(B-76) Enabling Large-Scale Three-Dimensional Models of the Adaptive Immune Response

Introduction:: Computational models of T cell motility and differentiation during an adaptive immune response, complemented with experimental methods, offer the potential for technological advancement in vaccine design, autoimmune disorders, and cancer therapy. However, resolving the interactions among densely packed T lymphocytes and dendritic cells (DCs) in the presence of chemical gradients in a 3D virtual lymph node microenvironment over long temporal domains is computationally burdensome. Moreover, it can quickly become intractable on shared memory machines. Current computational models of T cell priming are based on shared memory implementations, limiting their use to problem sizes that fit within a single compute node [1], [2], [3]. Moreover, the majority of these same models are 2D in nature, which adds artificial constraints on cell trajectories. Our objective is to overcome these challenges by introducing a 3D, distributed memory parallel model of T cell priming in the lymph node that can resolve the entire course of clonal expansion in the presence of an inflammation signal.

Materials and Methods:: The underlying mathematics of our model, describing T cell-DC contact dynamics, T cell jump probabilities, and chemotactic gradients, were adapted from the shared memory model described in [1] with modifications. Our model assumes a six-neighbor 3D von Neumann neighborhood for T cell motility. In contrast, a 26-neighbor 3D Moore neighborhood resolves interactions among cognate T cells, representing a fraction of the total T cell population and DCs. During each time step, new T cells arrive in the domain at random sites at an influx rate determined by the current inflammation level; stimulated T cells can divide and form new daughter cells, T cells can move about the simulated lattice domain, interact with a nearby DC and accumulate stimulus signal, or exit the LN upon reaching an exit zone representing a high endothelial venule. DC-presented antigen surface densities follow an exponential decay, and eventually, the stationary DCs die and are replaced by new DCs that spawn at random sites. We parallelized our code following a domain decomposition approach in which processors manage cells within their prescribed 3D bounding box and exchange pertinent information to resolve cell interactions between neighboring processors at task boundaries. Performance was evaluated with starting populations of 100,000 T cells with a cognate cell fraction of 10% and 2,000 DCs over 25 simulated hours.

Results, Conclusions, and Discussions:: Figure 1A displays the clonal expansion of cognate naïve T lymphocytes over the duration of an inflammation signal, using a starting population of 1,000 T cells (10% cognate fraction), 200 DCs. Figure 1B shows the performance results of the parallel implementation, demonstrating a peak speedup of approximately 230x in time-to-solution using 1,728 processors of Stampede2 Knights Landing compute nodes.



Overall, our 3D distributed parallel model of T cell dynamics during an adaptive immune response provided orders of magnitude speedup gains, enabling investigators to study T cell-DC contact dynamics at a physiological scale in minutes instead of hours. With these time-saving benefits, our platform facilitates high-throughput parameter studies and drug screens of intranodal biology that will inform bench side experiments and clinical applications.

Acknowledgements (Optional): : This work used the Extreme Science and Engineering Discovery Environment (XSEDE) Stampede2 at TACC through allocation MDE210001.

References (Optional): : [1] Bogle, Gib, and P. Rod Dunbar. "On-lattice simulation of T cell motility, chemotaxis, and trafficking in the lymph node paracortex." (2012): e45258. https://doi.org/10.1371/journal.pone.0045258

[2] Azarov, Ivan, et al. "Role of T cell-to-dendritic cell chemoattraction in T cell priming initiation in the lymph node: An agent-based modeling study." Frontiers in immunology 10 (2019): 1289.

[3] Grebennikov, Dmitry, et al. "Spatial lymphocyte dynamics in lymph nodes predicts the cytotoxic T cell frequency needed for HIV infection control." Frontiers in Immunology 10 (2019): 1213.