Adaptive Immune Recognition in InfectiousDisease and Cancer

Background and Objectives: Epstein–Barr virus (EBV) infection is strongly implicated in multiple sclerosis (MS), yet the mechanisms linking EBV to disease pathogenesis remain unclear. EBV-specific T cells are hypothesized to contribute to MS through molecular mimicry. We aimed to investigate the relationship between EBV-specific T cells and MS relapses using longitudinal T-cell receptor (TCR) repertoire profiling. Methods: We analyzed TCR repertoires from sequential peripheral blood samples collected before and after relapses in five MS patients. CD4⁺ and CD8⁺ T cells were sorted into naive, memory, and regulatory subsets. TCR antigen specificity was determined using exact matches to public databases and GLIPH2-based clustering, covering EBV, cytomegalovirus, influenza, Mycobacterium tuberculosis, and self-antigens. Results were compared to those from healthy controls. Results: At baseline, MS patients showed a significant enrichment of EBV-specific TCRs within CD8⁺ T cells, comparable to patterns seen in other autoimmune diseases. Following relapses, EBV-specific TCR frequencies markedly increased across all T cell subsets. TCRs recognizing other viral antigens, including polyspecific TCRs with broad reactivity, displayed similar baseline enrichment and relapse-associated expansion patterns. Discussion: Our findings indicate a temporal association between MS relapses, global immune activation, and EBV-specific T-cell expansions, which is consistent with a two-step model of MS relapse immunopathogenesis. An initial, unidentified trigger induces broad immune activation, promoting the expansion of antiviral T cells, including those specific to EBV. These EBV-specific and polyspecific T cells with heightened cross-reactivity may then contribute to MS relapses. Larger studies are needed to further define the role of T-cell dynamics in MS pathogenesis.

Regulatory T cells (Tregs) are central to immune tolerance, yet the antigenic determinants shaping their T cell receptor (TCR) repertoires, particularly in autoimmunity, remain poorly defined. In Type 1 Diabetes (T1D), Treg insufficiency permits pathogenic CD8⁺ T cells to destroy pancreatic β cells. While the TCRs of diabetogenic CD8⁺ T cells are increasingly characterized, it remains debated whether Tregs primarily recognize tissue-specific or ubiquitous antigens. Here, we profiled TCR repertoires from Treg and CD8⁺ T cells isolated from pancreatic (PLN), renal (RLN), and brachial (BLN) lymph nodes of non-obese diabetic (NOD) mice using high-throughput sequencing and advanced bioinformatics tools. Convergent TCR clusters significantly enriched in diabetic versus diabetes-resistant mice were identified using GLIPH2. Generation probabilities (Pgen) were estimated with OLGA, selection pressure (logQ) was modeled with SoNNia, and antigen specificities were queried against public databases. Deep learning tools such as TCRPeg were also employed to evaluate the discriminatory power of TCR signatures and their generalizability across tissues. Treg repertoires exhibited signatures shared across lymph nodes, composed of CDR3 sequences with relatively high Pgen and strong positive selection (logQ > 1). Deep learning classifiers confirmed the generalizability of Treg signatures across tissues (AUC ≈ 0.6–0.75). In contrast, pathogenic CD8⁺ T cell signatures from PLN were rare, tissue-restricted, composed of CDR3 sequences with low Pgen, and poorly transferable across tissues (AUC ≤ 0.5). Together, these results suggest that Treg repertoires are shaped by shared, positively selected TCRs likely directed toward conserved self-antigens. This supports a model in which Treg function is driven by recognition of ubiquitous, rather than organ-specific, antigens.

We evaluate whether TCR repertoires can be used to predict responses to cancer therapy. Immune checkpoint inhibitors (ICIs) have transformed cancer therapy, yet only a subset of patients benefit from these therapies. We analyzed TCR repertoires from 223 ICI-treated patients, including 122 responders and 101 non-responders, across both CD4⁺ and CD8⁺ subsets. After preprocessing more than 900,000 sequenced TCR Sequences, we derived repertoire-level features, such as diversity, clonality, and expansion metrics. To capture sequence similarity, we applied TCRdist to construct distance-based representations and defined meta-clones as radius-limited clusters of similar TCRs. This approach not only reduced repertoire complexity while preserving functional information but also enabled the identification of clusters of similar TCRs that reflect identical activity. We further incorporated graph-based features from bipartite patient–sequence networks. Using these representations, we trained a machine learning model to predict patient responsiveness based on TCR repertoire. Our analyses revealed that responders exhibited greater diversity and lower prevalence of dominant clones than non-responders. Moreover, responders were enriched for shared TCRs, which are clonotypes shared across multiple patients. Notably, highly shared TCRs enriched in responders have also been reported in the literature as being associated with cancer. A combination of clone, meta-clone, and repertoire features achieved predictive performance (ROC AUC = 0.7). These findings suggest that TCR repertoires capture biologically meaningful signals of therapeutic response and can serve as biomarkers of ICI treatment responsiveness.

The adaptive immune receptor repertoire (AIRR) encompasses an immense diversity of antibody and T-cell receptor sequences, whose collective organization – how receptors are distributed, clustered, and interrelated across sequence and functional (e.g., antigen-binding) dimensions – remains poorly characterized. Representing AIRRs in continuous representation spaces that capture sequence, biochemical, and structural similarity between receptors may enable comparisons beyond discrete sequence features. Using sequence only and protein language model (PLM) embeddings, we developed a quantitative framework to map immune receptor organization at single-sequence-level scales. We found that immune receptor sequences annotated with different antigen specificities occupy distinct regions of representation space. To resolve local relationships, we introduce a spatial homogeneity metric that quantifies the extent of functional clustering. We found higher spatial homogeneity in embedding spaces than in sequence space for diverse antigen-specific datasets. Our framework establishes a foundation for quantitative mapping of adaptive immune repertoire organization.

Adaptive immune recognition is a key determinant in host defense against infectious diseases and cancer progression. This doctoral research investigates the immunomodulatory potential of bioactive compounds derived from Ephedra alata using an in vivo Wistar rat model. By combining phytochemical characterization with physiological and immunological assessments, the study aims to evaluate the effects of plant extracts on adaptive immune responses, particularly T- and B-cell–mediated mechanisms. The experimental approach focuses on identifying immune modulation patterns that may enhance protective responses or regulate pathological immune activation. Preliminary observations indicate that Ephedra alata compounds may influence immune homeostasis, suggesting a promising role for plant-based modulators in infectious and oncological contexts. This work aligns with the workshop’s focus on adaptive immune recognition by providing experimental insights into natural immunomodulators and their potential application in disease prevention and therapy.

B-cells are an integral part of adaptive immunity. By producing immunoglobulins, they help in swift elimination of different pathogens. The extraordinary diversity of the B-cell repertoire is achieved not only through initial V(D)J recombination, but also through somatic hypermutation (SHM). During this process, B-cells have their receptors randomly mutated, and then B-cells with the highest affinity to an antigen are selected. Several approaches have been introduced to study the patterns of mutations in Ig genes during SHM, however, our understanding of this process remains limited. Here, we infer the parameters that would correspond to the selection processes occurring in the Ig genes during SHM. By reconstructing phylogenies in B-cell clonal families and analyzing mutation patterns in the nucleotide sequences of productive BCRs, we infer the selection factors that reflect the intensity of selection across the BCR heavy-chain sequence. These results can help model B-cell repertoires in the context of SHM.

Autoimmune disorders arise from aberrant adaptive immune responses against self tissues and affect millions worldwide. B cells play a central role in many autoimmune diseases, as autoantibody production can directly drive disease pathogenesis. Each B cell expresses a unique B cell receptor (BCR), whose secreted form is an antibody. B cell tolerance is established during B cell development by eliminating or silencing cells with self-reactive BCRs before they reach maturity. However, this process is impaired in several autoimmune diseases. Thus, high-throughput profiling of the BCR repertoire can potentially reveal disease-associated alterations in BCR composition. Although many studies have sequenced BCR repertoires in autoimmune conditions, integrative analyses across diseases remain limited due to heterogeneous protocols and inconsistent metadata. Here, we conducted a large-scale meta-analysis of public BCR sequencing datasets to identify shared repertoire signatures of autoimmunity across multiple diseases. We utilized nf-core/airrflow, an automated data analysis workflow we previously developed to homogeneously process BCR sequencing data using the Immcantation framework. Relevant clinical metadata were curated and harmonized from the original studies. The standardized analysis workflow enables extension of this meta-analysis to additional public datasets as they become available. We integrated BCR repertoires from 736 individuals with myasthenia gravis, systemic lupus erythematosus, type 1 diabetes, rheumatoid arthritis, and healthy controls from 13 publicly available studies. Across diseases, we evaluated repertoire features previously implicated in autoimmunity, including BCR heavy chain CDR3 length, CDR3 physicochemical properties, and IGHV gene usage. This study demonstrates the feasibility and value of harmonised re-analysis of public BCR repertoire datasets, resulting in a unified resource for the study of BCR repertoires across autoimmune diseases.

Type I interferons (IFN) act as an alarm signal during viral infections by triggering intra- and intercellular processes. Plasmacytoid dendritic cells (pDC), the main producers of IFNa, initiate an early immune response even at very low viral concentrations and thereby activate other immune cells. How this system remains sufficiently sensitive at low viral load while avoiding overstimulation at higher viral concentrations remains unclear. Using a data-based mathematical modeling framework, we address this question and mechanistically explain how pDC supports sensitivity and prevents excessive signal amplification after virus detection. Our theoretical modelling framework captures non-Markovian cell state transitions and is directly parameterized by in vitro IFNa kinetic data, demonstrating that an independently measured optimal response curve arises in a physiologically plausible parameter regime across viral concentrations. We mechanistically explain this behavior using a sensing and amplification mechanism, known as priming, which results in increased IFNα production by cells treated with IFNα prior to cell activation. In detailed network motif analyses, we illustrate that a state-locking property, where cells can only be primed if not yet activated, is crucial for this behavior, preventing excessive IFNα production. Due to the positive feedback of IFNα being produced by activated pDCs, the amount of IFNα in the system is amplified at intermediate viral load. Together, our results demonstrate how heterogeneous pDC responses can be mechanistically understood and leveraged to predict optimal immune activation, with potential implications for therapeutic modulation of antiviral responses.

Effective influenza vaccine formulation and development require a mechanistic, sequence-level understanding of how B-cell receptors (BCRs) recognize and bind viral antigens. Specifically, there is a need to understand the BCR repertoire features that drives antigen recognition and, conversely, how changes in the antigen affect the generation of new antigen-specific BCRs. Towards this goal, we have collected PBMC from healthcare workers following influenza vaccination to generate a large-scale resource of paired chain BCR sequences linked to different influenza antigens. By integrating BCR repertoire profiles at both bulk and single-cell levels, together with orthogonal information on gene expression and serological responses, this resource will enable fine-grained analyses of characteristics guiding BCR repertoire development, including somatic hypermutation trajectories, clonal lineage reconstruction, germline gene usage biases, and convergent sequence motifs across influenza types and lineages. From this foundation, we will develop machine learning models capable of predicting antigen-specific BCR sequences from unlabelled repertoire data, with the aim to recover specificity determinants from sequence alone, generalize across cohorts and influenza strains, and provide interpretable features. The resulting resource and models have the potential to accelerate antigen selection, anticipate variability in vaccine responses, and ultimately enhance the effectiveness and durability of next-generation influenza vaccines.

Many pathogens evolve to escape immunity, yet it remains difficult to predict whether immune pressure will lead to diversification, serial replacement of one variant by another, or more complex patterns. Pathogen strain dynamics are mediated by cross-protective immunity, whereby exposure to one strain partially protects against infection by antigenically diverged strains. There is growing evidence that this protection is influenced by early exposures, a phenomenon referred to as original antigenic sin (OAS) or imprinting. In this paper, we derive constraints on the emergence of the pattern of successive strain replacements demonstrated by influenza, SARS-CoV-2, seasonal coronaviruses, and other pathogens. We find that OAS implies that the limited diversity found with successive strain replacement can only be maintained if is less than a threshold set by the characteristic antigenic distances for cross-protection and for the creation of new immune memory. This bound implies a “speed limit” on the evolution of new strains and a minimum variance of the distribution of infecting strains in antigenic space at any time. To carry out this analysis, we develop a theoretical model of pathogen evolution in antigenic space that implements OAS by decoupling the antigenic distances required for protection from infection and strain-specific memory creation. Our results demonstrate that OAS can play an integral role in the emergence of strain structure from host immune dynamics, preventing highly transmissible pathogens from maintaining serial strain replacement without diversification.

Understanding and rationally designing antibody-antigen (Ab-Ag) binding remains a fundamental challenge in immunology. While machine learning offers promise, current immune receptor datasets have not enabled models that generalize across diverse Ab-Ag pairs. Deep mutational scanning (DMS) coupled with functional assays represents a promising high-throughput approach for generating the annotated data needed to advance this goal. However, optimal strategies for constructing DMS datasets that enable model generalization remain unclear. Here, we systematically investigate DMS affinity dataset design using synthetic data to guide experimental efforts. We examine trade-offs between dataset diversity and depth, quantify the value of incorporating structural information, and establish scaling laws for model performance on out-of-distribution Ab-Ag systems. Our results demonstrate that paired Ab-Ag DMS provides a practical experimental strategy for generating training data that enables generalizable computational binding prediction. Furthermore, paired DMS uniquely captures complex epistatic interactions between Ab and Ag mutations, revealing biophysical constraints governing binding evolution. While these epistatic patterns consistently challenge ML models, incorporating them into training data significantly improves predictive performance.

We develop a dynamic model capturing the interplay between Leishmania parasites and host T cell immunity. Each infectious bite introduces local parasites and stimulates effector activation, with memory cells capable of reactivation upon subsequent exposures. The model, expressed as ordinary differential equations, captures key processes including immune activation, migration, differentiation, and parasite clearance. By exploring bite frequency and prior immune history, it provides insights into conditions leading to parasite persistence or elimination. This framework offers a quantitative tool to connect immune mechanisms with infection outcomes, supporting experimental and epidemiological studies of vector-borne diseases.

In this investigation, we delve into the dynamics of an eco-epidemic model, considering the intertwined influences of fear, refuge-seeking behavior, and alternative food sources for predators with selective predation. We extend our model to incorporate the impact of fluctuating environmental noise on system dynamics. The deterministic model undergoes thorough scrutiny to ensure the positivity and boundedness of solutions, with equilibria derived and their stability properties meticulously examined. Furthermore, we explore the potential for Hopf bifurcation within the system dynamics. In the stochastic counterpart, we prioritize discussions on the existence of a globally positive solution. Through simulations, we unveil the stabilizing effect of the fear factor on susceptible prey reproduction, juxtaposed against the destabilizing roles of prey refuge behavior and disease prevalence intensity. Notably, when disease prevalence intensity is too low, the infection can be eradicated from the eco-system. Our deterministic analysis reveals a complex interplay of factors: the system destabilizes initially but then stabilizes as the fear factor suppressing disease prevalence intensifies, or as predators exhibit a stronger preference for infected prey over susceptible ones, or as predators are provided with more alternative food sources.

Background: Regulatory T cells (Treg) are central to immune tolerance and are quantitatively and/or functionally altered in autoimmune diseases, including systemic lupus erythematosus (SLE). Low-dose IL-2 (IL-2LD) is a Treg-targeting immunotherapy with demonstrated clinical efficacy in SLE, yet the clonal identity of Tregs mobilized by IL-2LD remains poorly defined. T-cell receptor (TCR) repertoire profiling provides a direct readout of clonal structure and can reveal whether IL-2LD drives global Treg activation or preferential expansion of a constrained subset of clones. Method: In the randomized, double-blind, placebo-controlled LUPIL-2 trial, moderate-to-severe SLE patients received IL-2LD or placebo. Peripheral blood Treg and CD4 effector T cells (Teff) were isolated at baseline and after treatment (day 5, month 1, month 3) and TCRαβ repertoires were generated by bulk sequencing. Healthy volunteers (HV), treated or not with comparable IL-2LD regimens, served as references. Repertoires were analyzed at the CDR3 level after quality control and harmonization. Treatment-associated TCR features were identified using sparse partial least squares discriminant analysis (sPLS-DA) and classified with penalized logistic regression (glmnet), with cross-validation and permutation testing. Results: First, SLE Treg and Teff repertoires displayed marked V-gene usage biases compared with HV and a significant reduction in diversity in both compartments, consistent with clonal expansions. Second, IL-2LD induced a rapid remodeling of the Treg repertoire by day 5, characterized by a significant increase in diversity toward an HV-like profile; this effect persisted through month 1 and was not observed under placebo. Third, we identified a robust Treg-specific day-5 signature comprising 150 TCRs that discriminated IL-2LD from placebo (AUC = 0.83), with signal abolishment under 100 permutations (AUC ≈ 0.5). Signature TCRs were highly shared across treated individuals, showed high generation probabilities and strong sequence similarity, and 98% were detectable pre-treatment at low frequency before rapidly expanding after IL-2LD and persisting to month 1. Notably, the same signature was observed in IL-2LD-treated HV under comparable dosing, and was absent at lower or higher doses, supporting a direct, dose-dependent association with IL-2LD independent of disease context. Discussion: These data indicate that IL-2LD does not act indiscriminately on the Treg compartment; instead, it preferentially amplifies a restricted, shared set of pre-existing Treg clones while transiently restoring Treg repertoire diversity in SLE. The public-like, sequence-convergent nature of the signature supports an antigen-constrained process and provides a mechanistic entry point to understand IL-2LD-driven tolerance restoration.

Mapping of T cell receptors (TCRs) to their cognate MHC-presented peptides (pMHC) is central for the development of precision immunotherapies and vaccine design. However, accurate prediction of TCR affinity to peptide antigens remains an open challenge. Most approaches rely solely on sequence information, although increasing evidence suggests that TCR-pMHC binding is primarily determined by three-dimensional structural interactions within the entire TCR-pMHC complex. Consequently, sequence-based methods often fail to generalize to peptides not included in the training data (unseen peptides). Here we introduce t2pmhc, a structure-based graph neural network framework for predicting TCR-pMHC binding using predicted structures of the entire TCR-pMHC complex. We evaluated a Graph Convolutional Network (GCN) and a Graph Attention Network, both demonstrating improved generalization to unseen peptides compared to state-of-the-art models across a variety of public datasets. Evaluation with crystallographic structures yields high-confidence predictions, indicating that current limitations of structure-based models are largely driven by the accuracy of structure prediction. Analysis of node attention patterns in t2pmhc-GCN reveals biologically consistent patterns, assigning high attention to the peptide and the CDR3 regions. Within the peptide sequence, canonical MHC anchor residues are consistently downweighted, whereas potential TCR-binding residues are upweighted. These findings establish t2pmhc as a structure-informed framework for robust TCR-pMHC binding prediction, enabling improved generalization to unseen antigens and providing a foundation for integrating TCR repertoire sequencing into vaccine design and immunotherapy.

Introduction: CD8+ T cells (T cells) play an important role in identifying and destroying infected and cancerous cells. They are activated upon binding of their T cell Receptor (TCR) to cognate peptide Major Histocompatibility Complex I (pMHC) on the antigen presenting cell surface, leading to a myriad of downstream responses including cytokine and granzyme B secretion. The activity of the TCR is regulated by a complex array of inhibitory and stimulatory co-signalling receptors, which help ensure CD8+ T cells respond appropriately[1]. Stimulatory co-signalling receptors, including CD28 and CD2, help sensitise T cells to the presence of antigenic pMHC, while inhibitory co-signalling receptors, including PD-1 and BTLA, constrain T cell activation to guard against excessive inflammation. This regulatory system can be hijacked to drive disease processes. For example, tumours may evade T cell recognition through expressing PD-L1[2]. Likewise checkpoint inhibitors which disrupt the PD-1:PD-L1 interaction have found success, treating tumours in the clinic[3,4]. However, the mechanisms of action of several co-signalling receptors remain hotly debated. For example, there are conflicting reports of PD-1 acting directly on the TCR or primarily through inhibition of CD28 [5,6]. This work leverages a large in vitro dataset of CD8+ T cell activation in response to a variety of co-signalling ligands, generated using the CombiCell platform[7–9], and mathematical modelling to investigate the regulatory interactions between co-signalling ligands and the TCR. Methods: The CombiCell platform utilises the SpyTag-SpyCatcher system to display specified combinations and concentrations of co-signalling ligands on the surface of cells. Using this platform, hamster CHO-K1 cells were labelled with varying concentrations of pMHC, from a multi-affinity panel, and co-signalling ligands including CD80, CD58, HVEM, PD-L1 and PD-L2, both individually and in different pairwise combinations (PD-L1/CD80, PD-L2/CD80, PD-L1/CD58, HVEM/CD58, HVEM/CD80), and then cocultured with Jurkat T cells. This generated a large dataset of T cell activation at steady state, measured by CD69 activation using flow cytometry[10]. To understand where within the TCR signalling pathway co-signalling ligands were likely acting, and whether they exert control independently or cooperatively, the TCR signalling pathway was mathematically described using the Kinetic Proofreading with Limited Signalling ODE model[11]. According to this model, the TCR undergoes a series of modification events ,upon binding pMHC, to enter a productive signalling state before transitioning to a non-productive state. pMHC dissociation at any stage aborts this process. A high-affinity pMHC with a slow dissociation rate is more likely to produce a signal than a low-affinity pMHC with a fast dissociation rate. This kinetic proofreading contributes to TCR discriminatory power. Possible co-signalling ligand interactions with the TCR pathway were simulated by selective parameter perturbations across conditions when fitting this model to the data. In this way, a large space of putative interaction networks was explored and ranked according to mean squared error. Results: Networks where co-signalling ligands targeted parameters associated with events either upstream or downstream of kinetic proofreading - including pMHC association, TCR surface availability, and cytoplasmic signalling sensitivity - outperformed networks where co-signalling ligands targeted parameters associated with kinetic proofreading such as the number of kinetic proofreading steps or the proofreading rate. Further, some pairs of co-signalling ligands including CD80/PD-L1 exhibit strong concentration-dependent cooperativity. Conclusion: Mathematical modelling of empirical T cell activation data suggests that co-signalling ligands regulate the sensitivity and potency of the T cell activation response to pMHC, independent of pMHC affinity, by targeting processes upstream or downstream of kinetic proofreading. Further, certain pairs of ligands such as CD80/PD-L1 act cooperatively, which may suggest that CD80 can modulate the efficacy of PD-1:PD-L1 checkpoint inhibitors. References 1. Chen, L. & Flies, D. B. Two-signal model Molecular mechanisms of T cell co-stimulation and co-inhibition. Nat. Rev. Immunol. 13, (2013). 2. Han, Y., Liu, D. & Li, L. PD-1/PD-L1 pathway: current researches in cancer. Am. J. Cancer Res. 10, (2020). 3. Liu, J. et al. PD-1/PD-L1 Checkpoint Inhibitors in Tumor Immunotherapy. Frontiers in Pharmacology vol. 12 Preprint at https://doi.org/10.3389/fphar.2021.731798 (2021). 4. Alsaab, H. O. et al. PD-1 and PD-L1 checkpoint signaling inhibition for cancer immunotherapy: mechanism, combinations, and clinical outcome. Frontiers in Pharmacology vol. 8 Preprint at https://doi.org/10.3389/fphar.2017.00561 (2017). 5. Mizuno, R. et al. PD-1 primarily targets TCR signal in the inhibition of functional T cell activation. Front. Immunol. 10, (2019). 6. Hui, E. et al. T cell costimulatory receptor CD28 is a primary target for PD-1-mediated inhibition. Science (1979). 355, 1428–1433 (2017). 7. Patel, A. et al. Using CombiCells, a platform enabling titration and combinatorial display of cell surface ligands, to study T cell antigen sensitivity by TCRs, CARs, and BiTEs. https://doi.org/10.1101/2023.06.15.545075v1 doi:10.1101/2023.06.15.545075v1. 8. Patel, A. et al. Using CombiCells, a platform for titration and combinatorial display of cell surface ligands, to study T-cell antigen sensitivity modulation by accessory receptors. EMBO J. 43, 132–150 (2024). 9. Bustamante Eguiguren, S. CombiCells allow combinatorial display of cell surface ligands.Nat Rev Immunol 25, 317 (2025). https://doi.org/10.1038/s41577-025-01168-z 10. Data generated by Bustamante Eguiguren, S. (Unpublished) 11. Lever, M., Maini, P. K., Van Der Merwe, P. A. & Dushek, O. Phenotypic models of T cell activation. Nat. Rev. Immunol. 14, 619–629 (2014).

Negative thymic selection eliminates T cells with excessive self-reactivity, yet it remains unclear how well TCR amino acid sequence features predict deletion. We model negative selection as a stochastic filtering process in which survival probabilities depend on sequence similarity to known self-reactive TCRs. Large $\alpha \beta$ repertoires are generated with a trained generative model, and selection outcomes are simulated using Monte Carlo sampling. To test how learnable this selection rule is from TCR sequences alone, we train logistic regression, random forest, and neural network classifiers on embeddings of sequence features. Their performance is compared to theoretical limits set by the intrinsic stochasticity of the selection process. Ultimately, this framework aims to identify the factors that constrain the learnability of thymic selection from TCR sequences.

Generating antibody competition data is now a central part of antibody discovery campaigns. Where available, described antibodies can be used for epitope mapping; for new targets, these experiments inform on the epitope diversity contained within a set of antibodies. Likewise, antibody-antigen complex structure prediction has greatly advanced in the last few years. However, these two powerful tools are not currently used in conjunction with one another. We introduce results from a new computational tool, GAX, that combines antibody-antigen structural prediction (Chai-1) with cross-blocking data to enable mutual information between the two.

Antibody repertoires emerge from a constrained stochastic process shaped by V(D)J recombination, somatic hypermutation (SHM), and selection. Whether data-driven language models capture these mechanisms remains an open question. We evaluate how well antibody sequences generated by the Immunoglobulin Language Model (IgLM) reflect fundamental immunological mechanisms underlying B-cell receptor diversity. Using datasets from natural human repertoires, we compare biologically interpretable metrics, including gene usage frequencies, mutation patterns, and CDRH3 generative probabilities, between natural and IgLM-generated sequences. We further benchmark IgLM likelihoods against established mechanistic models for recombination and SHM, OLGA and ARMADiLLO, respectively. To control for repertoire-level effects, we generate a synthetic repertoire matched to a natural repertoire from Le Quy et al. (2024) by matching the distribution of three-residue N-terminal prompts. Synthetic sequences diverge substantially from the natural repertoire, particularly in J gene usage and distance from germline. Consistent with other antibody language models, IgLM exhibits a germline bias and shows reduced coherence toward sequence termini, resulting in an overrepresentation of J gene calls that are rare in natural human repertoires from OAS. In the CDRH3 region, IgLM likelihoods correlate well with OLGA generation probabilities for natural sequences, but this agreement is significantly reduced for synthetic sequences, indicating that while IgLM likelihoods reflect constraints of V(D)J recombination, sequence generation does not fully recapitulate these mechanisms. Notably, IgLM fails to assign higher likelihoods to natural CDRH3 sequences than to scrambled variants with identical amino-acid composition, suggesting that the model has not learned higher-order statistical structure characteristic of human CDRH3 organization. To probe SHM at the nucleotide level, we introduce an Ig-specific reverse translation approach based on germline alignment and SHM-derived statistics and use it to compare IgLM-derived sequences with ARMADiLLO. Comparisons with ARMADiLLO and IMGT reference sequences reveal partial learning of V gene family–specific features, but little evidence that IgLM captures AID-driven SHM context. Together, these results delineate where IgLM captures biologically meaningful features of adaptive immune repertoires and where key immunological mechanisms remain missing, providing guidance for improving antibody language models and developing evaluation metrics that better reflect biological realism.

Acute respiratory viral infections (ARVIs) are a leading cause of global morbidity. While asymptomatic infections are known drivers of transmission—highlighted extensively during the COVID-19 pandemic—the specific pre-existing immune features that determine whether an infection becomes symptomatic or remains silent are poorly understood beyond SARS-CoV-2. We aimed to characterize the immunological "immune history" that predisposes individuals to asymptomatic outcomes across a broad range of respiratory viruses We utilized data from a longitudinal study with a cohort of 110 healthy adults followed for 9 months. The study protocol included weekly nasal/saliva sample collection for multiplex RT-PCR detection of 13 common respiratory viruses (including Influenza, RSV, hMPV, Adenovirus, and seasonal HCoVs) and SARS-CoV-2, alongside weekly symptom questionnaires. Monthly blood samples were profiled using high-throughput Antigen Microarrays (AMs) to quantify IgG and IgA titers against a panel of respiratory viral antigens4. We compared viral loads (Ct values) and antibody dynamics between symptomatic and asymptomatic infection events. Serological profiling revealed that individuals who developed asymptomatic infections possessed significantly higher baseline (pre-infection) IgG titers compared to those who became symptomatic. Longitudinally, a significant post-infection boost in antibody titers was observed only in symptomatic individuals. In contrast, antibody levels in asymptomatic individuals remained stable pre- and post-infection, showing no significant boosting effect. Our findings suggest that high pre-existing humoral immunity effectively limits viral replication and prevents clinical symptoms. However, this protection comes with a trade-off: symptomatic illness appears necessary to drive significant antibody boosting, while asymptomatic infections may fail to renew immunological memory magnitude. This study provides critical insights into the maintenance of immune memory and the "use it or lose it" dynamics of host-pathogen interactions

The progression of postmenopausal breast cancer is driven by interactions between estrogen signalling, immune regulation, and treatment effects. In this study, we develop a deterministic non-linear mathematical model to examine the combined impact of Letrozole and Palbociclib therapy. A key feature of this work is the estimation of model parameters directly from time-series clinical data using a Markov Chain Monte Carlo approach based on the Delayed Rejection Adaptive Metropolis (DRAM) algorithm. Parameter influence was subsequently quantified using Partial Rank Correlation Coefficient (PRCC) analysis to identify the most effective drivers of treatment response. Our analysis shows that the system exhibits bistable behavior, separated by a critical threshold that distinguishes cancer control from uncontrolled growth. This structure leads to two important clinical observations. First, cytoreductive surgery can act as a dynamical switch by pushing the system below this threshold, allowing immune control of the cancer to be restored. Second, reducing adiposity accelerates treatment response, as lower estrogen availability shortens the time required for cancer clearance by nearly 30\%. Overall, these results support treatment strategies that combine surgery, metabolic management, and targeted drug therapy.

Cytotoxic CD8 T cells are essential for antiviral immunity. T-cell development in infancy is characterized by high thymic output and rapid peripheral expansion, leading to the establishment of relatively stable numbers of naive and circulating memory cells in adulthood. The relative combinations of de novo production and peripheral expansion over the life course plays a key role in shaping T-cell receptor (TCR) diversity within peripheral subsets. In this study, we combined multiple fate-mapping systems with mathematical modelling to quantify the population dynamics and lineage relationships of naive, central memory, and effector CD8 T-cell subsets in mice from early life into adulthood. We systematically quantify how the contributions of influx, self-renewal, differentiation and loss shift with age, providing a unified description of CD8 T-cell dynamics across the mouse lifespan.

A major challenge in both autoimmune disease and cancer immunology is to connect predictive biomarkers of treatment response with a mechanistic understanding of immune interactions within tissue. Bridging machine learning–based patient stratification with spatially resolved immune profiling offers a powerful framework to link clinical outcomes to underlying cellular communication and tissue organization. In IBD, where up to 30% of patients do not respond to Vedolizumab, we applied a machine learning framework to multimodal immune profiling data comprising 126 features across CyTOF, flow cytometry (FACS), OLINK proteomics, and clinical parameters to predict treatment response. Models optimized for small patient cohorts identified FACS and OLINK data as the most informative modalities, with Ki67 emerging as a robust, clinically applicable, and externally validated predictor. Feature importance analysis highlighted chemokine receptors, implicating altered T-cell communication in treatment resistance and motivating further mechanistic studies to dissect T-cell–mediated resistance. Importantly, chemokine-driven immune communication is inherently spatial, raising the question of how these predictive immune states are organized and maintained within tissue. Building on this, we have applied cell–cell communication analyses in autoimmune disease to identify disease-specific immune interaction patterns, providing a mechanistic entry point toward spatially resolved approaches. Incorporating spatial transcriptomics will further enable the mapping of cytokine niches, the discovery of spatial cytokine networks, and the characterization of immune interactions in situ. Beyond IBD, we are extending these approaches to dissect early T–B cell dynamics in lymph nodes using an adoptive transfer model. By labeling transgenic lymphocytes and sampling lymph nodes over the first six days following immunization, we profile early stages of the adaptive immune response using single-cell transcriptomics, complemented by T-cell receptor sequencing of selected day-0 populations and high-resolution spatial mapping with Xenium. This integrated framework enables a mechanistic characterization of early T–B interactions and provides a foundation for understanding immune regulation across autoimmune settings.

\abstract{Glioblastoma remains one of the most lethal brain cancers with limited treatment options. Combination therapy using CAR-T cells and oncolytic viruses shows promise for enhanced tumor control. We develop and analyze mathematical models to quantify therapeutic synergy between these modalities. Using experimental data from patient-derived glioblastoma cells treated with IL-13R$\alpha$2-targeting CAR-T cells and oncolytic virus C134, we construct ordinary differential equation models that capture tumor-immune-virus dynamics. Through systematic model reduction based on timescale separation, we derive a simplified quasi-steady-state model that retains essential biological mechanisms while reducing parameter complexity. We compare our framework with existing models, demonstrating improved predictive accuracy across multiple experimental conditions. Parameter estimation reveals that viral amplification and density-dependent interactions are critical for therapeutic efficacy. }

B cell selection and evolution are key processes in regulating successful adaptive immune responses. Recent advances in single-cell sequencing and deep learning strategies have unlocked new potential to study affinity maturation of B cells at unprecedented scale and resolution. To unravel the complex dynamics of B cell repertoire evolution during immune responses and to facilitate PLM-guided antibody engineering, we created the R package AntibodyForests (1). AntibodyForests encompasses pipelines to infer B cell lineages, quantify inter- and intra-antibody repertoire evolution, and analyze somatic hypermutation (SHM) using protein language models (PLM) and protein structure. Using AntibodyForests, we explore how general and antibody-specific PLM-generated likelihoods relate to features of in vivo B cell selection, evolution, antigen specificity and binding affinity (2). We find that PLM likelihoods correlate with biologically relevant features including isotype and V-gene usage, mutational load, and SHM patterns. Additionally, we observed that mutating residues along evolutionary trajectories tend to have lower PLM likelihoods than conserved residues. These results indicate that PLMs could predict to what amino acid SHM will most likely mutate and at which position. Interestingly, our findings challenge in vitro observations (3) by revealing a negative correlation between PLM likelihoods and antigen binding affinity in in vivo repertoires. In our exploitation of these discoveries using six different PLMs and varying sequence regions, we uncovered that the region of antibody sequence (CDR3 or full-length V(D)J) provided to the PLM, as well as the type of PLM used, influences the resulting likelihoods. These comparisons emphasize the importance of PLM long-range interaction, potential training data biases, and pairing heavy and light chains. Together, these studies highlight the power of combining repertoire-wide phylogenetic inference with PLMs to better understand the principles governing antibody evolution and selection, and offer new tools for therapeutic antibody discovery and engineering. 1. van Ginneken D, Tromp V, Stalder L, Cotet TS, Bakker S, Samant A, et al. Delineating inter- and intra-antibody repertoire evolution with AntibodyForests. Bioinformatics. 2025 Oct 9;41(10). 2. van Ginneken D, Samant A, Daga-Krumins K, Glänzer W, Agrafiotis A, Kladis E, et al. Protein language model pseudolikelihoods capture features of in vivo B cell selection and evolution. Brief Bioinform. 2025 Jul 2;26(4). 3. Hie BL, Shanker VR, Xu D, Bruun TUJ, Weidenbacher PA, Tang S, et al. Efficient evolution of human antibodies from general protein language models. Nature Biotechnology. 2023 Apr 24;42(2):275–83.

Women represent about 80% of patients with autoimmune diseases. This may partly result from sex-based differences in T cell receptor (TCR) selection during thymocyte development, potentially influenced by hormones and the lower expression of the Autoimmune Regulator (AIRE) transcription factor in females. To investigate this, we analyzed sex-specific differences in TCR generation and selection. We examined TCR repertoires in double-positive thymocytes and single-positive thymic cells, including CD8⁺ and CD4⁺ effector T cells and regulatory T cells (Tregs), derived from male and female organ donors. Minimal sex-based differences were observed in V and J gene usage, and there were no notable differences in TCR repertoire diversity, complementarity-determining region 3 (CDR3) length, amino acid composition, or network structure. No TCR sequences were exclusive to either sex. However, female effector T cells exhibited a significantly higher prevalence of TCRs specific to self-antigens implicated in autoimmunity compared to males, while female Tregs showed a reduced frequency of such TCRs. These differences were not observed for TCRs targeting self-antigens unrelated to autoimmunity or antigens associated with cancer or viruses. Our findings identify a sex-specific imbalance in thymic selection of TCRs with autoimmunity-associated specificities, providing mechanistic insight into the increased susceptibility of women to autoimmune diseases. Please let me know if any further information is required.

High‑throughput discovery and engineering of therapeutic antibodies require precise affinity quantification across large variant libraries, yet existing platforms force a trade‑off between throughput and accuracy. Here, we introduce a set of multiplexed assays that enable quantitative, high‑throughput affinity measurements that require only common flow cytometers and widely available reagents, minimizing cost and complexity. By arraying either antibodies or antigens on an established microsphere‑assisted proteomics (MAP) platform, we are able to perform parallel equilibrium‑binding experiments across entire libraries on a single plate. Controlled immobilization on microspheres yields homogeneous antibody display and eliminates surface‑expression and avidity artifacts that limit the accuracy and controllability of yeast surface display-based methods such as Tite‑Seq. By reducing the number of titrations and combining direct-binding with competitive-binding assays we extended the scope from one-dimensional library screenings to a two-dimensional library-on-library readout. We demonstrate our system's capability by rapidly inferring binding affinities for multiple viral proteins across a library of ~70 affinity-matured antibodies (> 400 affinity measurements) while maintaining strong correlation with SPR and BLI reference data (r > 0.9). Our assays can be run in a single day, demonstrating order-of-magnitude higher throughput than conventional biosensor-based measurements.

The humoral immune response emerges from germinal center (GC) selection, plasma cell differentiation, and antibody secretion into serum. While B cell receptor (BCR) sequencing has enabled detailed characterization of GC and plasma cell repertoires, it remains difficult to determine which clones ultimately contribute to circulating antibodies. Here we integrate single-cell (GC) BCR sequencing (scBCR-seq), bulk bone marrow plasma cell BCR sequencing (bulkBCR-seq), and serum antibody mass spectrometry (Ab-seq) to directly connect clonal lineages across these compartments. Using full-length VDJ sequence matching, we map serum antibody peptides to their corresponding B cell clones and reconstruct lineage trajectories from GC B cells to plasma cells and secreted antibodies. This approach enables clone-resolved analysis of repertoire selection across multiple mice and infection conditions. In addition, we annotate antigen-specificity by tracking scBCR-seq serum-linked clones (scAb-seq), and by bead-based antigen-coupled antibody mass-spectrometry (Ag-Ab-seq). We quantify key bottlenecks in the humoral response by measuring GC-to-plasma seeding efficiency and plasma-to-serum recruitment probability. Serum-detected clones exhibit distinct repertoire features compared with non-serum clones, including increased somatic hypermutation and enrichment for antigen-binding GC cells. By integrating cellular and proteomic measurements, our framework provides a direct link between GC selection dynamics and the functional antibody pool present in serum.

There is an urgent need for an effective HIV- vaccine, but the durability and clonal diversity of HIV vaccine-elicited T-cells are poorly understood. To characterize the HIV-specific TCR repertoire elicited by vaccination, we developed a novel and high-throughput method for efficient TCR cataloging of rare antigen-specific T-cells. We applied this approach to generate HIV-ENV and GAG-specific TCR libraries using PBMC samples of healthy vaccinees. We also performed bulk TCR-sequencing on longitudinal samples. We identified 1680 unique ENV- and GAG-specific TCR-clonotypes and found the magnitude of responses correlated with intracellular cytokine staining (ICS) data on the same cohort, except for higher frequency of CD8 T-cell response to GAG, as measured by TCR sequencing than by ICS. Through detailed TCR repertoire analyses, we were able to delineate the major clonotypic and kinetic differences between ENV- and GAG-CD4 and CD8 T-cells in response to prime and boost. Specifically, we observed boosting expanded short-lived GAG-specific CD8 T-cell clonotypes that likely contributed to under sampling of the response by ICS. In an unbiased approached, we clustered clonotypes by the magnitude of their trajectories and identified 6099 additional vaccine-responsive TCRs. Three clusters displayed stable memory trajectories, and two clusters, which contains GAG-CD8 T-cell clones, resembled short-lived effectors. Lastly, our data was further validated when we observed concordance in cell-type and antigen for matched T-cell clones between our data and published HIV datasets. We developed a new pipeline to identify antigen-specific TCRs from peripheral blood samples. To our knowledge, this dataset represents the largest HIV-specific TCR catalogue to date, which can be used for discovery and comparison in other HIV contexts. Our longitudinal data provides an in-depth characterization of the cellular immune response to HIV vaccination.

This study explores the inter-allelic diversity and intra-allelic divergence of antigen presentation across human leukocyte antigens class-I (HLA-I) alleles to advance immunotherapy. The highly polymorphic nature of HLA-I genes leads to significant allelic variations, shaping individual immune responses. We quantified peptide diversity via entropy and divergence via Jensen-Shannon Divergence, leveraging probabilistic models like Position Weight Matrices and Restricted Boltzmann Machines (RBM). These models were trained on data from the Immune Epitope Database, with RBM optimized through hyperparameter search. Different confidence intervals around these estimates were computed using bootstrapping techniques, validated on synthetic datasets, and applied to the true datasets of 68 HLA-I alleles. Findings reveal relationship between peptide repertoire diversity and HLA-I allelic variations, including under-characterized alleles. We refined the classification of HLA-I alleles into super-types and provided insights into disease susceptibility and protective immunity. Integrating these results with tools like NetMHCpan, we identified SARS-CoV-2 vaccine targets. This framework enhances understanding of HLA-I diversity, aiding therapeutic strategies tailored to individual genetic profiles.

Protein language models (PLMs) have demonstrated significant potential in capturing the complex interaction patterns between amino acids in protein sequences. These models, trained on large datasets of protein sequences, can generate embeddings, high-dimensional numerical representations that encode valuable information about the structure, function, and evolution of proteins. However, conventional usage has largely been based on arbitrarily determined variable sets (embedding modes), including the choice of embedding layer, pooling method, and padding, which can potentially lead to a suboptimal representation with low information content for a given downstream task. The scalability of protein embedding mode extraction is limited by inefficiencies in both space (memory) and time (computation). (i) Accumulating all outputs in memory and writing them to disk in a single operation leads to a memory bottleneck. (ii) Additionally, repeated embedding of the same sequence to extract different embedding modes introduces unnecessary computational overhead and reduces throughput significantly. Here, we present PEPE (Parallel Extraction for Protein Embeddings), a command-line tool designed for high-throughput multi-modal protein sequence embedding extraction. We demonstrate that PEPE’s parallel process achieves a total run time several orders of magnitude faster than sequential approaches. We also demonstrate how, for a state-of-the-art (SOTA) method, peak memory usage scales with output size and fails once the memory capacity is exceeded, whereas PEPE’s peak memory usage remains consistently below the critical limit, allowing the extraction of multimodal embeddings that exceed the available memory. PEPE supports a wide range of publicly available and custom protein language models, providing a simple command-line interface for researchers. PEPE enables the generation of protein embedding datasets at previously unfeasible scales, facilitating the identification of optimal protein embedding settings for downstream analyses without requiring additional resources for fine-tuning.

Although the overall CD8+ T cell immune response is reproducible, individual naïve CD8+ T cells vary significantly in their contributions. A single naïve CD8+ T cell and all its descendants constitute a "family". Interestingly, while the size of a family in the primary response shows little correlations with its size in the secondary response, there is a strong correlation between family sizes across recall responses. Recent findings using the DivisionRecorder system have revealed that recall responses are primarily driven by lowly divided memory T cells with stem-cell like properties. Understanding how cells responsible for immunological memory are generated and maintained is crucial for improving immune protection post-vaccination. We developed a model in which activated CD8+ T cells initially undergo a few rapid, highly synchronized burst divisions. After a burst, the resulting daughter cells have diversified in branches. Some branches continue to divide and combat the current infection, while others return to quiescence and provide future protection by generating recall responses upon re-infection. The inclusion of these quiescent cells in the model accounts for the observed improved correlation between family sizes across recall responses. This is because the number of quiescent cells formed during the primary response is stochastic but correlated with the number formed during the subsequent response. In conclusion, a growing body of research highlights the crucial role of lowly divided memory T cells in long-term protection against pathogens.