PaperlayerView exact revision on arXiv

Source-linked AI summary

Multimodal reasoning for broadly neutralizing antibody discovery from label-free human B cell repertoires across virus families

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, Xuetao Cao

arXiv:2610.03160v1q-bio.QMcs.AIcs.CEq-bio.CB

TL;DR

ImmuneAgent addresses the challenge of finding broadly neutralizing antibodies in natural human repertoires, where bnAbs are rare and computational predictions may not reflect neutralization. It combines multimodal reasoning with continual learning and wet-lab feedback; among cloned candidates, it reports 60 neutralizing antibodies and 12 bnAbs.

  • Problem

    Fewer than 0.01% of B cells are described as broadly neutralizing, and existing models can memorize known antibody-antigen pairs or predict confident binders that fail to neutralize.

    Source: Introduction · p. 3

  • Method

    ImmuneAgent is a closed-loop system combining multimodal reasoning, continual meta-learning, and wet-lab feedback to identify antibodies from natural BCR repertoires.

    Source: Introduction · p. 4, continual-evolution for immunological… · p. 6, Discussion · p. 31

  • Results

    Among cloned candidates, ImmuneAgent yielded 60 neutralizing antibodies (55%) and 12 bnAbs (11%).

    Source: Discussion · p. 31

  • Takeaways & Limitations

    The reported findings support using ImmuneAgent to discover neutralizing antibodies from natural repertoires, including bnAbs.

    Source: Discussion · p. 31

    The reported antibody yields are not compared with the cited 0.01– reference.

    Source: Discussion · p. 31

Abstract

from arXiv · show

Discovering broadly neutralizing antibodies (bnAbs) from human natural immune repertoires remains a fundamental challenge in immunology, hindered by: the extreme rarity of bnAb, incomplete understanding of their cellular origins across pathogens, and the inability of existing computational tools to generalize across emerging viral threats. Here we present ImmuneAgent, a closed-loop AI system that integrates multimodal reasoning with continual meta-learning and wet-lab feedback to overcome these barriers. Applied to screen the natural BCR repertoires from vaccinated or infected cohorts, the system achieves a ~55% neutralization antibody discovery rate (60 of 110 cloned candidates) and a ~11% bnAb yield (12 of 110), substantially outperforming a state-of-the-art sequence-based neutralization predictor or cofolding models evaluated at the same cloning budget. Five ImmuneAgent-discovered antibodies conferred 100% in vivo protection against lethal influenza challenge, comparable to the clinical-stage therapeutic MEDI8852. The system recovered the cellular and structural determinants of bnAb activity and identified FCRL5+CD27+ atypical memory B cells as a conserved bnAb reservoir and hydrophobic interface enrichment as a cross-viral structural signature, which generalized to unseen antigens, discovering human metapneumovirus (hMPV) cross-neutralizing and human papillomavirus (HPV)-neutralizing antibodies without antigen-specific sorting. These results validate that ImmuneAgent is a generalizable framework for rapid therapeutic antibody discovery against emerging viral threats.

Introduction

The introduction presents ImmuneAgent as a framework connecting immune-receptor data and computational predictions to therapeutic discovery through multidisciplinary reasoning and wet-lab feedback. It describes a continual-learning workflow that refines its models and adapts to unseen datasets and emerging pathogens.

Introduction

  • Fewer than 0.01% of sequenced receptors have experimentally validated antigen specificity, while models can fall to random guessing (AUPRC ~0.59) on unseen epitopes.Source: Introduction · p. 3
  • De novo models can predict structurally confident binders that fail neutralization in vitro or in vivo, while hard negatives and systematic experimental feedback remain scarce.Source: Introduction · p. 3
  • ImmuneAgent coordinates repertoire, structural, molecular, and bioinformatics tools with meta-learning and LoRA updates from experiments to generalize under data-scarce conditions.Source: Introduction · p. 4
  • Five ImmuneAgent-discovered antibodies conferred 100% in vivo protection against lethal influenza challenge, matching clinical-stage therapeutic antibodies.Source: Introduction · p. 4

continual-evolution for immunological discovery

  • Its lab-in-the-loop Meta-BCR engine combines mean-teacher meta-learning and LoRA fine-tuning to evolve predictions with experimental data.Source: continual-evolution for immunological… · p. 6
  • Wet-lab feedback continuously updates ImmuneAgent’s models, enabling adaptation to unseen datasets and emerging pathogens and supporting generalizable, functionally validated discoveries.The system feeds experimental success and failures back into the agent and uses updated models to guide subsequent experimental validations.Source: continual-evolution for immunological… · p. 6
  • A six-stage reasoning pipeline integrates cell biology, structural biology, chemistry, and molecular biology through query decomposition, retrieval, reasoning, hypothesis generation, planning, and evaluation.The orchestration agent coordinates operations through a dynamic task list, memory, and a reasoning loop, delegating work to multidisciplinary sub-agents.Source: continual-evolution for immunological… · p. 7
  • The orchestration agent delegates tasks to sub-agents for bioinformatics, model training, code generation, simulation, and structural analysis.Sub-agents use CodeAct loops for code execution, result transformation, and debugging; interfaces include persistent memory, optional human oversight, and quality evaluation.Source: continual-evolution for immunological… · pp. 7–8
  • The Meta-BCR engine combines ProtBERT-based sequence encoding with mean-teacher meta-learning, training on HIV, influenza, and coronavirus BCR data and updating from experimental feedback.Feedback assays include cryo-EM structural determination, antibody cloning, epitope mapping, and virus-neutralization assays.Source: continual-evolution for immunological… · p. 8

ImmuneAgent outperforms state-of-the-art biomedical agents and general-purpose

ImmuneAgent outperformed biomedical agents and standalone language models on general biomedical and immunology benchmarks. Continual evolution further improved performance on both benchmark types.

language models on demanding benchmarks

  • With the DeepSeek-R1 backbone, ImmuneAgent attained 20% accuracy, outperforming the no-agent baseline by 135% and Biomni by 21%, and had the highest accuracy across all backbone LLMs.Source: language models on demanding benchmarks · p. 9
  • ImmuneBench comprised 46 tasks spanning adaptive and innate immunity and computational immunology, drawing questions from four datasets across influenza, SARS-CoV-2, EBV, and tumour contexts.The task areas included B- and T-cell biology, functional immunology, and computational bioinformatics.Source: language models on demanding benchmarks · p. 9
  • ImmuneAgent achieved a mean F1 score of 0.78 across immunological task categories, versus Biomni’s 0.43 and Virtual Lab’s 0.10.The results also report scores above 0.90 on clonality analysis, V(D)J recombination interpretation, and virus-specific B-cell identification.Source: language models on demanding benchmarks · pp. 9–10, Figure 2 · p. 11
  • ImmuneAgent’s specialised architecture and immunological tools were reported to advantage tasks requiring cross-disciplinary integration, where general-purpose scientific dialogue alone was insufficient.The section highlights integration of repertoire-level statistics with single-cell phenotype data and tasks such as tumour-specific T-cell characterization and cross-reactivity prediction.Source: language models on demanding benchmarks · p. 10
  • With continual evolution enabled, GLM4.5 accuracy rose to 27.9% from 23% and was 66% above Biomni’s 17%.DeepSeek-R1 accuracy rose from 20% to 26%, a 30% relative gain; seven related immunological tasks showed an average F1 uplift of 38%.Source: language models on demanding benchmarks · p. 10

Tool augmented multimodal reasoning improve the experimental planning

ImmuneAgent’s expert-assessed research outputs scored highly, while Meta-BCR showed strong prediction performance with semi-supervised adaptation on independent SARS-CoV-2 and influenza datasets.

Tool augmented multimodal reasoning improve the experimental planning

  • ImmuneAgent’s hypotheses included mechanistic pathways, quantitative predictions, and confidence intervals, whereas baseline LLM responses were described as generic and lacking testable predictions.Its automated experimental plans were also described as feasible and budget-conscious.Source: Tool augmented multimodal reasoning… · p. 12
  • ImmuneAgent with GPT-4o scored highest across all eight expert-assessed criteria, with no criterion below 4.00/5.00.The criteria included scientific rigor, practical utility, systematic validations, hypothesis quality, planning quality, computational readiness, and biological feasibility.Source: Tool augmented multimodal reasoning… · pp. 13–14
  • Meta-BCR-Sc achieved the highest performance across metrics on an independent SARS-CoV-2 cohort; adaptation with 30 labelled and about 16,000 unlabelled sequences improved accuracy by about 30% over Meta-BCR-MT.The benchmark compared Meta-BCR transfer and semi-supervised models with CNN and DeepAAI baselines.Source: Tool augmented multimodal reasoning… · p. 14
  • On the Jackson Turner influenza dataset, semi-supervised adaptation yielded about a 16% accuracy improvement.The dataset comprised 35,579 B cells.Source: Tool augmented multimodal reasoning… · p. 14
  • Meta-BCR-MT embeddings separated BCRs by pathogen and distinguished H1N1 HA binders from non-binders, while predicted influenza antigen correlations recapitulated group-level relationships.The MI15–CA09 correlation was 0.62 and was reported to agree with established antigenic mapping studies.Source: Tool augmented multimodal reasoning… · p. 14

Multi-modal reasoning with meta-learning enhance the generalization to the unseen

These results assess ImmuneAgent’s generalization through feedback-driven meta-learning and repertoire screening across held-out antigens and sorted and non-sorted datasets. They report experimentally validated cross-reactive antibodies and influenza activity, alongside neutralization and bnAb yields and co-folding confidence across datasets.

datasets

  • After feedback adaptation, Meta-BCR reached 0.71 F1 and 0.89 AUC-ROC on Guo et al.’s SARS-CoV-2 set and 0.85 F1 and 0.87 AUC-ROC on Turner et al.’s influenza set.These results represented a cumulative 11-25% improvement over the pre-adaptation baseline and significantly outperformed DeepAAI, Mason’s CNN, and other state-of-the-art methods across metrics (P < 0.0001).Source: datasets · p. 15
  • Meta-BCR embeddings clustered HIV-, SARS-CoV-2-, and influenza-specific BCRs, separated binders from non-binders within pathogen classes, and recapitulated known antigen relationships across 12 influenza strains.The influenza analysis recovered strong intra-Group 1 correlations and the expected H5N1-H1N1 similarity.Source: datasets · p. 15

SARS-CoV-2 variants using ImmuneAgent

  • Of 60 SARS-CoV-2 candidates tested, 39 bound ancestral and Omicron spike trimers with EC₅₀ <10 μg/mL; six also neutralized XBB.1.5 and JN.1 despite donors lacking prior exposure.All 39 spike-binding antibodies neutralized the tested variants at 5 μg/ml.Source: SARS-CoV-2 variants using ImmuneAgent · p. 17
  • Cryo-EM showed BA5-12 and BA5-39 engaging a conserved JN.1 receptor-binding motif core through heavy-chain CDR interactions, with distinct contacts mediating epitope recognition.BA5-12 contacted G416 and A477 through HCDR1, whereas BA5-39 used HCDR2 and HCDR3.Source: SARS-CoV-2 variants using ImmuneAgent · p. 17
  • Of 50 top influenza candidates, 33 bound all four tested H1N1 strains, 20 of those neutralized at least one strain, nine neutralized all four, and six neutralized H5N1.The tested strains included one historical heterologous strain and three recent circulating strains.Source: SARS-CoV-2 variants using ImmuneAgent · pp. 17–18
  • INF-15 and INF-31 had H5N1 IC₅₀ values of 1.01 and 1.60 μg/mL, respectively, versus 1.74 μg/mL for MEDI8852.All six tested ImmuneAgent-discovered antibodies neutralized the four evaluated influenza strains.Source: SARS-CoV-2 variants using ImmuneAgent · p. 18
  • ImmuneAgent-discovered antibodies showed superior neutralization potency against three H1N1 strains compared with MEDI8852.The reported comparison is against three H1N1 strains.Source: SARS-CoV-2 variants using ImmuneAgent · p. 18
  • Around 55% of candidates were neutralizing (60/110) and 11% were bnAbs (12/110), exceeding the other neutralization prediction model at the same cloning budget across both conditions.The results from sorted and non-sorted datasets support use when optimal sorting conditions are not feasible.Source: SARS-CoV-2 variants using ImmuneAgent · p. 18
  • Median AlphaFold3 predicted iptm was 0.547 for sorted BCR complexes and 0.345 for non-sorted BCR complexes.Source: SARS-CoV-2 variants using ImmuneAgent · p. 20
  • ImmuneAgent-discovered antibodies demonstrated in vivo protection against influenza A.Source: SARS-CoV-2 variants using ImmuneAgent · p. 20

viruses

In a lethal influenza challenge, all five tested ImmuneAgent antibodies protected mice as effectively as MEDI8852. They also reduced lung viral loads and lung pathology compared with control groups.

viruses

  • 100% survival was observed with all five tested ImmuneAgent antibodies, whose weight-loss profiles were indistinguishable from MEDI8852.The mice received antibody before challenge with 5× LD₅₀ of A/California/07/2009 (H1N1).Source: viruses · p. 21
  • Four ImmuneAgent antibodies had lung viral titers statistically indistinguishable from MEDI8852, and all five achieved comparable or superior reductions.Lung viral titers were measured at day 5 post-challenge.Source: viruses · p. 21
  • Lungs from MEDI8852- and ImmuneAgent-treated mice showed reduced inflammation and preserved alveolar architecture compared with PBS and D25 controls.Control lungs showed inflammatory infiltrates, alveolar wall thickening, cellular debris, and hemorrhage.Source: viruses · p. 21

ImmuneAgent recovers the cellular and structural determinants of broad neutralization

ImmuneAgent integrates cellular, sequence, and structural evidence to prioritize broadly neutralizing antibodies. The analyses identify cellular and structural features associated with breadth and test the contribution of antibody mutations.

ImmuneAgent recovers the cellular and structural determinants of broad neutralization

  • FCRL5 was the only marker consistently enriched across all datasets; CD27⁺ B cells (n = 48) showed higher median neutralization breadth than CD27⁻ cells, linking phenotype to functional breadth.Differentially expressed markers converged on an atypical/age-associated memory signature.Source: and integrates them to discover… · p. 23
  • HCDR1 reversion raised BA5-12’s JN.1 IC₅₀ from 9.3 to 89.5 ng mL⁻¹, while combined HCDR1+HCDR2 reversion abolished JN.1 and XBB.1.5 neutralization.For BA5-57, HCDR1 reversion increased the JN.1 IC₅₀ from 8.1 to 213.2 ng mL⁻¹.Source: and integrates them to discover… · p. 24
  • BA5-39 was minimally affected by HCDR1 reversion alone and lost breadth only after combined HCDR1+HCDR2 reversion.The authors describe this as an internal on-mechanism control.Source: and integrates them to discover… · p. 24
  • Cryo-EM showed hydrophobic contacts between BA5-12 CDR loops and conserved spike receptor-binding-motif residues; reverting relevant residues diminished these contacts.The reported contacts involved HCDR1 residues V29/S30/S31 and HCDR2 residues Y52/P53/G54/G55.Source: and integrates them to discover… · p. 24
  • Across 109 cloned coronavirus and influenza heavy chains, CDR-H2 hydrophobicity discriminated broadly from narrowly neutralizing antibodies (AUC 0.77); hydrophobic-to-hydrophilic contact ratio increased monotonically with neutralization breadth across complexes spanning viral families.CDR-H1 and CDR-H3 hydrophobicity did not discriminate broad from narrow neutralizers.Source: and integrates them to discover… · pp. 24–25
  • At matched selection budget, combined strategies reached approximately 60% bnAb yield against a ~17% base rate, while DeepAAI recovered functional antibodies in 8% to 29% of cases.The reported DeepAAI range was 8% (5 of 63 predicted) to 29% (5 of 17 tested).Source: and integrates them to discover… · p. 25

neutralization from single-cell repertoires and structures and integrates them to discover

The figures depict model-predicted antigen-specific FCRL5⁺ B cells and measures of antibody breadth, structure, and prioritization. They also show how neutralization and predicted antibody yield are represented across variants and selection budgets.

neutralization from single-cell repertoires and structures and integrates them to discover

  • The cellular panels show predicted antigen-specific FCRL5⁺ B cells colored by antigen, shared differentially expressed genes across datasets, and neutralization breadth of cloned antibodies split by CD27 status.The caption identifies RSV-F, SARS-CoV-2 RBD, and influenza HA as antigens and FCRL5 as the gene shared across three datasets.Source: neutralization from single-cell… · pp. 26–27
  • The structural panels show the BA5-12–JN.1 spike interface and representative RBD-antibody neutralization across WT, BA.1, BA.2, BA.5, XBB, EG5, and JN.1.The caption lists HCDR1 contact residues and two resolved states at 4.18 Å and 3.91 Å.Source: neutralization from single-cell… · p. 27
  • The yield panel plots bnAb fraction among top-k cloned antibodies against cloning budget for combined strategies, individual signals, random base rate, and DeepAAI.The caption also describes CDR-loop hydrophobicity and interface-contact ratio by neutralization breadth.Source: neutralization from single-cell… · p. 27

ImmuneAgent generalize to unseen antigen targets and discover neutralizing antibody

ImmuneAgent discovered neutralizing antibodies to untrained antigens from unsorted B-cell repertoires, including RSV-A/B–hMPV cross-neutralizers and HPV-neutralizing antibodies. The results span antibody activity, cell-of-origin patterns, and predicted binding interfaces.

lineage without antigen sorting

  • The two measured hMPV cross-neutralizers were predicted to bind pre-F through CDR-H3-anchored, predominantly hydrophobic contacts, with 14 and 7 hydrophobic contacts and one hydrogen bond each.Source: lineage without antigen sorting · p. 28
  • 13 of 30 nominated antibodies neutralized all five tested RSV-A/B strains and hMPV.The RSV-A strains were A2 and Long; the RSV-B strains were B9320 and B18537.Source: lineage without antigen sorting · p. 28
  • Nine of 30 nominated antibodies came from FCRL5lowCD27+ cells, while none cloned from FCRL5- cells neutralized hMPV.The nine from FCRL5lowCD27+ cells cross-neutralized RSV-A/B and hMPV.Source: lineage without antigen sorting · p. 28
  • ImmuneAgent combined five modalities in an equal-weighted ranking of HPV candidates despite lacking enough reference antibodies to fine-tune Meta-BCR.The five modalities included cell of origin, SHM, clonal expansion, Chai-1 ipTM, and Meta-BCR prediction.Source: lineage without antigen sorting · p. 29
  • Among the 50 selected HPV candidates, 35 met all four immunological criteria, and the set included more than 21 gene usages and more than 6 clusters.The criteria were class switching, somatic hypermutation above 1%, clonal expansion, and FCRL5+ or plasma-cell origin.Source: lineage without antigen sorting · p. 29
  • 9 of 50 cloned candidates neutralized at least one HPV variant, and two cross-neutralized HPV6, HPV11, HPV33, and HPV58.HPV33 and HPV58 are Group 1 high-risk types.Source: lineage without antigen sorting · p. 29
  • For HPV, FCRL5 expression, ipTM, and clonal expansion were positively correlated with neutralization, whereas Meta-BCR prediction and SHM failed to predict it.Source: lineage without antigen sorting · p. 29

Discussion

ImmuneAgent is presented as a closed-loop, multimodal framework for discovering therapeutically relevant antibodies from human B-cell repertoires. The authors report high discovery yields, protection in an influenza mouse challenge, and findings on bnAb cellular and structural features, while noting limits on some comparisons.

Discussion

  • The platform is described as a closed-loop, lab-in-the-loop system that uses experimental feedback and meta-learning to refine domain-specialized models.Its workflow formulates hypotheses, generates and executes computational plans, directs targeted validation, and iteratively refines models.Source: Discussion · pp. 31, 33
  • ImmuneAgent prioritization recovered 12 bnAbs (11% yield) and 60 neutralizing antibodies (55% hit rate) against SARS-CoV-2 and influenza A, exceeding a state-of-the-art sequence-based neutralization predictor under matched cloning budgets on the same repertoires.The paper notes that antigen-sorted screening frequencies (0.01–0.5% bnAbs) have unmatched denominators and are provided for context only.Source: Discussion · pp. 31–32
  • All five tested antibodies conferred 100% prophylactic protection against lethal influenza A challenge in mice, with profiles comparable to MEDI8852, but equivalence was not established.The experiment was not powered to establish equivalence.Source: Discussion · p. 32
  • ImmuneAgent operates on existing public single-cell V(D)J datasets without antigen sorting or large-scale screening, supporting rapid deployment against emerging threats.The discussion contrasts this with conventional antigen-sorted cloning workflows that can take weeks to months.Source: Discussion · p. 32
  • The authors frame repertoire-level mining as complementary to de novo design, using existing human memory B-cell repertoires to seek pre-existing breadth.They describe bnAb discovery as potentially beginning within days from accessible sequencing data.Source: Discussion · pp. 32, 34
  • ImmuneAgent generalized to unseen antigens, recovering HPV-neutralizing and RSV–hMPV cross-neutralizing antibodies from unsorted B-cell data.The discussion notes that receptor predictors can perform worse on unseen antigens than on training pairs.Source: Discussion · p. 32
  • ImmuneAgent identified FCRL5 expression as correlated with higher bnAb percentages and SHM-induced hydrophobic CDR contacts with conserved epitopes as a determinant of breadth.The FCRL5+CD27+ bnAb-source signal was independently assessed across SARS-CoV-2, influenza, HPV and RSV–hMPV.Source: Discussion · p. 33
  • ImmuneAgent produces explicit evidence chains integrating V(D)J usage, SHM patterns, B-cell phenotypes, Meta-BCR predictions and structural models.The authors distinguish this reasoning architecture from black-box approaches.Source: Discussion · p. 33

Methods

ImmuneAgent organizes multimodal immunological findings in a typed knowledge graph and uses retrieval, memory, and hierarchical task execution to plan scientific analyses. The framework also supports auditable coordination and expert oversight.

Knowledge Graph Construction

  • The graph represents entities across analytical modalities and their relationships, with type ontologies and confidence-scored relationship labels.Entities include genes, proteins, cell types, and phenotypes.Source: Knowledge Graph Construction · p. 34
  • Queries combine entity-specific retrieval with thematic clusters, while multi-hop graph traversal connects findings across modalities.Thematic clusters are identified through community detection.Source: Knowledge Graph Construction · pp. 34–35
  • ImmuneAgent retains successful execution trajectories and abstracts successful patterns into semantic memory for later planning.Retrieved trajectories provide examples of tool chains, parameter settings, and reasoning patterns.Source: Knowledge Graph Construction · p. 35
  • The system decomposes complex queries into ordered subtasks and treats scientific subtasks as options selected by a ReAct loop and executed through CodeAct.Examples include retrieving BCR sequences, annotating them, predicting binding, analyzing lineage evolution, and modeling structures.Source: Knowledge Graph Construction · p. 36
  • Shared reports provide downstream stages with prior results and a provenance chain, while interrupt-based human review supports oversight of critical decisions.The system also uses execution outcomes as signals for planning-model improvement.Source: Knowledge Graph Construction · p. 37

Meta-BCR architecture and training

Meta-BCR combines sequence-based antibody–antigen modeling with meta-learning and semi-supervised adaptation. Its evaluations report performance against state-of-the-art predictors and audit possible training-data overlap with external cohorts.

Meta-BCR architecture and training

  • Meta-BCR predicts BCR antigen specificity and neutralization breadth using pretrained sequence embeddings, cross-attention interaction features, and multitask classification outputs.Its design integrates meta-learning with semi-supervised learning to enhance adaptability.Source: Meta-BCR architecture and training · p. 37
  • Training proceeds through BCR-sequence transfer learning, antigen-antibody meta-learning, and mean-teacher fine-tuning with labeled and unlabeled sequences.The meta-learning phase uses a hyperparameter-adaptive REPTILE variant across antigen-antibody sequence pairs.Source: Meta-BCR architecture and training · p. 38
  • Meta-BCR with mean-teacher training consistently matched or exceeded state-of-the-art performance, with its largest gains in PR-AUC for influenza neutralization prediction.Comparisons covered accuracy, AUC-ROC, F1-score, and PR-AUC for SARS-CoV-2 and influenza prediction.Source: Meta-BCR architecture and training · p. 38

Training-data auditing

  • Published DeepAAI and Mason weights yielded approximately 0.3 AUC on both baselines, so both comparators were retrained on the study corpus for performance comparisons.The reported sequence identities are approximate because the analysis used normalized matching rather than gapped global alignment.Source: Training-data auditing · p. 39
  • The training-data audit found no exact training CDR-H3 match among 23 influenza evaluation sequences, while SARS-CoV-2 evaluation included one exact CDR-H3 match, TH70.The SARS-CoV-2 cohort also had an 18.88% tail at 90% identity, which the authors identify as a residual risk for absolute performance.Source: Training-data auditing · p. 39

Benchmark Design and Datasets

ImmuneBench standardizes evaluation across antibody and T-cell receptor datasets, with varied modalities, sizes, and biological scopes. Its questions define task-specific ground truth, and its scoring uses metrics suited to classification, enrichment, and knowledge retrieval.

Benchmark Design and Datasets

  • ImmuneBench assembles antibody, TCR, interface-mutation, and immunological knowledge tasks across datasets that vary in size, modality, and biological scope.The framework includes 46 questions and evaluates ImmuneAgent alongside Biomni and another AI agent.Source: Benchmark Design and Datasets · pp. 39–40

Question Design and Ground Truth Computation

  • Questions span binding, neutralization, breadth, cross-reactivity, clonal expansion, polyreactivity, enrichment, phenotype, and knowledge retrieval, with labels derived from experimental columns and defined thresholds.Broadly neutralizing status requires positive results for at least 60% of five H1N1 neutralization columns.Source: Question Design and Ground Truth… · pp. 40–41

Scoring Metrics

  • Scoring combines classification metrics with enrichment agreement, entity recall, and exact-match accuracy for knowledge tasks.Enrichment uses a ground-truth-derived baseline, and groups with fewer than five members are excluded.Source: Scoring Metrics · pp. 41–42

CD19+ B-cell isolation

The experimental methods cover B-cell isolation and sequencing, cell analyses, antibody production, and in vitro binding and neutralization assays. The section specifies the workflows used for each stage.

CD19+ B-cell isolation

  • PBMCs were enriched for CD19-positive cells, stained with fluorescent SARS-CoV-2 spike and live-dead markers, then sorted as live cells.The cells were rested before staining and sorting.Source: CD19+ B-cell isolation · p. 42

Single-cell V(D)J repertoire library preparation and sequencing

Differential expression analysis

Cell trajectory inference

  • B-cell developmental trajectories were examined with Monocle3, Diffusion Map, and Slingshot-based pseudotime inference.Monocle3 used a principal graph, while Slingshot constructed smooth curves from a minimum spanning tree.Source: Cell trajectory inference · p. 43

Antibody cloning and purification

  • Paired heavy- and light-chain variable regions were synthesized, cloned into human IgG1 expression vectors, and expressed in 293F cells before protein A purification.Source: Antibody cloning and purification · p. 43

Antibody sequence analysis

  • Antibody sequences were annotated for V-gene usage and CDR and framework regions, then analyzed for gene usage, CDR-H3 clustering, and sequence dominance.CDR-H3 clusters required equal length and more than 80% sequence similarity.Source: Antibody sequence analysis · pp. 43–44

Anti-RBD ELISA

  • Anti-RBD ELISA measured binding to SARS-CoV-2, Omicron, and SARS-CoV RBDs, defining positivity as OD450 at 1 μg/mL at least three times the negative control.Source: Anti-RBD ELISA · p. 44

Pseudovirus neutralization assay

  • Pseudovirus neutralization was assessed by mixing serially diluted antibodies with pseudovirus, adding Huh7 cells, and measuring luciferase after incubation.Each plasma or serum sample group was tested in two replicates.Source: Pseudovirus neutralization assay · p. 44

Influenza virus microneutralization assay

The influenza microneutralization assay measures antibody neutralization using a focus-forming assay, with virus titrated in focus-forming units and antibody–virus mixtures tested on MDCK cells. The section also specifies IC₅₀ analysis and reports supporting data and figure materials.

Influenza virus microneutralization assay

  • The assay uses focus-forming units to predetermine virus titre and tests antibody–virus mixtures on MDCK cells after overnight cell incubation.The listed setup uses 0.03 million MDCK cells per well and 100 FFU/50 μL virus mixed equally with antibody solution.Source: Influenza virus microneutralization… · pp. 44–45
  • After antibody–virus incubation, the assay proceeds through cell exposure, overlay incubation, fixation, staining, and spot counting to assess neutralization.The protocol specifies a 36-hour incubation under overlay and spot counting with an enzyme-linked spot analyser.Source: Influenza virus microneutralization… · p. 45

Statistical analysis

  • IC₅₀ values are determined by four-parameter nonlinear regression using GraphPad Prism 9.0.Statistical analysis is specified in the figure legends.Source: Statistical analysis · p. 45

Data and code availability

  • The single-cell RNA-seq data are publicly available under BioProject PRJCA038672, and cryo-EM structures are deposited under PDB accessions 9UHF, 9UHG, 9UHH, 9UHO, and 9UHQ.The passage also states that original code and models are available on GitHub.Source: Data and code availability · p. 46

Supporting information

Extended Data Fig. 1-8

  • ImmuneAgent’s extended figures include a benchmark visualization in which DeepAAI and CNN fail to separate H1N1 binders from non-binders, unlike ImmuneAgent.The comparison uses 1,847 sequences; the caption reports silhouette scores of 0.23 for DeepAAI and 0.18 for CNN.Source: Figure. S3 · p. 49

Figure. S3

  • Supplementary figures cover single-cell repertoire analyses, antibody structures and neutralization, lung histopathology, and tests on untrained antigens.The captions include influenza and SARS-CoV-2 repertoire analyses and RSV–hMPV and HPV task materials.Source: Figure S4 · p. 50, Figure. S3 · pp. 51–53, Figure S7 · p. 53

Related papers

Loading 2610.03160v1…