Paper deep dive
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
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract: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.
Tags
Links
- Source: https://arxiv.org/abs/2610.03160v1
- Canonical: https://arxiv.org/abs/2610.03160v1
Trouble viewing inline? Open PDF directly →
Full Text
118,254 characters extracted from source content.
Expand or collapse full text
Multimodal reasoning for broadly neutralizing antibody discovery from label-free human B cell repertoires across virus families Hantao Lou 1,*, # , Jianqing Zheng 2,# , Can Yue 3,# , Meihan Zhang 1,# , Yuanchao Bao 2,# , Yu Chen 2,# , Mengting Huang 4,7 , Yupeng Yang 4 , Qianyu Pan 4 , Nana Fu 1 , Yansong Shi 1 , Hongli Li 4,7 , Yangyang Chai 1 , Ruyi Chen 4,7 , Wansheng Li 4,7 , Zhu Liang 2 , Rongmei Yao 5 , Yuanhan Mo 2 , Lei Wang 3 , Chunmei Wang 2,6 , Yun Quan 4 , Qiong Zhang 4,7 , Xiangxi Wang 3,* , Xuetao Cao 1,2,6,* 1 State Key Laboratory of Medicinal Chemical Biology, Institute of Immunology, College of Life Sciences, Nankai University, Tianjin 300071, China 2 Department of Immunology, Center for Immunotherapy, Peking Union Medical College, Chinese Academy of Medical Sciences, Beijing 100005, China 3 CAS Key Laboratory of Infection and Immunity, National Laboratory of Macromolecules, Institute of Biophysics, Chinese Academy of Sciences, Beijing 100101, China 4 Guangzhou National Laboratory, No. 9 XingDaoHuanBei Road, Guangzhou International Bio Island, Guangzhou 510005, Guangdong Province, China. 5 Biological Safety Laboratory, Tianjin University of Traditional Chinese Medicine, Tianjin 301617, China 6 National Key Laboratory of Immunity and Inflammation, Suzhou Institute of Systems Medicine, Chinese Academy of Medical Sciences & Peking Union Medical College, Suzhou, Jiangsu, 215123, China 7 Guangzhou National Laboratory, Guangzhou, GuangDong 510220, China # These authors contributed equally to this work. *Co-correspondence: hantao.lou@nankai.edu.cn, xiangxi@ibp.ac.cn, caoxt@immunol.org Abstract 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 adaptive immune system functions as nature’s most sophisticated molecular recognition engine, deploying immune receptors (BCRs and TCRs) that span a combinatorial diversity exceeding 10 15 unique sequences to detect and eliminate pathogens and malignancies (Bod et al., 2023; Moir and Fauci, 2009; Phad et al., 2022; Sharonov et al., 2020). Harnessing this machinery has become the central endeavour of therapeutic development, underpinning monoclonal antibody therapeutics, CAR-T therapies, and mRNA vaccines (Jabbour et al., 2021; Laidlaw and Ellebedy, 2022). As the past two decades has witnessed an unprecedented expansion of immunological data from immune receptors research and multi-omics sequencing. Single-cell multi-omics technologies have generated billions of immune receptor sequences across diverse healthy and patient cohorts (King et al., 2021; Lou et al., 2026; Verstegen et al., 2023; Zehner et al., 2023). Yet a fundamental “data-to-therapy” gap persists: fewer than 0.01% of sequenced receptors have experimentally validated antigen specificity, leaving the vast majority of this sequencing diversity therapeutically unmined. Current artificial intelligence approaches have not resolved this gap. While deep learning has transformed protein structure prediction, it faces a persistent failure of generalization in immunology (Ishaan Gulrajani, 2020; Lu et al., 2026). Predictive models for TCR–epitope and BCR–antigen binding perform well on receptor–antigen pairs present in their training data, but degrade dramatically when confronted with novel epitopes. Benchmarks across 50 state-of-the- art models reveal a "memorization trap," where models achieve high accuracy on known pairs but degrade to random guessing (AUPRC ~0.59) on unseen epitopes (Lu et al., 2026). Second, de novo design models frequently predict structurally confident binders that fail to neutralize in vitro or in vivo, reflecting incomplete understanding of the cellular and physical constraints governing immune recognition (Bennett et al., 2026; He et al., 2024). These limitations stem in part from the scarcity of hard negative data and the absence of systematic experimental feedback capable of refining model inductive biases. Without exposure to non-functional sequences that essentially resemble binders, models cannot learn the subtle physical rules distinguishing neutralization antibodies or functional TCRs. Immunological discovery poses a further challenge that goes beyond single-task prediction as it requires simultaneous reasoning across disciplines. Identifying a broadly neutralizing antibody, for instance, demands integration of clonal evolution, structural biology and cellular immunology. This multidisciplinary reasoning distinguishes the problem from coding accuracy or chemistry tasks on which most AI agent systems are currently evaluated (Jimenez et al., 2024; Laurent et al., 2024). Current agentic systems like Virtual Lab and Biomni operate as static, open-loop architectures. These agents primarily rely on retrieval from published datasets or internal memory from the LLM, with limited domain specific tool orchestration for biological tasks (Huang et al., 2025; Swanson et al., 2025). Virtual Lab’s multi-agent meetings coordinate LLM reasoning roles (PI, specialist, critic) to construct computational pipelines and successfully designed nanobodies against SARS-CoV-2. However, the system lacks a mechanism to incorporate experimental failure or iteratively update its reasoning based on validation outcomes (Swanson et al., 2025). Without a feedback loop to incorporate experimental failure, these systems remain perpetually frozen in their training distributions, and struggled to generalize to out of distribution data. Consequently, these systems are vulnerable to hallucination and multi-step biological task failures, as evidenced by the persistent performance gap between computational benchmarks and experimental validation rates across the field. Here, we present ImmuneAgent, a multi-agent system designed to bridge the gap between data driven computational predictions and biological function. Unlike monodisciplinary models, ImmuneAgent employs a hierarchical architecture that coordinates multidisciplinary specialized tools spanning repertoire analysis, structural biology, and molecular biology, bioinformatics, model training, function prediction and data-mining, enabling reasoning across sequence evolution, structural constraints and cellular context to augment the immunological tasks. The ImmuneAgent tackle the generalization issue by adapting a meta-learner that facilitate the active evolving of the agentic system by feeding experimental data to the loop by meta-learning and Low-Rank Adaptation (LoRA) strategy. This continual-learning capability allows the agent to update its inductive biases in real-time, enabling rapid generalization to previously unseen viral variants and emerging therapeutic targets under data-scarce conditions, such as emerging H5N1 strains or tumour epitopes. We evaluate ImmuneAgent on the challenge of broadly neutralizing antibody discovery across influenza, respiratory syncytial virus and SARS-CoV-2. On independently published external datasets, ImmuneAgent achieved a substantial improvement over existing antibody discovery models and other agentic systems, demonstrating robust generalization beyond its training distribution. Crucially, five ImmuneAgent-discovered antibodies conferred 100% in vivo protection against lethal influenza challenge, matching clinical-stage therapeutic antibodies. Additionally, ImmuneAgent discovered the cellular and structural mechanisms behind the broadly neutralizing antibodies, which was then validated experimentally. By closing the loop between multimodal-reasoning and wet-lab validation, ImmuneAgent establishes a dynamically scalable and generalizable framework for therapeutic discovery. Results ImmuneAgent integrates deterministic multi-agent orchestration and "lab-in-the-loop" continual-evolution for immunological discovery ImmuneAgent is an autonomous, closed-loop multi-agent framework designed to bridge the gap between computational prediction and biological function. To address the multidisciplinary complexity inherent to modern immunological challenges, the system executes rigorous interdisciplinary reasoning across computational biology, molecular biology, immunology, structural biology, and molecular dynamics. It integrates literature review, hypothesis generation, experimental planning, multi-domain specialized tool execution augmented data analysis, and continual learning driven by experimental results to comprehensively address a wide spectrum of immunological research questions (Fig 1A, Supplementary Fig 1A, B). Prior to execution, the system performs comprehensive reasoning and tool augmented knowledge retrieval to formulate testable hypotheses and experimental plans. This planning phase is driven by an advanced Retrieval-Augmented Generation (RAG) system and tool calls that query an extensive knowledge base comprising over 20 datasets and online resources. System-wide operations are subsequently coordinated by a central Orchestration Agent, which utilizes a dynamic to-do list, a knowledge graph, and a Task Router to intelligently distribute context to individual sub-domain experts (Fig 1A, Supplementary Fig 1A). Following this orchestration, sub-domain experts deployed over 214 specialized tools (Supplementary Table 1) within these sub-domain experts. These sub-domain experts operate through a dynamic dual-loop execution pattern, which fuses ReAct (Reason + Act — interleaved natural-language reasoning and tool invocation) for high-level strategic planning with a CodeAct loop (in which the agent's actions are emitted as executable code rather than as fixed-schema tool calls) for sandboxed, autonomous code execution and debugging. This dual-loop structure allows the overall system to reason dynamically through augmented sub- domain expert execution (Supplementary Fig 1B). To ensure fault tolerance and logically inform future reasoning, the system maintains persistent semantic and episodic memory via its knowledge graph and long-term memory storage (Supplementary Fig 1B). The successfully executed workflows and usable scripts generated is saved for future sessions. The key conceptual advance of ImmuneAgent is its "lab-in-the-loop" continual-evolution mechanism. At the core of its predictive pipeline is the Meta-BCR learning engine, which integrates a meta-learning framework utilizing mean-teacher and LoRA finetune (Fig 1B, Supplementary Fig 1C). To actively address the scarcity of negative data in immunology, the system feeds experimental success and failures back into its agent with (LoRA) strategy (Fig 1B). This dynamic refinement architecture physically connects computational predictions to wet-lab validations, incorporating outputs from experimental results to continuously update the models (Fig 1B). The updated models will be utilized by the execution agent and suggest the experimental validations (Fig 1C). Ultimately, by continually learning directly from these wet- lab experimental results, the framework dynamically adapts to unseen datasets and emerging pathogens which enables highly generalizable and functionally validated biological discoveries. Figure 1. Architecture and workflow of ImmuneAgent. A. Overview of the ImmuneAgent multi-agent system for autonomous immunological discovery. (Top left) The six-stage reasoning pipeline integrates knowledge from cell biology, structural biology, chemistry, and molecular biology domains through sequential query decomposition, retrieval-augmented generation (RAG), multidisciplinary reasoning, hypothesis generation, experimental planning, and evaluation. (Top centre) The Orchestration Agent coordinates system-wide operations through a dynamic to-do list, knowledge graph (memory), ReAct reasoning loop, multidisciplinary reasoning module, and debug and reasoner component. The orchestration agent iteratively delegates tasks and aggregates results from downstream sub-agents. (Top right) Multidisciplinary sub-agents include bioinformatics, model training, code generation, simulation, and structural analysis modules. Each sub-agent operates within a CodeAct loop for autonomous code execution with result transformation and debugging capabilities. (Middle) External interfaces include the LightRAG and mem0 knowledge graph for persistent memory, a human-in-the-loop interrupt mechanism for optional expert oversight, and an evaluator agent for quality assessment. B. The Meta-BCR learning engine combines transfer learning via ProtBERT-based sequence encoding with a meta- learning framework employing mean-teacher optimization. BCR sequences from HIV, influenza, and coronavirus training data are processed through classification losses with gradient-based model updates (model θ), hyperparameter optimization (β), and regularization (λ(θ′–θ)) to enable rapid adaptation to unseen targets. C. Experimental validation outputs including cryo-EM structural determination, antibody cloning, epitope mapping, and virus neutralization assays fed back to the agent to update the model by meta-learning module. ImmuneAgent outperforms state-of-the-art biomedical agents and general-purpose language models on demanding benchmarks To rigorously evaluate ImmuneAgent’s reasoning and problem-solving capabilities, we benchmarked the system against biomedical AI agents and large language models (LLMs) on two complementary assessment frameworks: the Humanity’s Last Exam (HLE) biomedicine subset (n = 200) and a purpose-built immunology benchmark comprised of questions spanning TCR/BCR specificity, tumour specific clones, pathogen specific clones, functional gene usage, B/T cell clonality and phenotype enrichment (n = 46 across four independent datasets). ImmuneAgent achieved up to 5-fold improvement over no-agent baselines on the HLE biomedicine subset (Fig 2A), and significantly higher F1 scores than competing biomedical agent on the immunology benchmarks (Fig 2B). The HLE biomedicine subset comprises expert-curated, open-ended questions designed to test the frontier of scientific reasoning, requiring integration of domain knowledge, quantitative analysis, and multi-step inference. We evaluated ImmuneAgent against two comparator systems: Biomni, a recently reported general-purpose biomedical agent, and no-agent baselines consisting of standalone LLMs without agentic capabilities. To assess robustness across model architectures, we tested each system with three backbone LLMs: Qwen3-max, DeepSeek-R1, and GLM4.5, which are all released before the HLE benchmark to prevent data leakage. In the standard configuration (without continual-evolution), ImmuneAgent consistently achieved the highest accuracy across all backbone LLMs. With the DeepSeek-R1 backbone, ImmuneAgent attained 20% accuracy, outperforming the no-agent baseline by 135% and Biomni by 21%. Similar performance gains were observed with Qwen3-max (ImmuneAgent: 14% versus no- agent: 2.5% and Biomni: 9.5%) and GLM4.5 (ImmuneAgent: 24.5% versus no-agent: 7.4% and Biomni: 17%) (Fig 2A). The ImmuneAgent performance is further improved by the continual-evolving mechanism that update the agent memory and tool-set by accumulated executions (Supplementary Fig 2A, B). To evaluate domain-specific reasoning capabilities beyond general biomedical knowledge, we curated “ImmuneBench”, a comprehensive immunological task benchmark comprising questions that require integrated reasoning across adaptive immunity, innate immunity, and computational immunology. The benchmark encompasses 46 tasks spanning T-cell biology (antigen-specific T-cell identification, T-cell receptor repertoire analysis, tumour-specific T- cell characterisation), B-cell biology (virus-specific B-cell identification, B-cell receptor clonality analysis, broadly neutralizing antibody prediction), functional immunology (phenotypic characterization of immune cells, neutralization assessment, cross-reactivity evaluation), and computational bioinformatics (single-cell data interpretation, V(D)J recombination analysis, immune repertoire statistics). To ensure biological relevance and data leakage prevention, we curated questions from four independent datasets spanning distinct pathological contexts: influenza (quadrivalent vaccine cohort), SARS-CoV-2 (mRNA and adenoviral vaccine recipients), Epstein–Barr virus (EBV; natural infection cohort), and tumour immunology (tumour-infiltrating lymphocyte datasets), covering both B-cell and T-cell compartments. ImmuneAgent achieved a mean F1 score of 0.78 across all immunological task categories, outperforming Biomni by 81% (F1 = 0.43) and Virtual Lab by 680% (F1 = 0.10) (Fig 2B). The performance differential was particularly pronounced for tasks requiring multi-step reasoning across sequence, structure, and function: ImmuneAgent achieved the highest scores (F1 > 0.90) on clonality analysis, V(D)J recombination interpretation, and virus-specific B-cell identification (Supplementary Fig 2C), whereas Biomni’s performance degraded significantly on tasks requiring integration of repertoire-level statistics with single-cell phenotypic data. Virtual Lab, which relies primarily on multi-agent discussion without specialised immunological tools or domain-specific models, failed to produce meaningful outputs for most tasks (F1 < 0.20 for 8 of 12 categories). This highlights that general-purpose scientific dialogue alone is insufficient for domain-specific immunological reasoning. The disparity was most striking for tumour-specific T-cell characterisation and cross-reactivity prediction, where ImmuneAgent’s access to TCR specificity predictions, structural modelling, and repertoire analysis tools enabled integrative reasoning that neither comparator system could replicate. These results demonstrate that ImmuneAgent’s specialised multi-agent architecture and purpose-built immunological tool suite provide a decisive advantage for domain-specific reasoning tasks that demand cross-disciplinary integration. We next assessed the impact of ImmuneAgent’s continual-evolving mechanism on benchmark performance. With continual-evolution enabled, ImmuneAgent with GLM4.5 achieved 27.9% accuracy, representing a significant improvement over the non-evolving configuration (23%) and a 66% improvement over Biomni (17%). The DeepSeek-R1 backbone similarly benefited, with accuracy rising from 20% to 26%, a 30% relative gain (Fig 2C). Activating the lab-in- the-loop continual-evolution mechanism in which the Meta-BCR learner and episodic memory incorporate prior experimental outcomes into the reasoning pipeline. The continual-evolving mechanism conferred improvement in the ImmuneBench, resulting in average of 38% uplift for the F1 score of seven related immunological tasks, ranging from bioinformatics, b cell and t cell phenotypes and immune receptor function (Fig 2D). These results establish that the continual-evolving mechanism provides additive performance gains beyond static agentic architectures, consistent with the hypothesis that integrating experimental feedback into the reasoning loop sharpens the system’s ability to resolve ambiguous biomedical queries. Figure 2. Benchmarking ImmuneAgent against state-of-the-art AI systems. A. Performance on the Humanity’s Last Exam (HLE) biomedicine subset. Bar plots showing accuracy (%) of no-agent baselines (Qwen3-max, DeepSeek-R1 and GLM4.5), ImmuneAgent, and Biomni across three backbone LLMs. ImmuneAgent achieves the highest accuracy across both conditions, substantially outperforming no-agent configurations and the Biomni biomedical agent. B. Immunological task performance measured by F1 score. ImmuneAgent outperforms Biomni and Virtual Lab on domain-specific immunological reasoning tasks. n = 46 benchmark questions; single run, dispersion shown is between-question spread C. Performance of the ImmuneAgent on HLE tests with continual-evolving mechanism, compared with the SOTA methods. D. Performance of ImmuneAgent with meta-learner on ImmuneBench tasks compared without meta-learner. Tool augmented multimodal reasoning improve the experimental planning Beyond task-level accuracy, we assessed ImmuneAgent’s capabilities across five core performance dimensions: hypothesis quality, planning quality, tool selection, biological feasibility, and innovation (Fig 3A). ImmuneAgent with GPT-4o backbone consistently occupied the outermost regions of the radar plot across all dimensions (hypothesis quality: 91.2±4.6%; planning quality: 92.9±5.2%; tool selection: 90.2±4.8%; biological feasibility: 84.6±4.8%; innovation: 78.0±3.4%; n = 50 queries per model, mean±s.d.), substantially exceeding standalone GPT-4o (57.9–74.7%), DeepSeek V3.1 (59.7–81.4%), and GPT-3.5- Turbo (36.5–68.2%) (Fig 3A; Supplementary Note S1). The hypotheses featured detailed mechanistic pathways with quantitative predictions and confidence intervals, whereas baseline LLM responses were generic without testable predictions (Supplementary Fig 3A). The automated experimental plans from ImmuneAgent were highly feasible and budget-conscious, prioritising low-cost computational experiments in initial phases as validation for the hypotheses. Furthermore, ImmuneAgent integrated multidisciplinary methods to address major knowledge gaps, yielding higher discovery potential than baseline models (Supplementary Note S2). To ensure practical relevance, human expert evaluation by three single-blinded immunologists (Supplementary Note S1) validated these automated assessments using an eight-criterion rubric scored on a 1–5 Likert scale. ImmuneAgent with GPT-4o achieved the highest scores across all criteria: Scientific Rigor (4.90/5.00), Innovation (4.00/5.00), Practical Utility (5.00/5.00), Systematic Validations (4.90/5.00), Hypothesis Quality (5.00/5.00), Planning Quality (5.00/5.00), Computational Readiness (4.90/5.00), and Biological Feasibility (5.00/5.00) (Fig 3B). The heatmap visualisation demonstrates ImmuneAgent’s balanced excellence across all dimensions, with no criterion scoring below 4.00/5.00 compared to GPT- 4o’s maximum of 3.80/5.00 and DeepSeek V3.1’s maximum of 4.00/5.00. ImmuneAgent utilized knowledge graph to retrieve the key mechanisms and pathways related to the queries and the deep research findings, thereby enhancing the hypothesis generation and planning (Supplementary Fig 3B). Together, these results establish ImmuneAgent as a comprehensive immunological reasoning platform that significantly surpasses both specialised biomedical agents and frontier general-purpose LLMs across general biomedical knowledge, domain- specific immunological tasks, and expert-assessed research quality. Figure 3. Meta-BCR prediction performance and interpretability on independent external datasets. A. Radar plot of multi-dimensional agent capability assessment across five criteria: hypothesis quality, planning quality, tool selection, biological feasibility, and innovation. ImmuneAgent with GPT-4o backbone (green) and DeepSeek V3.1 backbone (blue) consistently occupy the outermost regions across all axes, whereas standalone GPT-4o (orange), DeepSeek V3.1 (purple), and GPT-3.5-Turbo (red) show progressively contracted profiles. B. Expert criteria evaluation heatmap. Three single-blinded immunologists scored system outputs on an eight- criterion Likert scale (1–5): scientific rigor, innovation, practical utility, systematic validations, hypothesis quality, planning quality, computational readiness, and biological feasibility. ImmuneAgent with GPT-4o achieved the highest scores across all criteria (range 4.0–5.0), with no criterion below 4.0/5.0, whereas the best standalone model (DeepSeek V3.1) peaked at 4.0/5.0. Colour scale: green (high) to red (low). C. Benchmarking on the Guo Yu et al. SARS- CoV-2 single-cell V(D)J dataset. Left: UMAP visualisation of memory B and naïve B cell populations from the independent cohort. Right: performance comparison of Meta-BCR-MT (meta-transfer), Meta-BCR-Sc (semi-supervised adapted), CNN baseline (Mason et al.), and DeepAAI across seven metrics (accuracy, precision, recall, AUC-ROC, PR-AUC, MCC, F1). Meta-BCR-Sc achieves the highest performance across metrics (****P < 0.0001, n = 5 model- initialisation seeds; these are computational replicates of the same data, not independent biological samples). Domain adaptation with 30 labelled and ~16,000 unlabelled sequences improved accuracy by ~30% relative to Meta-BCR-MT. D. Benchmarking on the Jackson Turner et al. influenza dataset comprising 35,579 B cells. Left: UMAP visualisation showing distinct B cell cluster architecture. Right: performance metrics demonstrating consistent Meta- BCR superiority, with semi-supervised adaptation yielding ~16% accuracy improvement (****P < 0.0001). E. UMAP projection of Meta-BCR-MT latent space embeddings. Left: pathogen-aware clustering reveals distinct separation of BCRs targeting HIV (green), SARS- CoV-2 RBD (blue), and influenza HA (orange), demonstrating learned antigen-class discrimination. Right: clear segregation of H1N1 HA binders (label 1, red/orange) from non- binders (label 0, grey), confirming embedding quality for downstream prediction tasks. F. Antigen correlation matrix derived from Meta-BCR predictions across 12 influenza strains (7 H1N1, 1 H2N2, 1 H5N1, 2 H3N2). The matrix recapitulates known group-level antigenic relationships, with strong intra-Group 1 correlations (warm colours) and low cross-group correlations (H3N2 vs. H1N1). Notably, the strong MI15–CA09 correlation (0.62) agrees with established antigenic mapping studies. H5N1 shows moderate correlation with H1N1 strains, consistent with shared Group 1 lineage. Colour scale: red (positive correlation) to blue (negative correlation). Multi-modal reasoning with meta-learning enhance the generalization to the unseen datasets A central challenge in BCR-antigen specificity prediction is the generalization gap: models trained on published antigen-antibody complexes struggles to predict specificity when applied to novel antigens or independent cohorts for immune receptors. To address this, we developed a two-phase adaptation strategy for lab-in-the-loop learning. In the post-experimental phase, validated experiment results are labelled and concatenated with unlabelled data to feedback to ImmuneAgent, refining the model's predictions for subsequent agent workflow via LoRA- based meta-learning (Fig 1B, C). This strategy transforms each experimental cycle into learning data, progressively improving prediction accuracy as the system accumulates experimental validations. We evaluated the prediction performance of the Meta-BCR across this adaptation pipeline on two independent external datasets that were fully held out during training. Prior to adaptation, Meta-BCR achieved 0.81 F1 score and 0.85 AUC-ROC on validation data but degraded to below 0.60 F1 on unseen datasets, consistent with the "memorization trap" observed across existing models (Fig 3C). Following LoRA-based meta-learning with experimental feedback (post-experimental phase), the model achieved 0.71 F1 and 0.89 AUC-ROC on the Guo et al. SARS-CoV-2 dataset for antibody neutralization prediction and 0.85 F1 and 0.87 AUC-ROC on the Turner et al. influenza dataset (35,579 B cells) for antigen specificity prediction, representing a cumulative 11-25% improvement over the pre-adaptation baseline and significantly outperforming DeepAAI, Mason’s CNN, and other state-of-the-art methods across metrics (Fig 3C, D; P < 0.0001) (Mason et al., 2021; Turner et al., 2020; Zhang et al., 2022). The updated models are actively selected by the ImmuneAgent for execution steps to enhance the adaptation of the framework to unseen datasets (Supplementary Fig 3C). Critically, Meta-BCR's latent space embeddings revealed pathogen specific clustering of HIV, SARS-CoV-2, and influenza specific BCRs (Fig 3E). The learned representations also discriminated binders from non-binders within each pathogen class, confirming that the model captures biologically meaningful features rather than superficial sequence patterns (Supplementary Fig 3D). Furthermore, antigen correlation analysis across 12 influenza strains recapitulated known group-level antigenic relationships, including strong intra-Group 1 correlations and the expected H5N1-H1N1 similarity (Fig 3F). These results demonstrated the continual-evolution coupled with multimodal reasoning could enhance the generalization of the framework to unseen datasets, which reduces the influence of generalization gap in immunology studies. High hit-rate discovery of broadly neutralizing antibodies against influenza A viruses and SARS-CoV-2 variants using ImmuneAgent Having established ImmuneAgent's superior performance over state-of-the-art (SOTA) language models and neutralization prediction algorithms, we assessed its utility for bnAb discovery against emerging viral threats from both antigen-sorted (SARS-CoV-2) and non- sorted (influenza) B cell datasets. We applied ImmuneAgent to identify bnAbs against SARS-CoV-2 variants from ChAdOx1 adenoviral vaccine recipients. Following the autonomous multi-agent workflow (Fig 4A), ImmuneAgent the system initiated a comprehensive discovery pipeline by knowledge retrieval followed by activating the planning agent and execution agent (Supplementary Note S3). ImmuneAgent executed Meta-BCR predictions across an antigen sorted single-cell V(D)J data, revealing enrichment of predicted broadly neutralizing antibodies within FCRL5 + memory and ITGB1 + memory B cell subsets (Fig 4B; Supplementary Fig 4A). The broadly reactive antibodies were found to have biased VH3-66 paired with VK1-39 usage (Supplementary Fig 4B). ImmuneAgent also performed clonotype analysis of the single-cell V(D)J data, showed that three medium-sized clonotypes (clone size ~10-100) exhibited predicted cross-reactivity to more than five variants, while the large clonotypes (clone size > 10 1.5 ) displayed low neutralization breadth (Fig 4C). ImmuneAgent then applied AlphaFold3 as a structural plausibility filter, modelling of top candidates that were predicted by Meta-BCR and bioinformatics analysis. The non-sorted BCRs had a median ipTM of 0.345 and the sorted BCRs a median ipTM of 0.547. ImmuneAgent used this ipTM axis as a low-confidence triage filter to deprioritise candidates least likely to form a modelled complex with the target antigen; ipTM is a co-folding confidence score and not evidence of binding, and no binding claim is made on its basis (Fig 4D, Supplementary Fig 4C). From these reasoning steps, ImmuneAgent prioritized the top 60 candidates for experimental validation based on predicted neutralization breadth, B cell subset (excluding naïve B cell and unswitched memory B cells), and structural modelling (ipTM > 0.35) (Supplementary Table 2, Supplementary Note S3). To validate the findings, we cloned the 60 ImmuneAgent-predicted candidates, we found that 39 antibodies bound spike trimers from ancestral WT, D614G, and Omicron variants (BA.1, BA.4, BA.5) with EC₅₀ <10 μg/mL, while all bound antibodies neutralized these variants at 5 μg/ml (Fig 4E; Supplementary Table 2). Critically, 6 antibodies neutralized highly immune- evasive XBB.1.5 and JN.1 variants despite donor samples lacking prior exposure (Fig 4E). Cryo-EM structural analysis at near atomic resolution of two representative bnAbs (BA5-12 and BA5-39) in complex with the JN.1 spike revealed their molecular basis for cross-variant neutralization (Fig 4F, Supplementary Fig 4D; Supplementary Table 2). Both antibodies engaged a conserved receptor-binding motif (RBM) core region spanning residues G416–Y489 through extensive heavy-chain CDR1, CDR2, and CDR3 interactions. BA5-12 formed critical hydrophobic contacts with G416 and A477 via HCDR1, while BA5-39 utilized HCDR2 and HCDR3 for epitope recognition (Fig 4F, Supplementary Fig 4D). To assess ImmuneAgent's performance under more challenging conditions without antigen enrichment, we queried ImmuneAgent: "Please design a budget friendly method by leveraging public datasets to identify H5N1 broadly neutralizing antibodies." Following the knowledge retrieval, literature review, hypothesis generation and experimental design, ImmuneAgent processed a public single-cell V(D)J data from quadrivalent influenza vaccine (QIV) recipients, identifying the predicted broadly reactive BCRs enriched within ITGB1 + memory, FCRL5 + memory, and GC B cell subsets (Fig 4G, Supplementary Fig 4E, Supplementary Note S3). Meta-BCR predictions identified cells with neutralization breadth spanning up to seven H1N1 strains (Fig 4G), with broad reactivity distributed across clonotypes of varying sizes, which are considerably smaller than antigen sorted datasets (Fig 4H). In contrast to SARS-CoV-2, a set of VH gene usage showed weak correlation with predicted binding breadth (Supplementary Fig 4F). As in the SARS-CoV-2 pipeline, AlphaFold3 served as a structural plausibility filter (ipTM threshold 0.35) to deprioritise candidates with low-confidence predicted complex geometries (Supplementary Table 2, Supplementary Fig 4G). From the top 50 ImmuneAgent-predicted broadly neutralizing antibodies, 33 recombinant antibodies bound all four tested H1N1 strains, including one heterologous historical strain (H1N1/A/Puerto Rico/8/34) and three recent circulating strains (H1N1/A/Michigan/45/2015, H1N1/A/Victoria/4897/2022, and H1N1/A/Wisconsin/67/2022) (Supplementary Table 2). Microneutralization assays revealed that 20 of these 33 antibodies neutralized at least one strain (Supplementary Fig 4H, Supplementary Table 2). Remarkably, nine antibodies achieved pan-H1N1 neutralization across all four strains, with six demonstrating H5N1 neutralization (Fig 4I, Supplementary Fig 4H). We evaluated in vitro neutralization potency against a panel of H1N1 and H5N1 strains, with CR6261 and MEDI8852 serving as benchmark controls. As expected, CR6261 showed robust neutralization of H1N1 strains (IC₅₀ <1 μg/mL for CA09, and WS67) but weaker activity against H5N1/A/Ostrich/2024 (IC₅₀ = 4.62 μg/mL), while MEDI8852 demonstrated moderate potency against H5N1/A/Ostrich (IC₅₀ = 1.74 μg/mL) but much weaker against H1N1/ A/Michigan/45/2015 (IC₅₀ = 8.10 μg/mL). Notably, all six tested ImmuneAgent-discovered antibodies neutralized the four strains evaluated (Fig 4I). Among them, INF-15 and INF-31 exhibited superior H5N1 neutralization potency (IC₅₀ = 1.01 and 1.60 μg/mL, respectively) compared to the benchmark therapeutic antibody MEDI8852 (IC₅₀ = 1.74 μg/mL) (Fig 4I, Supplementary Table 2). Furthermore, INF-14, INF-15 and INF-31 demonstrated superior neutralization potency against the three H1N1 strains than MEDI8852 (Fig 4I, Supplementary Table 2). Across both conditions, ImmuneAgent achieved around a 55% neutralization hit rate (60/110 candidates) and 11% bnAb yield (12/110), much higher than the other neutralizing antibody prediction model given the same cloning budget (Supplementary Table 2). This enhancement from both sorted and non-sorted datasets demonstrates ImmuneAgent's practical utility for rapid pandemic response when optimal sorting conditions are not feasible. Figure 4. Discovery and functional validation of broadly neutralizing antibodies by ImmuneAgent. A. Schematic of the autonomous ImmuneAgent multi-agent workflow. Single- cell V(D)J datasets from SARS-CoV-2 adenovirus-vectored vaccination (drawn on the artwork as “SARS-CoV-2 ADV vaccine”; ChAdOx1) and from influenza quadrivalent influenza vaccine (QIV) vaccination are supplied to ImmuneAgent together with Meta-BCR (drawn as “MetaBCR”) in response to the user query “Please design a budget friendly method to identify bnAb”; arrows trace the closed loop of analysis, hypothesis generation, planning and execution. Schematic only; no data are plotted. B. UMAP of B cells from the SARS-CoV-2 ChAdOx1 vaccination dataset (n = 12,847 cells), coloured by Meta-BCR-predicted neutralization breadth against SARS-CoV-2 variants (key, pale = 0 variant to dark red = 5 variants). C. Clonotype analysis of the SARS-CoV-2 dataset: clone size (y, log scale, 10⁰ to 10¹·⁸) versus Meta-BCR- predicted neutralization breadth (x, “Predicted number of variants (neu)”, 0–10). Grey, all clonotypes; D. AlphaFold3 co-folding confidence (“AF3 predicted iptm”, y, 0.0–1.0) for antibody–antigen complexes modelled from antigen-sorted BCRs (sorted BCR) and from non- sorted BCRs. Each symbol is one modelled complex; the horizontal bar is the group median (0.547 sorted, 0.345 non-sorted). E. Measured IC₅₀ (μg/mL) for six ImmuneAgent-discovered SARS-CoV-2 antibodies (BA5-39, BA5-57, BA5-30, BA5-5, BA5-12 and BA5-31) against variants BA.1, BA.5, XBB.1.5 and JN.1; the value is printed in each cell. Colour scale, pale (low IC₅₀, high potency) to dark blue (high IC₅₀, low potency). Data are mean IC₅₀ from three biological repeats. F. Cryo-EM structure of antibody BA5-12 bound to the SARS-CoV-2 JN.1 receptor-binding domain (RBD), shown in two views related by a 180° rotation. BA5-12 heavy chain (keyed on the artwork as “BA.5-12 H”, green), BA5-12 light chain (“BA.5-12 L”, grey- blue), RBD (tan); all six complementarity-determining regions are labelled (HCDR1–HCDR3, LCDR1–LCDR3). G. UMAP of B cells from the influenza QIV vaccination dataset (n = 9,412 cells), coloured by Meta-BCR-predicted neutralization breadth (key, dark purple = 0 variant to yellow = 5 variants; the key is printed on the artwork as “Neutralization breath Prediction”). Values are predictions (predicted ≠ measured). H. Clonotype analysis of the influenza dataset: clone size (y, log scale) versus Meta-BCR-predicted binding breadth (x, “Predicted number of variants (bind)”, 0–8). Grey, all clonotypes; purple, highlighted clonotypes, of which four are labelled by clonotype ID (34879, 8156, 12235 and 27686) and were taken forward for experimental validation. I. Measured IC₅₀ (μg/mL) for six ImmuneAgent-discovered influenza antibodies (INF-14, INF-31, INF-9, INF-15, INF-44 and INF-12) together with the benchmark control antibodies CR6261 and MEDI8852, against three H1N1 strains (PR8, CA09, WS67) and one H5N1 strain (OS24); the value is printed in each cell. Colour scale, pale (low IC₅₀, high potency) to dark blue (high IC₅₀, low potency). Data are mean IC₅₀ from three biological repeats. ImmuneAgent discovered antibodies demonstrate in vivo protection against influenza A viruses Having established ImmuneAgent's capability to computationally identify broadly neutralizing antibodies with in vitro neutralization potency compared to clinical-stage therapeutics, we next evaluated whether these antibodies translate to protective efficacy in vivo against viral challenge. We assessed the prophylactic potential of ImmuneAgent-discovered antibodies in a mouse model of influenza A infection, comparing their performance directly against MEDI8852. Female BALB/c mice received a single intraperitoneal dose of 10 mg/kg antibody one day before intranasal challenge with 5× LD₅₀ of A/California/07/2009 (H1N1) and were monitored for 14 days for survival and weight changes (Fig 5A). MEDI8852 served as the positive control and RSV-specific antibody D25 served as the negative control. As expected, all D25-treated mice reached the predefined humane endpoint and were euthanised by day 6 post-challenge, and PBS-treated animals reached 0% survival by day 7. Both groups exhibited progressive weight loss exceeding 25% by the humane endpoint (Fig 5B, C). In contrast, MEDI8852 conferred complete protection (100% survival) with weight loss remaining below 10% on day 4 post-infection, followed by recovery to baseline by day 7 (Fig 5B, C). Notably, all five tested ImmuneAgent-discovered antibodies (INF-9, INF-12, INF-14, INF-15 and INF-31) provided 100% survival with weight loss profiles indistinguishable from MEDI8852, demonstrating that ImmuneAgent-predicted antibodies confer robust prophylactic protection comparable to the clinical benchmark (Fig 5B, C). To quantify viral replication in the respiratory tract, we performed a parallel experiment measuring lung viral titers at day 5 post-challenge. PBS and D25 groups exhibited similarly viral loads (>10 5 FFU/mL/g in lung homogenates), whereas MEDI8852 reduced viral titers by approximately one log (P < 0.01 versus PBS and D25, Fig 5D). All five ImmuneAgent antibodies achieved comparable or superior reductions in lung viral load, with INF-31, INF- 12, INF-14, and INF-15 showing viral titers statistically indistinguishable from MEDI8852 (Fig 5D). Histopathological analysis by H&E staining revealed extensive inflammatory infiltrates, alveolar wall thickening, cellular debris and hemorrhage in PBS and D25 groups, consistent with severe viral pneumonia (Fig 5E, Supplementary Fig 5A, B). In contrast, lungs from MEDI8852 and ImmuneAgent-discovered antibody treated mice exhibited markedly reduced inflammation and preserved alveolar architecture, with minimal hemorrhage and focal lesions (Fig 5E, Supplementary Fig 5C - H). Together, these data establish that ImmuneAgent-discovered antibodies provide protection against influenza challenge in vivo, with efficacy matching the clinical-stage therapeutic antibody MEDI8852. The convergent in vivo protection observed across multiple ImmuneAgent-discovered antibodies establishes that multi-agent reasoning process reliably translate to functional efficacy. Critically, the ability to achieve clinical-grade protection from antibodies computationally mined from non-sorted single-cell datasets demonstrates ImmuneAgent's capacity to overcome a major bottleneck in therapeutic antibody discovery. Figure 5. ImmuneAgent discovered antibodies demonstrated in vivo protection A. Experimental timeline for prophylactic efficacy studies. Female BALB/c mice received single intraperitoneal injection of antibodies (10 mg/kg) on day 0, followed by intranasal challenge with 5× LD₅₀ of A/California/07/2009 (H1N1) on day 1. Lung viral load was assessed at day 5 post-infection, while body weight monitoring and survival analysis continued through day 14. B. Survival curve of the positive control, negative control and ImmuneAgent discovered antibodies over the course of 14 days. C. Longitudinal body weight changes expressed as percentage of initial weight for the positive control, negative control and ImmuneAgent discovered antibodies over the course of 14 days. Data represent mean ± s.d., n=6 mice per group. D. Pulmonary viral load at day 5 post-infection quantified by focus- forming unit (FFU) assay. Box plots show median (centre line), interquartile range (box), and range (whiskers). Statistical analysis by one-way ANOVA: *P < 0.05, **P < 0.01. n = 6. E. Representative H&E-stained lung sections at day 5 post-infection. Top row shows low magnification (scale bar: 400 μm); bottom row shows high magnification of boxed regions. ImmuneAgent recovers the cellular and structural determinants of broad neutralization and integrates them to discover functional antibodies To establish whether an agentic system can reason with proof instead of hallucinations, we posed two queries to ImmuneAgent: the transcriptional phenotype of broadly reactive B cells, and the structural features that distinguish broadly neutralizing antibodies. For each query the orchestrator decomposed the task and coordinated the bioinformatics, deep learning and structural sub-agents, producing an explicit, human-auditable chain of evidence linking each conclusion to its supporting analysis (Fig 6A; Supplementary Note S3). In response to the cellular origin query, ImmuneAgent integrated paired single-cell transcriptomes and V(D)J sequences across seven public datasets (Supplementary Fig 6A). It then applied the Meta-BCR neutralization predictor to nominate broadly reactive clones, and, by differential expression, converged on an atypical/age-associated memory signature (ZEB2, ITGAX and FCRL5; Supplementary Fig 6B), of which FCRL5 was the only marker consistently enriched across all datasets (Fig 6C). Projected onto the FCRL5⁺ compartment, the model-predicted RSV-F, RBD and HA broadly reactive cells are enriched within a transcriptionally conserved memory B cell subset, plasma cells and germinal centre B cells (Fig 6B). To validate the ImmuneAgent finding, we cloned antibodies derived from FCRL4/5⁺ B cells that were CD27⁺ (n = 48), which showed a higher median neutralization breadth than those from CD27⁻ cells, linking the phenotype to functional breadth (Fig 6D, Supplementary Fig 6C). In terms of the structural determinants, ImmuneAgent hypothesized that accumulated somatic hypermutations (SHMs) in broadly neutralizing antibodies enable hydrophobic paratope contacts. Combined with immunological reasoning and Meta-BCR predictions, the HCDR1 substitutions F28I/L, T29I and S32R/N were predicted to enhance SARS-CoV-2 XBB.1.5 and JN.1 neutralization, and germline reversion of these positions was predicted to abrogate it (Supplementary Fig 6D). We tested these predictions directly by germline-reverting HCDR1, or HCDR1 together with HCDR2, in three lineage members and measuring pseudovirus neutralization (Supplementary Fig 6E). Reversion weakened neutralization potency by approximately 6- to 25-fold against the most immune-evasive variants: for BA5-12, HCDR1 reversion raised the JN.1 IC₅₀ from 9.3 to 89.5 ng mL⁻¹ (≈10-fold) and combined HCDR1+HCDR2 reversion abolished JN.1 and XBB.1.5 neutralization (IC₅₀ > 10 000 ng mL⁻¹); for BA5-57, HCDR1 reversion increased the JN.1 IC₅₀ from 8.1 to 213.2 ng mL⁻¹ (≈26- fold). Consistent with the prediction that this lineage member engages through HCDR2/HCDR3 rather than HCDR1, BA5-39 was minimally affected by HCDR1 reversion alone (JN.1 IC₅₀ 95.2 → 35.7 ng mL⁻¹) and lost breadth only upon combined HCDR1+HCDR2 reversion, providing an internal on-mechanism control (Supplementary Fig 6E). To validate the findings by ImmuneAgent, cryo-EM structures of the BA5-12–JN.1 spike complex were resolved. In the wild-type paratope, HCDR1 residues V29/S30/S31 and HCDR2 residues Y52/P53/G54/G55 formed hydrophobic contacts with the conserved receptor-binding-motif residues A475/G476 and A421/G416, respectively. The germline-reverting substitutions I28T (HCDR1) and P53S/G55S (HCDR2) each diminished these contacts, validating the findings by ImmuneAgent (Fig 6E). The CDRH2 region of the bnAbs against the coronavirus RBD and influenza virus HA demonstrated enhanced hydrophobicity, which further supports the notion that hydrophobic interaction by CDRH2 contribute to neutralization breadth (Fig 6H). At the repertoire level this signal was carried specifically by CDR-H2: across the 109 cloned coronavirus and influenza heavy chains, CDR-H2 hydrophobicity discriminated broadly from narrowly neutralizing antibodies (area under the receiver-operating-characteristic curve 0.77), whereas CDR-H1 and CDR-H3 hydrophobicity did not, and because in the IGHV3-53/3-66 lineage this hydrophobicity is somatic-hypermutation-acquired rather than germline-encoded as in VH1-69 stem antibodies, this per-loop resolution departs from the CDR-H3-centric literature. ImmuneAgent quantified interface interactions across 66 antibody–antigen complexes spanning viral families: the hydrophobic-to-hydrophilic contact ratio increased monotonically with neutralization breadth (Fig 6H). Representative broadly neutralizing antibodies against the influenza HA stalk and the SARS-CoV-2 RBD illustrate this hydrophobic-contact interface (Fig 6F, Supplementary Fig 6F, G). Finally, because no single modality was individually decisive, we tested whether ImmuneAgent's integrated prioritization converts prediction into function. Within the single cell repertoire and the cloned antibodies that were ranked by the combined multimodal score enriched genuine broadly neutralizing antibodies far above the random base rate and above every individual signal or LLM alone (Fig 6G, Fig 2B): at a matched selection budget the combined strategies reached approximately 60% bnAb yield against a ~17% base rate, whereas a state-of-the-art single predictor (DeepAAI) selecting by its own high-confidence threshold recovered functional antibodies in only 8% (5 of 63 predicted) to 29% (5 of 17 tested) of cases. Integration across cellular, sequence, structural and predictive evidence, rather than any individual signal, is what converts computational prediction into functional broadly neutralizing antibodies. Collectively, Fig 6 shows an agentic workflow that recovers an established cellular anchor (Fig 6A–D) and, where it extends prior work, contributes a cross-viral, epitope-dependent quantification of interface chemistry, a hydrophobic-to-hydrophilic contact ratio that scales with neutralization breadth (Fig 6H), and a demonstration that multimodal integration, rather than any single predictor is what converts computational prediction into functional broadly neutralizing antibodies (Fig 6G). Figure 6. ImmuneAgent recovers the cellular and structural determinants of broad neutralization from single-cell repertoires and structures and integrates them to discover functional antibodies. (A) ImmuneAgent orchestrator and its bioinformatics and immune- receptor sub-agents resolving the query “what is the phenotype of broadly reactive B cells in response to vaccination?” across seven single-cell V(D)J and RNA datasets (SARS-CoV-2 BNT and adenovirus-vectored vaccination and infection, influenza QIV vaccination, tonsil). (B) UMAP of model-predicted antigen-specific FCRL5⁺ B cells, coloured by antigen (RSV-F, SARS-CoV-2 RBD, influenza HA). (C) Venn diagram of genes differentially expressed in broadly-reactive B cells across the adenovirus-vectored vaccination, QIV vaccination and SARS-CoV-2 infection datasets; FCRL5 is the single gene shared by all three. (D) Measured SARS-CoV-2 neutralization breadth of cloned antibodies from FcRL4/5⁺ B cells, split by CD27 status (CD27⁺, n = 48; CD27⁻, n = 32). (E) Cryo-EM interface of the BA5-12–JN.1 spike complex: HCDR1 residues (L27/I28/V29/S30/S31) contacting the conserved receptor-binding- motif residues (A475/G476/Y473) in two resolved states (4.18 Å and 3.91 Å). (F) Measured neutralization of representative RBD antibodies across variants (WT, BA.1, BA.2, BA.5, XBB, EG5, JN.1) with their epitope surfaces (XGv302, BA5-57, BD55-1205). (G) Broadly- neutralizing-antibody yield (fraction of the top-k cloned antibodies) versus cloning budget k for the ImmuneAgent combined strategies (rank-sum and staged) and each individual signal (Meta-BCR breadth, ipTM, CDR-H2 hydrophobicity, cell-type score), against the random base rate and the DeepAAI predictor (29% and 8% at its high-confidence thresholds). (H) Left, per- loop CDR hydrophobicity for broadly versus narrowly neutralizing antibodies. Right, hydrophobic-to-hydrophilic interface-contact ratio by neutralization breadth. ImmuneAgent generalize to unseen antigen targets and discover neutralizing antibody lineage without antigen sorting To test the generalization of ImmuneAgent beyond influenza, coronavirus and HIV, we challenged the system with two antigens outside the antigens MetaBCR was trained on: RSV– human metapneumovirus (hMPV) cross-neutralization and human papillomavirus (HPV) neutralizing antibodies from unsorted single cell V(D)J data. For each task we queried with a single prompt and let ImmuneAgent decompose the query, retrieve literature by deep research, formulate a hypothesis and execute a CodeAct plan in a closed lab-in-the-loop (Supplementary Fig 7A). For RSV–hMPV, the deep-research stage established the prefusion- F (pre-F) antigenic-site organisation of RSV and hMPV and the existence of pre-F cross- neutralizing antibodies. Because of the lack of neutralization model for hMPV or HPV, ImmuneAgent combined the RSV-A/B Meta-BCR neutralization predictor with clonotyping, marker gene analysis and Chai-1 co-folding to shortlist the cross-neutralizers, and the deep- research stage induced a hypothesis that guided the multidisciplinary reasoning (Supplementary Fig 7A). ImmuneAgent validated the discovery chain by testing each sub- agent (that execute a set of tools) with published neutralizing antibodies and non-neutralizing antibodies and score the accuracy of different combinations of tools and sub-agents. Across a non-sorted paired single-cell V(D)J dataset, ImmuneAgent filtered the BCRs according to the execution plan by considering the marker gene expression, Meta-BCR prediction, clonotyping, hypermutation rate and class switching, removing the BCRs that do not meet at least three criteria (for example clonal expanded, class-switched with SHMs but predicted not to be binder is kept as candidate). The structure of 180 candidate antibodies with hMPV were then predicted and iptm of antibody-hMPV complexes higher than threshold are selected (Fig 7A, Supplementary Fig 7B). As in the SARS-CoV-2 and influenza tasks, co- folding ipTM served as a low-confidence signal, so the RSV and hMPV ipTM axes were used to tier candidates rather than to exclude them, and ranked antibodies were further evaluated by ImmuneAgent by their predicted interfaces on both antigens (Fig 7A, B). ImmuneAgent plotted the predicted hMPV pre-F interfaces of the two measured cross-neutralizers engaged the antigen through CDR-H3-anchored, predominantly hydrophobic contacts (14 and 7 hydrophobic atomic contacts and one hydrogen bond each) (Fig 7B). ImmuneAgent then selected 30 BCRs by an composite score combining cell type, transcriptional-module, MetaBCR-prediction, co-folding iptm, SHM, clonal expansion and interface features, some BCRs with negative FCRL5 expression but high composite scores were also selected to increase diversity (Fig 7C). Selected antibodies were tested for neutralization against RSV-A (A2, Long), RSV-B (B9320, B18537) and hMPV pre-F, and 13 neutralized all five viruses(Fig 7C, D). Nine of the 30 nominated antibodies derive from FCRL5 low CD27 + cells also cross- neutralized RSV-A/B and hMPV while none of the antibodies cloned from FCRL5 - cells neutralize hMPV (Fig 7C, D). We next applied ImmuneAgent to an even stricter generalization test, HPV, for which neither HPV antigens nor HPV-specific antibodies appear in the training data. The deep-research stage established, from primary literature, that the L1 major-capsid surface loops are the HPV neutralization target; it then curated an HPV L1 antigen and antibody database, which is composed of a collection of public monoclonal antibodies (15 sequenced B-cell receptors from IEDB) together with epitopes mapped to the L1 hypervariable loops (BC 49–80, DE 110–154, EF 170–190, FG 260–290, HI 347–360 in HPV16 numbering). However, this collection is not sufficient for Meta-BCR finetune although the structural knowledge for downstream reasoning could be used as prior knowledge. ImmuneAgent completed a CodeAct plan whose execution log records seven completed tasks spanning single cell V(D)J integration, subset annotation, clonal-expansion analysis, SHM with marker gene expression, Chai-1 co-folding predictions, Meta-BCR predictions (low confidence as it was not trained on HPV or related antigens) and MD simulation (Supplementary Fig 7C). ImmuneAgent then reasoned across cell of origin, clonotype, trajectory and SHM and yielded a candidate BCR lists (which is curated by selecting class-switched B cells that is either clonal expanded or/and high SHM) for MetaBCR prediction, cofolding prediction and MD simulations (Fig 7E). The five modalities were then combined as an equal-weighted mean into a single ranking, and the top 50 of that ranking were cloned, expressed and tested for HPV neutralization (Fig 7E). Within the selected 50 candidates, 35 satisfy all four immunological criteria which are class-switch, somatic hypermutation above 1%, clonal expansion, and FCRL5 + or plasma-cell origin while 15 BCRs failed on one of the criteria (Fig 7E).The diversity of the candidates were also evaluated and the 50 candidates demonstrated decent diversity with more than 21 different gene usages and more than 6 distinct clusters (Fig 7F). Of the 50 antibodies cloned, 9 candidates were found to neutralize at least one HPV variants and Fcrl5 expression and iptm with clonal expansion is positively correlated with the neutralization, while MetaBCR prediction and SHM failed to predict the neutralization (Supplementary Fig 7D, E). The selected antibodies’ co-folded complexes couple hydrogen bonds with hydrophobic contacts across the range seen in the pool (Supplementary Fig 7F). Two neutralizing antibodies cross-neutralize HPV6, HPV11, HPV33 and HPV58, two of which (HPV33 and HPV58) are Group 1 high-risk types (Fig 7G, Supplementary Fig 7E). These findings demonstrate the importance of multimodal reasoning across cellular, immunological, structural and sequence model for functional antibody discovery. Together, these results establish that ImmuneAgent generalizes to antigen it was never trained on, recovering RSV- A/B/hMPV cross-neutralizers and HPV neutralizing antibodies from unsorted repertoires. Figure 7. The single-cell antibody-discovery pipeline generalizes to untrained antigens (A) Each point is one antibody of the RSV–hMPV arm, positioned by Chai-1 co-folding confidence against RSV prefusion F (x, ipTM = maximum of the RSV-A and RSV-B co-folds) and against hMPV prefusion F (y, hMPV ipTM). Grey, all 180 co-folded antibodies; coloured, the 30 antibodies nominated by ImmuneAgent, partitioned by cell-of-origin phenotype into FCRL5 high CD27⁺ (blue, n = 21) and FCRL5 low CD27⁺ (orange, n = 9). (B) Two Chai-1 co- folding solutions against hMPV pre-F, each drawn as a Cα-backbone trace: hMPV pre-F (grey), antibody heavy-chain variable domain V_H (light blue), light-chain variable domain V_L (purple). Interface residues are marked at Cα — hydrogen-bond contact (blue), hydrophobic contact (gold), antigen epitope on pre-F (red), CDR-H3 paratope (black). (C) Left, all 30 nominated antibodies (hMPV-1 to hMPV-30, IGHV gene in brackets), ordered by the ImmuneAgent hMPV-cross composite score which is composed of FCRL5⁺CD27⁺ phenotype; ABC and memory transcriptional-module scores; Meta-BCR predicted neutralization probability against RSV-A and against RSV-B; Chai-1 RSV ipTM and hMPV ipTM; the ImmuneAgent hMPV-cross composite; hydrogen-bond and hydrophobic interface contact counts on RSV F and on hMPV F; and CDR-H3 hydrophobicity. Cell colour is the per-row min–max-scaled value (0–1; the colourbar is labelled “scaled within row”); Right, measured neutralization potency for the antibodies that were measured against RSV-A (A2, Long), RSV- B (B9320, B18537) and hMPV pre-F. (D) Neutralization curves for five representative cross neutralizing antibodies, x, log₁₀ antibody concentration (ng mL⁻¹); y, per cent neutralization. (E) Each horizontal strip represents a modality with all 213 candidates, 50 cloned candidates are highlighted in blue. Strips are cell of origin, somatic hypermutation, clonal expansion, Chai- 1 ipTM and Meta-BCR HPV binding, with the equal-weighted composite of the five shown as the bottom row. (F) Each point is one candidate antibody in an ESM2 sequence-embedding space (PC1 and PC2; explained variance not shown). Grey, all candidates; open circles, the antibodies cloned and expressed; filled circles, neutralizing antibodies. (G) Neutralization curves of HPV antibodies cloned from the ImmuneAgent nominated BCRs. Discussion In this study, we developed and validated ImmuneAgent, a closed-loop multi-agent AI platform that generalises across immunological tasks and bridges the gap between computational prediction and experimental validation. ImmuneAgent accelerates broadly neutralizing antibody (bnAb) discovery while revealing where bnAb potential resides in the human B-cell landscape and the structural mechanisms underpinning bnAb. From 110 cloned candidates prioritised by ImmuneAgent, we recovered 12 bnAbs (11% yield) and 60 neutralising antibodies (55% hit rate) against SARS-CoV-2 and influenza A. Under a matched cloning budget on the same repertoires, this exceeded the yield of a state-of-the-art sequence-based neutralization predictor (Supplementary Table 2). We do not compare these yields to the 0.01– 0.5% bnAb frequencies reported for antigen-sorted screening (Pinto et al., 2021; Rappazzo et al., 2022) or to de novo design pipelines (Bennett et al., 2026; Swanson et al., 2025), because those denominators are not matched to ours; the citations are given for context only. Critically, all five tested ImmuneAgent-discovered antibodies conferred 100% prophylactic protection against lethal influenza A challenge in mice, with survival and weight-loss profiles comparable to the clinical-stage therapeutic MEDI8852 in this experiment; the design was not powered to establish equivalence (Kallewaard et al., 2016). These results establish that tool augmented multi-modal AI reasoning with meta-learning can discover therapeutically relevant B cell repertoires from unsorted single-cell datasets. A central capability of ImmuneAgent is its capacity for rapid deployment against unseen pathogens (Graham and Sullivan, 2018). Unlike conventional antibody discovery workflows that require weeks to months for antigen sorted BCR cloning, hybridoma generation, or large- scale library screening (Corti et al., 2021), ImmuneAgent operates on existing public single- cell V(D)J datasets without requiring antigen sorting, pathogen-specific antigens, or large scale screening results. This capability is particularly urgent given the ongoing H5N1 epizootic (Damodaran et al., 2026), where the pace of antigenic evolution in circulating avian influenza strains outstrips conventional therapeutic development timelines (Abu-Shmais et al., 2025). ImmuneAgent provides a framework for bnAb discovery that is immediately deployable against emerging threats, without waiting for antigen characterisation or sorting panel development. Repertoire level screening by ImmuneAgent offers a complementary route to de novo design for therapeutic antibody discovery through mining existing human memory B cell repertoires for pre-existing breadth. ImmuneAgent also generalize to antigens not included in the training dataset despite the fact that immune receptor predictors are known to fall into memorization trap. The immune receptor predictors perform well on receptor-antigen pairs in the training but degrading substantially on unseen antigens (Lu et al., 2026). From unsorted B cell atlas, ImmuneAgent recovered HPV neutralizing antibodies and RSV-hMPV cross-neutralizing antibodies through hypothesis driven multidisciplinary reasoning. A therapeutic HPV bnAb would be complementary to the highly effective prophylactic L1 virus-like-particle vaccines, whose cross-type protection is variable (Schellenbacher et al., 2009), addressing post-exposure reservoirs or vaccine-escape variants. Furthermore, ImmuneAgent facilitate the identification of cellular and structural mechanisms behind the bnAb against pathogens. The FCRL5 expression was found to correlate with higher percentage of bnAb, which is consistent with the better recall response in influenza infection and mRNA vaccination (Kochayoo et al., 2025; Nellore et al., 2023; Spangler et al., 2025). Our independent assessment of the FCRL5⁺CD27⁺ bnAb-source signal in SARS-CoV-2, Influenza, HPV and RSV–hMPV extends an established compartment across pathogens for future bnAb discovery. The SHM-induced hydrophobic interaction of the antibody antigen complexes was recognized by ImmuneAgent to be one of the determinants of neutralization breadth. This finding indicates the rare clones that accumulated SHMs enhancing hydrophobic CDR contacts with conserved epitope residues displayed the broadest reactivity profiles. These mechanistic insights were generated through ImmuneAgent's transparent reasoning architecture, which produces explicit evidence chains integrating V(D)J usage, SHM patterns, B-cell phenotypes, Meta-BCR predictions, and structural models, distinguishing it from black-box approaches. ImmuneAgent’s architecture addresses fundamental limitations of existing AI systems for biological discovery. While systems like Virtual Lab and CRISPR-GPT demonstrate powerful task-specific automation, they are often constrained by predefined workflows (Qu et al., 2026; Swanson et al., 2025). Similarly, emerging general biology agents like Biomni showcase impressive breadth across biomedical subfields but lack the deep domain specialization required for hypothesis-driven discovery in complex, data-rich disciplines like immunology (Huang et al., 2025). The past year has seen a rapid proliferation of autonomous and "agentic- physical" laboratory systems that couple LLM planners to experimental execution (Jin et al., 2025; Luo et al., 2025; Wang et al., 2025); ImmuneAgent is distinguished within this landscape less by automation per se than by its lab-in-the-loop, meta-learning continual evolution and its domain-specialized predictive models, which together let it not only execute but improve its inductive biases as experimental feedback accrues. ImmuneAgent formulates a hypothesis, generates and executes computational plans, directs targeted experimental validation, and iteratively refines its models—yielding consistent gains across general biomedical knowledge , immunological tasks and expert-assessed research questions (Center for AI Safety et al., 2026). By closing the loop between computational results and experiments, ImmuneAgent enables the rapid identification and mechanistic characterisation of therapeutic antibodies from existing datasets without antigen sorting or large-scale library screening. This capability positions ImmuneAgent as a deployable framework for rapid response to emerging pathogens, as the bnAb discovery can begin within days by accessible sequencing data, accelerating the discovery therapeutic candidate Methods Knowledge Graph Construction Let M = M₁, M₂, ..., Mₖ denote the set of analytical modalities (sequence analysis, structure prediction, expression profiling, binding measurements). Each modality Mᵢ produces a document dᵢ containing extracted findings. LightRAG processes these documents to construct a heterogeneous knowledge graph: 퐺 = ( 푉,퐸,휑,휓 ) • V = entities extracted across all modalities (genes, proteins, cell types, phenotypes, experimental conditions) • E = relationships discovered between entities (encodes, binds, expresses, correlates_with, inhibits) • φ: V → T_V maps entities to type ontology (molecular, cellular, clinical) • ψ: E → T_E maps edges to relationship types with confidence scores Entity-Relationship Extraction. For each document chunk c from modality Mᵢ, an LLM extracts entity-relationship triplets: 퐸푥푡푟푎푐푡(푐) → ( 푒 ℎ ,푟,푒 푡 ,푐표푛푓 )| 푒 ℎ ,푒 푡 ∈ 퐸푛푡푖푡푖푒푠,푟 ∈ 푅푒푙푎푡푖표푛푠,푐표푛푓∈[0,1] Dual-Level Retrieval. Query processing combines local (entity-specific) and global (thematic) retrieval: 푅푒푡푟푖푒푣푒(푞) = 훼 · 푅푒푡푟푖푒푣푒 푙표푐푎푙 (푞) + (1−훼) · 푅푒푡푟푖푒푣푒 푔푙표푏푎푙 ( 푞 ) where Retrieve_local identifies specific entities and direct connections, while Retrieve_global captures broader thematic clusters through community detection on G. Cross-Modal Reasoning. Multi-hop queries traverse the graph to connect findings across modalities: 푃푎푡ℎ(푒 푠표푢푟푐푒 ,푒 푡푎푟푔푒푡 ) = 푎푟푔푚푎푥 휋 ∏푐표푛푓(푒ᵢ,푟,푒ⱼ) ( 푒ᵢ,푟,푒ⱼ ) ∈휋 This enables questions like: 'Which transcription factors (from scRNA-seq) are associated with the structural features (from AlphaFold) of broadly neutralizing antibodies (from Meta-BCR binding predictions)?' Memory Architecture. ImmuneAgent maintains two complementary memory systems via mem0 and LightRAG: Episodic Memory (mem0): Stores successful execution trajectories as retrievable episodes: 휏 = (푠₀,푎₀,푟₀),(푠₁,푎₁,푟₁),...,(푠ₙ,푎ₙ,푟ₙ) where sᵢ = state (query context, intermediate results), aᵢ = action (tool call, parameter choice), rᵢ = outcome (success/failure, output quality). Selective Persistence: Only trajectories with final status = SUCCESS persisted: Memory.add(τ) if Status(τ) = SUCCESS Semantic Memory (LightRAG): Abstracts successful patterns into generalizable knowledge: 퐺 푠푒푚푎푛푡푖푐 =퐸푥푡푟푎푐푡 푝푎푡푒푟푛푠 (휏 푠푢푐푒푠푓푢푙 )→(푡표푙 푐표푚푏푖푛푎푡푖표푛푠 ,푝푎푟푎푚푒푡푒푟 푟푎푛푔푒푠 ,푞푢푒푟푦 푡푦푝푒푠 ) Memory-Augmented Planning. When generating a new plan, the system retrieves relevant prior successes: 푃푙푎푛(푞)=퐿푀(푞,푅푒푡푟푖푒푣푒 푚푒푚0 (푞,푘=3),푅푒푡푟푖푒푣푒 퐿푖푔ℎ푡푅퐴퐺 (푞)) The retrieved trajectories provide concrete examples of successful tool chains, parameter configurations, and reasoning patterns for similar queries. Meta-Learning Signal. Execution outcomes generate signals for model improvement: 휃 푡+1 = 휃 푡 + 휂 · 훻 휃 퐿 ( 휏,표푢푡푐표푚푒 ) where L measures alignment between predicted and actual execution success. This enables continuous improvement of the planning model based on accumulated experience. Hierarchical Decision Process. The dual-loop architecture can be formalized as a hierarchical Semi-Markov Decision Process (Semi-MDP) following Sutton, Precup & Singh's options framework: Options (Temporally Extended Actions). An option o = ⟨I, π, β⟩ consists of: • 퐼 ⊆ 푆: initiation set (states where the option can begin) • 휋: 푆 × 퐴 → [0,1]: intra-option policy (how to act while executing) • 훽: 푆 → [0,1]: termination condition (when to stop) In ImmuneAgent, each scientific subtask (e.g., 'analyze BCR sequences', 'predict structures', 'correlate with phenotypes') is an option. The ReAct loop selects which option to execute; the CodeAct loop implements the intra-option policy. Multi-Agent Coordination Domain. The orchestration system is defined as: Orch = ⟨A, S, K, M, R⟩ • A = bioinformatics, Executor, orchestration, ...: specialized sub-agents • S = MCP_tools: skill/tool library • K: A → P(S): capability mapping (which agent can use which tools) • M: message passing protocol (state updates, task results) • R: resource constraints (parallel execution limits, timeouts) Hierarchical Task Network (HTN) Decomposition. Complex immunological queries are decomposed into executable subtasks: 퐷푒푐표푚푝표푠푒(푄)→푇₁,푇₂,...,푇ₙ with partial order ≺ and dependency graph D For example: 'Identify broadly neutralizing antibody candidates' decomposes into: (1) retrieve BCR sequences, (2) run IgBLAST annotation, (3) predict binding with Meta-BCR, (4) analyze lineage evolution, (5) model structures with AlphaFold. Skill Composition Operators. Atomic skills combine through: • Sequential: s₁ ▷ s₂ (execute s₁ then s₂, pass outputs) • Parallel: s₁ ∥ s₂ (concurrent execution, merge results) • Conditional: s₁ ◁^c s₂ (if condition c then s₁ else s₂) • Loop: s* (repeat until termination condition) File-Based State Synchronization. Sub-agents coordinate through a shared sandbox filesystem: sandbox/logs/, outputs/, temp/, uploads/ Each stage writes persistent reports (retrieval_report.md, hypothesis_report.md, planning_report.md) that subsequent stages can query. This provides both coordination (later stages see earlier results) and auditability (complete provenance chain). Human-in-the-Loop Integration. The orchestrator supports interrupt-based HITL at planning and evaluation stages: if requires_approval(plan): write(hitl_request.json); poll(hitl_response.json); incorporate_feedback() This enables expert oversight of critical decisions without blocking automated execution of routine steps. Meta-BCR architecture and training Meta-BCR was designed to predict BCR antigen specificity and neutralization breadth from single-cell V(D)J repertoires while maintaining performance on datasets that differ substantially from its training distribution. Unlike conventional convolutional neural network (CNN) or graph neural network (GNN) approaches, Meta-BCR integrates meta-learning with semi-supervised learning to enhance adaptability. The network consists of four modules: feature embedding, feature processing, feature selection, and pattern classification. Input data follow the format of XBCR-net and comprise three sequences, immunoglobulin heavy chains, light chains, and antigen sequences, which are encoded into one-hot vectors and index vectors. In the feature embedding module, a ProteinBERT model pre-trained on 400 million immunoglobulin heavy- and light-chain sequences generates contextual embeddings that capture sequence-level BCR features. In the feature processing and selection module, a cross-attention layer integrated into the ACNN-type XBCR-net structure extracts antibody–antigen interaction patterns, followed by global max- pooling for feature selection. In the classification module, residual fully connected layers with multiple parallel outputs enable multitask prediction. The model was trained in three sequential phases. In the transfer-learning phase, the ProteinBERT encoder was pre-trained on masked BCR sequences to initialize the feature embeddings (Supplementary Fig 1C). In the meta-learning phase, a hyperparameter-adaptive variant of the REPTILE algorithm (HA-meta-learning) was applied to antigen-antibody sequence pairs from public databases, producing aggregated gradient updates optimised for generalisation across diverse antigen specificities. In the fine-tuning phase, a mean-teacher semi-supervised learning strategy incorporated both labelled sequences (with experimentally validated specificity or neutralization) and unlabelled sequences from independent single-cell datasets. Performance was benchmarked against state-of-the-art (SOTA) methods, including DeepAAI, for predicting neutralization of SARS-CoV-2 and influenza antibodies. Across metrics (accuracy, area under the receiver-operating characteristic curve [AUC-ROC], F1-score, and precision–recall AUC [PR-AUC]), Meta-BCR with mean-teacher training (Meta-BCR-mt) consistently matched or exceeded SOTA performance, with the largest gains in PR-AUC for influenza neutralization prediction Training-data auditing Sequence overlap between model training data and the two external evaluation cohorts was quantified before any performance comparison was interpreted. Both baselines were retrained on the study corpus rather than applied with published weights: DeepAAI was retrained per five-fold split with epoch selection on validation F1, and the Mason convolutional architecture was retrained per antigen on concatenated heavy and light chain variable sequences, which differs from the CDR-H3 window used in the original report. All three models therefore share a single training corpus (6,631 influenza and 47,845 SARS-CoV-2 heavy chains) and were audited against it jointly; the published DeepAAI corpora (3,051 unique SARS-CoV-2 and 254 unique HIV antibodies, from the authors' repository) were audited separately. Because the per- fold splits were unavailable, overlap was computed against the combined corpus, which upper- bounds the overlap of any constituent fold. CDR-H3 regions were extracted by conserved- junction motif matching, and exact amino-acid identity of the evaluation CDR-H3 to any training CDR-H3 was tabulated. Full heavy-chain similarity was estimated by a 6-mer prefilter followed by normalised sequence-matching against the sixty highest-scoring training candidates, retaining the maximum per evaluation sequence; every evaluation sequence was scored, without sampling. Published DeepAAI and Mason convolutional-network weights were evaluated first and gave ≈0.3 AUC on both baselines, consistent with these weights not having been trained on a dataset comparable to the present study's cohorts; both baselines were therefore retrained on this study's corpus under the same protocol described above. In the influenza (Turner) cohort no sequence matched any training CDR-H3 exactly (0 of 23), none reached 90% full-chain identity, and the highest observed identity was 88.5%. In the SARS-CoV-2 (Guo) cohort one antibody, TH70, carried a CDR-H3 (AREAYGMDV, 9 residues) present verbatim in the training corpus and also in the published DeepAAI SARS- CoV-2 corpus (1 of 285, 0.35%); 54 of 286 sequences (18.88%) reached 90% full-chain identity and one (TH137, 0.35%) reached 95.12%, with none at 99%. Against the published DeepAAI HIV corpus, neither cohort showed any exact CDR-H3 match or any sequence above 90% identity. Because contamination is identical across the three compared models by construction, it cannot account for their relative ranking; the residual risk is that the 18.88% tail at 90% identity in the SARS-CoV-2 cohort inflates absolute performance for all three alike. Similarity was computed as a normalised matching ratio rather than a gapped global alignment, since no alignment toolchain was available in the analysis environment; reported identities therefore approximate, and are not interchangeable with, cluster-based identity. Benchmark Design and Datasets We developed ImmuneBench, a standardized benchmark framework for the systematic evaluation of AI agents on immunology tasks. ImmuneBench comprises 46 questions that collectively assess agent capabilities across antigen binding prediction, neutralization prediction, cross-reactivity identification, clonal expansion analysis, enrichment analysis, and immunological knowledge retrieval. Three AI agents were evaluated: ImmuneAgent, Biomni, and Virtual-Lab. The benchmark incorporated datasets of varying size, modality, and biological scope. The influenza antibody dataset (flu_benchmark) contained 100 antibodies with paired heavy and light chain sequences, for which ground truth binary binding labels were available across six influenza strains (H1N1 A/Michigan/45/2015, H1N1 A/Puerto Rico/8/1934, H1N1 A/Victoria/2570/2019, H1N1 A/Wisconsin/588/2019, H3N2 A/Hong Kong/45/2019, and H3N2 A/Singapore/INFIMH-16-0019/2016), neutralization measurements for five H1N1 strains, polyreactivity against LPS and dsDNA, immunoglobulin heavy chain (IGH) isotype, and B cell subtype annotations. The SARS-CoV-2 antibody dataset (sars_benchmark) comprised 80 unique antibodies (180 rows reflecting multiple measurements per antibody) with binary binding data for four variants (D614G, BA.1, BA.4, wild-type S1), neutralization data for five variants (BA.1, BA.4/5, JN.1, wild-type, XBB.1.5), polyreactivity against four antigens (cardiolipin, LPS, dsDNA, plasmid DNA), IGH isotype, and B cell subtype classifications. A smaller influenza panel (flu_turner_benchmark_v1) contained 23 antibodies from Turner et al. with binding measurements to seven hemagglutinin subtypes (four H1 strains, H2, H5, and H6), seven B cell subtype annotations, and IGH isotype data. Two T cell receptor (TCR) datasets were included. The TCR-ICON dataset (tcr_icon_benchmark) contained 2,080 TCRs with paired CDR3α and CDR3β sequences and binary epitope binding annotations for eight epitopes: MART-1 (melanoma-associated), Flu- MP (influenza matrix protein), five Epstein-Barr virus epitopes, and one cytomegalovirus epitope, along with TRB V-gene usage data. The Zhang TCR dataset (tcr_Zhang_benchmark) comprised 2,000 TCRs from SARS-CoV-2 patients with clonal expansion classifications (Single, Small, Medium, Large, and Hyperexpanded categories) and epitope specificity annotations for five SARS-CoV-2 peptides (VLNDILSRL, YLQPRTFLL, RLNEVAKNL, NLNESLIDL, and SPRRARSVA). Additional datasets provided complementary evaluation axes. The sars_guoyu_benchmark contained 286 SARS-CoV-2 antibodies with continuous ELISA OD450nm binding measurements for Omicron Spike, Omicron RBD, and wild-type Spike. The FDG_benchmark included 398 influenza antibodies with discretized binary binding data. The CoVAbDab_benchmark drew from 12,918 entries in the Coronavirus Antibody Database with text-based neutralization annotations in semicolon-delimited format. The Skempi_benchmark comprised 5,729 protein-protein interface mutations with experimentally determined changes in binding free energy (ΔG, kcal/mol). Question Design and Ground Truth Computation The benchmark questions were organized into 10 categories of increasing complexity. Binding questions assessed simple single-antigen binary binding prediction, where the ground truth label was derived by direct lookup of a single experimental column. Neutralization questions tested single- or multi-column binary neutralization prediction, including conjunctive (AND) logic for questions requiring simultaneous neutralization of multiple variants (e.g., Q07: BA.1 AND XBB.1.5). Neutralizing questions applied a breadth threshold, labeling an antibody as broadly neutralizing if at least 60% of the five H1N1 neutralization columns were positive. Cross-reactive questions required AND logic across subtype groups, such that an antibody was labeled cross-reactive only if it bound at least one strain in each of multiple subtype groups (e.g., any H1N1 binding AND any H3N2 binding). Broadly binding questions used breadth thresholds of 60% across six influenza strains and 50% across nine SARS-CoV-2 binding and neutralization columns, respectively. TCR binding questions assessed binary epitope-specific TCR binding prediction. Clonal expansion questions evaluated binary classification of clonal expansion status, where a TCR was labeled as expanded if its cloneType category was Large (proportion > 0.01 and ≤ 0.1) or Hyperexpanded (proportion > 0.1 and ≤ 1.0); a combinatorial variant required both expansion AND antigen specificity (any epitope column equal to 1). Polyreactivity questions used OR logic across polyreactivity columns, labeling an antibody as polyreactive if it tested positive against any of the measured autoantigens. Enrichment and phenotype questions assessed enrichment of categorical variables including B cell subtypes, IGH isotypes, or TRB V-genes— among functionally defined positive groups. Continuous prediction questions required quantitative predictions of ΔG or sequence similarity. Knowledge questionstested text-based entity recall, text matching, and exact-match accuracy on immunological knowledge retrieval and multiple-choice examination questions. Scoring Metrics Binary classification performance was assessed using five metrics. The F1 score served as the primary metric and was computed using scikit-learn (v1.6.1) f1_score with zero_division set to 0. Precision, recall, and balanced accuracy were computed as secondary metrics. The area under the receiver operating characteristic curve (AUC-ROC) was reported when agent probability scores were available, computed using scikit-learn's roc_auc_score. Enrichment and phenotype question performance was assessed using entity recall (or F1) and Spearman rank correlation. For each enrichment question, the ground truth’s lowest non-zero eligible group fraction was used as a universal enrichment baseline, anchoring the threshold to the biology of the dataset rather than an agent’s overall prediction rate. Specifically, for a given categorical variable (e.g., B cell subtype, IGH isotype, or V-gene), the fraction of broadly binding (or neutralizing) antibodies was computed for each group in both the ground truth and the agent’s predictions. Groups with fewer than 5 members (min_group_n = 5) were excluded from enrichment analysis to avoid noise from small samples, and groups with 0% positive rate in the ground truth were not used as baseline references, as depleted groups do not provide a meaningful enrichment anchor. A group was classified as “enriched” when its fraction of positive antibodies exceeded the baseline. Both ground truth enrichment and agent-predicted enrichment were evaluated against the same ground-truth-derived baseline, ensuring that agents must reach biologically meaningful enrichment levels to receive credit. Entity recall and F1 measured the overlap between ground-truth-enriched and agent-enriched groups, while enrichment Spearman correlation measured the rank-order agreement of per-group positive fractions between the agent and the ground truth. Chi-square tests were additionally applied as a global test for non-random association between the categorical variable and broadly binding status. Text and knowledge question performance was assessed by entity recall for feature identification tasks and exact-match accuracy for multiple-choice questions. CD19+ B-cell isolation PBMCs were isolated from donor blood by standard density gradient (Lymphoprep) separation with centrifugation at 1,000 × g for 20 minutes without braking at room temperature. The cells were washed twice with RPMI and centrifuged at 500 × g for 5 minutes at room temperature. They were then counted with CounterStar slides and an automated haemocytometer counter before another wash in FACS buffer. The PBMCs were then coincubated with anti-CD19 antibodies and magnetic beads for CD19-positive cell enrichment, as described previously. The enriched CD19-positive cells were then cultured in RPMI + 10% FBS at 37°C for 2 hours to rest. Anti-CD19, fluorescently labelled SARS-CoV-2 spike and live-dead stains were mixed and added to the CD19-positive cells, which were subsequently loaded into a Sony M800 sorter for the isolation of CD19+SARS-CoV-2-Spike+ live cells. Single-cell V(D)J repertoire library preparation and sequencing The sorted CD19-positive live cells were collected in PBS supplemented with 30% FBS, washed with PBS, and centrifuged at 400 × g for 5 minutes. The cells were then counted using a manual haemocytometer and adjusted to 1,000 cells/μl, with a total of 20 μl for loading into 10x Genomics Chromium device chips following the manufacturer’s protocol. The 10x Genomics libraries were prepared using the 10x Genomics Chromium Single Cell 5’ Library Construction Kit and sequenced on an Illumina NovaSeq instrument. Differential expression analysis We used the FindMarkers function to identify differentially expressed genes between two groups and applied the Wilcoxon test to evaluate the significance of each gene, with multiple hypothesis correction applied via the Benjamini‒Hochberg procedure. Genes with adjusted P values less than 0.05 and absolute values of log2-fold change (log2FC) greater than 0.25 were considered differentially expressed genes. Cell trajectory inference Monocle3, Diffusion Map and Slingshot were used to perform developmental trajectory inference of B cells. We applied trajectory inference analyses based on different algorithms to provide corroborative evidence. Monocle3 (v1.0.0) generated the trajectory using the principal graph algorithm after dimension reduction. Diffusion Map was used to perform nonlinear dimensionality reduction on the snRNA-seq data and capture the differentiation trajectories of classical monocytes via R package destiny (v3.14.0). Pseudotime was estimated using the R package slingshot (v2.8.0), which applies a simultaneous principal curve algorithm to construct smooth curves from a minimum spanning tree. Antibody cloning and purification The method was described in detail in previous studies. The variable regions of paired heavy and light chains were synthesized, cloned and inserted into vectors expressing the constant region of human IgG1. These vectors were then transfected into 293F cells with 1 mg/mL polyetherimide (PEI). The supernatant was collected on day 5 and purified using protein A- embedded columns. The antibody concentration was measured with a NanoDrop system. Antibody sequence analysis The antibody sequences were analysed using IMGT-high-V-quest to obtain detailed information on V gene usage and the CDR and FR region sequences. The sequences were then processed with ChangeO and VDJtools as described in earlier papers. V and J gene usage was plotted with the VDJtools package, and the statistical analyses were performed using Prism 8.0. The CDR-H3 length and sequence dominance were analysed and plotted using the VDJtools package. To cluster CDR-H3, antibodies with the same CDR-H3 length and more than 80% similarity were allocated to the same cluster. The antibodies in different clusters, therefore, should have more than a 20% difference in CDR-H3. Sequence logos were generated with an online Python-based tool (http://weblogo.berkeley.edu/logo.cgi). Anti-RBD ELISA NUNC 96-well plates were coated with 1 μg/mL RBD of SARS-CoV-2, SARS-CoV-2 Omicron or SARS-CoV overnight at 4°C. The plates were then washed with PBST (0.1% Tween) three times before being blocked with 3% BSA for one hour on a shaker. mAbs were diluted to 1 μg/mL and coincubated with the coated antigens for one hour before washing and subsequent coincubation with anti-human IgG-AP (Sigma). The colour was developed with PNPP tablets (Sigma), and the optical density at 450 nm (OD450) was recorded with an ELISA microplate reader. An antibody was defined as ELISA positive when the OD450 (1 μg mL−1 RBD) was three times greater than that of the negative control using an H7N9-specific human IgG1 antibody (HG1K, Sino Biological). Pseudovirus neutralization assay The pseudovirus neutralization procedure has been described in detail in previous studies11, 20. mAbs were diluted to six gradients in total DMEM and mixed with pseudovirus solution in a 96-well plate. Huh7 cells were cultured in high-glucose DMEM supplemented with 10% FBS, 100 U/mL penicillin‒streptomycin, and 20 mM HEPES in 37°C incubators. The Huh7 cells were trypsinized and seeded onto mAb-pseudovirus plates at 2 × 105 cells/mL. After 24 hours of incubation at 37°C, the culture supernatant was carefully aspirated such that 100 μl remained in each well. The luciferase substrate was then added to the wells, and 150 μl of cell lysate was transferred to a white opaque 96-well microplate for luciferase detection with a microplate luminometer. Two replicates were performed for each group of plasma/serum samples. Influenza virus microneutralization assay Focus-forming assay (FFA) is a methodology employed to quantify the neutralization ability of mAbs against H1N1 and H5N1 influenza viruses. Prior to initiating the assay, it is essential to predetermine the virus titre in focus-forming units (FFUs). Preparation involved plating MDCK cells in a 96-well flat-bottom cell culture plate at a density of 0.03 million cells per well and incubating them overnight in a 37°C incubator with 5% CO2. On the subsequent day, the mAbs were diluted to a concentration of 40 μg/mL in DMEM, while the virus was diluted to 100 FFU/50 μL in DMEM supplemented with 25 mM HEPES and 1 μg/mL TPCK trypsin. These diluted solutions were then mixed in equal proportions and incubated together at 37°C for one hour. Following this incubation, 100 μL of the mixed solution was added to the MDCK cell- containing 96-well plate (after the supernatant was aspirated and the cells were washed twice with PBS). The plate was then further incubated at 37°C for an additional hour. After the second incubation, the supernatant was aspirated, and the cells were washed once with PBS. Subsequently, 100 μL of overlay medium, prepared by mixing 2% CMC (2X) and MEM at a 1:1 volume ratio and supplemented with a final concentration of 2 μg/mL TPCK trypsin, was added to each well. The plate was then incubated at 37°C for 36 hours. Upon completion of the reaction, each well was fixed with 200 μL of 4% paraformaldehyde at room temperature for 30 minutes. The fixation solution was aspirated, and the plate was washed three times with PBS. The cells were then simultaneously blocked and permeabilized with 1% bovine serum albumin (BSA) containing 0.2% Triton at 50 μL per well for 30 minutes at room temperature, followed by three washes with PBS. Next, 50 μL of an anti-H1N1-N antibody (clone number: 4R5) diluted 1:4000 in 1% BSA was added to each well, and the plate was incubated at 37°C for one hour. The primary antibody was then aspirated, and the plate was washed three times with PBST (0.1% Tween-PBS). Goat anti-mouse IgG-HRP diluted 1:2000 in 1% BSA was then added to each well at 50 μL, and the plate was incubated at 37°C for another hour. Next, the secondary antibody was aspirated, and the plate was washed three times with PBST. TrueBlue (KPL, cat. 50-78-02) was used for colour development at 50 μL per well for 15 minutes at room temperature. The plate was then washed three times with double-distilled water (ddH2O) and allowed to dry for 4 hours, after which spot counting was performed using an enzyme-linked spot analyser (AT-Stop1100/CTL). This refined protocol ensures the precision and reproducibility of the FFA in assessing the neutralization reaction between H1N1 influenza viruses and mAbs. Statistical analysis Statistical analysis was performed and is specified in the figure legends. IC₅₀ was determined by four-parameter nonlinear regression using GraphPad Prism 9.0. Data and code availability • Single-cell RNA–seq data have been deposited in BioProject PRJCA038672 and are publicly available as of the date of publication. • Cryo-EM structures have been deposited in the PDB under accession numbers 9UHF, 9UHG, 9UHH, 9UHO and 9UHQ. • All original code and models are available on GitHub: https://github.com/hantaolou/ImmuneAgent and https://github.com/jianqingzheng/meta_bcr Supporting information Document S1. Figures S1–S8 and Table S1 Acknowledgements We would like to thank Prof. Xianwen Ren for reading and commenting on the manuscript and Prof. Jun Pu for sharing BCR sequences related to the single-cell RNA dataset. Ethics approval was granted under NKUIRB2023144. This work was supported by grants from the Tianjin Science and Technology Commission for grant 24ZXZSSS00490, National Key R&D Program of China (2023YFA1801400), the National Natural Science Foundation of China 32300780, the CAMS Innovation Fund for Medical Sciences (2024-I2M-ZD-005, 2021-I2M-1-017) and Major Projects of Guangzhou National Laboratory (Grant No. GZNL2023A01006, SRPG2200206). Author contributions Conceptualization, H.L., X.W., X.C. and J.Z.; methodology, H.L., J.Z., C.Y., Q.Z. and Y.B.; Investigation, H.L., J.Z., C.Y., Q.Z., M.Z., Y.B., Y.Y., B.L., Y.C., X.C., Z.L., R.Y., Y.M. and Y.S.; writing original draft, H.L., J.Z., C.Y., X.C., Q.Z. and X.W.; writing—review & editing, H.L., J.Z., Q.Z., X.W. and X.C.; funding acquisition, H.L., X.W. and X.C.; resources, L.W., Y.Q., S.L., Z.L., C.W., H.L., X.W. and X.C.; supervision, H.L., X.W. and X.C. Conflict of interest The authors declare that they have no conflicts of interest. Extended Data Extended Data Fig. 1-8 Supplementary Table S1 Supplementary Table S2 Figure. S1 | Detailed workflow architecture of ImmuneAgent’s deep research and orchestration modules. A. Deep Research Workflow. Schematic of the six-stage research pipeline from user query input to final experimental plan output. B. Framework of the orechestrator agent and sub- agent communications C. Meta-BCR architecture for pan-viral BCR specificity training. Fig. S2 | Extended benchmarking of ImmuneAgent and baseline embedding comparisons. A, B. Reasoning process by ImmuneAgent for HLE benchmark questions C, D. Results plotted by execution agent for ImmuneBench. Figure. S3 A, Ranking table displaying human expert evaluation scores (1-5 Likert scale, normalized to percentage) across eight criteria assessed by three blinded immunologists. Criteria include Scientific Rigor, Innovation Score, Validations, Hypothesis Quality, Planning Quality, Readiness, and Biological Feasibility. B, Knowledge Graph generated by ImmuneAgent deep research and hypothesis according to query. C, Planning and execution log by ImmuneAgent for immunological questions. D, UMAP projections showing poor BCR clustering by baseline models. DeepAAI (top, silhouette=0.23) and CNN (bottom, silhouette=0.18) fail to separate H1N1 binders (red) from non-binders (gray), unlike ImmuneAgent, n=1,847 sequences. Fig. S4 | ImmuneAgent identifies broadly neutralizing antibodies from single-cell repertoires. A, UMAP of longitudinal B cell responses from ChAdOx1-nCoV-19 (ADV) vaccine recipients (n=8 individuals, 12,847 cells across three timepoints). B, Gene-usage analysis of the influenza specific antibodies predicted by ImmuneAgent. C, Alphafold3 modelled complexes of the representative antibody with RBD of SARS-CoV-2. D, Cryo-EM structure of BA5-39 in complex with SARS-CoV-2 JN.1 E, UMAP of integrated scRNA-seq and scV(D)J-seq data from PBMCs and lymph nodes of quadrivalent influenza vaccine (QIV) recipients (n=12 individuals, 9,412 B cells). F, V(D)J gene usage of SARS-CoV-2 antibodies. Dot size represents clone frequency; color indicates neutralization breadth against ten variants including W.T, Alpha, Beta, Delta, and Omicron variants. G, Alphafold3 modelled complex of HA trimer with representative anti-HA antibodies from the single cell dataset. H, Neutralization of the ImmuneAgent predicted antibodies against four H1N1 strains. Fig. S5| Histopathological analysis of lung tissue following prophylactic antibody treatment and influenza challenge. A-H, Representative whole-lung H&E-stained sections at day 5 post-infection with A/California/07/2009 (H1N1). Boxed regions highlight representative areas of pathological changes (where present) or preserved tissue architecture. All sections from left lung lobe fixed in 4% paraformaldehyde for 12 hours. Scale bars: 500 μm (low magnification). Fig. S6 | Cellular and Structural Characteristics for broadly neutralizing antibodies. A, UMAP of the integrated B cell compartment (459,128 B cells) pooled across the paired single-cell V(D)J and RNA-sequencing datasets analysed by ImmuneAgent. B, Volcano plot of genes differentially expressed in ImmuneAgent-predicted broadly reactive B cells relative to non-reactive B cells, titled "Upregulated DEGs in bnab B cell. C, Flow cytometry of B cells stained in parallel with wild-type and XBB spike probes. D, Lineage tree of the broadly neutralizing clonotype comprising BA5-12, sc_197, sc_931, sc_527 and BA5-39, rooted on the inferred germline. E, Measured pseudovirus neutralization IC₅₀ (ng mL⁻¹, printed as ng/ml) for three lineage members and their germline-reverted variants against four SARS- CoV-2 pseudoviruses. F, Structural comparison of two published anti-influenza haemagglutinin antibodies bound to HA1, the strain-specific D1 H1-17/H3-14 (PDB 6XPQ) and the broadly neutralizing D2 H1-1/H3-1 (PDB 6XPR). G, Neutralization matrix for the same two antibodies across twelve influenza A haemagglutinin subtypes, arranged by HA phylogenetic group. Fig. S7 | ImmuneAgent generalize to untrained antigens A, Schematic of the ImmuneAgent closed loop for the RSV–hMPV task, titled: ImmuneAgent agentic workflow: online deep research followed by hypothesis that initiated planning and CodeAct execution. B, UMAP of the integrated B cell atlas assembled for this task (n = 45,857 cells), coloured by the 16 annotated subsets. C, ImmuneAgent CodeAct execution log for the HPV run. D, Neutralizing antibodies’ cell of origin, SHM, clonal expansion, cofolding results. E. Neutralizing results of the 9 antibodies. F. Interaction analysis of the representative antibodies Supplementary Table S1 | Statistics for cryo-EM data collection, refinement, and validation. Data collection Complex JN.1 RBD+BA5- 12 JN.1 RBD+BA5- 12-CDR1 D614G RBD+BA5- 12-GR JN.1 Spike+BA5- 39 JN.1 RBD+BA5- 39 (Local refinement) Microscope FEI Titan FEI Talos Arctica FEI Titan FEI Titan FEI Titan Camera Gatan K3 Gatan K2 Gatan K2 Gatan K3 Gatan K3 Voltage (kV) 300 200 300 300 300 Total dose (e - /A 2 ) 60 60 60 60 60 Micrographs (total) 4,481 3,209 2,850 4,897 4,897 Micrographs (used) 4,481 3,209 2,850 4,897 4,897 Particles selected 252,119 294,014 147,454 1,074,974 1,055,814 Particles included in final reconstruction 94,590 183,315 109,000 250,932 161,866 Sampling, Å per pixel 1.07 1.0 1.04 1.07 1.07 Defocus range (μm) -1.2 to -2.0 -1.2 to -2.0 -1.2 to -2.0 -1.2 to -2.0 -1.2 to -2.0 Symmetry C1 C1 C1 C1 C1 Resolution (Å) (FSC=0.143 criterion) 4.18 3.91 4.05 3.69 4.62 Model refinement Ramachandran statistics (%) Most favoured 93.30 92.11 94.70 96.76 93.78 Allowed 6.70 7.89 5.30 3.21 6.22 Outliers 0.00 0.00 0.00 0.03 0.00 Bonds (RMSD) Bond length (Å) 0.003 0.005 0.002 0.003 0.004 Bond angle (°) 0.656 0.807 0.605 0.580 0.836 MolProbity score 2.03 2.22 1.78 1.59 2.04 Clash score 12.10 17.31 7.62 7.01 12.93 Rama-Z Rotamer outliers (%) 0 0.27 0 0.03 0 Cβ outliers (%) 0 0 0 0 0 Deposit PDB 9UHF 9UHG 9UHH 9UHO 9UHQ EMDB 64162 64163 64164 64171 64172 Supplementary Table S2 ELISA (OD) 1 g/ml Microneutralization 10 g/ml mAbs H1N1 Victoria H1N1 Wisconsin H1N1 A/Puerto Rico/8/34 H1N1 A/Michigan /45/2015 WS67 MCG45 CA07 PR8 INF-1 2.4814 2.3121 0.7107 0.3946 0.0 0 0 0 INF-2 0.5075 0.3986 0.1305 0.1372 NB NB NB NB INF-3 0.2628 0.2960 0.0818 0.1685 NB NB NB NB INF-4 3.1538 3.2536 1.6335 1.2840 84.2 65 81 27 INF-5 1.3295 0.3148 0.1509 0.1957 0.0 0 0 0 INF-6 0.0700 0.0616 0.0446 0.0821 NB NB NB NB INF-7 0.4391 1.3974 0.0580 1.1510 NB NB NB NB INF-8 0.4511 0.5553 0.1545 0.3712 NB NB NB NB INF-9 3.3296 2.9813 1.9076 1.3814 106.6 98 92 83 INF-10 3.4133 3.1272 1.7651 1.1250 93.7 96 99 73.8 INF-11 3.3740 2.9816 1.5831 1.1008 99.9 95 99 72.4 INF-12 3.5019 3.3111 1.8003 0.8233 109.4 98 96 68.5 INF-13 3.2481 3.0972 1.6337 1.1601 82.0 97 78 48.3 INF-14 3.4771 2.8805 1.6752 1.1408 107.5 99 100 83.7 INF-15 3.2534 2.8042 1.5974 0.9855 83.7 96 100 80.15 INF-16 0.3525 0.3323 0.0806 0.2319 NB NB NB NB INF-17 3.5206 2.7788 1.4114 1.1356 95.4 92 97 81 INF-18 1.7338 1.9617 0.2530 0.4527 12.1 0 0 12 INF-19 1.9187 1.7169 0.2165 1.2465 0.0 0 0 23.12 INF-20 0.3919 0.6456 0.0624 0.1013 NB NB NB NB INF-21 3.3593 2.9777 1.8285 1.2362 0.0 0 0 9 INF-22 0.4340 2.7984 0.1429 0.1447 6.2 0 0 0 INF-23 1.6181 1.8343 0.3369 0.3000 4.1 0 0 0 INF-24 1.0788 1.2828 0.3313 0.2354 6.2 0 0 0 INF-25 3.1775 2.6901 0.7705 1.1272 0.0 0 0 23 INF-26 1.0851 0.9468 0.2436 0.1806 NB NB NB NB INF-27 1.5761 1.1139 0.2046 0.2429 1.4 0 0 19.7 INF-28 2.3903 2.7568 0.2602 0.4397 45.8 0 0 11 INF-29 1.6378 1.6428 0.4440 0.4957 0.0 0 0 12 INF-30 2.8687 2.7053 1.0234 0.9504 21.1 0 0 9 INF-31 3.3653 3.3557 1.7400 1.0360 99.4 99 100 91 INF-32 0.4968 0.6506 0.0744 0.1681 NB NB NB NB INF-33 2.4920 2.7787 1.5634 0.9184 85.2 76.4 85.29 52.9 INF-34 2.9246 2.9277 1.6503 1.0228 65.0 75 41 52.6 INF-35 2.3866 2.7162 0.6283 1.1412 68.0 33 0 35.5 INF-36 3.4341 3.1896 1.6162 1.2377 98.9 100 100 96 INF-37 3.0285 3.0771 1.6342 1.0689 93.1 100 78.57 80.3 INF-38 2.8838 3.2036 1.6598 1.0218 92.7 84.2 88.24 80.6 INF-39 2.2881 2.2811 0.7224 0.3394 66.0 0 0 6 INF-40 1.7030 1.9078 0.5464 0.2904 12.0 0 0 0 INF-41 3.0577 2.5319 0.8828 1.1582 0.0 0 0 0 INF-42 3.0078 3.1272 2.5629 2.4441 31.0 21 83 30 INF-43 2.2502 1.9087 3.0235 3.4277 18.12 0 0 0 INF-44 3.1074 3.5740 2.4368 2.4308 94.8 100 100 82.35 INF-45 3.3705 3.9746 2.4999 2.7700 49.1 73.5 77 18.6 INF-46 0.2749 0.2642 NT NT NB NB NB NB INF-47 3.2837 3.7496 2.9093 2.6398 43.8 53.6 79.4 21.1 INF-48 0.2017 0.1632 NT NT NB NB NB NB INF-49 3.1615 2.4392 2.8195 2.6596 70.4 61.0 88.2 56.7 INF-50 3.2711 2.6747 2.6767 2.4505 75.2 38.0 77.0 0.0 Binding (ELISA) 1 g/ml Neutralization5 g/ml mAbs D614G WT BA.1 BA.4/5 WT BA.1 BA.4/5 XBB.1. 5 BA5_1 0.19 0.34 0.75 0.95 100% 12% 100% 20% BA5_2 0.65 4.98 1.03 4.14 100% 96% 92% 70% BA5_3 3.70 5.45 4.31 2.49 100% 11% 23% 18% BA5_4 3.45 5.29 3.74 2.02 100% 12% 31% 11% BA5_5 2.87 4.32 0.79 0.52 100% 100% 100% 75% BA5_6 2.11 4.62 1.08 1.69 40% 100% 100% 60% BA5_7 2.18 4.60 0.87 0.40 100% 90% 41% 72% BA5_8 2.52 4.88 1.63 1.57 100% 81% 100% 100% BA5_9 2.46 4.50 0.80 0.68 100% 92% 100% 15% BA5_10 2.67 5.12 2.53 1.40 100% 100% 100% 15% BA5_11 2.96 4.89 1.91 1.26 100% 90% 67% 20% BA5_12 2.84 4.81 1.01 1.80 100% 100% 100% 100% BA5_13 2.73 4.72 1.84 1.76 39% 96% 87% 71% BA5_15 3.06 2.29 2.90 3.14 100% 100% 100% 49% BA5_16 2.94 2.16 1.32 3.13 100% 87% 86% 55% BA5_17 0.07 0.09 0.19 0.30 43% 19% 34% 56% BA5_18 2.86 2.13 0.16 2.80 100% 59% 38% 48% BA5_19 2.94 2.00 1.85 2.97 100% 95% 99% 57% BA5_20 2.76 2.11 1.78 3.42 100% 95% 100% 83% BA5_21 2.72 1.97 3.31 3.41 82% 98% 0% 82% BA5_22 2.87 2.11 0.78 2.79 100% 56% 62% 49% BA5_23 2.92 2.12 2.68 3.17 100% 100% 93% 47% BA5_24 2.89 2.10 2.87 3.15 100% 100% 99% 60% BA5_25 2.86 2.24 2.87 2.80 100% 93% 52% 65% BA5_26 3.09 2.02 1.50 3.51 100% 92% 100% 64% BA5_27 0.07 0.08 0.13 0.16 39% 52% 48% 69% BA5_28 3.00 2.24 2.76 3.40 100% 100% 99% 0% BA5_29 2.97 2.14 3.04 3.66 83% 98% 75% 29% BA5_30 2.92 2.07 2.57 3.46 100% 100% 100% 98% BA5_31 3.01 2.17 3.25 3.13 100% 100% 99% 80% BA5_32 0.16 0.16 0.24 0.44 33% 49% 40% 63% BA5_33 3.00 2.25 2.78 3.13 100% 99% 90% 50% BA5_34 2.95 2.14 1.33 2.88 100% 96% 89% 55% BA5_35 2.94 2.17 3.13 3.30 83% 96% 65% 60% BA5_36 2.94 2.14 2.48 3.19 100% 100% 100% 78% BA5_37 0.54 0.44 0.67 0.86 30% 59% 31% 63% BA5_38 3.09 2.12 3.11 3.21 84% 98% 31% 68% BA5_39 2.85 2.27 2.10 3.14 100% 95% 100% 97% BA5_40 0.12 0.14 0.23 0.41 78% 55% 45% 63% BA5_41 0.26 0.16 0.28 0.34 75% 53% 39% 66% BA5_42 2.69 1.90 3.06 3.25 95% 99% 99% 82% BA5_43 0.19 0.28 0.14 0.16 78% 43% 25% 56% BA5_44 2.54 2.09 2.17 2.99 99% 54% 12% 60% BA5_45 0.39 0.43 0.46 0.79 78% 64% 25% 73% BA5_46 2.64 1.72 2.29 3.03 100% 100% 100% 69% BA5_47 2.65 1.96 2.56 3.20 100% 100% 100% 74% BA5_48 2.70 2.06 3.35 3.29 95% 96% 100% 79% BA5_49 0.10 0.19 0.24 3.42 85% 61% 38% 62% BA5_50 0.06 0.54 0.11 0.15 82% 73% 93% 99% BA5_51 0.07 0.09 0.10 0.14 87% 61% 20% 54% BA5_52 2.99 2.23 1.48 3.23 100% 98% 99% 65% BA5_53 0.25 0.18 0.22 0.38 82% 56% 63% 44% BA5_54 1.11 0.41 0.38 2.44 93% 73% 41% 66% BA5_55 2.98 2.05 3.33 3.30 93% 96% 80% 83% BA5_56 2.87 2.13 3.23 3.55 95% 100% 88% 59% BA5_57 2.69 1.88 3.00 3.29 100% 100% 100% 99% BA5_59 2.76 1.97 3.21 3.40 97% 97% 68% 92% BA5_60 3.00 2.21 3.45 3.32 96% 95% 46% 69% References 1. Abu-Shmais, A. A., Freeman, G., Creanga, A., Vukovich, M. J., Malla, T., Mantus, G. E., Shimberg, G. D., Gillespie, R. A., Guerra Canedo, V., Dadonaite, B., Rodgers, M. D., Chopde, A. J., Bardwil-Lugones, E., Bylund, T., Henry, A. R., Roberts-Torres, J., Johnston, T. S., Smith, S., Yang, E. S.,...Andrews, S. F. 2025. Cross-neutralizing and potent human monoclonal antibodies against historical and emerging h5nx influenza viruses. Nat Microbiol. 10, 2903-2918. https://doi.org/10.1038/s41564-025-02137-x 2. Bennett, N. R., Watson, J. L., Ragotte, R. J., Borst, A. J., See, D. L., Weidle, C., Biswas, R., Yu, Y., Shrock, E. L., Ault, R., Leung, P. J. Y., Huang, B., Goreshnik, I., Tam, J., Carr, K. D., Singer, B., Criswell, C., Wicky, B. I. M., Vafeados, D.,...Baker, D. 2026. Atomically accurate de novo design of antibodies with rfdiffusion. Nature. 649, 183-193. https://doi.org/10.1038/s41586-025-09721-5 3. Bod, L., Kye, Y. C., Shi, J., Torlai Triglia, E., Schnell, A., Fessler, J., Ostrowski, S. M., Von-Franque, M. Y., Kuchroo, J. R., Barilla, R. M., Zaghouani, S., Christian, E., Delorey, T. M., Mohib, K., Xiao, S., Slingerland, N., Giuliano, C. J., Ashenberg, O., Li, Z.,...Kuchroo, V. K. 2023. B-cell-specific checkpoint molecules that regulate anti-tumour immunity. Nature. 619, 348-356. https://doi.org/10.1038/s41586-023-06231-0 4. Center for Ai Safety, Scale Ai, & Consortium., H. C. 2026. A benchmark of expert-level academic questions to assess ai capabilities. Nature. 649, 1139-1146. https://doi.org/10.1038/s41586-025-09962-4 5. Corti, D., Purcell, L. A., Snell, G., & Veesler, D. 2021. Tackling covid-19 with neutralizing monoclonal antibodies. Cell. 184, 4593-4595. https://doi.org/10.1016/j.cell.2021.07.027 6. Damodaran, L., Jaeger, A. S., & Moncla, L. H. 2026. Ecology and spread of the north american h5n1 epizootic. Nature. 649, 432-441. https://doi.org/10.1038/s41586-025- 09737-x 7. Graham, B. S., & Sullivan, N. J. 2018. Emerging viral diseases from a vaccinology perspective: Preparing for the next pandemic. Nat Immunol. 19, 20-28. https://doi.org/10.1038/s41590-017-0007-9 8. He, H., He, B., Guan, L., Zhao, Y., Jiang, F., Chen, G., Zhu, Q., Chen, C. Y., Li, T., & Yao, J. 2024. De novo generation of sars-cov-2 antibody cdrh3 with a pre-trained generative large language model. Nat Commun. 15, 6867. https://doi.org/10.1038/s41467- 024-50903-y 9. Huang, K., Zhang, S., Wang, H., Qu, Y., Lu, Y., Roohani, Y., Li, R., Qiu, L., Li, G., Zhang, J., Yin, D., Marwaha, S., Carter, J. N., Zhou, X., Wheeler, M., Bernstein, J. A., Wang, M., He, P., Zhou, J.,...Leskovec, J. 2025. Biomni: A general-purpose biomedical ai agent. bioRxiv. 10. Ishaan Gulrajani, D. L.-P. 2020. In search of lost domain generalization. arXiv. https://doi.org/10.48550/arXiv.2007.01434 11. Jabbour, E., Paul, S., & Kantarjian, H. 2021. The clinical development of antibody-drug conjugates - lessons from leukaemia. Nat Rev Clin Oncol. 18, 418-433. https://doi.org/10.1038/s41571-021-00484-2 12. Jimenez, C. E., Yang, J., Wettig, A., Yao, S., Pei, K., Press, O., & Narasimhan, K. (2024). Swe-bench: Can language models resolve real-world github issues? https://proceedings.iclr.c/paper_files/paper/2024/file/edac78c3e300629acfe6cbe9ca88fb 84-Paper-Conference.pdf 13. Jin, R., Guo, Y.-C., Qu, Y., Yang, M., Shang, C., Yang, Q., Chao, L., Zhou, Y., Xu, R., Xu, Z., Zhou, R., Zhang, Z., Wang, M., Zhang, X., & Cong, L. 2025. Biolab: End-to-end autonomous life sciences research with multi-agents system integrating biological foundation models. bioRxiv. 14. Kallewaard, N. L., Corti, D., Collins, P. J., Neu, U., Mcauliffe, J. M., Benjamin, E., Wachter-Rosati, L., Palmer-Hill, F. J., Yuan, A. Q., Walker, P. A., Vorlaender, M. K., Bianchi, S., Guarino, B., De Marco, A., Vanzetta, F., Agatic, G., Foglierini, M., Pinna, D., Fernandez-Rodriguez, B.,...Skehel, J. J. 2016. Structure and function analysis of an antibody recognizing all influenza a subtypes. Cell. 166, 596-608. https://doi.org/10.1016/j.cell.2016.05.073 15. King, H. W., Orban, N., Riches, J. C., Clear, A. J., Warnes, G., Teichmann, S. A., & James, L. K. 2021. Single-cell analysis of human b cell maturation predicts how antibody class switching shapes selection dynamics. Science Immunology. 6, eabe6291. https://doi.org/doi:10.1126/sciimmunol.abe6291 16. Kochayoo, P., Moriyama, S., Kotaki, R., Thawornpan, P., Malee, C., Leepiyasakulchai, C., Ntumngia, F. B., Adams, J. H., Takahashi, Y., & Chootong, P. 2025. Atypical memory b cells from natural malaria infection produced broadly neutralizing antibodies against plasmodium vivax variants. PLOS Pathogens. 21, e1012866. https://doi.org/10.1371/journal.ppat.1012866 17. Laidlaw, B. J., & Ellebedy, A. H. 2022. The germinal centre b cell response to sars-cov-2. Nat Rev Immunol. 22, 7-18. https://doi.org/10.1038/s41577-021-00657-1 18. Laurent, J. M., Janizek, J. D., Ruzo, M., Hinks, M. M., Hammerling, M. J., Narayanan, S., Ponnapati, M., White, A. D., & Rodriques, S. G. (2024). Lab-bench: Measuring capabilities of language models for biology research. In. 19. Lou, H., Zhang, M., Zhang, B., Lu, Q., Zheng, J., & Cao, X. 2026. Multimodal computational framework identifies b cell convergence in autoimmunity and ageing. bioRxiv. 20. Lu, Y., Wang, Y., Xu, M., Xie, B., Yang, Y., Xu, H., & Suo, S. 2026. Assessment of computational methods in predicting tcr-epitope binding recognition. Nat Methods. 23, 248-259. https://doi.org/10.1038/s41592-025-02910-0 21. Luo, Y., Shi, L., Li, Y., Zhuang, A., Gong, Y., Liu, L., & Lin, C. 2025. From intention to implementation: Automating biomedical research via llms. Science China Information Sciences. 68, https://doi.org/10.1007/s11432-024-4485-0 22. Mason, D. M., Friedensohn, S., Weber, C. R., Jordi, C., Wagner, B., Meng, S. M., Ehling, R. A., Bonati, L., Dahinden, J., Gainza, P., Correia, B. E., & Reddy, S. T. 2021. Optimization of therapeutic antibodies by predicting antigen specificity from antibody sequence via deep learning. Nature Biomedical Engineering. 5, 600-612. https://doi.org/10.1038/s41551-021-00699-9 23. Moir, S., & Fauci, A. S. 2009. B cells in hiv infection and disease. Nat Rev Immunol. 9, 235-245. https://doi.org/10.1038/nri2524 24. Nellore, A., Zumaquero, E., Scharer, C. D., Fucile, C. F., Tipton, C. M., King, R. G., Mi, T., Mousseau, B., Bradley, J. E., Zhou, F., Mutneja, S., Goepfert, P. A., Boss, J. M., Randall, T. D., Sanz, I., Rosenberg, A. F., & Lund, F. E. 2023. A transcriptionally distinct subset of influenza-specific effector memory b cells predicts long-lived antibody responses to vaccination in humans. Immunity. 56, 847-863 e848. https://doi.org/10.1016/j.immuni.2023.03.001 25. Phad, G. E., Pinto, D., Foglierini, M., Akhmedov, M., Rossi, R. L., Malvicini, E., Cassotta, A., Fregni, C. S., Bruno, L., Sallusto, F., & Lanzavecchia, A. 2022. Clonal structure, stability and dynamics of human memory b cells and circulating plasmablasts. Nat Immunol. 23, 1076-1085. https://doi.org/10.1038/s41590-022-01230-1 26. Pinto, D., Sauer, M. M., Czudnochowski, N., Low, J. S., Tortorici, M. A., Housley, M. P., Noack, J., Walls, A. C., Bowen, J. E., Guarino, B., Rosen, L. E., Di Iulio, J., Jerak, J., Kaiser, H., Islam, S., Jaconi, S., Sprugasci, N., Culap, K., Abdelnabi, R.,...Veesler, D. 2021. Broad betacoronavirus neutralization by a stem helix-specific human antibody. Science. 373, 1109-1116. https://doi.org/10.1126/science.abj3321 27. Qu, Y., Huang, K., Yin, M., Zhan, K., Liu, D., Yin, D., Cousins, H. C., Johnson, W. A., Wang, X., Shah, M., Altman, R. B., Zhou, D., Wang, M., & Cong, L. 2026. Crispr-gpt for agentic automation of gene-editing experiments. Nat Biomed Eng. 10, 245-258. https://doi.org/10.1038/s41551-025-01463-z 28. Rappazzo, C. G., Hsieh, C. L., Rush, S. A., Esterman, E. S., Delgado, T., Geoghegan, J. C., Wec, A. Z., Sakharkar, M., Mas, V., Mclellan, J. S., & Walker, L. M. 2022. Potently neutralizing and protective anti-human metapneumovirus antibodies target diverse sites on the fusion glycoprotein. Immunity. 55, 1710-1724 e1718. https://doi.org/10.1016/j.immuni.2022.07.003 29. Schellenbacher, C., Roden, R., & Kirnbauer, R. 2009. Chimeric l1-l2 virus-like particles as potential broad-spectrum human papillomavirus vaccines. J Virol. 83, 10085-10095. https://doi.org/10.1128/jvi.01088-09 30. Sharonov, G. V., Serebrovskaya, E. O., Yuzhakova, D. V., Britanova, O. V., & Chudakov, D. M. 2020. B cells, plasma cells and antibody repertoires in the tumour microenvironment. Nat Rev Immunol. 20, 294-307. https://doi.org/10.1038/s41577-019-0257-x 31. Spangler, A., Shimberg, G. D., Mantus, G. E., Malek, R., Cominsky, L. Y., Tsybovsky, Y., Li, N., Gillespie, R. A., Ravichandran, M., Creanga, A., Raab, J. E., Gajjala, S. R., Mendoza, F., Houser, K. V., Dropulic, L., Mcdermott, A. B., Kanekiyo, M., & Andrews, S. F. 2025. Early influenza virus exposure shapes the b cell response to influenza vaccination in individuals 50 years later. Immunity. 58, 728-744 e729. https://doi.org/10.1016/j.immuni.2025.02.004 32. Swanson, K., Wu, W., Bulaong, N. L., Pak, J. E., & Zou, J. 2025. The virtual lab of ai agents designs new sars-cov-2 nanobodies. Nature. 646, 716-723. https://doi.org/10.1038/s41586-025-09442-9 33. Turner, J. S., Zhou, J. Q., Han, J., Schmitz, A. J., Rizk, A. A., Alsoussi, W. B., Lei, T., Amor, M., Mcintire, K. M., Meade, P., Strohmeier, S., Brent, R. I., Richey, S. T., Haile, A., Yang, Y. R., Klebert, M. K., Suessen, T., Teefey, S., Presti, R. M.,...Ellebedy, A. H. 2020. Human germinal centres engage memory and naive b cells after influenza vaccination. Nature. 586, 127-132. https://doi.org/10.1038/s41586-020-2711-0 34. Verstegen, N. J. M., Pollastro, S., Unger, P. A., Marsman, C., Elias, G., Jorritsma, T., Streutker, M., Bassler, K., Haendler, K., Rispens, T., Schultze, J. L., Ten Brinke, A., Beyer, M., & Van Ham, S. M. 2023. Single-cell analysis reveals dynamics of human b cell differentiation and identifies novel b and antibody-secreting cell intermediates. Elife. 12, https://doi.org/10.7554/eLife.83578 35. Wang, W., Swain, S., Lee, J., Lin, Z., Canales, B., Aljović, A., Liu, Y., Li, Q., Marin- Llobet, A., Liu, M., Gao, Z., Liu, R., Alvarez-Dominguez, J. R., & Liu, J. 2025. Agentic lab: An agentic-physical ai system for cell and organoid experimentation and manufacturing. bioRxiv. 36. Zehner, M., Alt, M., Ashurov, A., Goldsmith, J. A., Spies, R., Weiler, N., Lerma, J., Gieselmann, L., Stohr, D., Gruell, H., Schultz, E. P., Kreer, C., Schlachter, L., Janicki, H., Laib Sampaio, K., Stegmann, C., Nemetchek, M. D., Dahling, S., Ullrich, L.,...Klein, F. 2023. Single-cell analysis of memory b cells from top neutralizers reveals multiple sites of vulnerability within hcmv trimer and pentamer. Immunity. 56, 2602-2620 e2610. https://doi.org/10.1016/j.immuni.2023.10.009 37. Zhang, J., Du, Y., Zhou, P., Ding, J., Xia, S., Wang, Q., Chen, F., Zhou, M., Zhang, X., Wang, W., Wu, H., Lu, L., & Zhang, S. 2022. Predicting unseen antibodies’ neutralizability via adaptive graph neural networks. Nature Machine Intelligence. 4, 964- 976. https://doi.org/10.1038/s42256-022-00553-w