The core result
An AI-based antibody-discovery system reported a roughly 55% neutralization hit rate when 110 selected candidates were tested across its SARS-CoV-2 and influenza programs. Twelve candidates, or 11%, met the study's measure of broadly neutralizing activity, and the paper reports a higher yield than another neutralization predictor run with the same cloning budget. Here, broad neutralization means activity across multiple tested strains or variants, not protection against every virus.
The system, called ImmuneAgent, is described as a lab-in-the-loop continual-evolution system. It centers on Meta-BCR meta-learning and uses mean-teacher and LoRA fine-tuning, an approach for adapting a model from new feedback. The design links computational ranking with experimental testing so laboratory results can feed into later selections.
On the reported benchmark comparisons, ImmuneAgent scored 20% with DeepSeek-R1, 14% with Qwen3-max and 24.5% with GLM4.5. Its mean F1 score was 0.78, compared with 0.43 for Biomni and 0.10 for Virtual Lab. F1 is a single score that balances how many positive calls are correct with how many positives the system finds. The supplied analysis gives no confidence intervals for these figures.
External tests were tougher before the model received feedback: pre-adaptation Meta-BCR F1 fell below 0.60 on unseen data. After LoRA-based updating, the reported scores were 0.71 F1 and 0.89 AUC-ROC on SARS-CoV-2, and 0.85 F1 and 0.87 AUC-ROC on influenza, described as an 11% to 25% cumulative improvement. AUC-ROC is a ranking measure, so these results describe how well the system separated candidates in the test sets, not clinical effectiveness.
From model scores to laboratory tests
In the SARS-CoV-2 laboratory validation, 60 candidates were cloned. Thirty-nine bound spike trimers at the study's reported EC50 threshold of below 10 micrograms per milliliter, and six neutralized both the XBB.1.5 and JN.1 variants. The EC50 here is the paper's binding measurement, while micrograms per milliliter is the concentration unit. This was an in-vitro assay outcome, not a test in people.
For influenza, 20 of 33 tested antibodies neutralized at least one strain. Nine neutralized all four H1N1 strains tested, and six showed H5N1 neutralization. The study measured influenza neutralization with a focus-forming assay and estimated IC50 values using four-parameter nonlinear regression in GraphPad Prism 9.0. Those proportions describe the selected antibodies that were tested, not the frequency of broadly neutralizing antibodies in an unselected human repertoire.
The animal test reported a strong prophylactic signal in female BALB/c mice. Animals received 10 mg/kg of antibody before a challenge dose equal to five times the study's reported LD50, then were monitored for 14 days. All five tested ImmuneAgent antibodies and the MEDI8852 control had 100% survival. The D25 control reached its humane endpoint by day 6, while the PBS control had no survivors by day 7. The experiment was not powered to establish equivalence with MEDI8852, and the report did not give animal-allocation randomization procedures.
The clues behind breadth
The researchers also looked for biological features associated with breadth. Across seven public datasets, FCRL5 was the only marker consistently enriched in cells predicted to be broadly reactive. In a cloned-antibody comparison, antibodies from FCRL4/5-positive CD27-positive cells, 48 in total, had higher median neutralization breadth than antibodies from CD27-negative cells. This is an association: the result does not show that either cell population causes broad neutralization or that the marker will predict it prospectively.
Sequence and structure analyses pointed to a second possible signal. In reversion experiments, the reported JN.1 IC50 for BA5-12 shifted from 9.3 to 89.5 ng/mL, while BA5-57 shifted from 8.1 to 213.2 ng/mL. Across 109 cloned coronavirus and influenza heavy chains, hydrophobicity in a heavy-chain segment called CDR-H2 distinguished broad from narrow neutralization with an AUC of 0.77. The study presents these as associations and model-generated hypotheses, not a demonstrated causal mechanism.
The model was then tested against virus targets not used in the main training sets. In an RSV-hMPV analysis, 180 candidate structures were predicted and 30 antibodies nominated. Thirteen neutralized all five tested viruses, while nine antibodies from FCRL5-low CD27-positive cells cross-neutralized RSV-A, RSV-B and hMPV. None from FCRL5-negative cells neutralized hMPV. The result came from a selected antibody subset and had no randomized comparison.
For HPV, the ranking averaged five modalities with equal weight before cloning 50 antibodies. Nine neutralized at least one HPV variant, and two cross-neutralized HPV6, HPV11, HPV33 and HPV58. The authors describe this validation as exploratory, so it is evidence that the workflow produced testable candidates, not evidence that it will generalize to all HPV types.
Why the findings remain early
Several features of the study make the headline numbers provisional. Candidate yield depends on the ranking strategy and the fixed cloning budget. The training-data audit found no exact CDR-H3 match among 23 Turner-cohort sequences, but found one exact match in the Guo cohort and 54 of 286 sequences at 90% full-chain identity. The authors said that this residual similarity could inflate absolute performance for all three models.
Taken together, the study is best read as a candidate-prioritization framework and a set of hypotheses about where broad antibodies may be found. It does not establish human efficacy or safety, prove that FCRL5-positive CD27-positive cells or hydrophobic CDR-H2 contacts cause breadth, or show that the method will work for every unseen virus or cohort. Independent prospective tests across other cohorts, pathogens, laboratories and cloning budgets would be needed to answer those questions.
Paper data and sources
Original title: Multimodal reasoning for broadly neutralizing antibody discovery from label-free human B cell repertoires across virus families
Authors: Hantao Lou, Jianqing Zheng, Can Yue et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-10-02
DOI: Not available
Original paper · Full text