• Gregoire Altan-Bonnet

We present an experimental/theoretical pipeline to build quantitative models of leukocyte activation. We introduce a robotic platform to quantify the dynamics of cell differentiation and cytokine production/consumption by T cells ex vivo. These high-dimensional dynamics can be compressed into a 2D model using tools from machine learning. Our model highlights two modalities of T cell activation that enforce adaptive kinetic proofreading of antigen-TCR interactions, and that encode antigen discrimination. We test our model of antigen discrimination across varied immunological settings, including CAR-T and signaling-impaired T cells. To conclude, we highlight the power of lab automation, data integration, machine learning and theoretical modeling to usher new insights in systems immunology.

• Duygu Ucar

• Kenneth Hoehn, Kenneth Hoehn, Dartmouth College
• Cole Jensen, Cole Jensen, Yale School of Medicine
• Jake Sumner, Jake Sumner, Yale School of Medicine
• Steven Kleinstein, Steven Kleinstein, Yale School of Medicine

Antibodies are vital to human immune responses and are composed of genetically variable heavy and light chains. These structures are initially expressed as B cell receptors (BCRs). BCR diversity is shaped through somatic hypermutation and selection during immune responses. This evolutionary process produces B cell clones, cells that descend from a common ancestor but differ by mutations. Phylogenetic trees inferred from BCR sequences can reconstruct the history of mutations within a clone. Until recently, BCR sequencing technologies separated heavy and light chains, but advancements in single cell sequencing now pair heavy and light chains from individual cells. However, it is unclear how these separate genes should be combined to infer B cell phylogenies. In this study, we investigated strategies for using paired heavy and light chain sequences to build phylogenetic trees. We found incorporating light chains significantly improved tree accuracy and reproducibility across all methods tested. This improvement was greater than the difference between tree building methods and persisted even when mixing bulk and single cell sequencing data. However, we also found that many phylogenetic methods estimated significantly biased branch lengths when some light chains were missing, such as when mixing single cell and bulk BCR data. This bias was eliminated using maximum likelihood methods with separate branch lengths for heavy and light chain gene partitions. Thus, we recommend using maximum likelihood methods with separate heavy and light chain partitions, especially when mixing data types. We implemented these methods in the R package Dowser: https://dowser.readthedocs.io.

• Hatice Osmanbeyoglu, Hatice Osmanbeyoglu, University of Pittsburgh

• Alireza Karbalayghareh, Alireza Karbalayghareh, Memorial Sloan Kettering Cancer Center
• Darko Barisic, Darko Barisic, Weill Cornell Medicine
• Christopher Chin, Christopher Chin, Weill Cornell Medicine
• Ari Melnick, Ari Melnick, Weill Cornell Medicine
• Christina Leslie, Christina Leslie, Memorial Sloan Kettering Cancer Center

The epigenome and transcriptome influence each other in differentiation trajectories. We use single cell multiomic data, with RNA expression and chromatin accessibility (ATAC) readouts, to model the interplay of epigenome and transcriptome in dynamic germinal center (GC) B cells. We have developed a model, DynaVelo, to learn the latent time and joint dynamics of cells in both RNA expression and TF motif accessibility spaces. In addition to RNA velocities that inform us of cell trajectories in RNA space, we further define an analogous velocity in the motif accessibility space. We adopt techniques from variational autoencoders and neural ODEs and leverage RNA expression and velocity as well as motif accessibility to learn the joint dynamics of cells. Finally, we apply these models to the GC B cells to learn the dynamics of wildtype samples as well as samples with mutations in epigenetic regulators ARID1A and CTCF, modeling somatic alterations seen in B cell lymphomas. The learned latent times demonstrate plausible starting and end points that are consistent with the velocities. In silico perturbation of these models provide predictions for the impact of experimental interventions on the trajectories of B cells. We perform in silico perturbations of critical B cell TFs to dissect their roles in GC dynamics and identify potential perturbation targets for rescuing the loss of function in mutant samples.

• Alisa Omelchenko, Alisa Omelchenko, University of Pittsburgh
• Jane Siwek, Jane Siwek, University of Pittsburgh
• Prabal Chhibbar, Prabal Chhibbar, University of Pittsburgh
• Sanya Arshad, Sanya Arshad, University of Pittsburgh
• Iliyan Nazarali, Iliyan Nazarali, University of Pittsburgh
• Kiran Nazarali, Kiran Nazarali, University of Pittsburgh
• Annaelaine Rosengart, Annaelaine Rosengart, University of Pittsburgh
• Javad Rahimikollu, Javad Rahimikollu, University of Pittsburgh
• Jeremy Tilstra, Jeremy Tilstra, University of Pittsburgh
• Mark Shlomchik, Mark Shlomchik, University of Pittsburgh
• David Koes, David Koes, University of Pittsburgh
• Alok Joglekar, Alok Joglekar, University of Pittsburgh
• Jishnu Das, Jishnu Das, University of Pittsburgh

Language models (LMs) and protein LMs have now been employed in many frameworks. Traditionally for protein-protein or peptide-protein interactions (PPIs), corresponding sequences are concatenated and co-embedded. However, these are highly limited as no method utilizes a language representation of the interaction itself. We developed Sliding Window Interaction Grammar (SWING), a novel interaction LM (iLM) that leverages differences in amino acid properties to generate an interaction vocabulary. This is then input into an LM, and corresponding features are used for different downstream supervised prediction steps.
SWING was first applied to predicting peptide:MHC (pMHC) interactions. Currently, existing approaches have separate Class I and Class II pMHC prediction models as these interactions are distinct structurally and functionally. SWING however was not only successful at generating Class I and Class II models that have comparable prediction to state-of-the-art approaches, but a SWING model trained only on Class I alleles was predictive for Class II, a complex task not attempted by any existing approach. For de-novo data, using only Class I or Class II data, SWING also accurately predicted Class II pMHC interactions in murine models of SLE and T1D, that were validated experimentally.
SWING also predicted the disruption of specific interactions by missense mutations. Modern methods like AlphaMissense and ESM1b can predict pathogenicity per mutation but are unable to predict interaction-specific disruptions. SWING accurately predicted the impact of Mendelian mutations and population variants on PPIs, outperforming AlphaMissense/ESM1b. Overall, SWING is a first-in-class few-shot iLM that learns the language of PPIs using sequence alone.

• Michael Klausen, Michael Klausen, Evaxion Biotech A/S
• Christian Garde, Christian Garde, Evaxion Biotech A/S
• Thomas Trolle, Thomas Trolle, Evaxion Biotech A/S
• Jens Kringelum, Jens Kringelum, Evaxion Biotech A/S

In this study, we address the challenge of accurately predicting peptide presentation by Major Histocompatibility Complex (MHC) molecules, a key feature in developing efficacious personalized and precision vaccines. Our approach introduces a novel deep learning framework that enhances prediction accuracy by utilizing the growing wealth of available immunopeptidomic data. Our methodology introduces three new strategies: The creation of a unified representation for both MHC class I and -II molecules, the implementation of a deep transformer encoder-decoder architecture, and the adoption of a generative adversarial network (GAN) pretraining mechanism.

The unified representation allows our model to leverage MHC binding cores across MHC classes and the deep learning architecture facilitates direct training on peptide sequences without the need for pre-processing, such as the extraction or alignment of binding cores. The GAN pretraining stabilizes the training phase and enhances overall performance by generating synthetic peptide data for training enhancement.

Our results demonstrate significant improvements in the prediction of peptide-MHC interactions, particularly for MHC class II molecules, which have historically been difficult to predict accurately. This advancement offers a more reliable and efficient tool for the design of personalized cancer immunotherapies and the development of vaccines against a wide range of pathogens. Through these technical achievements, our work contributes to the advancement of precision medicine, highlighted by a significant improvement when deployed in our AI-immunology™ platform. An improvement shown in preclinical studies and now being used in clinical trials.

• Ha Young Kim, Ha Young Kim, Korea Advanced Institute of Science and Technology
• Sungsik Kim, Sungsik Kim, GENINUS Inc.
• Woong-Yang Park, Woong-Yang Park, GENINUS Inc.
• Dongsup Kim, Dongsup Kim, Korea Advanced Institute of Science and Technology

Prediction of T-cell receptor (TCR)-epitope interactions is important for many applications in biomedical research, such as cancer immunotherapy and vaccine design. Although numerous computational methods have been developed, the prediction of TCR-epitope interactions remains challenging. The prediction is known to be particularly difficult for novel epitopes, due to the scarcity of available data. Here, we propose TSpred, a new deep learning approach for the pan-specific prediction of TCR binding specificity based on paired chain TCR data. We develop a robust model that generalizes well to unseen epitopes by utilizing an ensemble model of a CNN-based model and an attention-based model. In particular, we design a reciprocal attention mechanism which is specifically designed to extract the patterns underlying TCR-epitope interactions. Upon a comprehensive evaluation of our model, we find that TSpred achieves state-of-the-art performances in both seen and unseen epitope specificity prediction tasks. Also, compared to other predictors, TSpred is more robust to bias related to peptide imbalance in the dataset. Furthermore, the reciprocal attention component of our model allows for model interpretability by capturing structurally interacting residue pairs that contribute to TCR-epitope binding. Results indicate that TSpred is a robust and reliable method for the task of TCR-epitope binding prediction.

• Susanna Kiwala, Susanna Kiwala, Department of Medicine
• Huiming Xia, Huiming Xia, Department of Medicine
• Megan M. Richters, Megan M. Richters, Department of Medicine
• Evelyn Schmidt, Evelyn Schmidt, Department of Medicine
• My Hoang, My Hoang, Department of Medicine
• Kartik Singhal, Kartik Singhal, Department of Medicine
• Mariam Khanfar, Mariam Khanfar, Department of Medicine
• Jasreet Hundal, Jasreet Hundal, McDonnell Genome Institute
• Alexander

Personalized neoantigen vaccines utilize immunogenomics and immuno-oncology strategies to combat cancer. Somatic variants in tumor cells generate neoantigens that may bind to MHC molecules and get presented on the tumor cell’s surface. Immunotherapies, such as checkpoint blockade therapies, personalized cancer vaccines and other T cell based therapies, target these neoantigens to stimulate a tumor-specific immune response.

We have developed a computational framework for neoantigen identification and prioritization, pVACtools (pVACtools.org), that integrates tumor variant and expression data (DNA- and RNA-Seq) into an end-to-end solution for design of neoantigen targeting therapies including personalized vaccines. pVACtools consists of multiple command line tools for neoantigen prediction from somatic alterations (pVACseq, pVACfuse, pVACsplice, and pVACbind), a tool for designing DNA vector–based constructs (pVACvector), as well as a web application (pVACview) for visualizing, reviewing, prioritizing, and selecting neoantigen candidates for peptide or nucleic acid vaccine manufacturing platforms.

The full pVACtools suite seamlessly allows users to: 1) Identify altered peptides from different mechanisms (i.e. point mutations, indels, gene fusions, splice site alterations, or frameshift variants). 2) Predict peptide binding affinity, elution, and immunogenicity metrics via an ensemble of MHC Class I and II algorithms. 3) Filter based on variant allele expression, read-counts, variant allele fractions, transcript support level, and peptide binding affinities. 4) Visualize results to allow further interactive neoantigen prioritization. 5) Design DNA-vector vaccines.

pVACtools rapidly and efficiently identifies potentially immunogenic neoepitopes, and is being used in both basic and translational research, as well as over 10 clinical trials on clinicialtrials.gov to date.

• Ayshwarya Subramanian, Ayshwarya Subramanian, Cornell University
• Kimitoshi Kimura, Kimitoshi Kimura, Kyoto University
• Vijay Kuchroo, Vijay Kuchroo, Harvard Medical School

Microglia are the major resident innate immune cells of the central nervous system (CNS) playing essential roles in CNS homeostasis and disease. Although strongly implicated in Alzheimer’s disease (AD) pathology, the regulatory mechanism of microglial activation remains to be fully elucidated. By combining population and single-cell (scRNA-seq) transcriptomics, IP-MS, neuroinflammation assays, and behavioral studies, we investigated the role of the AD risk gene (Wightman et al., Nat Genet, 2021) and immune checkpoint molecule HAVCR2/TIM-3 in microglial homeostasis and activation. Tim-3 (Havcr2) is highly and specifically expressed in microglia compared to other cell types in both mouse and human brain. Gene expression profiles of Tim-3 deficient microglia resemble those of phagocytic microglia, and microglia in neurodegeneration (MGnD). Mechanistically, Tim-3 enhances TGF-b signaling by promoting phosphorylation of Smad2, thus contributing to the homeostasis of microglia, in a mechanism independent of its activity in T-cells. Importantly, we report that microglia-specific Tim-3 deficiency reduced Amyloid beta (Ab) plaque load and neuronal damage, and resulted in improved cognitive function in a 5XFAD mouse model of AD. Single-nucleus and single-cell RNA sequencing identified that microglia in Havcr2-deficient 5xFAD mice are characterized by increased pro-phagocytic and anti-inflammatory gene expression together with decreased proinflammatory gene expression. Thus, Tim-3 may serve to regulate phagocytotic and inflammatory functions of activated microglia in AD. Collectively, these results hold promise for a potential new therapeutic strategy targeting the checkpoint molecule Tim-3 in AD.

• Sarah Walker, Sarah Walker, Princeton University
• Joris van der Veeken, Joris van der Veeken, Research Institute of Molecular Pathology
• Yuri Pritykin, Yuri Pritykin, Princeton University

Mechanisms of gene expression regulation in T cell activation and differentiation across lineages and functional states are incompletely understood. To address this, we generated and analyzed single-cell multi-omic scATAC+scRNA-seq data in splenic T cells (including both CD4+ and CD8+) responding to acute and chronic mouse infection with lymphocytic choriomeningitis virus (LCMV). Computational scATAC-seq data analysis is challenging due to extreme sparsity and incompletely understood statistical properties of the data, as well as high level of cell type specificity and lack of available epigenomic references. Thus, we collected a compendium of published bulk ATAC-seq datasets for mouse CD4 and CD8 T cells, and then built a comprehensive atlas of chromatin accessibility peaks and epigenomic signatures of functional T cell states. This allowed us to analyze our new scATAC-seq data at unprecedented level of detail. We were able to clearly distinguish between CD4 and CD8 T cells, as well as between naive and memory cells in both lineages. We observed that most cell states, including both effector and exhausted cells, as well as progenitor stem-cell-like Tcf1+ CD8 T cells, were found in both acute and chronic responses, though sometimes with different frequency. Interestingly, progenitor Tcf1+ CD8 T cells were most similar epigenetically and transcriptionally to CD4 follicular helper T cells. We hypothesize that similar molecular regulation, potentially driven by their direct or shared physical cell-cell interactions or shared microenvironment, may be associated with their shared regulatory program.

• Sara Capponi, Sara Capponi, IBM Research
• Kahn Rhrissorrakrai, Kahn Rhrissorrakrai, IBM Research
• Meltem Tolunay, Meltem Tolunay, IBM Quantum
• Tanvi Gujarati, Tanvi Gujarati, IBM Quantum
• Jie Shi, Jie Shi, IBM Research
• Filippo Utro, Filippo Utro, IBM Research
• Roshan M Regy, Roshan M Regy, IBM Quantum
• Joseph Morrone, Joseph Morrone, IBM Research
• Laxmi Parida, Laxmi Parida, IBM Research

Chimeric antigen receptor (CAR) T-cells are promising new medicines that apply to the treatment of many cancers, and potentially represent novel approaches to treat infectious diseases and autoimmunity. A central challenge in expanding and enhancing CAR T-cell functions is identifying beneficial combinations of intracellular costimulatory motifs to elicit desired phenotypes. This is due to the large motif combinatorial space and high experimental costs, in time and resources, to generate and measure CAR performance. This results in a highly data-constrained problem where developing accurate predictive models of CAR T-cell behavior is difficult. State-of-the-art machine learning (ML) models based on convolutional neural networks (CNNs) combined with long-short term memory have been shown to reach an accuracy of 70% when predicting CAR T-cell stemness and cytotoxicity. With the clear need for more accurate predictive models, we investigated the performance of quantum convolutional neural networks (QCNNs) as a novel ML model for improved predictive performance. QCNNs carry advantages over classical CNNs, including good generalization in underdetermined problems and usage of significantly fewer training parameters as these scale logarithmically with the number of qubits. QCNNs also have advantages over certain other quantum ML methods, including the absence of barren plateaus. Our study showed that QCNN matches and occasionally exceeds the performance of CNNs in classifying CAR T cells by cytotoxicity levels. Employing larger and more expressive QCNN models may further enhance performance, potentially resulting in a superior predictive tool for CAR T-cell phenotypes.

• Seda Arat, Seda Arat, Pfizer
• Athena Mustakis, Athena Mustakis, Pfizer
• Dac-Trung Nguyen, Dac-Trung Nguyen, Pfizer
• Grigorios Koulouras, Grigorios Koulouras, Pfizer
• Helena Qi, Helena Qi, Pfizer
• Brajesh Rai, Brajesh Rai, Pfizer
• Matt Martin, Matt Martin, Pfizer

SARS-CoV-2 virus is a member of a large family of viruses called coronaviruses that causes a respiratory disease COVID-19. There are almost 775 million cases and more than 7 million deaths worldwide. While there are short-term effects of SARS-CoV-2 infection such as fever, fatigue, headache and loss of taste/smell, the long-term effects can be very serious, including organ damage. In some people, lasting health effects may include, heart complications, chronic kidney impairment, stroke, and Guillain-Barre syndrome (GBS). Myocarditis is also a well-known adverse event associated with both SARS-CoV-2 infection and mRNA-based COVID-19 vaccines. This study focuses on (1) building a computational pipeline based on protein sequence and 3D structure similarities, and (2) determining if these adverse events are occurring due to molecular mimicry. Initially, we identified ~2500 human proteins that contains peptides (9 and 15 amino acids) with at least 50% similarity within the SARS-CoV-2 protein sequences. These proteins were triaged to more than 50 human endogenous proteins highly expressed in the heart (for myocarditis), kidney (for chronic kidney impairment) and brain (for GBS) that could potentially lead to immune-mediated adverse events based on 3D structure analysis using Alphafold. The presentation will describe our computational approach that can be instrumental for infectious disease, vaccine and drug safety scientists for auto-immunity risk assessment.

• Joshua Welch

• Joseph Ng, Joseph Ng, University College London
• Guillem Montamat Garcia, Guillem Montamat Garcia, University College London
• Alexander Stewart, Alexander Stewart, University of Surrey
• Paul Blair, Paul Blair, University College London
• Claudia Mauri, Claudia Mauri, University College London
• Deborah Dunn-Walters, Deborah Dunn-Walters, University of Surrey
• Franca Fraternali, Franca Fraternali, University College London

Class-switch recombination (CSR) is an integral part of B cell maturation. Whilst tools such as pseudotime and RNA velocity exists to infer cellular dynamics in single-cell RNA sequencing (scRNA-seq) data, they ignore processes such as CSR which are specific to B cells. Here we present sciCSR (pronounced ‘scissor’, single-cell inference of class-switch recombination), a computational pipeline that analyzes CSR events and dynamics of B cells from scRNA-seq experiments. Validated on both simulated and real data, sciCSR re-analyzes scRNA-seq alignments to differentiate productive heavy-chain immunoglobulin transcripts from germline ‘sterile’ transcripts. From a snapshot of B cell scRNA-seq data, a Markov state model is built to infer the dynamics and direction of CSR. Applying sciCSR on SARS-CoV-2 time-course scRNA-seq data, we observe that sciCSR predicts, using data from an earlier time point in the collected time-course, the isotype distribution of B cell receptor repertoires of subsequent time points with high accuracy (cosine similarity ~0.9). Using processes specific to B cells, sciCSR identifies transitions that are often missed by conventional RNA velocity analyses and can reveal insights into the dynamics of B cell CSR during immune response. We believe sciCSR offers a starting point to model B cell maturation in scRNA-seq data; these models can be further analyzed to understand the molecular cues of CSR and different steps of the maturation process, their regulation in situ within tissues, and their dysregulation in diseases.

• Lingyu Li, Lingyu Li, The University of Hong Kong
• Yuanhua Huang, Yuanhua Huang, The University of Hong Kong

Cell-cell communication (CCC) is crucial for understanding complex tumor microenvironments, such as interactions between tumor and immune cells. Many of these intricate intercellular interactions are facilitated by ligand-receptor interactions (LRIs) within limited spatial ranges. Spatial Transcriptomics (ST) has advanced LRI identification by detecting spatially co-expressed ligand-receptor pairs (LR pairs). However, the widely used 10x Visium ST platform measures the transcriptome in low-resolution spots and low-capture rate tissues, which limits its ability to detect fine-grained detection of intricate communication patterns. Here, we propose FineST (Fine-grained Spatial Transcriptomics), a bi-modal deep fusion framework that combines spatial RNA-seq with histological images for super-resolution ST expression prediction and detailed LRI discovery at sub-spot and single-cell levels. We evaluated FineST's performance on various datasets, including Visium nasopharyngeal carcinoma (NPC), Xenium breast cancer (BRCA), and Visium HD colorectal cancer (CRC) from 10x Genomics, and compared its accuracy with state-of-the-art methods TESLA and iStar. As a computational toolbox, FineST further detects refined LRIs to provide insights into context-specific cellular interactions and signaling, including fine-grained interacting visualizations of spatially co-expressed LR pairs, distinct CCC pattern clustering, pattern consistency analysis at sub-spot and single-cell levels, and functional enrichment analysis of putative pathways. Our results reveal that FineST's prediction for tumor-restricted genes more closely resembles the ground truth than TESLA and iStar. Moreover, FineST's predicted super-resolved ST achieves superior power in identifying interacting LRIs and uniquely discerns CCC patterns across immune and tumor cells, enabling the enormous potential for high-throughput analysis of histology images in both research and clinical applications.

• Chang Lu, Chang Lu, Institute for Computational Biomedicine
• Camilla Engblom, Camilla Engblom, SciLifeLab
• Kim Thrane, Kim Thrane, SciLifeLab
• Qirong Lin, Qirong Lin, Department of Cell and Molecular Biology
• Joakim Lundeberg, Joakim Lundeberg, SciLifeLab
• Jovan Tanevski, Jovan Tanevski, Institute for Computational Biomedicine
• Julio Saez-Rodriguez, Julio Saez-Rodriguez, Institute for Computational Biomedicine

Understanding the role of adaptive immune cells (i.e. T and B lymphocytes) in the immune system is key to treating a range of diseases from autoimmune disorders to cancer and improving vaccination. Recent biotechnological advancements enable spatial mapping of T and/or B cell receptor sequences (VDJ sequences) within tissues, offering insights into spatially resolving clonal diversity with respect to tissue morphology and gene expression. However, existing computational tools for VDJ analysis primarily focus on single-cell or bulk VDJ sequencing data, necessitating novel methodologies to comprehensively understand how these antigen receptors relate to cellular patterns and signaling.
In this study, we propose a computational strategy to elucidate immune-related sets of intra- and intercellular signaling relationships at a spatial resolution, leveraging both spatial and sequence-based similarities inherent in the adaptive immune repertoires of T and B lymphocytes. Cell deconvolution techniques were first employed to infer the cell-subset-specific expression profiles of genes within the spatially resolved gene expression data. Next, gene co-expression networks were constructed based on the gene expression profiles of individual cell subsets, alongside T-cell and B-cell clonal networks derived from the similarities in clonal sequences. These networks were subsequently integrated into multilayer networks based on the spatial co-localization patterns of tissues. Finally, cross-network community associations were calculated by employing deep multi-graph clustering, thereby unraveling both intracellular signaling dynamics (intralayer relationships) and intercellular communication modalities (interlayer relationships).
This approach promises profound insights into immune dysregulation, facilitating the exploitation of antigen-specific clones for tailored cellular and antibody-based therapies.