 — are the central performance figures given in the abstract for screening of natural BCR repertoires from vaccinated or infected cohorts.
The authors state that the result substantially outperformed a state-of-the-art sequence-based neutralization predictor evaluated at the same cloning budget.
The authors also state that the result substantially outperformed cofolding models evaluated at the same cloning budget.
The comparison at the same cloning budget is explicitly framed as a like-for-like assessment of discovery efficiency, not as an unbounded screen.
In practical terms, the reported numbers mean more than half of the cloned selections neutralized in the assays used, and more than one in ten met the broader neutralization criterion applied by the team.
The rarity context makes those fractions notable: where bnAbs are described as extremely rare in natural repertoires, a 12-of-110 yield represents concentrated recovery from label-free repertoires.
The vaccinated-or-infected cohort detail matters because the starting material was natural human immune repertoires shaped by vaccination or infection, not synthetic libraries.
Why the same-budget comparison matters
The abstract explicitly anchors the comparison to the same 110-candidate cloning budget.
That design controls for wet-lab effort across ImmuneAgent, the sequence-based neutralization predictor, and the cofolding models.
The state-of-the-art sequence-based neutralization predictor is identified as one baseline.
Cofolding models are identified as the other baseline category.
The preprint states that ImmuneAgent substantially outperformed both baselines under that matched-budget condition.
No additional baseline scores, model names, or numerical margins beyond that superiority statement are provided in the abstract text available on the arXiv abstract page.
Protection test in influenza challenge
Five antibodies with 100% in vivo protection
Beyond in vitro neutralization, the preprint reports an in vivo protection result against lethal influenza challenge.
Five ImmuneAgent-discovered antibodies conferred 100% in vivo protection against lethal influenza challenge in the experiments described.
The abstract compares that protection outcome to the clinical-stage therapeutic MEDI8852, describing the five-antibody result as comparable.
MEDI8852 is therefore the named clinical-stage reference therapeutic for the influenza protection comparison.
The 100% figure applies to the five antibodies singled out for the lethal challenge test, not to all 60 neutralizers or all 12 bnAbs.
The lethal influenza challenge wording indicates a stringent in vivo test in which untreated controls would not be expected to survive.
The result connects computational selection from vaccinated or infected BCR repertoires to survival outcome in a live-virus influenza model.
The comparison to a clinical-stage therapeutic provides a benchmark for the magnitude of protection observed, without stating that the new antibodies are approved therapeutics.
Cellular clue recovered by the system
FCRL5+CD27+ atypical memory B cells as a conserved reservoir
The preprint states that the system recovered the cellular and structural determinants of bnAb activity.
One cellular determinant named is FCRL5+CD27+ atypical memory B cells.
Those cells are identified as a conserved bnAb reservoir in the abstract wording.
The conserved-reservoir language indicates that the association was observed across more than one pathogen context examined by the team.
The incomplete understanding of cellular origins across pathogens was one of the three starting barriers, so recovery of a cellular determinant directly addresses that gap as framed by the authors.
The FCRL5 marker and CD27 marker together define the atypical memory B cell subset highlighted in the findings.
The atypical memory B cell label distinguishes this subset from classical memory B cell populations discussed in B cell literature.
The finding reframes where to look in label-free repertoires when seeking bnAbs: toward the FCRL5+CD27+ atypical memory compartment identified by ImmuneAgent.
Structural clue recovered by the system
Hydrophobic interface enrichment as a cross-viral signature
The second determinant named is hydrophobic interface enrichment.
That enrichment is described as a cross-viral structural signature.
The cross-viral wording indicates relevance beyond a single virus family in the datasets analyzed.
The structural-determinant language parallels the cellular-determinant result, showing that ImmuneAgent recovered both levels of explanation for bnAb activity.
Hydrophobic interface enrichment points to the binding interface chemistry rather than to any single sequence motif named in the abstract.
The abstract presents this signature alongside the FCRL5+CD27+ reservoir finding as recovered determinants, not as pre-imposed filters.
The inability of existing computational tools to generalize across emerging viral threats was the third starting barrier, so a cross-viral structural signature directly addresses that generalization problem as framed by the authors.
How the two determinants fit together
The cellular result identifies which B cells tend to harbor bnAb activity.
The structural result identifies what binding-interface property tends to mark cross-viral activity.
Together, they provide a cell-plus-structure explanation recovered from the same closed-loop workflow that produced the 60-of-110 and 12-of-110 yields.
The abstract links both determinants to generalization to unseen antigens, which is tested in the human metapneumovirus and human papillomavirus work described below.
Generalization to unseen antigens
hMPV cross-neutralizing and HPV-neutralizing antibodies without antigen-specific sorting
The preprint states that the learned determinants generalized to unseen antigens.
The first generalization case involved human metapneumovirus, abbreviated hMPV.
For hMPV, the system discovered cross-neutralizing antibodies.
The second generalization case involved human papillomavirus, abbreviated HPV.
For HPV, the system discovered neutralizing antibodies.
Both discoveries were made without antigen-specific sorting, in the wording of the abstract.
The without-antigen-specific-sorting detail is central because it means the hMPV and HPV antibodies were recovered from repertoires without a target-specific pre-enrichment step for those antigens.
The unseen-antigen wording indicates that these viral targets were not part of the original training or selection context that produced the influenza-focused determinants.
The hMPV result demonstrates cross-neutralization within that generalization test.
The HPV result extends neutralization discovery to a distinct virus family from respiratory RNA viruses.
Across virus families, in the title wording, therefore reflects demonstrated activity spanning influenza, hMPV and HPV contexts reported in the abstract.
These results are presented as validation that ImmuneAgent is a generalizable framework for rapid therapeutic antibody discovery against emerging viral threats.
What this means for rapid therapeutic discovery
A generalizable framework claim tied to specific reported outputs
The authors validate ImmuneAgent as a generalizable framework for rapid therapeutic antibody discovery against emerging viral threats, in the abstract's concluding statement.
That framework claim rests on the specific outputs listed together: the 60-of-110 neutralization rate, the 12-of-110 bnAb yield, the five antibodies with 100% influenza protection comparable to MEDI8852, the FCRL5+CD27+ reservoir, the hydrophobic interface signature, and the hMPV and HPV antibodies found without antigen-specific sorting.
The vaccinated-or-infected cohort source material supports the natural-repertoire relevance of the framework.
The matched-budget outperformance over the sequence-based predictor and cofolding models supports the efficiency claim within the tested scale.
The influenza in vivo result supports the functional relevance beyond binding or in vitro neutralization alone.
The hMPV and HPV results support the cross-viral generalization claim beyond the initial antigen context.
The work was posted as arXiv:2610.03160v1 on 2 Oct 2026, so the reported findings reflect the preprint record available on that date.
Readers tracking antibody discovery should watch for the full manuscript, peer review, sequence deposits, assay protocols and independent replication that would normally follow an arXiv posting in q-bio.QM, cs.AI, cs.CE and q-bio.CB.