Early and accurate disease diagnosis is crucial for preventing disease development and defining therapy strategies. Due to predominantly unspecific symptoms, diagnosis of autoimmune diseases is notoriously challenging. However, multiple types of data are not leveraged for precision diagnostics due to the difficulties of integrating and encoding multi-omics data with clinical and laboratory values, as well as a lack of standardization; clinical decision support systems are often limited to only certain data types. Accordingly, even sophisticated data models fall short when supporting accurate diagnoses and presenting data analyses in a user-friendly form. Therefore, the integration of various data types is not only an opportunity but also a competitive advantage in research and for the industry. We have developed an integration pipeline to enable the use of machine learning for patient classification based on multi-omics data such as genetics, immunomics and metabolomics, in combination with clinical values and laboratory results. Machine learning models resulted in 95% prediction accuracy of autoimmune diseases using integrated data. Our results deliver insights into autoimmune disease research and have the potential to be adapted for applications across disease conditions.

We profiled blood and draining lymph node (LN) samples from human volunteers after influenza vaccination over two years to define evolution in the T follicular helper cell (TFH) response. We show LN TFH cells expanded in a clonal-manner during the first two weeks after vaccination and persisted within the LN for up to six months. LN and circulating TFH (cTFH) clonotypes overlapped but had distinct kinetics. LN TFH cell phenotypes were heterogeneous and mutable, first differentiating into pre-TFH during the month after vaccination before maturing into GC and IL-10+ TFH cells. TFH expansion, upregulation of glucose metabolism, and redifferentiation into GC TFH cells occurred with faster kinetics after re-vaccination in the second year. We identified several influenza-specific TFH clonal lineages, including multiple responses targeting internal influenza proteins, and show each TFH state is attainable within a lineage. This study demonstrates that human TFH cells form a durable and dynamic multi-tissue network.

Colonization by the microbiota causes a marked stimulation of B cells and induction of immunoglobulin, but mammals colonized with many taxa have highly complex and individualized immunoglobulin repertoires. To study this, we opted for a simplified model of defined transient exposures to different microbial taxa in germ-free mice to deconstruct how the microbiota shapes the B cell pool and its functional responsiveness. We previously showed that microbial exposure at the intestinal mucosa generated oligoclonal responses that differed from those in germ-free mice, and from the diverse repertoire that was generated after intravenous systemic exposure to microbiota. Our results reflected a contrast between a flexible response to systemic exposure with the need to avoid fatal sepsis, and a restricted response to mucosal exposure that reflects the generic nature of host–microbial mutualism in the mucosa. A hallmark of mucosal IgA is the high mutational load that accumulates throughout life. This has been mainly interpreted in terms of differences in microbial induction, although once established in early life, the microbiome of humans and experimental mice is relatively stable and mutational activity has been shown to be independent of B cell receptor signaling. Using germ-free and colonized mice provided with different diets formulated with proprietary grain-based processing or from purified chemicals with different principal macronutrient calorie sources, we show that diet affects IgA induction, its repertoire and mutational diversification independently of microbial exposure.

The recognition of a T-cell epitope target is driven by the unique sequence of the T-cell receptor (TCR). As the TCR sequence theoretically contains all the information that determines its target, it must be possible to infer its target from this sequence. Indeed, within specific settings, we now have performant machine learning models that can address this challenge. It has therefore become necessary to start considering how we can apply these models to gain novel immunological understanding.

In this talk, I want to focus on the application of TCR-epitope prediction models to identify epitope-specific T-cells in full repertoire data, and the additional challenges that are often missed by current evaluation efforts. I also aim to highlight the biggest advancements and milestones that we can expect in the coming years in the field, especially in regard to the unseen epitope prediction problem.

Pathogens and cancer cells can escape recognition by T cell receptors (TCRs) through mutations of immunogenic epitopes. TCR cross-reactivity, i.e., recognition of multiple epitopes with sequence similarities, can counteract such mutational escape, but also may cause severe side effects in cell-based immunotherapies. To predict the effect of epitope mutations on T cell functionality in silico, we present “Predicting T cell Epitope-specific Activation against Mutant versions” (P-TEAM). The Random Forest based model was developed on two comprehensive datasets of murine and human TCRs in response to systematic single-amino acid mutations of their target epitopes to predict T cell reactivity for unobserved mutations, or even unseen TCRs. P-TEAM is complemented with an active learning framework to guide experimental design to minimize primary data acquisition costs. Overall, P-TEAM provides an effective computational tool to study T cell responses against mutated epitopes.

T cells orchestrate the adaptive immune response against pathogens and cancer by recognizing epitopes presented on MHC molecules. The heterogeneity of the MHC peptidome, including the high polymorphism of MHC genes, is influencing TCR repertoires and represents an important challenge towards accurate prediction and identification of T-cell epitopes in different individuals and different species. Here we generated and curated a dataset of more than a million unique MHC-I and MHC-II ligands identified by mass spectrometry. This enabled us to precisely determine the binding motifs of >200 MHC alleles across human, mouse, cattle and chicken. Analysis of these binding specificities combined with X-ray crystallography refined our understanding of the molecular determinants of MHC motifs and revealed alternative binding modes of MHC ligands. We then developed machine learning frameworks to accurately predict binding specificities and ligands of any MHC-I (MixMHCpred) and MHC-II (MixMHC2pred) allele, and further integrated TCR recognition into our epitope prediction pipeline (PRIME). Prospectively applying our tools to SARS-CoV-2 proteins identified several epitopes and TCR sequencing revealed a monoclonal response in effector/memory CD8+ T cells against one of these epitopes with cross-reactivity against the homologous peptides from other coronaviruses. Overall, our work shows how in depth characterization of MHC motifs can help mapping the targets of T cells and understanding TCR cross-reactivity.

Despite the vast diversity of the antibody repertoire, infected individuals often mount antibody responses to precisely the same epitopes within antigens. The immunological mechanisms underpinning this phenomenon remain unknown. By mapping 376 immunodominant “public epitopes” at high resolution and characterizing several of their cognate antibodies, we concluded that germline-encoded sequences in antibodies drive recurrent recognition. Systematic analysis of antibody-antigen structures uncovered 18 human and 21 partially overlapping mouse germline-encoded amino acid–binding (GRAB) motifs within heavy and light V gene segments that in case studies proved critical for public epitope recognition. GRAB motifs represent a fundamental component of the immune system’s architecture that promotes recognition of pathogens and leads to species-specific public antibody responses that can exert selective pressure on pathogens.

As individual researchers, we are well-trained in designing research projects that address specific and moderately challenging questions. However, when faced with a grand challenge such as understanding antigen recognition by T- and B-cell receptors, how can we, both individually and collectively, effectively tackle this problem? We currently have antigen binding data for only a minuscule proportion of the immune receptor sequence space, and despite the rapid emergence of new machine learning methods, consensus on what constitutes the most promising directions forward remains scarce.

In my talk, I will offer my perspective on this strategic question and discuss some recently published and ongoing work aligned with this strategy. Briefly, I believe the first step should be a rigorous characterisation of the computational challenges of the immune receptor-antigen binding prediction problem. Second, we must ensure that sufficient data is available to guide methodology development, where we in the foreseeable future need to rely on the combined use of experimental and simulated data. Third, we must prioritise interoperability and reproducibility of methods, along with the development of standardised benchmarks, to effectively compare performance and identify limitations of current approaches.

A new approach that utilizes the Lempel-Ziv 76 algorithm (LZ-76) to encode and represent AIRRs without relying on sequence annotation. The approach involves creating a graph-like model, which enables a wide range of potential applications, including generation probability inference, informative feature vector derivation, sequence generation, sequence analysis, and a new measure for repertoire diversity estimation.

T cells play a crucial role in our immunity by recognizing and eliminating infected and abnormal cells. T cell response is triggered upon T cell recognition of specific antigens presented by MHC molecules on the surface of target cells. However, the underlying rules of the interactions are incompletely understood. Over the past few years, we have been investigating T cell cross-reactivity and common specificity as two key drivers of T cell antigen specificity. In this seminar, I will focus on T cell cross-reactivity, and will introduce a new deep neural network model that we have developed for accurate and reliable prediction of CD8 T cell targets. I will illustrate how we have been using this model to model the dynamics of CD8 T cell response to emerging SARS-CoV-2 mutant variants.

Antibodies result from the competition of B cell lineages evolving under selection for improved antigen recognition, a process known as affinity maturation. High-affinity antibodies to pathogens such as HIV, influenza, and SARS-CoV-2 are frequently reported to arise from B cells whose receptors are encoded by particular immunoglobulin genes. This raises the possibility that the presence of particular germline genes in the B cell repertoire is a major determinant of the quality of the antibody response. Alternatively, initial differences in germline genes’ propensities to form high-affinity receptors might be overcome by chance events during affinity maturation. We first show how this can happen in simulations: even when fitness differences between germline genes lead to similar gene usage across individuals early on, gene usage can become increasingly dissimilar over time. We next find that mice experimentally infected with influenza virus demonstrate the same pattern of divergence in the weeks following infection. We investigated whether affinity maturation might nonetheless strongly select for particular amino acid motifs across diverse genetic backgrounds, but we found no evidence of convergence to similar CDR3 sequences or amino acid substitutions. These results suggest germline-encoded specificities might enable fast recognition of specific antigens early in the response, but diverse evolutionary routes to high affinity limit the genetic predictability of responses to infection and vaccination in the long term.

Early and accurate disease diagnosis is crucial for preventing disease development and defining therapy strategies. Due to predominantly unspecific symptoms, diagnosis of autoimmune diseases is notoriously challenging. However, multiple types of data are not leveraged for precision diagnostics due to the difficulties of integrating and encoding multi-omics data with clinical and laboratory values, as well as a lack of standardization; clinical decision support systems are often limited to only certain data types. Accordingly, even sophisticated data models fall short when supporting accurate diagnoses and presenting data analyses in a user-friendly form. Therefore, the integration of various data types is not only an opportunity but also a competitive advantage in research and for the industry. We have developed an integration pipeline to enable the use of machine learning for patient classification based on multi-omics data such as genetics, immunomics and metabolomics, in combination with clinical values and laboratory results. Machine learning models resulted in 95% prediction accuracy of autoimmune diseases using integrated data. Our results deliver insights into autoimmune disease research and have the potential to be adapted for applications across disease conditions.

We profiled blood and draining lymph node (LN) samples from human volunteers after influenza vaccination over two years to define evolution in the T follicular helper cell (TFH) response. We show LN TFH cells expanded in a clonal-manner during the first two weeks after vaccination and persisted within the LN for up to six months. LN and circulating TFH (cTFH) clonotypes overlapped but had distinct kinetics. LN TFH cell phenotypes were heterogeneous and mutable, first differentiating into pre-TFH during the month after vaccination before maturing into GC and IL-10+ TFH cells. TFH expansion, upregulation of glucose metabolism, and redifferentiation into GC TFH cells occurred with faster kinetics after re-vaccination in the second year. We identified several influenza-specific TFH clonal lineages, including multiple responses targeting internal influenza proteins, and show each TFH state is attainable within a lineage. This study demonstrates that human TFH cells form a durable and dynamic multi-tissue network.

Colonization by the microbiota causes a marked stimulation of B cells and induction of immunoglobulin, but mammals colonized with many taxa have highly complex and individualized immunoglobulin repertoires. To study this, we opted for a simplified model of defined transient exposures to different microbial taxa in germ-free mice to deconstruct how the microbiota shapes the B cell pool and its functional responsiveness. We previously showed that microbial exposure at the intestinal mucosa generated oligoclonal responses that differed from those in germ-free mice, and from the diverse repertoire that was generated after intravenous systemic exposure to microbiota. Our results reflected a contrast between a flexible response to systemic exposure with the need to avoid fatal sepsis, and a restricted response to mucosal exposure that reflects the generic nature of host–microbial mutualism in the mucosa. A hallmark of mucosal IgA is the high mutational load that accumulates throughout life. This has been mainly interpreted in terms of differences in microbial induction, although once established in early life, the microbiome of humans and experimental mice is relatively stable and mutational activity has been shown to be independent of B cell receptor signaling. Using germ-free and colonized mice provided with different diets formulated with proprietary grain-based processing or from purified chemicals with different principal macronutrient calorie sources, we show that diet affects IgA induction, its repertoire and mutational diversification independently of microbial exposure.

The recognition of a T-cell epitope target is driven by the unique sequence of the T-cell receptor (TCR). As the TCR sequence theoretically contains all the information that determines its target, it must be possible to infer its target from this sequence. Indeed, within specific settings, we now have performant machine learning models that can address this challenge. It has therefore become necessary to start considering how we can apply these models to gain novel immunological understanding.

In this talk, I want to focus on the application of TCR-epitope prediction models to identify epitope-specific T-cells in full repertoire data, and the additional challenges that are often missed by current evaluation efforts. I also aim to highlight the biggest advancements and milestones that we can expect in the coming years in the field, especially in regard to the unseen epitope prediction problem.

Pathogens and cancer cells can escape recognition by T cell receptors (TCRs) through mutations of immunogenic epitopes. TCR cross-reactivity, i.e., recognition of multiple epitopes with sequence similarities, can counteract such mutational escape, but also may cause severe side effects in cell-based immunotherapies. To predict the effect of epitope mutations on T cell functionality in silico, we present “Predicting T cell Epitope-specific Activation against Mutant versions” (P-TEAM). The Random Forest based model was developed on two comprehensive datasets of murine and human TCRs in response to systematic single-amino acid mutations of their target epitopes to predict T cell reactivity for unobserved mutations, or even unseen TCRs. P-TEAM is complemented with an active learning framework to guide experimental design to minimize primary data acquisition costs. Overall, P-TEAM provides an effective computational tool to study T cell responses against mutated epitopes.

T cells orchestrate the adaptive immune response against pathogens and cancer by recognizing epitopes presented on MHC molecules. The heterogeneity of the MHC peptidome, including the high polymorphism of MHC genes, is influencing TCR repertoires and represents an important challenge towards accurate prediction and identification of T-cell epitopes in different individuals and different species. Here we generated and curated a dataset of more than a million unique MHC-I and MHC-II ligands identified by mass spectrometry. This enabled us to precisely determine the binding motifs of >200 MHC alleles across human, mouse, cattle and chicken. Analysis of these binding specificities combined with X-ray crystallography refined our understanding of the molecular determinants of MHC motifs and revealed alternative binding modes of MHC ligands. We then developed machine learning frameworks to accurately predict binding specificities and ligands of any MHC-I (MixMHCpred) and MHC-II (MixMHC2pred) allele, and further integrated TCR recognition into our epitope prediction pipeline (PRIME). Prospectively applying our tools to SARS-CoV-2 proteins identified several epitopes and TCR sequencing revealed a monoclonal response in effector/memory CD8+ T cells against one of these epitopes with cross-reactivity against the homologous peptides from other coronaviruses. Overall, our work shows how in depth characterization of MHC motifs can help mapping the targets of T cells and understanding TCR cross-reactivity.

Despite the vast diversity of the antibody repertoire, infected individuals often mount antibody responses to precisely the same epitopes within antigens. The immunological mechanisms underpinning this phenomenon remain unknown. By mapping 376 immunodominant “public epitopes” at high resolution and characterizing several of their cognate antibodies, we concluded that germline-encoded sequences in antibodies drive recurrent recognition. Systematic analysis of antibody-antigen structures uncovered 18 human and 21 partially overlapping mouse germline-encoded amino acid–binding (GRAB) motifs within heavy and light V gene segments that in case studies proved critical for public epitope recognition. GRAB motifs represent a fundamental component of the immune system’s architecture that promotes recognition of pathogens and leads to species-specific public antibody responses that can exert selective pressure on pathogens.

As individual researchers, we are well-trained in designing research projects that address specific and moderately challenging questions. However, when faced with a grand challenge such as understanding antigen recognition by T- and B-cell receptors, how can we, both individually and collectively, effectively tackle this problem? We currently have antigen binding data for only a minuscule proportion of the immune receptor sequence space, and despite the rapid emergence of new machine learning methods, consensus on what constitutes the most promising directions forward remains scarce.

In my talk, I will offer my perspective on this strategic question and discuss some recently published and ongoing work aligned with this strategy. Briefly, I believe the first step should be a rigorous characterisation of the computational challenges of the immune receptor-antigen binding prediction problem. Second, we must ensure that sufficient data is available to guide methodology development, where we in the foreseeable future need to rely on the combined use of experimental and simulated data. Third, we must prioritise interoperability and reproducibility of methods, along with the development of standardised benchmarks, to effectively compare performance and identify limitations of current approaches.

A new approach that utilizes the Lempel-Ziv 76 algorithm (LZ-76) to encode and represent AIRRs without relying on sequence annotation. The approach involves creating a graph-like model, which enables a wide range of potential applications, including generation probability inference, informative feature vector derivation, sequence generation, sequence analysis, and a new measure for repertoire diversity estimation.

T cells play a crucial role in our immunity by recognizing and eliminating infected and abnormal cells. T cell response is triggered upon T cell recognition of specific antigens presented by MHC molecules on the surface of target cells. However, the underlying rules of the interactions are incompletely understood. Over the past few years, we have been investigating T cell cross-reactivity and common specificity as two key drivers of T cell antigen specificity. In this seminar, I will focus on T cell cross-reactivity, and will introduce a new deep neural network model that we have developed for accurate and reliable prediction of CD8 T cell targets. I will illustrate how we have been using this model to model the dynamics of CD8 T cell response to emerging SARS-CoV-2 mutant variants.

Antibodies result from the competition of B cell lineages evolving under selection for improved antigen recognition, a process known as affinity maturation. High-affinity antibodies to pathogens such as HIV, influenza, and SARS-CoV-2 are frequently reported to arise from B cells whose receptors are encoded by particular immunoglobulin genes. This raises the possibility that the presence of particular germline genes in the B cell repertoire is a major determinant of the quality of the antibody response. Alternatively, initial differences in germline genes’ propensities to form high-affinity receptors might be overcome by chance events during affinity maturation. We first show how this can happen in simulations: even when fitness differences between germline genes lead to similar gene usage across individuals early on, gene usage can become increasingly dissimilar over time. We next find that mice experimentally infected with influenza virus demonstrate the same pattern of divergence in the weeks following infection. We investigated whether affinity maturation might nonetheless strongly select for particular amino acid motifs across diverse genetic backgrounds, but we found no evidence of convergence to similar CDR3 sequences or amino acid substitutions. These results suggest germline-encoded specificities might enable fast recognition of specific antigens early in the response, but diverse evolutionary routes to high affinity limit the genetic predictability of responses to infection and vaccination in the long term.

Early and accurate disease diagnosis is crucial for preventing disease development and defining therapy strategies. Due to predominantly unspecific symptoms, diagnosis of autoimmune diseases is notoriously challenging. However, multiple types of data are not leveraged for precision diagnostics due to the difficulties of integrating and encoding multi-omics data with clinical and laboratory values, as well as a lack of standardization; clinical decision support systems are often limited to only certain data types. Accordingly, even sophisticated data models fall short when supporting accurate diagnoses and presenting data analyses in a user-friendly form. Therefore, the integration of various data types is not only an opportunity but also a competitive advantage in research and for the industry. We have developed an integration pipeline to enable the use of machine learning for patient classification based on multi-omics data such as genetics, immunomics and metabolomics, in combination with clinical values and laboratory results. Machine learning models resulted in 95% prediction accuracy of autoimmune diseases using integrated data. Our results deliver insights into autoimmune disease research and have the potential to be adapted for applications across disease conditions.

We profiled blood and draining lymph node (LN) samples from human volunteers after influenza vaccination over two years to define evolution in the T follicular helper cell (TFH) response. We show LN TFH cells expanded in a clonal-manner during the first two weeks after vaccination and persisted within the LN for up to six months. LN and circulating TFH (cTFH) clonotypes overlapped but had distinct kinetics. LN TFH cell phenotypes were heterogeneous and mutable, first differentiating into pre-TFH during the month after vaccination before maturing into GC and IL-10+ TFH cells. TFH expansion, upregulation of glucose metabolism, and redifferentiation into GC TFH cells occurred with faster kinetics after re-vaccination in the second year. We identified several influenza-specific TFH clonal lineages, including multiple responses targeting internal influenza proteins, and show each TFH state is attainable within a lineage. This study demonstrates that human TFH cells form a durable and dynamic multi-tissue network.

Colonization by the microbiota causes a marked stimulation of B cells and induction of immunoglobulin, but mammals colonized with many taxa have highly complex and individualized immunoglobulin repertoires. To study this, we opted for a simplified model of defined transient exposures to different microbial taxa in germ-free mice to deconstruct how the microbiota shapes the B cell pool and its functional responsiveness. We previously showed that microbial exposure at the intestinal mucosa generated oligoclonal responses that differed from those in germ-free mice, and from the diverse repertoire that was generated after intravenous systemic exposure to microbiota. Our results reflected a contrast between a flexible response to systemic exposure with the need to avoid fatal sepsis, and a restricted response to mucosal exposure that reflects the generic nature of host–microbial mutualism in the mucosa. A hallmark of mucosal IgA is the high mutational load that accumulates throughout life. This has been mainly interpreted in terms of differences in microbial induction, although once established in early life, the microbiome of humans and experimental mice is relatively stable and mutational activity has been shown to be independent of B cell receptor signaling. Using germ-free and colonized mice provided with different diets formulated with proprietary grain-based processing or from purified chemicals with different principal macronutrient calorie sources, we show that diet affects IgA induction, its repertoire and mutational diversification independently of microbial exposure.

The recognition of a T-cell epitope target is driven by the unique sequence of the T-cell receptor (TCR). As the TCR sequence theoretically contains all the information that determines its target, it must be possible to infer its target from this sequence. Indeed, within specific settings, we now have performant machine learning models that can address this challenge. It has therefore become necessary to start considering how we can apply these models to gain novel immunological understanding.

In this talk, I want to focus on the application of TCR-epitope prediction models to identify epitope-specific T-cells in full repertoire data, and the additional challenges that are often missed by current evaluation efforts. I also aim to highlight the biggest advancements and milestones that we can expect in the coming years in the field, especially in regard to the unseen epitope prediction problem.

Pathogens and cancer cells can escape recognition by T cell receptors (TCRs) through mutations of immunogenic epitopes. TCR cross-reactivity, i.e., recognition of multiple epitopes with sequence similarities, can counteract such mutational escape, but also may cause severe side effects in cell-based immunotherapies. To predict the effect of epitope mutations on T cell functionality in silico, we present “Predicting T cell Epitope-specific Activation against Mutant versions” (P-TEAM). The Random Forest based model was developed on two comprehensive datasets of murine and human TCRs in response to systematic single-amino acid mutations of their target epitopes to predict T cell reactivity for unobserved mutations, or even unseen TCRs. P-TEAM is complemented with an active learning framework to guide experimental design to minimize primary data acquisition costs. Overall, P-TEAM provides an effective computational tool to study T cell responses against mutated epitopes.

T cells orchestrate the adaptive immune response against pathogens and cancer by recognizing epitopes presented on MHC molecules. The heterogeneity of the MHC peptidome, including the high polymorphism of MHC genes, is influencing TCR repertoires and represents an important challenge towards accurate prediction and identification of T-cell epitopes in different individuals and different species. Here we generated and curated a dataset of more than a million unique MHC-I and MHC-II ligands identified by mass spectrometry. This enabled us to precisely determine the binding motifs of >200 MHC alleles across human, mouse, cattle and chicken. Analysis of these binding specificities combined with X-ray crystallography refined our understanding of the molecular determinants of MHC motifs and revealed alternative binding modes of MHC ligands. We then developed machine learning frameworks to accurately predict binding specificities and ligands of any MHC-I (MixMHCpred) and MHC-II (MixMHC2pred) allele, and further integrated TCR recognition into our epitope prediction pipeline (PRIME). Prospectively applying our tools to SARS-CoV-2 proteins identified several epitopes and TCR sequencing revealed a monoclonal response in effector/memory CD8+ T cells against one of these epitopes with cross-reactivity against the homologous peptides from other coronaviruses. Overall, our work shows how in depth characterization of MHC motifs can help mapping the targets of T cells and understanding TCR cross-reactivity.

Despite the vast diversity of the antibody repertoire, infected individuals often mount antibody responses to precisely the same epitopes within antigens. The immunological mechanisms underpinning this phenomenon remain unknown. By mapping 376 immunodominant “public epitopes” at high resolution and characterizing several of their cognate antibodies, we concluded that germline-encoded sequences in antibodies drive recurrent recognition. Systematic analysis of antibody-antigen structures uncovered 18 human and 21 partially overlapping mouse germline-encoded amino acid–binding (GRAB) motifs within heavy and light V gene segments that in case studies proved critical for public epitope recognition. GRAB motifs represent a fundamental component of the immune system’s architecture that promotes recognition of pathogens and leads to species-specific public antibody responses that can exert selective pressure on pathogens.

As individual researchers, we are well-trained in designing research projects that address specific and moderately challenging questions. However, when faced with a grand challenge such as understanding antigen recognition by T- and B-cell receptors, how can we, both individually and collectively, effectively tackle this problem? We currently have antigen binding data for only a minuscule proportion of the immune receptor sequence space, and despite the rapid emergence of new machine learning methods, consensus on what constitutes the most promising directions forward remains scarce.

In my talk, I will offer my perspective on this strategic question and discuss some recently published and ongoing work aligned with this strategy. Briefly, I believe the first step should be a rigorous characterisation of the computational challenges of the immune receptor-antigen binding prediction problem. Second, we must ensure that sufficient data is available to guide methodology development, where we in the foreseeable future need to rely on the combined use of experimental and simulated data. Third, we must prioritise interoperability and reproducibility of methods, along with the development of standardised benchmarks, to effectively compare performance and identify limitations of current approaches.

A new approach that utilizes the Lempel-Ziv 76 algorithm (LZ-76) to encode and represent AIRRs without relying on sequence annotation. The approach involves creating a graph-like model, which enables a wide range of potential applications, including generation probability inference, informative feature vector derivation, sequence generation, sequence analysis, and a new measure for repertoire diversity estimation.

T cells play a crucial role in our immunity by recognizing and eliminating infected and abnormal cells. T cell response is triggered upon T cell recognition of specific antigens presented by MHC molecules on the surface of target cells. However, the underlying rules of the interactions are incompletely understood. Over the past few years, we have been investigating T cell cross-reactivity and common specificity as two key drivers of T cell antigen specificity. In this seminar, I will focus on T cell cross-reactivity, and will introduce a new deep neural network model that we have developed for accurate and reliable prediction of CD8 T cell targets. I will illustrate how we have been using this model to model the dynamics of CD8 T cell response to emerging SARS-CoV-2 mutant variants.

Antibodies result from the competition of B cell lineages evolving under selection for improved antigen recognition, a process known as affinity maturation. High-affinity antibodies to pathogens such as HIV, influenza, and SARS-CoV-2 are frequently reported to arise from B cells whose receptors are encoded by particular immunoglobulin genes. This raises the possibility that the presence of particular germline genes in the B cell repertoire is a major determinant of the quality of the antibody response. Alternatively, initial differences in germline genes’ propensities to form high-affinity receptors might be overcome by chance events during affinity maturation. We first show how this can happen in simulations: even when fitness differences between germline genes lead to similar gene usage across individuals early on, gene usage can become increasingly dissimilar over time. We next find that mice experimentally infected with influenza virus demonstrate the same pattern of divergence in the weeks following infection. We investigated whether affinity maturation might nonetheless strongly select for particular amino acid motifs across diverse genetic backgrounds, but we found no evidence of convergence to similar CDR3 sequences or amino acid substitutions. These results suggest germline-encoded specificities might enable fast recognition of specific antigens early in the response, but diverse evolutionary routes to high affinity limit the genetic predictability of responses to infection and vaccination in the long term.

Early and accurate disease diagnosis is crucial for preventing disease development and defining therapy strategies. Due to predominantly unspecific symptoms, diagnosis of autoimmune diseases is notoriously challenging. However, multiple types of data are not leveraged for precision diagnostics due to the difficulties of integrating and encoding multi-omics data with clinical and laboratory values, as well as a lack of standardization; clinical decision support systems are often limited to only certain data types. Accordingly, even sophisticated data models fall short when supporting accurate diagnoses and presenting data analyses in a user-friendly form. Therefore, the integration of various data types is not only an opportunity but also a competitive advantage in research and for the industry. We have developed an integration pipeline to enable the use of machine learning for patient classification based on multi-omics data such as genetics, immunomics and metabolomics, in combination with clinical values and laboratory results. Machine learning models resulted in 95% prediction accuracy of autoimmune diseases using integrated data. Our results deliver insights into autoimmune disease research and have the potential to be adapted for applications across disease conditions.

We profiled blood and draining lymph node (LN) samples from human volunteers after influenza vaccination over two years to define evolution in the T follicular helper cell (TFH) response. We show LN TFH cells expanded in a clonal-manner during the first two weeks after vaccination and persisted within the LN for up to six months. LN and circulating TFH (cTFH) clonotypes overlapped but had distinct kinetics. LN TFH cell phenotypes were heterogeneous and mutable, first differentiating into pre-TFH during the month after vaccination before maturing into GC and IL-10+ TFH cells. TFH expansion, upregulation of glucose metabolism, and redifferentiation into GC TFH cells occurred with faster kinetics after re-vaccination in the second year. We identified several influenza-specific TFH clonal lineages, including multiple responses targeting internal influenza proteins, and show each TFH state is attainable within a lineage. This study demonstrates that human TFH cells form a durable and dynamic multi-tissue network.

Colonization by the microbiota causes a marked stimulation of B cells and induction of immunoglobulin, but mammals colonized with many taxa have highly complex and individualized immunoglobulin repertoires. To study this, we opted for a simplified model of defined transient exposures to different microbial taxa in germ-free mice to deconstruct how the microbiota shapes the B cell pool and its functional responsiveness. We previously showed that microbial exposure at the intestinal mucosa generated oligoclonal responses that differed from those in germ-free mice, and from the diverse repertoire that was generated after intravenous systemic exposure to microbiota. Our results reflected a contrast between a flexible response to systemic exposure with the need to avoid fatal sepsis, and a restricted response to mucosal exposure that reflects the generic nature of host–microbial mutualism in the mucosa. A hallmark of mucosal IgA is the high mutational load that accumulates throughout life. This has been mainly interpreted in terms of differences in microbial induction, although once established in early life, the microbiome of humans and experimental mice is relatively stable and mutational activity has been shown to be independent of B cell receptor signaling. Using germ-free and colonized mice provided with different diets formulated with proprietary grain-based processing or from purified chemicals with different principal macronutrient calorie sources, we show that diet affects IgA induction, its repertoire and mutational diversification independently of microbial exposure.

The recognition of a T-cell epitope target is driven by the unique sequence of the T-cell receptor (TCR). As the TCR sequence theoretically contains all the information that determines its target, it must be possible to infer its target from this sequence. Indeed, within specific settings, we now have performant machine learning models that can address this challenge. It has therefore become necessary to start considering how we can apply these models to gain novel immunological understanding.

In this talk, I want to focus on the application of TCR-epitope prediction models to identify epitope-specific T-cells in full repertoire data, and the additional challenges that are often missed by current evaluation efforts. I also aim to highlight the biggest advancements and milestones that we can expect in the coming years in the field, especially in regard to the unseen epitope prediction problem.

Pathogens and cancer cells can escape recognition by T cell receptors (TCRs) through mutations of immunogenic epitopes. TCR cross-reactivity, i.e., recognition of multiple epitopes with sequence similarities, can counteract such mutational escape, but also may cause severe side effects in cell-based immunotherapies. To predict the effect of epitope mutations on T cell functionality in silico, we present “Predicting T cell Epitope-specific Activation against Mutant versions” (P-TEAM). The Random Forest based model was developed on two comprehensive datasets of murine and human TCRs in response to systematic single-amino acid mutations of their target epitopes to predict T cell reactivity for unobserved mutations, or even unseen TCRs. P-TEAM is complemented with an active learning framework to guide experimental design to minimize primary data acquisition costs. Overall, P-TEAM provides an effective computational tool to study T cell responses against mutated epitopes.

T cells orchestrate the adaptive immune response against pathogens and cancer by recognizing epitopes presented on MHC molecules. The heterogeneity of the MHC peptidome, including the high polymorphism of MHC genes, is influencing TCR repertoires and represents an important challenge towards accurate prediction and identification of T-cell epitopes in different individuals and different species. Here we generated and curated a dataset of more than a million unique MHC-I and MHC-II ligands identified by mass spectrometry. This enabled us to precisely determine the binding motifs of >200 MHC alleles across human, mouse, cattle and chicken. Analysis of these binding specificities combined with X-ray crystallography refined our understanding of the molecular determinants of MHC motifs and revealed alternative binding modes of MHC ligands. We then developed machine learning frameworks to accurately predict binding specificities and ligands of any MHC-I (MixMHCpred) and MHC-II (MixMHC2pred) allele, and further integrated TCR recognition into our epitope prediction pipeline (PRIME). Prospectively applying our tools to SARS-CoV-2 proteins identified several epitopes and TCR sequencing revealed a monoclonal response in effector/memory CD8+ T cells against one of these epitopes with cross-reactivity against the homologous peptides from other coronaviruses. Overall, our work shows how in depth characterization of MHC motifs can help mapping the targets of T cells and understanding TCR cross-reactivity.

Despite the vast diversity of the antibody repertoire, infected individuals often mount antibody responses to precisely the same epitopes within antigens. The immunological mechanisms underpinning this phenomenon remain unknown. By mapping 376 immunodominant “public epitopes” at high resolution and characterizing several of their cognate antibodies, we concluded that germline-encoded sequences in antibodies drive recurrent recognition. Systematic analysis of antibody-antigen structures uncovered 18 human and 21 partially overlapping mouse germline-encoded amino acid–binding (GRAB) motifs within heavy and light V gene segments that in case studies proved critical for public epitope recognition. GRAB motifs represent a fundamental component of the immune system’s architecture that promotes recognition of pathogens and leads to species-specific public antibody responses that can exert selective pressure on pathogens.

As individual researchers, we are well-trained in designing research projects that address specific and moderately challenging questions. However, when faced with a grand challenge such as understanding antigen recognition by T- and B-cell receptors, how can we, both individually and collectively, effectively tackle this problem? We currently have antigen binding data for only a minuscule proportion of the immune receptor sequence space, and despite the rapid emergence of new machine learning methods, consensus on what constitutes the most promising directions forward remains scarce.

In my talk, I will offer my perspective on this strategic question and discuss some recently published and ongoing work aligned with this strategy. Briefly, I believe the first step should be a rigorous characterisation of the computational challenges of the immune receptor-antigen binding prediction problem. Second, we must ensure that sufficient data is available to guide methodology development, where we in the foreseeable future need to rely on the combined use of experimental and simulated data. Third, we must prioritise interoperability and reproducibility of methods, along with the development of standardised benchmarks, to effectively compare performance and identify limitations of current approaches.

A new approach that utilizes the Lempel-Ziv 76 algorithm (LZ-76) to encode and represent AIRRs without relying on sequence annotation. The approach involves creating a graph-like model, which enables a wide range of potential applications, including generation probability inference, informative feature vector derivation, sequence generation, sequence analysis, and a new measure for repertoire diversity estimation.

T cells play a crucial role in our immunity by recognizing and eliminating infected and abnormal cells. T cell response is triggered upon T cell recognition of specific antigens presented by MHC molecules on the surface of target cells. However, the underlying rules of the interactions are incompletely understood. Over the past few years, we have been investigating T cell cross-reactivity and common specificity as two key drivers of T cell antigen specificity. In this seminar, I will focus on T cell cross-reactivity, and will introduce a new deep neural network model that we have developed for accurate and reliable prediction of CD8 T cell targets. I will illustrate how we have been using this model to model the dynamics of CD8 T cell response to emerging SARS-CoV-2 mutant variants.

Antibodies result from the competition of B cell lineages evolving under selection for improved antigen recognition, a process known as affinity maturation. High-affinity antibodies to pathogens such as HIV, influenza, and SARS-CoV-2 are frequently reported to arise from B cells whose receptors are encoded by particular immunoglobulin genes. This raises the possibility that the presence of particular germline genes in the B cell repertoire is a major determinant of the quality of the antibody response. Alternatively, initial differences in germline genes’ propensities to form high-affinity receptors might be overcome by chance events during affinity maturation. We first show how this can happen in simulations: even when fitness differences between germline genes lead to similar gene usage across individuals early on, gene usage can become increasingly dissimilar over time. We next find that mice experimentally infected with influenza virus demonstrate the same pattern of divergence in the weeks following infection. We investigated whether affinity maturation might nonetheless strongly select for particular amino acid motifs across diverse genetic backgrounds, but we found no evidence of convergence to similar CDR3 sequences or amino acid substitutions. These results suggest germline-encoded specificities might enable fast recognition of specific antigens early in the response, but diverse evolutionary routes to high affinity limit the genetic predictability of responses to infection and vaccination in the long term.

Early and accurate disease diagnosis is crucial for preventing disease development and defining therapy strategies. Due to predominantly unspecific symptoms, diagnosis of autoimmune diseases is notoriously challenging. However, multiple types of data are not leveraged for precision diagnostics due to the difficulties of integrating and encoding multi-omics data with clinical and laboratory values, as well as a lack of standardization; clinical decision support systems are often limited to only certain data types. Accordingly, even sophisticated data models fall short when supporting accurate diagnoses and presenting data analyses in a user-friendly form. Therefore, the integration of various data types is not only an opportunity but also a competitive advantage in research and for the industry. We have developed an integration pipeline to enable the use of machine learning for patient classification based on multi-omics data such as genetics, immunomics and metabolomics, in combination with clinical values and laboratory results. Machine learning models resulted in 95% prediction accuracy of autoimmune diseases using integrated data. Our results deliver insights into autoimmune disease research and have the potential to be adapted for applications across disease conditions.

We profiled blood and draining lymph node (LN) samples from human volunteers after influenza vaccination over two years to define evolution in the T follicular helper cell (TFH) response. We show LN TFH cells expanded in a clonal-manner during the first two weeks after vaccination and persisted within the LN for up to six months. LN and circulating TFH (cTFH) clonotypes overlapped but had distinct kinetics. LN TFH cell phenotypes were heterogeneous and mutable, first differentiating into pre-TFH during the month after vaccination before maturing into GC and IL-10+ TFH cells. TFH expansion, upregulation of glucose metabolism, and redifferentiation into GC TFH cells occurred with faster kinetics after re-vaccination in the second year. We identified several influenza-specific TFH clonal lineages, including multiple responses targeting internal influenza proteins, and show each TFH state is attainable within a lineage. This study demonstrates that human TFH cells form a durable and dynamic multi-tissue network.

Colonization by the microbiota causes a marked stimulation of B cells and induction of immunoglobulin, but mammals colonized with many taxa have highly complex and individualized immunoglobulin repertoires. To study this, we opted for a simplified model of defined transient exposures to different microbial taxa in germ-free mice to deconstruct how the microbiota shapes the B cell pool and its functional responsiveness. We previously showed that microbial exposure at the intestinal mucosa generated oligoclonal responses that differed from those in germ-free mice, and from the diverse repertoire that was generated after intravenous systemic exposure to microbiota. Our results reflected a contrast between a flexible response to systemic exposure with the need to avoid fatal sepsis, and a restricted response to mucosal exposure that reflects the generic nature of host–microbial mutualism in the mucosa. A hallmark of mucosal IgA is the high mutational load that accumulates throughout life. This has been mainly interpreted in terms of differences in microbial induction, although once established in early life, the microbiome of humans and experimental mice is relatively stable and mutational activity has been shown to be independent of B cell receptor signaling. Using germ-free and colonized mice provided with different diets formulated with proprietary grain-based processing or from purified chemicals with different principal macronutrient calorie sources, we show that diet affects IgA induction, its repertoire and mutational diversification independently of microbial exposure.

The recognition of a T-cell epitope target is driven by the unique sequence of the T-cell receptor (TCR). As the TCR sequence theoretically contains all the information that determines its target, it must be possible to infer its target from this sequence. Indeed, within specific settings, we now have performant machine learning models that can address this challenge. It has therefore become necessary to start considering how we can apply these models to gain novel immunological understanding.

In this talk, I want to focus on the application of TCR-epitope prediction models to identify epitope-specific T-cells in full repertoire data, and the additional challenges that are often missed by current evaluation efforts. I also aim to highlight the biggest advancements and milestones that we can expect in the coming years in the field, especially in regard to the unseen epitope prediction problem.

Pathogens and cancer cells can escape recognition by T cell receptors (TCRs) through mutations of immunogenic epitopes. TCR cross-reactivity, i.e., recognition of multiple epitopes with sequence similarities, can counteract such mutational escape, but also may cause severe side effects in cell-based immunotherapies. To predict the effect of epitope mutations on T cell functionality in silico, we present “Predicting T cell Epitope-specific Activation against Mutant versions” (P-TEAM). The Random Forest based model was developed on two comprehensive datasets of murine and human TCRs in response to systematic single-amino acid mutations of their target epitopes to predict T cell reactivity for unobserved mutations, or even unseen TCRs. P-TEAM is complemented with an active learning framework to guide experimental design to minimize primary data acquisition costs. Overall, P-TEAM provides an effective computational tool to study T cell responses against mutated epitopes.

T cells orchestrate the adaptive immune response against pathogens and cancer by recognizing epitopes presented on MHC molecules. The heterogeneity of the MHC peptidome, including the high polymorphism of MHC genes, is influencing TCR repertoires and represents an important challenge towards accurate prediction and identification of T-cell epitopes in different individuals and different species. Here we generated and curated a dataset of more than a million unique MHC-I and MHC-II ligands identified by mass spectrometry. This enabled us to precisely determine the binding motifs of >200 MHC alleles across human, mouse, cattle and chicken. Analysis of these binding specificities combined with X-ray crystallography refined our understanding of the molecular determinants of MHC motifs and revealed alternative binding modes of MHC ligands. We then developed machine learning frameworks to accurately predict binding specificities and ligands of any MHC-I (MixMHCpred) and MHC-II (MixMHC2pred) allele, and further integrated TCR recognition into our epitope prediction pipeline (PRIME). Prospectively applying our tools to SARS-CoV-2 proteins identified several epitopes and TCR sequencing revealed a monoclonal response in effector/memory CD8+ T cells against one of these epitopes with cross-reactivity against the homologous peptides from other coronaviruses. Overall, our work shows how in depth characterization of MHC motifs can help mapping the targets of T cells and understanding TCR cross-reactivity.

Despite the vast diversity of the antibody repertoire, infected individuals often mount antibody responses to precisely the same epitopes within antigens. The immunological mechanisms underpinning this phenomenon remain unknown. By mapping 376 immunodominant “public epitopes” at high resolution and characterizing several of their cognate antibodies, we concluded that germline-encoded sequences in antibodies drive recurrent recognition. Systematic analysis of antibody-antigen structures uncovered 18 human and 21 partially overlapping mouse germline-encoded amino acid–binding (GRAB) motifs within heavy and light V gene segments that in case studies proved critical for public epitope recognition. GRAB motifs represent a fundamental component of the immune system’s architecture that promotes recognition of pathogens and leads to species-specific public antibody responses that can exert selective pressure on pathogens.

As individual researchers, we are well-trained in designing research projects that address specific and moderately challenging questions. However, when faced with a grand challenge such as understanding antigen recognition by T- and B-cell receptors, how can we, both individually and collectively, effectively tackle this problem? We currently have antigen binding data for only a minuscule proportion of the immune receptor sequence space, and despite the rapid emergence of new machine learning methods, consensus on what constitutes the most promising directions forward remains scarce.

In my talk, I will offer my perspective on this strategic question and discuss some recently published and ongoing work aligned with this strategy. Briefly, I believe the first step should be a rigorous characterisation of the computational challenges of the immune receptor-antigen binding prediction problem. Second, we must ensure that sufficient data is available to guide methodology development, where we in the foreseeable future need to rely on the combined use of experimental and simulated data. Third, we must prioritise interoperability and reproducibility of methods, along with the development of standardised benchmarks, to effectively compare performance and identify limitations of current approaches.

A new approach that utilizes the Lempel-Ziv 76 algorithm (LZ-76) to encode and represent AIRRs without relying on sequence annotation. The approach involves creating a graph-like model, which enables a wide range of potential applications, including generation probability inference, informative feature vector derivation, sequence generation, sequence analysis, and a new measure for repertoire diversity estimation.

T cells play a crucial role in our immunity by recognizing and eliminating infected and abnormal cells. T cell response is triggered upon T cell recognition of specific antigens presented by MHC molecules on the surface of target cells. However, the underlying rules of the interactions are incompletely understood. Over the past few years, we have been investigating T cell cross-reactivity and common specificity as two key drivers of T cell antigen specificity. In this seminar, I will focus on T cell cross-reactivity, and will introduce a new deep neural network model that we have developed for accurate and reliable prediction of CD8 T cell targets. I will illustrate how we have been using this model to model the dynamics of CD8 T cell response to emerging SARS-CoV-2 mutant variants.

Antibodies result from the competition of B cell lineages evolving under selection for improved antigen recognition, a process known as affinity maturation. High-affinity antibodies to pathogens such as HIV, influenza, and SARS-CoV-2 are frequently reported to arise from B cells whose receptors are encoded by particular immunoglobulin genes. This raises the possibility that the presence of particular germline genes in the B cell repertoire is a major determinant of the quality of the antibody response. Alternatively, initial differences in germline genes’ propensities to form high-affinity receptors might be overcome by chance events during affinity maturation. We first show how this can happen in simulations: even when fitness differences between germline genes lead to similar gene usage across individuals early on, gene usage can become increasingly dissimilar over time. We next find that mice experimentally infected with influenza virus demonstrate the same pattern of divergence in the weeks following infection. We investigated whether affinity maturation might nonetheless strongly select for particular amino acid motifs across diverse genetic backgrounds, but we found no evidence of convergence to similar CDR3 sequences or amino acid substitutions. These results suggest germline-encoded specificities might enable fast recognition of specific antigens early in the response, but diverse evolutionary routes to high affinity limit the genetic predictability of responses to infection and vaccination in the long term.

Early and accurate disease diagnosis is crucial for preventing disease development and defining therapy strategies. Due to predominantly unspecific symptoms, diagnosis of autoimmune diseases is notoriously challenging. However, multiple types of data are not leveraged for precision diagnostics due to the difficulties of integrating and encoding multi-omics data with clinical and laboratory values, as well as a lack of standardization; clinical decision support systems are often limited to only certain data types. Accordingly, even sophisticated data models fall short when supporting accurate diagnoses and presenting data analyses in a user-friendly form. Therefore, the integration of various data types is not only an opportunity but also a competitive advantage in research and for the industry. We have developed an integration pipeline to enable the use of machine learning for patient classification based on multi-omics data such as genetics, immunomics and metabolomics, in combination with clinical values and laboratory results. Machine learning models resulted in 95% prediction accuracy of autoimmune diseases using integrated data. Our results deliver insights into autoimmune disease research and have the potential to be adapted for applications across disease conditions.

We profiled blood and draining lymph node (LN) samples from human volunteers after influenza vaccination over two years to define evolution in the T follicular helper cell (TFH) response. We show LN TFH cells expanded in a clonal-manner during the first two weeks after vaccination and persisted within the LN for up to six months. LN and circulating TFH (cTFH) clonotypes overlapped but had distinct kinetics. LN TFH cell phenotypes were heterogeneous and mutable, first differentiating into pre-TFH during the month after vaccination before maturing into GC and IL-10+ TFH cells. TFH expansion, upregulation of glucose metabolism, and redifferentiation into GC TFH cells occurred with faster kinetics after re-vaccination in the second year. We identified several influenza-specific TFH clonal lineages, including multiple responses targeting internal influenza proteins, and show each TFH state is attainable within a lineage. This study demonstrates that human TFH cells form a durable and dynamic multi-tissue network.

Colonization by the microbiota causes a marked stimulation of B cells and induction of immunoglobulin, but mammals colonized with many taxa have highly complex and individualized immunoglobulin repertoires. To study this, we opted for a simplified model of defined transient exposures to different microbial taxa in germ-free mice to deconstruct how the microbiota shapes the B cell pool and its functional responsiveness. We previously showed that microbial exposure at the intestinal mucosa generated oligoclonal responses that differed from those in germ-free mice, and from the diverse repertoire that was generated after intravenous systemic exposure to microbiota. Our results reflected a contrast between a flexible response to systemic exposure with the need to avoid fatal sepsis, and a restricted response to mucosal exposure that reflects the generic nature of host–microbial mutualism in the mucosa. A hallmark of mucosal IgA is the high mutational load that accumulates throughout life. This has been mainly interpreted in terms of differences in microbial induction, although once established in early life, the microbiome of humans and experimental mice is relatively stable and mutational activity has been shown to be independent of B cell receptor signaling. Using germ-free and colonized mice provided with different diets formulated with proprietary grain-based processing or from purified chemicals with different principal macronutrient calorie sources, we show that diet affects IgA induction, its repertoire and mutational diversification independently of microbial exposure.

The recognition of a T-cell epitope target is driven by the unique sequence of the T-cell receptor (TCR). As the TCR sequence theoretically contains all the information that determines its target, it must be possible to infer its target from this sequence. Indeed, within specific settings, we now have performant machine learning models that can address this challenge. It has therefore become necessary to start considering how we can apply these models to gain novel immunological understanding.

In this talk, I want to focus on the application of TCR-epitope prediction models to identify epitope-specific T-cells in full repertoire data, and the additional challenges that are often missed by current evaluation efforts. I also aim to highlight the biggest advancements and milestones that we can expect in the coming years in the field, especially in regard to the unseen epitope prediction problem.

Pathogens and cancer cells can escape recognition by T cell receptors (TCRs) through mutations of immunogenic epitopes. TCR cross-reactivity, i.e., recognition of multiple epitopes with sequence similarities, can counteract such mutational escape, but also may cause severe side effects in cell-based immunotherapies. To predict the effect of epitope mutations on T cell functionality in silico, we present “Predicting T cell Epitope-specific Activation against Mutant versions” (P-TEAM). The Random Forest based model was developed on two comprehensive datasets of murine and human TCRs in response to systematic single-amino acid mutations of their target epitopes to predict T cell reactivity for unobserved mutations, or even unseen TCRs. P-TEAM is complemented with an active learning framework to guide experimental design to minimize primary data acquisition costs. Overall, P-TEAM provides an effective computational tool to study T cell responses against mutated epitopes.

T cells orchestrate the adaptive immune response against pathogens and cancer by recognizing epitopes presented on MHC molecules. The heterogeneity of the MHC peptidome, including the high polymorphism of MHC genes, is influencing TCR repertoires and represents an important challenge towards accurate prediction and identification of T-cell epitopes in different individuals and different species. Here we generated and curated a dataset of more than a million unique MHC-I and MHC-II ligands identified by mass spectrometry. This enabled us to precisely determine the binding motifs of >200 MHC alleles across human, mouse, cattle and chicken. Analysis of these binding specificities combined with X-ray crystallography refined our understanding of the molecular determinants of MHC motifs and revealed alternative binding modes of MHC ligands. We then developed machine learning frameworks to accurately predict binding specificities and ligands of any MHC-I (MixMHCpred) and MHC-II (MixMHC2pred) allele, and further integrated TCR recognition into our epitope prediction pipeline (PRIME). Prospectively applying our tools to SARS-CoV-2 proteins identified several epitopes and TCR sequencing revealed a monoclonal response in effector/memory CD8+ T cells against one of these epitopes with cross-reactivity against the homologous peptides from other coronaviruses. Overall, our work shows how in depth characterization of MHC motifs can help mapping the targets of T cells and understanding TCR cross-reactivity.

Despite the vast diversity of the antibody repertoire, infected individuals often mount antibody responses to precisely the same epitopes within antigens. The immunological mechanisms underpinning this phenomenon remain unknown. By mapping 376 immunodominant “public epitopes” at high resolution and characterizing several of their cognate antibodies, we concluded that germline-encoded sequences in antibodies drive recurrent recognition. Systematic analysis of antibody-antigen structures uncovered 18 human and 21 partially overlapping mouse germline-encoded amino acid–binding (GRAB) motifs within heavy and light V gene segments that in case studies proved critical for public epitope recognition. GRAB motifs represent a fundamental component of the immune system’s architecture that promotes recognition of pathogens and leads to species-specific public antibody responses that can exert selective pressure on pathogens.

As individual researchers, we are well-trained in designing research projects that address specific and moderately challenging questions. However, when faced with a grand challenge such as understanding antigen recognition by T- and B-cell receptors, how can we, both individually and collectively, effectively tackle this problem? We currently have antigen binding data for only a minuscule proportion of the immune receptor sequence space, and despite the rapid emergence of new machine learning methods, consensus on what constitutes the most promising directions forward remains scarce.

In my talk, I will offer my perspective on this strategic question and discuss some recently published and ongoing work aligned with this strategy. Briefly, I believe the first step should be a rigorous characterisation of the computational challenges of the immune receptor-antigen binding prediction problem. Second, we must ensure that sufficient data is available to guide methodology development, where we in the foreseeable future need to rely on the combined use of experimental and simulated data. Third, we must prioritise interoperability and reproducibility of methods, along with the development of standardised benchmarks, to effectively compare performance and identify limitations of current approaches.

A new approach that utilizes the Lempel-Ziv 76 algorithm (LZ-76) to encode and represent AIRRs without relying on sequence annotation. The approach involves creating a graph-like model, which enables a wide range of potential applications, including generation probability inference, informative feature vector derivation, sequence generation, sequence analysis, and a new measure for repertoire diversity estimation.

T cells play a crucial role in our immunity by recognizing and eliminating infected and abnormal cells. T cell response is triggered upon T cell recognition of specific antigens presented by MHC molecules on the surface of target cells. However, the underlying rules of the interactions are incompletely understood. Over the past few years, we have been investigating T cell cross-reactivity and common specificity as two key drivers of T cell antigen specificity. In this seminar, I will focus on T cell cross-reactivity, and will introduce a new deep neural network model that we have developed for accurate and reliable prediction of CD8 T cell targets. I will illustrate how we have been using this model to model the dynamics of CD8 T cell response to emerging SARS-CoV-2 mutant variants.

Antibodies result from the competition of B cell lineages evolving under selection for improved antigen recognition, a process known as affinity maturation. High-affinity antibodies to pathogens such as HIV, influenza, and SARS-CoV-2 are frequently reported to arise from B cells whose receptors are encoded by particular immunoglobulin genes. This raises the possibility that the presence of particular germline genes in the B cell repertoire is a major determinant of the quality of the antibody response. Alternatively, initial differences in germline genes’ propensities to form high-affinity receptors might be overcome by chance events during affinity maturation. We first show how this can happen in simulations: even when fitness differences between germline genes lead to similar gene usage across individuals early on, gene usage can become increasingly dissimilar over time. We next find that mice experimentally infected with influenza virus demonstrate the same pattern of divergence in the weeks following infection. We investigated whether affinity maturation might nonetheless strongly select for particular amino acid motifs across diverse genetic backgrounds, but we found no evidence of convergence to similar CDR3 sequences or amino acid substitutions. These results suggest germline-encoded specificities might enable fast recognition of specific antigens early in the response, but diverse evolutionary routes to high affinity limit the genetic predictability of responses to infection and vaccination in the long term.

Early and accurate disease diagnosis is crucial for preventing disease development and defining therapy strategies. Due to predominantly unspecific symptoms, diagnosis of autoimmune diseases is notoriously challenging. However, multiple types of data are not leveraged for precision diagnostics due to the difficulties of integrating and encoding multi-omics data with clinical and laboratory values, as well as a lack of standardization; clinical decision support systems are often limited to only certain data types. Accordingly, even sophisticated data models fall short when supporting accurate diagnoses and presenting data analyses in a user-friendly form. Therefore, the integration of various data types is not only an opportunity but also a competitive advantage in research and for the industry. We have developed an integration pipeline to enable the use of machine learning for patient classification based on multi-omics data such as genetics, immunomics and metabolomics, in combination with clinical values and laboratory results. Machine learning models resulted in 95% prediction accuracy of autoimmune diseases using integrated data. Our results deliver insights into autoimmune disease research and have the potential to be adapted for applications across disease conditions.

We profiled blood and draining lymph node (LN) samples from human volunteers after influenza vaccination over two years to define evolution in the T follicular helper cell (TFH) response. We show LN TFH cells expanded in a clonal-manner during the first two weeks after vaccination and persisted within the LN for up to six months. LN and circulating TFH (cTFH) clonotypes overlapped but had distinct kinetics. LN TFH cell phenotypes were heterogeneous and mutable, first differentiating into pre-TFH during the month after vaccination before maturing into GC and IL-10+ TFH cells. TFH expansion, upregulation of glucose metabolism, and redifferentiation into GC TFH cells occurred with faster kinetics after re-vaccination in the second year. We identified several influenza-specific TFH clonal lineages, including multiple responses targeting internal influenza proteins, and show each TFH state is attainable within a lineage. This study demonstrates that human TFH cells form a durable and dynamic multi-tissue network.

Colonization by the microbiota causes a marked stimulation of B cells and induction of immunoglobulin, but mammals colonized with many taxa have highly complex and individualized immunoglobulin repertoires. To study this, we opted for a simplified model of defined transient exposures to different microbial taxa in germ-free mice to deconstruct how the microbiota shapes the B cell pool and its functional responsiveness. We previously showed that microbial exposure at the intestinal mucosa generated oligoclonal responses that differed from those in germ-free mice, and from the diverse repertoire that was generated after intravenous systemic exposure to microbiota. Our results reflected a contrast between a flexible response to systemic exposure with the need to avoid fatal sepsis, and a restricted response to mucosal exposure that reflects the generic nature of host–microbial mutualism in the mucosa. A hallmark of mucosal IgA is the high mutational load that accumulates throughout life. This has been mainly interpreted in terms of differences in microbial induction, although once established in early life, the microbiome of humans and experimental mice is relatively stable and mutational activity has been shown to be independent of B cell receptor signaling. Using germ-free and colonized mice provided with different diets formulated with proprietary grain-based processing or from purified chemicals with different principal macronutrient calorie sources, we show that diet affects IgA induction, its repertoire and mutational diversification independently of microbial exposure.

The recognition of a T-cell epitope target is driven by the unique sequence of the T-cell receptor (TCR). As the TCR sequence theoretically contains all the information that determines its target, it must be possible to infer its target from this sequence. Indeed, within specific settings, we now have performant machine learning models that can address this challenge. It has therefore become necessary to start considering how we can apply these models to gain novel immunological understanding.

In this talk, I want to focus on the application of TCR-epitope prediction models to identify epitope-specific T-cells in full repertoire data, and the additional challenges that are often missed by current evaluation efforts. I also aim to highlight the biggest advancements and milestones that we can expect in the coming years in the field, especially in regard to the unseen epitope prediction problem.

Pathogens and cancer cells can escape recognition by T cell receptors (TCRs) through mutations of immunogenic epitopes. TCR cross-reactivity, i.e., recognition of multiple epitopes with sequence similarities, can counteract such mutational escape, but also may cause severe side effects in cell-based immunotherapies. To predict the effect of epitope mutations on T cell functionality in silico, we present “Predicting T cell Epitope-specific Activation against Mutant versions” (P-TEAM). The Random Forest based model was developed on two comprehensive datasets of murine and human TCRs in response to systematic single-amino acid mutations of their target epitopes to predict T cell reactivity for unobserved mutations, or even unseen TCRs. P-TEAM is complemented with an active learning framework to guide experimental design to minimize primary data acquisition costs. Overall, P-TEAM provides an effective computational tool to study T cell responses against mutated epitopes.

T cells orchestrate the adaptive immune response against pathogens and cancer by recognizing epitopes presented on MHC molecules. The heterogeneity of the MHC peptidome, including the high polymorphism of MHC genes, is influencing TCR repertoires and represents an important challenge towards accurate prediction and identification of T-cell epitopes in different individuals and different species. Here we generated and curated a dataset of more than a million unique MHC-I and MHC-II ligands identified by mass spectrometry. This enabled us to precisely determine the binding motifs of >200 MHC alleles across human, mouse, cattle and chicken. Analysis of these binding specificities combined with X-ray crystallography refined our understanding of the molecular determinants of MHC motifs and revealed alternative binding modes of MHC ligands. We then developed machine learning frameworks to accurately predict binding specificities and ligands of any MHC-I (MixMHCpred) and MHC-II (MixMHC2pred) allele, and further integrated TCR recognition into our epitope prediction pipeline (PRIME). Prospectively applying our tools to SARS-CoV-2 proteins identified several epitopes and TCR sequencing revealed a monoclonal response in effector/memory CD8+ T cells against one of these epitopes with cross-reactivity against the homologous peptides from other coronaviruses. Overall, our work shows how in depth characterization of MHC motifs can help mapping the targets of T cells and understanding TCR cross-reactivity.

Despite the vast diversity of the antibody repertoire, infected individuals often mount antibody responses to precisely the same epitopes within antigens. The immunological mechanisms underpinning this phenomenon remain unknown. By mapping 376 immunodominant “public epitopes” at high resolution and characterizing several of their cognate antibodies, we concluded that germline-encoded sequences in antibodies drive recurrent recognition. Systematic analysis of antibody-antigen structures uncovered 18 human and 21 partially overlapping mouse germline-encoded amino acid–binding (GRAB) motifs within heavy and light V gene segments that in case studies proved critical for public epitope recognition. GRAB motifs represent a fundamental component of the immune system’s architecture that promotes recognition of pathogens and leads to species-specific public antibody responses that can exert selective pressure on pathogens.

As individual researchers, we are well-trained in designing research projects that address specific and moderately challenging questions. However, when faced with a grand challenge such as understanding antigen recognition by T- and B-cell receptors, how can we, both individually and collectively, effectively tackle this problem? We currently have antigen binding data for only a minuscule proportion of the immune receptor sequence space, and despite the rapid emergence of new machine learning methods, consensus on what constitutes the most promising directions forward remains scarce.

In my talk, I will offer my perspective on this strategic question and discuss some recently published and ongoing work aligned with this strategy. Briefly, I believe the first step should be a rigorous characterisation of the computational challenges of the immune receptor-antigen binding prediction problem. Second, we must ensure that sufficient data is available to guide methodology development, where we in the foreseeable future need to rely on the combined use of experimental and simulated data. Third, we must prioritise interoperability and reproducibility of methods, along with the development of standardised benchmarks, to effectively compare performance and identify limitations of current approaches.

A new approach that utilizes the Lempel-Ziv 76 algorithm (LZ-76) to encode and represent AIRRs without relying on sequence annotation. The approach involves creating a graph-like model, which enables a wide range of potential applications, including generation probability inference, informative feature vector derivation, sequence generation, sequence analysis, and a new measure for repertoire diversity estimation.

T cells play a crucial role in our immunity by recognizing and eliminating infected and abnormal cells. T cell response is triggered upon T cell recognition of specific antigens presented by MHC molecules on the surface of target cells. However, the underlying rules of the interactions are incompletely understood. Over the past few years, we have been investigating T cell cross-reactivity and common specificity as two key drivers of T cell antigen specificity. In this seminar, I will focus on T cell cross-reactivity, and will introduce a new deep neural network model that we have developed for accurate and reliable prediction of CD8 T cell targets. I will illustrate how we have been using this model to model the dynamics of CD8 T cell response to emerging SARS-CoV-2 mutant variants.

Antibodies result from the competition of B cell lineages evolving under selection for improved antigen recognition, a process known as affinity maturation. High-affinity antibodies to pathogens such as HIV, influenza, and SARS-CoV-2 are frequently reported to arise from B cells whose receptors are encoded by particular immunoglobulin genes. This raises the possibility that the presence of particular germline genes in the B cell repertoire is a major determinant of the quality of the antibody response. Alternatively, initial differences in germline genes’ propensities to form high-affinity receptors might be overcome by chance events during affinity maturation. We first show how this can happen in simulations: even when fitness differences between germline genes lead to similar gene usage across individuals early on, gene usage can become increasingly dissimilar over time. We next find that mice experimentally infected with influenza virus demonstrate the same pattern of divergence in the weeks following infection. We investigated whether affinity maturation might nonetheless strongly select for particular amino acid motifs across diverse genetic backgrounds, but we found no evidence of convergence to similar CDR3 sequences or amino acid substitutions. These results suggest germline-encoded specificities might enable fast recognition of specific antigens early in the response, but diverse evolutionary routes to high affinity limit the genetic predictability of responses to infection and vaccination in the long term.

Early and accurate disease diagnosis is crucial for preventing disease development and defining therapy strategies. Due to predominantly unspecific symptoms, diagnosis of autoimmune diseases is notoriously challenging. However, multiple types of data are not leveraged for precision diagnostics due to the difficulties of integrating and encoding multi-omics data with clinical and laboratory values, as well as a lack of standardization; clinical decision support systems are often limited to only certain data types. Accordingly, even sophisticated data models fall short when supporting accurate diagnoses and presenting data analyses in a user-friendly form. Therefore, the integration of various data types is not only an opportunity but also a competitive advantage in research and for the industry. We have developed an integration pipeline to enable the use of machine learning for patient classification based on multi-omics data such as genetics, immunomics and metabolomics, in combination with clinical values and laboratory results. Machine learning models resulted in 95% prediction accuracy of autoimmune diseases using integrated data. Our results deliver insights into autoimmune disease research and have the potential to be adapted for applications across disease conditions.

We profiled blood and draining lymph node (LN) samples from human volunteers after influenza vaccination over two years to define evolution in the T follicular helper cell (TFH) response. We show LN TFH cells expanded in a clonal-manner during the first two weeks after vaccination and persisted within the LN for up to six months. LN and circulating TFH (cTFH) clonotypes overlapped but had distinct kinetics. LN TFH cell phenotypes were heterogeneous and mutable, first differentiating into pre-TFH during the month after vaccination before maturing into GC and IL-10+ TFH cells. TFH expansion, upregulation of glucose metabolism, and redifferentiation into GC TFH cells occurred with faster kinetics after re-vaccination in the second year. We identified several influenza-specific TFH clonal lineages, including multiple responses targeting internal influenza proteins, and show each TFH state is attainable within a lineage. This study demonstrates that human TFH cells form a durable and dynamic multi-tissue network.

Colonization by the microbiota causes a marked stimulation of B cells and induction of immunoglobulin, but mammals colonized with many taxa have highly complex and individualized immunoglobulin repertoires. To study this, we opted for a simplified model of defined transient exposures to different microbial taxa in germ-free mice to deconstruct how the microbiota shapes the B cell pool and its functional responsiveness. We previously showed that microbial exposure at the intestinal mucosa generated oligoclonal responses that differed from those in germ-free mice, and from the diverse repertoire that was generated after intravenous systemic exposure to microbiota. Our results reflected a contrast between a flexible response to systemic exposure with the need to avoid fatal sepsis, and a restricted response to mucosal exposure that reflects the generic nature of host–microbial mutualism in the mucosa. A hallmark of mucosal IgA is the high mutational load that accumulates throughout life. This has been mainly interpreted in terms of differences in microbial induction, although once established in early life, the microbiome of humans and experimental mice is relatively stable and mutational activity has been shown to be independent of B cell receptor signaling. Using germ-free and colonized mice provided with different diets formulated with proprietary grain-based processing or from purified chemicals with different principal macronutrient calorie sources, we show that diet affects IgA induction, its repertoire and mutational diversification independently of microbial exposure.

The recognition of a T-cell epitope target is driven by the unique sequence of the T-cell receptor (TCR). As the TCR sequence theoretically contains all the information that determines its target, it must be possible to infer its target from this sequence. Indeed, within specific settings, we now have performant machine learning models that can address this challenge. It has therefore become necessary to start considering how we can apply these models to gain novel immunological understanding.

In this talk, I want to focus on the application of TCR-epitope prediction models to identify epitope-specific T-cells in full repertoire data, and the additional challenges that are often missed by current evaluation efforts. I also aim to highlight the biggest advancements and milestones that we can expect in the coming years in the field, especially in regard to the unseen epitope prediction problem.

Pathogens and cancer cells can escape recognition by T cell receptors (TCRs) through mutations of immunogenic epitopes. TCR cross-reactivity, i.e., recognition of multiple epitopes with sequence similarities, can counteract such mutational escape, but also may cause severe side effects in cell-based immunotherapies. To predict the effect of epitope mutations on T cell functionality in silico, we present “Predicting T cell Epitope-specific Activation against Mutant versions” (P-TEAM). The Random Forest based model was developed on two comprehensive datasets of murine and human TCRs in response to systematic single-amino acid mutations of their target epitopes to predict T cell reactivity for unobserved mutations, or even unseen TCRs. P-TEAM is complemented with an active learning framework to guide experimental design to minimize primary data acquisition costs. Overall, P-TEAM provides an effective computational tool to study T cell responses against mutated epitopes.

T cells orchestrate the adaptive immune response against pathogens and cancer by recognizing epitopes presented on MHC molecules. The heterogeneity of the MHC peptidome, including the high polymorphism of MHC genes, is influencing TCR repertoires and represents an important challenge towards accurate prediction and identification of T-cell epitopes in different individuals and different species. Here we generated and curated a dataset of more than a million unique MHC-I and MHC-II ligands identified by mass spectrometry. This enabled us to precisely determine the binding motifs of >200 MHC alleles across human, mouse, cattle and chicken. Analysis of these binding specificities combined with X-ray crystallography refined our understanding of the molecular determinants of MHC motifs and revealed alternative binding modes of MHC ligands. We then developed machine learning frameworks to accurately predict binding specificities and ligands of any MHC-I (MixMHCpred) and MHC-II (MixMHC2pred) allele, and further integrated TCR recognition into our epitope prediction pipeline (PRIME). Prospectively applying our tools to SARS-CoV-2 proteins identified several epitopes and TCR sequencing revealed a monoclonal response in effector/memory CD8+ T cells against one of these epitopes with cross-reactivity against the homologous peptides from other coronaviruses. Overall, our work shows how in depth characterization of MHC motifs can help mapping the targets of T cells and understanding TCR cross-reactivity.

Despite the vast diversity of the antibody repertoire, infected individuals often mount antibody responses to precisely the same epitopes within antigens. The immunological mechanisms underpinning this phenomenon remain unknown. By mapping 376 immunodominant “public epitopes” at high resolution and characterizing several of their cognate antibodies, we concluded that germline-encoded sequences in antibodies drive recurrent recognition. Systematic analysis of antibody-antigen structures uncovered 18 human and 21 partially overlapping mouse germline-encoded amino acid–binding (GRAB) motifs within heavy and light V gene segments that in case studies proved critical for public epitope recognition. GRAB motifs represent a fundamental component of the immune system’s architecture that promotes recognition of pathogens and leads to species-specific public antibody responses that can exert selective pressure on pathogens.

As individual researchers, we are well-trained in designing research projects that address specific and moderately challenging questions. However, when faced with a grand challenge such as understanding antigen recognition by T- and B-cell receptors, how can we, both individually and collectively, effectively tackle this problem? We currently have antigen binding data for only a minuscule proportion of the immune receptor sequence space, and despite the rapid emergence of new machine learning methods, consensus on what constitutes the most promising directions forward remains scarce.

In my talk, I will offer my perspective on this strategic question and discuss some recently published and ongoing work aligned with this strategy. Briefly, I believe the first step should be a rigorous characterisation of the computational challenges of the immune receptor-antigen binding prediction problem. Second, we must ensure that sufficient data is available to guide methodology development, where we in the foreseeable future need to rely on the combined use of experimental and simulated data. Third, we must prioritise interoperability and reproducibility of methods, along with the development of standardised benchmarks, to effectively compare performance and identify limitations of current approaches.

A new approach that utilizes the Lempel-Ziv 76 algorithm (LZ-76) to encode and represent AIRRs without relying on sequence annotation. The approach involves creating a graph-like model, which enables a wide range of potential applications, including generation probability inference, informative feature vector derivation, sequence generation, sequence analysis, and a new measure for repertoire diversity estimation.

T cells play a crucial role in our immunity by recognizing and eliminating infected and abnormal cells. T cell response is triggered upon T cell recognition of specific antigens presented by MHC molecules on the surface of target cells. However, the underlying rules of the interactions are incompletely understood. Over the past few years, we have been investigating T cell cross-reactivity and common specificity as two key drivers of T cell antigen specificity. In this seminar, I will focus on T cell cross-reactivity, and will introduce a new deep neural network model that we have developed for accurate and reliable prediction of CD8 T cell targets. I will illustrate how we have been using this model to model the dynamics of CD8 T cell response to emerging SARS-CoV-2 mutant variants.

Antibodies result from the competition of B cell lineages evolving under selection for improved antigen recognition, a process known as affinity maturation. High-affinity antibodies to pathogens such as HIV, influenza, and SARS-CoV-2 are frequently reported to arise from B cells whose receptors are encoded by particular immunoglobulin genes. This raises the possibility that the presence of particular germline genes in the B cell repertoire is a major determinant of the quality of the antibody response. Alternatively, initial differences in germline genes’ propensities to form high-affinity receptors might be overcome by chance events during affinity maturation. We first show how this can happen in simulations: even when fitness differences between germline genes lead to similar gene usage across individuals early on, gene usage can become increasingly dissimilar over time. We next find that mice experimentally infected with influenza virus demonstrate the same pattern of divergence in the weeks following infection. We investigated whether affinity maturation might nonetheless strongly select for particular amino acid motifs across diverse genetic backgrounds, but we found no evidence of convergence to similar CDR3 sequences or amino acid substitutions. These results suggest germline-encoded specificities might enable fast recognition of specific antigens early in the response, but diverse evolutionary routes to high affinity limit the genetic predictability of responses to infection and vaccination in the long term.

Early and accurate disease diagnosis is crucial for preventing disease development and defining therapy strategies. Due to predominantly unspecific symptoms, diagnosis of autoimmune diseases is notoriously challenging. However, multiple types of data are not leveraged for precision diagnostics due to the difficulties of integrating and encoding multi-omics data with clinical and laboratory values, as well as a lack of standardization; clinical decision support systems are often limited to only certain data types. Accordingly, even sophisticated data models fall short when supporting accurate diagnoses and presenting data analyses in a user-friendly form. Therefore, the integration of various data types is not only an opportunity but also a competitive advantage in research and for the industry. We have developed an integration pipeline to enable the use of machine learning for patient classification based on multi-omics data such as genetics, immunomics and metabolomics, in combination with clinical values and laboratory results. Machine learning models resulted in 95% prediction accuracy of autoimmune diseases using integrated data. Our results deliver insights into autoimmune disease research and have the potential to be adapted for applications across disease conditions.

We profiled blood and draining lymph node (LN) samples from human volunteers after influenza vaccination over two years to define evolution in the T follicular helper cell (TFH) response. We show LN TFH cells expanded in a clonal-manner during the first two weeks after vaccination and persisted within the LN for up to six months. LN and circulating TFH (cTFH) clonotypes overlapped but had distinct kinetics. LN TFH cell phenotypes were heterogeneous and mutable, first differentiating into pre-TFH during the month after vaccination before maturing into GC and IL-10+ TFH cells. TFH expansion, upregulation of glucose metabolism, and redifferentiation into GC TFH cells occurred with faster kinetics after re-vaccination in the second year. We identified several influenza-specific TFH clonal lineages, including multiple responses targeting internal influenza proteins, and show each TFH state is attainable within a lineage. This study demonstrates that human TFH cells form a durable and dynamic multi-tissue network.

Colonization by the microbiota causes a marked stimulation of B cells and induction of immunoglobulin, but mammals colonized with many taxa have highly complex and individualized immunoglobulin repertoires. To study this, we opted for a simplified model of defined transient exposures to different microbial taxa in germ-free mice to deconstruct how the microbiota shapes the B cell pool and its functional responsiveness. We previously showed that microbial exposure at the intestinal mucosa generated oligoclonal responses that differed from those in germ-free mice, and from the diverse repertoire that was generated after intravenous systemic exposure to microbiota. Our results reflected a contrast between a flexible response to systemic exposure with the need to avoid fatal sepsis, and a restricted response to mucosal exposure that reflects the generic nature of host–microbial mutualism in the mucosa. A hallmark of mucosal IgA is the high mutational load that accumulates throughout life. This has been mainly interpreted in terms of differences in microbial induction, although once established in early life, the microbiome of humans and experimental mice is relatively stable and mutational activity has been shown to be independent of B cell receptor signaling. Using germ-free and colonized mice provided with different diets formulated with proprietary grain-based processing or from purified chemicals with different principal macronutrient calorie sources, we show that diet affects IgA induction, its repertoire and mutational diversification independently of microbial exposure.

The recognition of a T-cell epitope target is driven by the unique sequence of the T-cell receptor (TCR). As the TCR sequence theoretically contains all the information that determines its target, it must be possible to infer its target from this sequence. Indeed, within specific settings, we now have performant machine learning models that can address this challenge. It has therefore become necessary to start considering how we can apply these models to gain novel immunological understanding.

In this talk, I want to focus on the application of TCR-epitope prediction models to identify epitope-specific T-cells in full repertoire data, and the additional challenges that are often missed by current evaluation efforts. I also aim to highlight the biggest advancements and milestones that we can expect in the coming years in the field, especially in regard to the unseen epitope prediction problem.

Pathogens and cancer cells can escape recognition by T cell receptors (TCRs) through mutations of immunogenic epitopes. TCR cross-reactivity, i.e., recognition of multiple epitopes with sequence similarities, can counteract such mutational escape, but also may cause severe side effects in cell-based immunotherapies. To predict the effect of epitope mutations on T cell functionality in silico, we present “Predicting T cell Epitope-specific Activation against Mutant versions” (P-TEAM). The Random Forest based model was developed on two comprehensive datasets of murine and human TCRs in response to systematic single-amino acid mutations of their target epitopes to predict T cell reactivity for unobserved mutations, or even unseen TCRs. P-TEAM is complemented with an active learning framework to guide experimental design to minimize primary data acquisition costs. Overall, P-TEAM provides an effective computational tool to study T cell responses against mutated epitopes.

T cells orchestrate the adaptive immune response against pathogens and cancer by recognizing epitopes presented on MHC molecules. The heterogeneity of the MHC peptidome, including the high polymorphism of MHC genes, is influencing TCR repertoires and represents an important challenge towards accurate prediction and identification of T-cell epitopes in different individuals and different species. Here we generated and curated a dataset of more than a million unique MHC-I and MHC-II ligands identified by mass spectrometry. This enabled us to precisely determine the binding motifs of >200 MHC alleles across human, mouse, cattle and chicken. Analysis of these binding specificities combined with X-ray crystallography refined our understanding of the molecular determinants of MHC motifs and revealed alternative binding modes of MHC ligands. We then developed machine learning frameworks to accurately predict binding specificities and ligands of any MHC-I (MixMHCpred) and MHC-II (MixMHC2pred) allele, and further integrated TCR recognition into our epitope prediction pipeline (PRIME). Prospectively applying our tools to SARS-CoV-2 proteins identified several epitopes and TCR sequencing revealed a monoclonal response in effector/memory CD8+ T cells against one of these epitopes with cross-reactivity against the homologous peptides from other coronaviruses. Overall, our work shows how in depth characterization of MHC motifs can help mapping the targets of T cells and understanding TCR cross-reactivity.

Despite the vast diversity of the antibody repertoire, infected individuals often mount antibody responses to precisely the same epitopes within antigens. The immunological mechanisms underpinning this phenomenon remain unknown. By mapping 376 immunodominant “public epitopes” at high resolution and characterizing several of their cognate antibodies, we concluded that germline-encoded sequences in antibodies drive recurrent recognition. Systematic analysis of antibody-antigen structures uncovered 18 human and 21 partially overlapping mouse germline-encoded amino acid–binding (GRAB) motifs within heavy and light V gene segments that in case studies proved critical for public epitope recognition. GRAB motifs represent a fundamental component of the immune system’s architecture that promotes recognition of pathogens and leads to species-specific public antibody responses that can exert selective pressure on pathogens.

As individual researchers, we are well-trained in designing research projects that address specific and moderately challenging questions. However, when faced with a grand challenge such as understanding antigen recognition by T- and B-cell receptors, how can we, both individually and collectively, effectively tackle this problem? We currently have antigen binding data for only a minuscule proportion of the immune receptor sequence space, and despite the rapid emergence of new machine learning methods, consensus on what constitutes the most promising directions forward remains scarce.

In my talk, I will offer my perspective on this strategic question and discuss some recently published and ongoing work aligned with this strategy. Briefly, I believe the first step should be a rigorous characterisation of the computational challenges of the immune receptor-antigen binding prediction problem. Second, we must ensure that sufficient data is available to guide methodology development, where we in the foreseeable future need to rely on the combined use of experimental and simulated data. Third, we must prioritise interoperability and reproducibility of methods, along with the development of standardised benchmarks, to effectively compare performance and identify limitations of current approaches.

A new approach that utilizes the Lempel-Ziv 76 algorithm (LZ-76) to encode and represent AIRRs without relying on sequence annotation. The approach involves creating a graph-like model, which enables a wide range of potential applications, including generation probability inference, informative feature vector derivation, sequence generation, sequence analysis, and a new measure for repertoire diversity estimation.

T cells play a crucial role in our immunity by recognizing and eliminating infected and abnormal cells. T cell response is triggered upon T cell recognition of specific antigens presented by MHC molecules on the surface of target cells. However, the underlying rules of the interactions are incompletely understood. Over the past few years, we have been investigating T cell cross-reactivity and common specificity as two key drivers of T cell antigen specificity. In this seminar, I will focus on T cell cross-reactivity, and will introduce a new deep neural network model that we have developed for accurate and reliable prediction of CD8 T cell targets. I will illustrate how we have been using this model to model the dynamics of CD8 T cell response to emerging SARS-CoV-2 mutant variants.

Antibodies result from the competition of B cell lineages evolving under selection for improved antigen recognition, a process known as affinity maturation. High-affinity antibodies to pathogens such as HIV, influenza, and SARS-CoV-2 are frequently reported to arise from B cells whose receptors are encoded by particular immunoglobulin genes. This raises the possibility that the presence of particular germline genes in the B cell repertoire is a major determinant of the quality of the antibody response. Alternatively, initial differences in germline genes’ propensities to form high-affinity receptors might be overcome by chance events during affinity maturation. We first show how this can happen in simulations: even when fitness differences between germline genes lead to similar gene usage across individuals early on, gene usage can become increasingly dissimilar over time. We next find that mice experimentally infected with influenza virus demonstrate the same pattern of divergence in the weeks following infection. We investigated whether affinity maturation might nonetheless strongly select for particular amino acid motifs across diverse genetic backgrounds, but we found no evidence of convergence to similar CDR3 sequences or amino acid substitutions. These results suggest germline-encoded specificities might enable fast recognition of specific antigens early in the response, but diverse evolutionary routes to high affinity limit the genetic predictability of responses to infection and vaccination in the long term.

Early and accurate disease diagnosis is crucial for preventing disease development and defining therapy strategies. Due to predominantly unspecific symptoms, diagnosis of autoimmune diseases is notoriously challenging. However, multiple types of data are not leveraged for precision diagnostics due to the difficulties of integrating and encoding multi-omics data with clinical and laboratory values, as well as a lack of standardization; clinical decision support systems are often limited to only certain data types. Accordingly, even sophisticated data models fall short when supporting accurate diagnoses and presenting data analyses in a user-friendly form. Therefore, the integration of various data types is not only an opportunity but also a competitive advantage in research and for the industry. We have developed an integration pipeline to enable the use of machine learning for patient classification based on multi-omics data such as genetics, immunomics and metabolomics, in combination with clinical values and laboratory results. Machine learning models resulted in 95% prediction accuracy of autoimmune diseases using integrated data. Our results deliver insights into autoimmune disease research and have the potential to be adapted for applications across disease conditions.

We profiled blood and draining lymph node (LN) samples from human volunteers after influenza vaccination over two years to define evolution in the T follicular helper cell (TFH) response. We show LN TFH cells expanded in a clonal-manner during the first two weeks after vaccination and persisted within the LN for up to six months. LN and circulating TFH (cTFH) clonotypes overlapped but had distinct kinetics. LN TFH cell phenotypes were heterogeneous and mutable, first differentiating into pre-TFH during the month after vaccination before maturing into GC and IL-10+ TFH cells. TFH expansion, upregulation of glucose metabolism, and redifferentiation into GC TFH cells occurred with faster kinetics after re-vaccination in the second year. We identified several influenza-specific TFH clonal lineages, including multiple responses targeting internal influenza proteins, and show each TFH state is attainable within a lineage. This study demonstrates that human TFH cells form a durable and dynamic multi-tissue network.

Colonization by the microbiota causes a marked stimulation of B cells and induction of immunoglobulin, but mammals colonized with many taxa have highly complex and individualized immunoglobulin repertoires. To study this, we opted for a simplified model of defined transient exposures to different microbial taxa in germ-free mice to deconstruct how the microbiota shapes the B cell pool and its functional responsiveness. We previously showed that microbial exposure at the intestinal mucosa generated oligoclonal responses that differed from those in germ-free mice, and from the diverse repertoire that was generated after intravenous systemic exposure to microbiota. Our results reflected a contrast between a flexible response to systemic exposure with the need to avoid fatal sepsis, and a restricted response to mucosal exposure that reflects the generic nature of host–microbial mutualism in the mucosa. A hallmark of mucosal IgA is the high mutational load that accumulates throughout life. This has been mainly interpreted in terms of differences in microbial induction, although once established in early life, the microbiome of humans and experimental mice is relatively stable and mutational activity has been shown to be independent of B cell receptor signaling. Using germ-free and colonized mice provided with different diets formulated with proprietary grain-based processing or from purified chemicals with different principal macronutrient calorie sources, we show that diet affects IgA induction, its repertoire and mutational diversification independently of microbial exposure.

The recognition of a T-cell epitope target is driven by the unique sequence of the T-cell receptor (TCR). As the TCR sequence theoretically contains all the information that determines its target, it must be possible to infer its target from this sequence. Indeed, within specific settings, we now have performant machine learning models that can address this challenge. It has therefore become necessary to start considering how we can apply these models to gain novel immunological understanding.

In this talk, I want to focus on the application of TCR-epitope prediction models to identify epitope-specific T-cells in full repertoire data, and the additional challenges that are often missed by current evaluation efforts. I also aim to highlight the biggest advancements and milestones that we can expect in the coming years in the field, especially in regard to the unseen epitope prediction problem.

Pathogens and cancer cells can escape recognition by T cell receptors (TCRs) through mutations of immunogenic epitopes. TCR cross-reactivity, i.e., recognition of multiple epitopes with sequence similarities, can counteract such mutational escape, but also may cause severe side effects in cell-based immunotherapies. To predict the effect of epitope mutations on T cell functionality in silico, we present “Predicting T cell Epitope-specific Activation against Mutant versions” (P-TEAM). The Random Forest based model was developed on two comprehensive datasets of murine and human TCRs in response to systematic single-amino acid mutations of their target epitopes to predict T cell reactivity for unobserved mutations, or even unseen TCRs. P-TEAM is complemented with an active learning framework to guide experimental design to minimize primary data acquisition costs. Overall, P-TEAM provides an effective computational tool to study T cell responses against mutated epitopes.

T cells orchestrate the adaptive immune response against pathogens and cancer by recognizing epitopes presented on MHC molecules. The heterogeneity of the MHC peptidome, including the high polymorphism of MHC genes, is influencing TCR repertoires and represents an important challenge towards accurate prediction and identification of T-cell epitopes in different individuals and different species. Here we generated and curated a dataset of more than a million unique MHC-I and MHC-II ligands identified by mass spectrometry. This enabled us to precisely determine the binding motifs of >200 MHC alleles across human, mouse, cattle and chicken. Analysis of these binding specificities combined with X-ray crystallography refined our understanding of the molecular determinants of MHC motifs and revealed alternative binding modes of MHC ligands. We then developed machine learning frameworks to accurately predict binding specificities and ligands of any MHC-I (MixMHCpred) and MHC-II (MixMHC2pred) allele, and further integrated TCR recognition into our epitope prediction pipeline (PRIME). Prospectively applying our tools to SARS-CoV-2 proteins identified several epitopes and TCR sequencing revealed a monoclonal response in effector/memory CD8+ T cells against one of these epitopes with cross-reactivity against the homologous peptides from other coronaviruses. Overall, our work shows how in depth characterization of MHC motifs can help mapping the targets of T cells and understanding TCR cross-reactivity.

Despite the vast diversity of the antibody repertoire, infected individuals often mount antibody responses to precisely the same epitopes within antigens. The immunological mechanisms underpinning this phenomenon remain unknown. By mapping 376 immunodominant “public epitopes” at high resolution and characterizing several of their cognate antibodies, we concluded that germline-encoded sequences in antibodies drive recurrent recognition. Systematic analysis of antibody-antigen structures uncovered 18 human and 21 partially overlapping mouse germline-encoded amino acid–binding (GRAB) motifs within heavy and light V gene segments that in case studies proved critical for public epitope recognition. GRAB motifs represent a fundamental component of the immune system’s architecture that promotes recognition of pathogens and leads to species-specific public antibody responses that can exert selective pressure on pathogens.

As individual researchers, we are well-trained in designing research projects that address specific and moderately challenging questions. However, when faced with a grand challenge such as understanding antigen recognition by T- and B-cell receptors, how can we, both individually and collectively, effectively tackle this problem? We currently have antigen binding data for only a minuscule proportion of the immune receptor sequence space, and despite the rapid emergence of new machine learning methods, consensus on what constitutes the most promising directions forward remains scarce.

In my talk, I will offer my perspective on this strategic question and discuss some recently published and ongoing work aligned with this strategy. Briefly, I believe the first step should be a rigorous characterisation of the computational challenges of the immune receptor-antigen binding prediction problem. Second, we must ensure that sufficient data is available to guide methodology development, where we in the foreseeable future need to rely on the combined use of experimental and simulated data. Third, we must prioritise interoperability and reproducibility of methods, along with the development of standardised benchmarks, to effectively compare performance and identify limitations of current approaches.

A new approach that utilizes the Lempel-Ziv 76 algorithm (LZ-76) to encode and represent AIRRs without relying on sequence annotation. The approach involves creating a graph-like model, which enables a wide range of potential applications, including generation probability inference, informative feature vector derivation, sequence generation, sequence analysis, and a new measure for repertoire diversity estimation.

T cells play a crucial role in our immunity by recognizing and eliminating infected and abnormal cells. T cell response is triggered upon T cell recognition of specific antigens presented by MHC molecules on the surface of target cells. However, the underlying rules of the interactions are incompletely understood. Over the past few years, we have been investigating T cell cross-reactivity and common specificity as two key drivers of T cell antigen specificity. In this seminar, I will focus on T cell cross-reactivity, and will introduce a new deep neural network model that we have developed for accurate and reliable prediction of CD8 T cell targets. I will illustrate how we have been using this model to model the dynamics of CD8 T cell response to emerging SARS-CoV-2 mutant variants.

Antibodies result from the competition of B cell lineages evolving under selection for improved antigen recognition, a process known as affinity maturation. High-affinity antibodies to pathogens such as HIV, influenza, and SARS-CoV-2 are frequently reported to arise from B cells whose receptors are encoded by particular immunoglobulin genes. This raises the possibility that the presence of particular germline genes in the B cell repertoire is a major determinant of the quality of the antibody response. Alternatively, initial differences in germline genes’ propensities to form high-affinity receptors might be overcome by chance events during affinity maturation. We first show how this can happen in simulations: even when fitness differences between germline genes lead to similar gene usage across individuals early on, gene usage can become increasingly dissimilar over time. We next find that mice experimentally infected with influenza virus demonstrate the same pattern of divergence in the weeks following infection. We investigated whether affinity maturation might nonetheless strongly select for particular amino acid motifs across diverse genetic backgrounds, but we found no evidence of convergence to similar CDR3 sequences or amino acid substitutions. These results suggest germline-encoded specificities might enable fast recognition of specific antigens early in the response, but diverse evolutionary routes to high affinity limit the genetic predictability of responses to infection and vaccination in the long term.

Early and accurate disease diagnosis is crucial for preventing disease development and defining therapy strategies. Due to predominantly unspecific symptoms, diagnosis of autoimmune diseases is notoriously challenging. However, multiple types of data are not leveraged for precision diagnostics due to the difficulties of integrating and encoding multi-omics data with clinical and laboratory values, as well as a lack of standardization; clinical decision support systems are often limited to only certain data types. Accordingly, even sophisticated data models fall short when supporting accurate diagnoses and presenting data analyses in a user-friendly form. Therefore, the integration of various data types is not only an opportunity but also a competitive advantage in research and for the industry. We have developed an integration pipeline to enable the use of machine learning for patient classification based on multi-omics data such as genetics, immunomics and metabolomics, in combination with clinical values and laboratory results. Machine learning models resulted in 95% prediction accuracy of autoimmune diseases using integrated data. Our results deliver insights into autoimmune disease research and have the potential to be adapted for applications across disease conditions.

We profiled blood and draining lymph node (LN) samples from human volunteers after influenza vaccination over two years to define evolution in the T follicular helper cell (TFH) response. We show LN TFH cells expanded in a clonal-manner during the first two weeks after vaccination and persisted within the LN for up to six months. LN and circulating TFH (cTFH) clonotypes overlapped but had distinct kinetics. LN TFH cell phenotypes were heterogeneous and mutable, first differentiating into pre-TFH during the month after vaccination before maturing into GC and IL-10+ TFH cells. TFH expansion, upregulation of glucose metabolism, and redifferentiation into GC TFH cells occurred with faster kinetics after re-vaccination in the second year. We identified several influenza-specific TFH clonal lineages, including multiple responses targeting internal influenza proteins, and show each TFH state is attainable within a lineage. This study demonstrates that human TFH cells form a durable and dynamic multi-tissue network.

Colonization by the microbiota causes a marked stimulation of B cells and induction of immunoglobulin, but mammals colonized with many taxa have highly complex and individualized immunoglobulin repertoires. To study this, we opted for a simplified model of defined transient exposures to different microbial taxa in germ-free mice to deconstruct how the microbiota shapes the B cell pool and its functional responsiveness. We previously showed that microbial exposure at the intestinal mucosa generated oligoclonal responses that differed from those in germ-free mice, and from the diverse repertoire that was generated after intravenous systemic exposure to microbiota. Our results reflected a contrast between a flexible response to systemic exposure with the need to avoid fatal sepsis, and a restricted response to mucosal exposure that reflects the generic nature of host–microbial mutualism in the mucosa. A hallmark of mucosal IgA is the high mutational load that accumulates throughout life. This has been mainly interpreted in terms of differences in microbial induction, although once established in early life, the microbiome of humans and experimental mice is relatively stable and mutational activity has been shown to be independent of B cell receptor signaling. Using germ-free and colonized mice provided with different diets formulated with proprietary grain-based processing or from purified chemicals with different principal macronutrient calorie sources, we show that diet affects IgA induction, its repertoire and mutational diversification independently of microbial exposure.

The recognition of a T-cell epitope target is driven by the unique sequence of the T-cell receptor (TCR). As the TCR sequence theoretically contains all the information that determines its target, it must be possible to infer its target from this sequence. Indeed, within specific settings, we now have performant machine learning models that can address this challenge. It has therefore become necessary to start considering how we can apply these models to gain novel immunological understanding.

In this talk, I want to focus on the application of TCR-epitope prediction models to identify epitope-specific T-cells in full repertoire data, and the additional challenges that are often missed by current evaluation efforts. I also aim to highlight the biggest advancements and milestones that we can expect in the coming years in the field, especially in regard to the unseen epitope prediction problem.

Pathogens and cancer cells can escape recognition by T cell receptors (TCRs) through mutations of immunogenic epitopes. TCR cross-reactivity, i.e., recognition of multiple epitopes with sequence similarities, can counteract such mutational escape, but also may cause severe side effects in cell-based immunotherapies. To predict the effect of epitope mutations on T cell functionality in silico, we present “Predicting T cell Epitope-specific Activation against Mutant versions” (P-TEAM). The Random Forest based model was developed on two comprehensive datasets of murine and human TCRs in response to systematic single-amino acid mutations of their target epitopes to predict T cell reactivity for unobserved mutations, or even unseen TCRs. P-TEAM is complemented with an active learning framework to guide experimental design to minimize primary data acquisition costs. Overall, P-TEAM provides an effective computational tool to study T cell responses against mutated epitopes.

T cells orchestrate the adaptive immune response against pathogens and cancer by recognizing epitopes presented on MHC molecules. The heterogeneity of the MHC peptidome, including the high polymorphism of MHC genes, is influencing TCR repertoires and represents an important challenge towards accurate prediction and identification of T-cell epitopes in different individuals and different species. Here we generated and curated a dataset of more than a million unique MHC-I and MHC-II ligands identified by mass spectrometry. This enabled us to precisely determine the binding motifs of >200 MHC alleles across human, mouse, cattle and chicken. Analysis of these binding specificities combined with X-ray crystallography refined our understanding of the molecular determinants of MHC motifs and revealed alternative binding modes of MHC ligands. We then developed machine learning frameworks to accurately predict binding specificities and ligands of any MHC-I (MixMHCpred) and MHC-II (MixMHC2pred) allele, and further integrated TCR recognition into our epitope prediction pipeline (PRIME). Prospectively applying our tools to SARS-CoV-2 proteins identified several epitopes and TCR sequencing revealed a monoclonal response in effector/memory CD8+ T cells against one of these epitopes with cross-reactivity against the homologous peptides from other coronaviruses. Overall, our work shows how in depth characterization of MHC motifs can help mapping the targets of T cells and understanding TCR cross-reactivity.

Despite the vast diversity of the antibody repertoire, infected individuals often mount antibody responses to precisely the same epitopes within antigens. The immunological mechanisms underpinning this phenomenon remain unknown. By mapping 376 immunodominant “public epitopes” at high resolution and characterizing several of their cognate antibodies, we concluded that germline-encoded sequences in antibodies drive recurrent recognition. Systematic analysis of antibody-antigen structures uncovered 18 human and 21 partially overlapping mouse germline-encoded amino acid–binding (GRAB) motifs within heavy and light V gene segments that in case studies proved critical for public epitope recognition. GRAB motifs represent a fundamental component of the immune system’s architecture that promotes recognition of pathogens and leads to species-specific public antibody responses that can exert selective pressure on pathogens.

As individual researchers, we are well-trained in designing research projects that address specific and moderately challenging questions. However, when faced with a grand challenge such as understanding antigen recognition by T- and B-cell receptors, how can we, both individually and collectively, effectively tackle this problem? We currently have antigen binding data for only a minuscule proportion of the immune receptor sequence space, and despite the rapid emergence of new machine learning methods, consensus on what constitutes the most promising directions forward remains scarce.

In my talk, I will offer my perspective on this strategic question and discuss some recently published and ongoing work aligned with this strategy. Briefly, I believe the first step should be a rigorous characterisation of the computational challenges of the immune receptor-antigen binding prediction problem. Second, we must ensure that sufficient data is available to guide methodology development, where we in the foreseeable future need to rely on the combined use of experimental and simulated data. Third, we must prioritise interoperability and reproducibility of methods, along with the development of standardised benchmarks, to effectively compare performance and identify limitations of current approaches.

A new approach that utilizes the Lempel-Ziv 76 algorithm (LZ-76) to encode and represent AIRRs without relying on sequence annotation. The approach involves creating a graph-like model, which enables a wide range of potential applications, including generation probability inference, informative feature vector derivation, sequence generation, sequence analysis, and a new measure for repertoire diversity estimation.

T cells play a crucial role in our immunity by recognizing and eliminating infected and abnormal cells. T cell response is triggered upon T cell recognition of specific antigens presented by MHC molecules on the surface of target cells. However, the underlying rules of the interactions are incompletely understood. Over the past few years, we have been investigating T cell cross-reactivity and common specificity as two key drivers of T cell antigen specificity. In this seminar, I will focus on T cell cross-reactivity, and will introduce a new deep neural network model that we have developed for accurate and reliable prediction of CD8 T cell targets. I will illustrate how we have been using this model to model the dynamics of CD8 T cell response to emerging SARS-CoV-2 mutant variants.

Antibodies result from the competition of B cell lineages evolving under selection for improved antigen recognition, a process known as affinity maturation. High-affinity antibodies to pathogens such as HIV, influenza, and SARS-CoV-2 are frequently reported to arise from B cells whose receptors are encoded by particular immunoglobulin genes. This raises the possibility that the presence of particular germline genes in the B cell repertoire is a major determinant of the quality of the antibody response. Alternatively, initial differences in germline genes’ propensities to form high-affinity receptors might be overcome by chance events during affinity maturation. We first show how this can happen in simulations: even when fitness differences between germline genes lead to similar gene usage across individuals early on, gene usage can become increasingly dissimilar over time. We next find that mice experimentally infected with influenza virus demonstrate the same pattern of divergence in the weeks following infection. We investigated whether affinity maturation might nonetheless strongly select for particular amino acid motifs across diverse genetic backgrounds, but we found no evidence of convergence to similar CDR3 sequences or amino acid substitutions. These results suggest germline-encoded specificities might enable fast recognition of specific antigens early in the response, but diverse evolutionary routes to high affinity limit the genetic predictability of responses to infection and vaccination in the long term.

Early and accurate disease diagnosis is crucial for preventing disease development and defining therapy strategies. Due to predominantly unspecific symptoms, diagnosis of autoimmune diseases is notoriously challenging. However, multiple types of data are not leveraged for precision diagnostics due to the difficulties of integrating and encoding multi-omics data with clinical and laboratory values, as well as a lack of standardization; clinical decision support systems are often limited to only certain data types. Accordingly, even sophisticated data models fall short when supporting accurate diagnoses and presenting data analyses in a user-friendly form. Therefore, the integration of various data types is not only an opportunity but also a competitive advantage in research and for the industry. We have developed an integration pipeline to enable the use of machine learning for patient classification based on multi-omics data such as genetics, immunomics and metabolomics, in combination with clinical values and laboratory results. Machine learning models resulted in 95% prediction accuracy of autoimmune diseases using integrated data. Our results deliver insights into autoimmune disease research and have the potential to be adapted for applications across disease conditions.

We profiled blood and draining lymph node (LN) samples from human volunteers after influenza vaccination over two years to define evolution in the T follicular helper cell (TFH) response. We show LN TFH cells expanded in a clonal-manner during the first two weeks after vaccination and persisted within the LN for up to six months. LN and circulating TFH (cTFH) clonotypes overlapped but had distinct kinetics. LN TFH cell phenotypes were heterogeneous and mutable, first differentiating into pre-TFH during the month after vaccination before maturing into GC and IL-10+ TFH cells. TFH expansion, upregulation of glucose metabolism, and redifferentiation into GC TFH cells occurred with faster kinetics after re-vaccination in the second year. We identified several influenza-specific TFH clonal lineages, including multiple responses targeting internal influenza proteins, and show each TFH state is attainable within a lineage. This study demonstrates that human TFH cells form a durable and dynamic multi-tissue network.

Colonization by the microbiota causes a marked stimulation of B cells and induction of immunoglobulin, but mammals colonized with many taxa have highly complex and individualized immunoglobulin repertoires. To study this, we opted for a simplified model of defined transient exposures to different microbial taxa in germ-free mice to deconstruct how the microbiota shapes the B cell pool and its functional responsiveness. We previously showed that microbial exposure at the intestinal mucosa generated oligoclonal responses that differed from those in germ-free mice, and from the diverse repertoire that was generated after intravenous systemic exposure to microbiota. Our results reflected a contrast between a flexible response to systemic exposure with the need to avoid fatal sepsis, and a restricted response to mucosal exposure that reflects the generic nature of host–microbial mutualism in the mucosa. A hallmark of mucosal IgA is the high mutational load that accumulates throughout life. This has been mainly interpreted in terms of differences in microbial induction, although once established in early life, the microbiome of humans and experimental mice is relatively stable and mutational activity has been shown to be independent of B cell receptor signaling. Using germ-free and colonized mice provided with different diets formulated with proprietary grain-based processing or from purified chemicals with different principal macronutrient calorie sources, we show that diet affects IgA induction, its repertoire and mutational diversification independently of microbial exposure.

The recognition of a T-cell epitope target is driven by the unique sequence of the T-cell receptor (TCR). As the TCR sequence theoretically contains all the information that determines its target, it must be possible to infer its target from this sequence. Indeed, within specific settings, we now have performant machine learning models that can address this challenge. It has therefore become necessary to start considering how we can apply these models to gain novel immunological understanding.

In this talk, I want to focus on the application of TCR-epitope prediction models to identify epitope-specific T-cells in full repertoire data, and the additional challenges that are often missed by current evaluation efforts. I also aim to highlight the biggest advancements and milestones that we can expect in the coming years in the field, especially in regard to the unseen epitope prediction problem.

Pathogens and cancer cells can escape recognition by T cell receptors (TCRs) through mutations of immunogenic epitopes. TCR cross-reactivity, i.e., recognition of multiple epitopes with sequence similarities, can counteract such mutational escape, but also may cause severe side effects in cell-based immunotherapies. To predict the effect of epitope mutations on T cell functionality in silico, we present “Predicting T cell Epitope-specific Activation against Mutant versions” (P-TEAM). The Random Forest based model was developed on two comprehensive datasets of murine and human TCRs in response to systematic single-amino acid mutations of their target epitopes to predict T cell reactivity for unobserved mutations, or even unseen TCRs. P-TEAM is complemented with an active learning framework to guide experimental design to minimize primary data acquisition costs. Overall, P-TEAM provides an effective computational tool to study T cell responses against mutated epitopes.

T cells orchestrate the adaptive immune response against pathogens and cancer by recognizing epitopes presented on MHC molecules. The heterogeneity of the MHC peptidome, including the high polymorphism of MHC genes, is influencing TCR repertoires and represents an important challenge towards accurate prediction and identification of T-cell epitopes in different individuals and different species. Here we generated and curated a dataset of more than a million unique MHC-I and MHC-II ligands identified by mass spectrometry. This enabled us to precisely determine the binding motifs of >200 MHC alleles across human, mouse, cattle and chicken. Analysis of these binding specificities combined with X-ray crystallography refined our understanding of the molecular determinants of MHC motifs and revealed alternative binding modes of MHC ligands. We then developed machine learning frameworks to accurately predict binding specificities and ligands of any MHC-I (MixMHCpred) and MHC-II (MixMHC2pred) allele, and further integrated TCR recognition into our epitope prediction pipeline (PRIME). Prospectively applying our tools to SARS-CoV-2 proteins identified several epitopes and TCR sequencing revealed a monoclonal response in effector/memory CD8+ T cells against one of these epitopes with cross-reactivity against the homologous peptides from other coronaviruses. Overall, our work shows how in depth characterization of MHC motifs can help mapping the targets of T cells and understanding TCR cross-reactivity.

Despite the vast diversity of the antibody repertoire, infected individuals often mount antibody responses to precisely the same epitopes within antigens. The immunological mechanisms underpinning this phenomenon remain unknown. By mapping 376 immunodominant “public epitopes” at high resolution and characterizing several of their cognate antibodies, we concluded that germline-encoded sequences in antibodies drive recurrent recognition. Systematic analysis of antibody-antigen structures uncovered 18 human and 21 partially overlapping mouse germline-encoded amino acid–binding (GRAB) motifs within heavy and light V gene segments that in case studies proved critical for public epitope recognition. GRAB motifs represent a fundamental component of the immune system’s architecture that promotes recognition of pathogens and leads to species-specific public antibody responses that can exert selective pressure on pathogens.

As individual researchers, we are well-trained in designing research projects that address specific and moderately challenging questions. However, when faced with a grand challenge such as understanding antigen recognition by T- and B-cell receptors, how can we, both individually and collectively, effectively tackle this problem? We currently have antigen binding data for only a minuscule proportion of the immune receptor sequence space, and despite the rapid emergence of new machine learning methods, consensus on what constitutes the most promising directions forward remains scarce.

In my talk, I will offer my perspective on this strategic question and discuss some recently published and ongoing work aligned with this strategy. Briefly, I believe the first step should be a rigorous characterisation of the computational challenges of the immune receptor-antigen binding prediction problem. Second, we must ensure that sufficient data is available to guide methodology development, where we in the foreseeable future need to rely on the combined use of experimental and simulated data. Third, we must prioritise interoperability and reproducibility of methods, along with the development of standardised benchmarks, to effectively compare performance and identify limitations of current approaches.

A new approach that utilizes the Lempel-Ziv 76 algorithm (LZ-76) to encode and represent AIRRs without relying on sequence annotation. The approach involves creating a graph-like model, which enables a wide range of potential applications, including generation probability inference, informative feature vector derivation, sequence generation, sequence analysis, and a new measure for repertoire diversity estimation.

T cells play a crucial role in our immunity by recognizing and eliminating infected and abnormal cells. T cell response is triggered upon T cell recognition of specific antigens presented by MHC molecules on the surface of target cells. However, the underlying rules of the interactions are incompletely understood. Over the past few years, we have been investigating T cell cross-reactivity and common specificity as two key drivers of T cell antigen specificity. In this seminar, I will focus on T cell cross-reactivity, and will introduce a new deep neural network model that we have developed for accurate and reliable prediction of CD8 T cell targets. I will illustrate how we have been using this model to model the dynamics of CD8 T cell response to emerging SARS-CoV-2 mutant variants.

Antibodies result from the competition of B cell lineages evolving under selection for improved antigen recognition, a process known as affinity maturation. High-affinity antibodies to pathogens such as HIV, influenza, and SARS-CoV-2 are frequently reported to arise from B cells whose receptors are encoded by particular immunoglobulin genes. This raises the possibility that the presence of particular germline genes in the B cell repertoire is a major determinant of the quality of the antibody response. Alternatively, initial differences in germline genes’ propensities to form high-affinity receptors might be overcome by chance events during affinity maturation. We first show how this can happen in simulations: even when fitness differences between germline genes lead to similar gene usage across individuals early on, gene usage can become increasingly dissimilar over time. We next find that mice experimentally infected with influenza virus demonstrate the same pattern of divergence in the weeks following infection. We investigated whether affinity maturation might nonetheless strongly select for particular amino acid motifs across diverse genetic backgrounds, but we found no evidence of convergence to similar CDR3 sequences or amino acid substitutions. These results suggest germline-encoded specificities might enable fast recognition of specific antigens early in the response, but diverse evolutionary routes to high affinity limit the genetic predictability of responses to infection and vaccination in the long term.

Early and accurate disease diagnosis is crucial for preventing disease development and defining therapy strategies. Due to predominantly unspecific symptoms, diagnosis of autoimmune diseases is notoriously challenging. However, multiple types of data are not leveraged for precision diagnostics due to the difficulties of integrating and encoding multi-omics data with clinical and laboratory values, as well as a lack of standardization; clinical decision support systems are often limited to only certain data types. Accordingly, even sophisticated data models fall short when supporting accurate diagnoses and presenting data analyses in a user-friendly form. Therefore, the integration of various data types is not only an opportunity but also a competitive advantage in research and for the industry. We have developed an integration pipeline to enable the use of machine learning for patient classification based on multi-omics data such as genetics, immunomics and metabolomics, in combination with clinical values and laboratory results. Machine learning models resulted in 95% prediction accuracy of autoimmune diseases using integrated data. Our results deliver insights into autoimmune disease research and have the potential to be adapted for applications across disease conditions.

We profiled blood and draining lymph node (LN) samples from human volunteers after influenza vaccination over two years to define evolution in the T follicular helper cell (TFH) response. We show LN TFH cells expanded in a clonal-manner during the first two weeks after vaccination and persisted within the LN for up to six months. LN and circulating TFH (cTFH) clonotypes overlapped but had distinct kinetics. LN TFH cell phenotypes were heterogeneous and mutable, first differentiating into pre-TFH during the month after vaccination before maturing into GC and IL-10+ TFH cells. TFH expansion, upregulation of glucose metabolism, and redifferentiation into GC TFH cells occurred with faster kinetics after re-vaccination in the second year. We identified several influenza-specific TFH clonal lineages, including multiple responses targeting internal influenza proteins, and show each TFH state is attainable within a lineage. This study demonstrates that human TFH cells form a durable and dynamic multi-tissue network.

Colonization by the microbiota causes a marked stimulation of B cells and induction of immunoglobulin, but mammals colonized with many taxa have highly complex and individualized immunoglobulin repertoires. To study this, we opted for a simplified model of defined transient exposures to different microbial taxa in germ-free mice to deconstruct how the microbiota shapes the B cell pool and its functional responsiveness. We previously showed that microbial exposure at the intestinal mucosa generated oligoclonal responses that differed from those in germ-free mice, and from the diverse repertoire that was generated after intravenous systemic exposure to microbiota. Our results reflected a contrast between a flexible response to systemic exposure with the need to avoid fatal sepsis, and a restricted response to mucosal exposure that reflects the generic nature of host–microbial mutualism in the mucosa. A hallmark of mucosal IgA is the high mutational load that accumulates throughout life. This has been mainly interpreted in terms of differences in microbial induction, although once established in early life, the microbiome of humans and experimental mice is relatively stable and mutational activity has been shown to be independent of B cell receptor signaling. Using germ-free and colonized mice provided with different diets formulated with proprietary grain-based processing or from purified chemicals with different principal macronutrient calorie sources, we show that diet affects IgA induction, its repertoire and mutational diversification independently of microbial exposure.

The recognition of a T-cell epitope target is driven by the unique sequence of the T-cell receptor (TCR). As the TCR sequence theoretically contains all the information that determines its target, it must be possible to infer its target from this sequence. Indeed, within specific settings, we now have performant machine learning models that can address this challenge. It has therefore become necessary to start considering how we can apply these models to gain novel immunological understanding.

In this talk, I want to focus on the application of TCR-epitope prediction models to identify epitope-specific T-cells in full repertoire data, and the additional challenges that are often missed by current evaluation efforts. I also aim to highlight the biggest advancements and milestones that we can expect in the coming years in the field, especially in regard to the unseen epitope prediction problem.

Pathogens and cancer cells can escape recognition by T cell receptors (TCRs) through mutations of immunogenic epitopes. TCR cross-reactivity, i.e., recognition of multiple epitopes with sequence similarities, can counteract such mutational escape, but also may cause severe side effects in cell-based immunotherapies. To predict the effect of epitope mutations on T cell functionality in silico, we present “Predicting T cell Epitope-specific Activation against Mutant versions” (P-TEAM). The Random Forest based model was developed on two comprehensive datasets of murine and human TCRs in response to systematic single-amino acid mutations of their target epitopes to predict T cell reactivity for unobserved mutations, or even unseen TCRs. P-TEAM is complemented with an active learning framework to guide experimental design to minimize primary data acquisition costs. Overall, P-TEAM provides an effective computational tool to study T cell responses against mutated epitopes.

T cells orchestrate the adaptive immune response against pathogens and cancer by recognizing epitopes presented on MHC molecules. The heterogeneity of the MHC peptidome, including the high polymorphism of MHC genes, is influencing TCR repertoires and represents an important challenge towards accurate prediction and identification of T-cell epitopes in different individuals and different species. Here we generated and curated a dataset of more than a million unique MHC-I and MHC-II ligands identified by mass spectrometry. This enabled us to precisely determine the binding motifs of >200 MHC alleles across human, mouse, cattle and chicken. Analysis of these binding specificities combined with X-ray crystallography refined our understanding of the molecular determinants of MHC motifs and revealed alternative binding modes of MHC ligands. We then developed machine learning frameworks to accurately predict binding specificities and ligands of any MHC-I (MixMHCpred) and MHC-II (MixMHC2pred) allele, and further integrated TCR recognition into our epitope prediction pipeline (PRIME). Prospectively applying our tools to SARS-CoV-2 proteins identified several epitopes and TCR sequencing revealed a monoclonal response in effector/memory CD8+ T cells against one of these epitopes with cross-reactivity against the homologous peptides from other coronaviruses. Overall, our work shows how in depth characterization of MHC motifs can help mapping the targets of T cells and understanding TCR cross-reactivity.

Despite the vast diversity of the antibody repertoire, infected individuals often mount antibody responses to precisely the same epitopes within antigens. The immunological mechanisms underpinning this phenomenon remain unknown. By mapping 376 immunodominant “public epitopes” at high resolution and characterizing several of their cognate antibodies, we concluded that germline-encoded sequences in antibodies drive recurrent recognition. Systematic analysis of antibody-antigen structures uncovered 18 human and 21 partially overlapping mouse germline-encoded amino acid–binding (GRAB) motifs within heavy and light V gene segments that in case studies proved critical for public epitope recognition. GRAB motifs represent a fundamental component of the immune system’s architecture that promotes recognition of pathogens and leads to species-specific public antibody responses that can exert selective pressure on pathogens.

As individual researchers, we are well-trained in designing research projects that address specific and moderately challenging questions. However, when faced with a grand challenge such as understanding antigen recognition by T- and B-cell receptors, how can we, both individually and collectively, effectively tackle this problem? We currently have antigen binding data for only a minuscule proportion of the immune receptor sequence space, and despite the rapid emergence of new machine learning methods, consensus on what constitutes the most promising directions forward remains scarce.

In my talk, I will offer my perspective on this strategic question and discuss some recently published and ongoing work aligned with this strategy. Briefly, I believe the first step should be a rigorous characterisation of the computational challenges of the immune receptor-antigen binding prediction problem. Second, we must ensure that sufficient data is available to guide methodology development, where we in the foreseeable future need to rely on the combined use of experimental and simulated data. Third, we must prioritise interoperability and reproducibility of methods, along with the development of standardised benchmarks, to effectively compare performance and identify limitations of current approaches.

A new approach that utilizes the Lempel-Ziv 76 algorithm (LZ-76) to encode and represent AIRRs without relying on sequence annotation. The approach involves creating a graph-like model, which enables a wide range of potential applications, including generation probability inference, informative feature vector derivation, sequence generation, sequence analysis, and a new measure for repertoire diversity estimation.

T cells play a crucial role in our immunity by recognizing and eliminating infected and abnormal cells. T cell response is triggered upon T cell recognition of specific antigens presented by MHC molecules on the surface of target cells. However, the underlying rules of the interactions are incompletely understood. Over the past few years, we have been investigating T cell cross-reactivity and common specificity as two key drivers of T cell antigen specificity. In this seminar, I will focus on T cell cross-reactivity, and will introduce a new deep neural network model that we have developed for accurate and reliable prediction of CD8 T cell targets. I will illustrate how we have been using this model to model the dynamics of CD8 T cell response to emerging SARS-CoV-2 mutant variants.

Antibodies result from the competition of B cell lineages evolving under selection for improved antigen recognition, a process known as affinity maturation. High-affinity antibodies to pathogens such as HIV, influenza, and SARS-CoV-2 are frequently reported to arise from B cells whose receptors are encoded by particular immunoglobulin genes. This raises the possibility that the presence of particular germline genes in the B cell repertoire is a major determinant of the quality of the antibody response. Alternatively, initial differences in germline genes’ propensities to form high-affinity receptors might be overcome by chance events during affinity maturation. We first show how this can happen in simulations: even when fitness differences between germline genes lead to similar gene usage across individuals early on, gene usage can become increasingly dissimilar over time. We next find that mice experimentally infected with influenza virus demonstrate the same pattern of divergence in the weeks following infection. We investigated whether affinity maturation might nonetheless strongly select for particular amino acid motifs across diverse genetic backgrounds, but we found no evidence of convergence to similar CDR3 sequences or amino acid substitutions. These results suggest germline-encoded specificities might enable fast recognition of specific antigens early in the response, but diverse evolutionary routes to high affinity limit the genetic predictability of responses to infection and vaccination in the long term.