The Immunodynamics Group (led by Grégoire Altan-Bonnet) within the Laboratory of Integrative Cancer Immunology of the National Cancer Institute is dedicated to developing quantitative models of the immune system responding to tumors and infections. We contribute to the development of systems immunology by combining novel experimental approaches and computational/theoretical modeling. Here is our manifesto:

Top-to-bottom and bottom-up approaches for systems immunology.

In top-to-bottom approaches, we leverage the explosion of novel experimental techniques to monitor the multifactorial nature of immune responses: we use single-cell analyses (e.g. mass cytometry, spectral cytometry and single-cell transcriptomics) to characterize millions

of leukocytes, their phenotype, their state of activation and their overall contribution to the dynamics of immune responses

In our bottom-up approaches, we seek to build “simple” biochemically-accurate models of T cell activation to make targeted predictions about new functions for immune cells. This approach takes a commitment to reducing the immune complexity to few phenomenological equations, towards deriving generalizable understanding of immune responses.

In our top-to-bottom approach, we assemble large datasets (e.g. multiplexing donors, activation conditions, timepoints etc.) to explore the combinatorial complexity of immune responses. In that context, we use machine learning to derive new low-dimensional representation and biomarkers that best capture how immune cells function (e.g. identifying a specific cell types whose abundance predicts clinical outcomes goal is ).

Ultimately, we aim at merging both top-to-bottom and bottom-up approaches: 1) to learn a latent space that best captures the immune system; 2) to build models that explain the dynamics and functional relevance of this latent space.

Immunology and machine learning

The large datasets generated represent the biggest challenge and opportunity for the field of systems immunology. A typical experiment in the lab accumulates 100 Million cells, characterized by more than 40 single-cell features (e.g. up/down regulation of surface receptors, expression of transcription factors, activation of kinases, production of cytokines) across multiple activation conditions and time. We tackle these large high-dimensional datasets to derive a better understanding (i.e. a model) using tools from machine learning. In particular, auto-encoders and multi-layer perceptrons (MLP) have been used to build quantitative models of ligand discrimination by T cells. We are actively seeking novel tools from machine learning to improve our modeling capacities.

Of note, this interplay between systems immunology and artificial intelligence goes back-and-forth: while tools of machine learning obviously help tackle the combinatorial nature of our experimental observations, novel approaches in machine learning can also get inspired by the dynamic architecture of the immune system (this analogy was introduced by Paul François & Colleagues). At a fundamental level, we are exploring how the immune system must be conceptualized as a liquid brain, relying on the the multiscale functionalities of its leukocytes in order to classify pathological challenges and to decide accurate self/non-self discrimination.

Adaptive Kinetic Proofreading: a central scheme for decision making in the immune system.

How does the immune system discriminate between self tissues (that shall be left untouched) and non-self (that shall be attacked and cleared)? If viruses, bacteria and parasites clearly generate non-self antigens that trigger T cell activation, what about tumors? We have been building models of lymphocytes' decision making to understand how molecules (e.g. short peptides taken from virus, bacteria or tumors) determine functional response by immune cells. Our recent study on antigen encoding (re)discovered the "simple" scheme that accounts for T cells' sharp, specific, sensitive and rapid ligand discrimination. In a nutshell, the competition between a positive  kinetic pruning of ligand/receptor activation and a negative feedback that limits the march towards activation is sufficient to recapitulate many aspects of T cell biology. In particular, our model predicts the existence of antagonism i.e. the nonlinear and often paradoxical response of T cells to mixture of ligands.  In collaboration with Paul François (McGill University), we are committed to further explore the theoretical and practical underpinning of such adaptive kinetic proofreading scheme

Immune responses: coordinating cell activation from local to global effects

Collaborations:

Due to the obvious multidisciplinary nature of our research projects, we systematically seek novel collaborations to expand the range of experimental systems and theoretical approaches in our systems immunology projects. We document here few of the bridges we establish with other groups

            We have a long-standing collaboration with Paul François’s group in the Department of Physics at McGill University to develop phenomenological and biochemically-accurate models of leukocyte activation. Our recent joint publication demonstrated the symbiotic nature of multiplexing dynamic measurements and machine learning to derive a compressed model of T cell activation and discrimination of ligands.

            We collaborate with the DeVoe lab in the Department of Mechanical Engineering of the University of Maryland @ College Park to develop novel microfluidic devices to assemble, monitor and manipulate immune reactions.

            The laboratories of the National Institutes of Health in Bethesda are extremely collaborative, allowing the deployment of our systems immunology/machine learning toolbox to many projects, including:

• Innate responses against viral infections (with Nihal Altan-Bonnet’s group at NHLBI)
• T cell development and signaling (with Paul Love’s group at NICHD)
• Cytokine signaling (with Carmen Molina Paris & Grant Lythe at the university of Leeds and Anton Zilman’s group at the university of Toronto)
• Cytokine-mediated regulations of self/non-self discrimination (with Ron Germain’s group at NIAID) and responses to tumors (with Nadège Bercovici’s group at Institut Gustave Roussy and Audrey Gérard’s group at Oxford University).

Finally, we have established collaborations with clinical labs to realize the promises of bench-to-bedside (and back!). Of note, we have a strong collaboration with Nirali Shah’s and Naomi Taylor’s groups within the Pediatric Oncology Branch at NCI to characterize, model and optimize Chimeric-Antigen Receptor (CAR) T cell therapies. We apply our robotic/machine learning pipeline to characterize CAR-T cells from clinical products as well mouse models. We aim at better understanding how CAR-T not only kill tumor targets, but also differentiate, proliferate and persist towards sustained eradication of tumors. This collaborative project with the Shah/Taylor groups has been supported by an intramural synergy grant of the National Cancer Institute.