4th London Quantitative Immunology Day

Immune memory systems exhibit prolonged but finite responses that underlie their efficacy. I will present our analysis of two distinct memory systems, plasma cell persistence following immune responses in mice and humans, and transgenerational gene silencing in C. elegans. Despite relying on different biological mechanisms, these systems share key similarities captured by a unifying mathematical framework, where memory units (cells or silenced genomic regions) are tuned near critical tipping points through competition over shared resources. This framework generates specific testable predictions, which we apply to analyse the buildup of memory, heterogeneity within and between responses, and the effects of perturbations. I will conclude by discussing how this resource competition mechanism may serve as an efficient quality control strategy for immunity.

Affinity maturation enhances B cell binding within germinal centers, where spatial structure preserves sequence diversity by restricting cell movement. While recent studies show that some B cell lineages span multiple germinal centers, the sources, rates and consequences of this spreading process remain unknown. Here, we show that the spatial arrangement of B cells in the human tonsil is driven by local migration during affinity maturation. Through an evolutionary re-analysis of spatial transcriptomics data, we demonstrate that these local migrations follow a clock-like process, in which cells migrate at an average rate of ~1/50 cell divisions that is consistent across lineages and time. Migrating cells continue to evolve and diversify in their new germinal centers at similar rates, such that the largest lineages in each germinal center often originate from another. These results suggest that affinity maturation operates in a regime of pervasive but intermediate migration, balancing diversity and selection.

Peptide presentation by human leukocyte antigen class I (HLA-I) molecules shapes CD8+ T-cell recognition and varies substantially across alleles. Probabilistic models trained on immunopeptidomics data provide allele-specific descriptions of peptide presentation landscapes. In this work, we compare two such models: position weight matrices (PWMs), which assume positional independence, and energy-based restricted Boltzmann machines (RBMs), which capture higher-order positional correlations. We quantify repertoire diversity within an allele using Shannon entropy and divergence between allelic repertoires using Jensen--Shannon divergence (JSD). These quantities provide interpretable measures of presentation breadth and overlap, but their empirical estimation is challenging in finite samples, where plug-in estimators are biased and model fitting introduces additional uncertainty beyond sampling noise. To address this, we develop a bootstrap-based framework for bias correction and confidence interval construction that accounts for both finite sampling and model misspecification. Using synthetic ground-truth distributions with known entropy and Jensen--Shannon divergence (JSD), we benchmark parametric and non-parametric bootstrap procedures together with five confidence interval constructions across multiple scenarios. We find that the parametric bootstrap-t interval achieves the most reliable near-nominal coverage under practical sample-size and bootstrap-replicate settings. We then apply this framework to immunopeptidomics data from 68 HLA-I alleles spanning HLA-A, HLA-B, and HLA-C, enabling uncertainty-aware comparisons of intra-allelic diversity and inter-allelic divergence. Our results reveal substantial heterogeneity in allele-specific presentation landscapes and robustly quantify peptide repertoire overlap within and across supertypes. Overall, this work provides a statistically grounded framework for comparing HLA-I peptide presentation profiles, with potential applications in quantitative immunology, immunotherapy, and vaccine design.

Crohn’s disease often presents with fistulae, abnormal tunnels that connect the intestine to the skin or other organs. Despite their profound effect on morbidity, the molecular basis of fistula formation remains unclear, largely owing to the challenge of capturing intact fistula tracts and their inherent heterogeneity. Here we construct a subcellular-resolution spatial atlas of 68 intestinal fistulae spanning diverse anatomical locations. We describe fistula-associated epithelial, immune and stromal cell states, revealing abnormal zonation of growth factors and morphogens linked to establishment of tunnelling anatomy. We identify fistula-associated stromal (FAS) fibroblasts, which are assembled in concentric layers: a proliferative, lumen-adjacent zone beneath neutrophil and macrophage-rich granulation tissue, an active lesion core of FAS cells and a quiescent, pro-fibrotic outer zone. We examine the architecture of the extracellular matrix in the fistula tract and demonstrate that FAS populations associate with distinct collagen structures, exhibiting properties ranging from proliferation, migration and extracellular matrix remodelling to dense collagen deposition and fibrosis. We define niches supporting epithelialization of fistula tunnels and a FAS-like population that is detected at the base of ulcers in non-penetrating Crohn’s disease. Our study demonstrates that common molecular pathways and cellular niches underpin fistulae across intestinal locations, revealing the cellular protagonists of fistula establishment and persistence. This resource will inform the development of model systems and interventions to mitigate aberrant fibroblast activity while preserving their regenerative properties in Crohn’s disease.

Protein language models can learn rich representations from large-scale data, but they often overlook covariate feature information like experimental batch, species of origin, or assay type that play a central role in shaping downstream data distributions. This omission limits their effectiveness in many applications, where fine-tuning data are limited and covariate-driven variation must be preserved for accurate prediction. To address this, we introduce Covariate-based Rank Adaptation (CoRA), a parameter-efficient fine-tuning approach applicable to any protein language model with categorical covariate information. CoRA augments shared low-rank matrices using a diagonal gating matrix Γ(c) that is a function of the covariate. CoRA offers a robust framework for integrating and managing covariate information. Notably, standard LoRA emerges as a special case when Γ= I. On ProteinGym, conditioning on assay type and protein identity, CoRA surpasses both unconditional LoRA and per-condition adapters. More broadly, CoRA provides a general framework for covariate-aware adaptation in any setting where fine-tuning data include structured metadata

T cells are a key part of the adaptive immune system. Using their surface-bound T cell antigen receptors (TCRs), these cells scan peptides and other antigens presented to them by major histocompatibility complex molecules (MHCs) on the surface of nucleated cells, searching for abnormal cells. Although determining the map between TCRs and their target antigens is of vital importance for the design of safe and effective T cell-based vaccines and therapeutics, decoding these interactions is challenging. Experimental methods are not scalable, and sequence-based computational methods have issues generalising to new antigens. The IMMREP25 benchmark of methods for predicting T cell antigen specificity showed that AlphaFold-based methods promise improved generalisation to novel antigens. However, the ability of structure prediction models to predict T cell antigen specificity has not been robustly evaluated previously. In this work, we characterise AlphaFold’s ability to predict T cell antigen specificity. We developed a platform to run AlphaFold predictions for thousands of interacting TCRs and peptide-MHCs (pMHCs) and benchmark AlphaFold 3 and like models at predicting T cell antigen specificity. We investigated the underlying correlates of AlphaFold-derived binding scores and found that its predictive power is related to the positioning of TCRs over the pMHC and not chemical interactions. Furthermore, we refine the AlphaFold-derived binding scores and investigate the model’s ability to uncover the clustering rule of TCR repertoires and recapitulate mutational scanning experiments. Our results highlight both the promise and the current limitations of structure-based approaches for predicting TCR specificity, guiding the development of more reliable immunological prediction methods.

The success of antitumour immunotherapy is based on the ability of immune cells to migrate in the stroma and establish contact with the tumour cells. Previous experimental studies demonstrated the influence of collagen fibres’ orientations on the migration of resident T cells, showing that aligned fibres may prevent immune cells from entering tumour islets. However, the exact mechanisms governing the migration are still poorly understood. Here we present a stochastic agent-based model of T-cell migration in the tumour microenvironment, in which cells move alternately forwards and backwards along collagen fibres. We apply this framework to 3D microprinted scaffolds designed to replicate the fibre architectures characteristic of the tumour microenvironment, encompassing a range of structural configurations with varying degrees of alignment and network connectivity. Different scaffold configurations are shown to be associated with distinct migration behaviours, providing a controlled setting in which to isolate the structural determinants of immune cell movement. These findings offer mechanistic insight into the factors governing T-cell infiltration and may inform the design of therapeutic strategies aimed at enhancing immune infiltration in the context of antitumour immunotherapy

Data-driven methods for predicting scientific phenomena pose a foundational dilemma: while black box models yield better predictive accuracy at the expense of interpretability, carefully tailored and interpretable scientific models often produce worse predictions. Here, I introduce a new statistical approach geared towards resolving this choice, and which is yields high predictive accuracy from interpretable models through the mathematical innovation of predictively oriented uncertainty—a breakthrough discovery that dramatically boosts predictive performance of interpretable scientific models. To explain how the approach works, I will give some illustrations from nuclear physics models.

Since launch, AlphaFold has underperformed in predicting immunological complexes. AlphaFold obtains the underlying signals for structure predictions from Multiple Sequence Alignments (MSAs), in which evolutionary-scale coevolutionary changes (such as complementary substitutions) provide information on the spatial relationships between positions in sequence/structure space. Immune system complexes are different, though. In addition to evolutionary-scale relationships, there are also relationships between epitope and paratope, or, in the case of MHC molecules, between polymorphisms and the bound peptide repertoire. This information is not represented in the default AlphaFold MSAs. We have recently developed a computational screen to understand how to provide this information as part of the MSAs,, in particular regarding their composition. Too little information leads to poor-quality predictions, while too much or too similar information reduces model confidence. Using these specific MSAs, which encode information describing peptide repertoires and polymorphisms, significantly improves the performance of the open-source AlphaFold2 model without retraining or code alteration, enabling very high-quality predictions at scale on commodity NVIDIA GPUs.

Autoencoders are a class of deep learning architectures with the ability to uncover the low-dimensional surfaces on which high dimensional datasets reside. These approaches have gained widespread use in single cell analyses to learn patterns of biological organization in immunity and diverse other contexts. However, current methods have limitations with generalizability and predictions on new datasets not represented in the training phase. Applying methods of differential geometry to the learned manifolds, we have developed a new class of autoencoders which make significantly more robust embeddings for out of distribution datasets and generative predictions in novel contexts.

Understanding variation in immune responses across tissues and over time, and the cell-specific mechanisms underlying these differences, remains a central challenge in human immunology. I present two case studies applying single-cell RNA-sequencing in distinct contexts. First, using paired peripheral blood and cerebrospinal fluid from patients with tuberculous, bacterial, viral, or cryptococcal meningitis, I identify compartment-specific immune programmes and pathogen-associated signatures that distinguish inflammatory states across tissues. Second, I analyse longitudinal samples from individuals receiving ChAdOx1 nCoV-19 or BNT162b2 vaccination, revealing early transcriptional responses shaped by vaccine platform and dosing interval. Together, these analyses show that single-cell approaches can resolve spatial and temporal heterogeneity in human immunity, providing insights into disease pathogenesis and vaccine responses.

High-affinity antibody–producing B cells arise through affinity maturation within germinal centers (GCs), guided by T follicular helper (Tfh) cells. Recent studies reveal that these helper cells also undergo antigen-dependent selection, dynamically reshaping their repertoire. Yet, how mutual feedback between the two populations shapes affinity maturation remains unclear. To address this gap, we developed a population dynamics model that captures key Tfh cell properties and their reciprocal interactions with B cells. We find that such reciprocal interactions enforce robust “homeostatic” control of GC composition, maintaining a constant B-to-T cell ratio. Moreover, preferential expansion of high-sensitivity clones prolongs the GC reaction. Together, these findings highlight dynamic reciprocity between B and T cells as a mechanism for stabilizing GC dynamics—a principle that may extend to other adaptive immune processes.