Bridging cell and organism scales to model viral, cell, and microenvironmental determinants of infection outcome

Duke University, Durham NC

Investigators

Abstract

Research Project 2 (RP2) proposes to develop equation learning and model-constrained neural network frameworks to bridge between spatially explicit cell scale models and individual scale models. Agent based models (ABM) will be used to model the cell scale, while differential equation and stochastic process models will be used to represent the individual scale. Crucially, we will then use modern deep learning approaches to build individual scale models that incorporate information from bulk, single-cell, and spatial assays. We will use these models to determine the viral, cellular, and spatial drivers generating atypical T-bet+ memory B cells (ABC) after EBV infection, and precisely quantify conditions for HIV-1 reactivation in latently infected T cells in the lymph node after administering latency reversing agent (LRA). The clinical importance of these questions is that ABCs are a likely causal factor in the strong association between EBV and autoimmune disease, especially multiple sclerosis, and our studies will contribute to insight about the pathogenesis of EBV-associated autoimmune disuse, and LRA reactivation is a part of a major strategy for HIV cure, and hence this work will help define the precise conditions under which latently infected cells can be reactivated in vivo. To develop the bridging framework, we propose to characterize viral infection dynamics in lymphoid tissue cultures (Aim 1), construct cell scale ABM to model spatiotemporal interactions between virus and host cells (Aim 2), and bridge cell to organism scale models using equation learning and Biologically Informed Neural Networks (BINN) (Aim 3). We will use tonsil explants for the EBV experiments, and a microfluidic lymph node on a chip system for HIV experiments. For both experiments, we will perform various assays across multiple time points, including viral loads, cytokine profiles, single-cell gene expression, and Co-Detection by Indexing (CODEX) spatial proteomics. Next, ABMs based on the Vivarium software platform will used to model the spatiotemporal interactions between the virus and host cells, seeking to replicate in silico the observed tissue dynamics. These models will be initialized using data generated by the lymphoid tissue culture experiments, including spatial single-cell CODEX data and dynamic viral loads. The calibrated model will be used to study outcome variation, explore perturbations, and generate training data, particularly regarding dynamic cellular trajectories and their multicellular interactions. Finally, we will use synthetic data sets generated by calibrated ABM models for equation learning and training the BINN. This involves automatically learning and calibrating mathematical models. Stochastic process models can be converted into partial differential equations or truncated ordinary differential equations using the master equation approach and learned in the same way. If successful, the bridging framework developed in RP3 will provide a process for automatic condensation of high-dimensional ABM spatiotemporal dynamics into a set of low-dimensional parameterized equations, a process that might otherwise take a graduate student in Mathematical Biology several years to accomplish.