|Parallel stochastic spines in NEURON reaction-diffusion simulations
|Year of Publication
|Mcdougal, R. A., Newton A. J. H., Patoary M. N. I., Tropper C., Hines M. L., & Lytton WW.
|Society for Neuroscience 2017 (SFN '17)
|SFN, Society for Neuroscience
The brain processes information across multiple spatial scales – from the communication between brain regions to the regulation of individual ions within neurons. The NEURON simulator provides a domain-specific way to develop predictive, reusable models of individual morphologically detailed neurons and networks of neurons. In NEURON 7.3, we extended NEURON from its electrophysiological roots to incorporate the role of intracellular reaction-diffusion kinetics (e.g. calcium dynamics or protein oscillations). As with the electrophysiology, the reaction-diffusion kinetics are modeled using parabolic partial differential equations. This numerical scheme implicitly assumes concentrations are sufficiently high and compartment volumes are sufficiently large; we have used it, for example, to simulate a propagating calcium wave in one dimension along an apical dendrite. These assumptions break down in the special cases of calcium in the spine (a typical spine head under typical conditions is estimated to contain only 30 calcium ions) or in comparably fine 3D discretizations of a neuronal process. To simulate these kinetics appropriately, we have previously coupled NEURON with the parallel Next Subvolume Method solver NTW. We have now increased the integration of NTW into NEURON to allow coupling between the stochastic and deterministic solvers e.g. stochastic spines with deterministic dendrites with all kinetics are specified through NEURON's existing reaction-diffusion specification. We compare implementation and performance of two approaches to parallelization in this problem: NTW's default using an optimistic time warp algorithm vs with a separate thread per-spine. This integrated multi-scale, multi-method approach to reaction-diffusion simulations now allows NEURON to more faithfully reproduce neuronal behavior with minimal computational and implementer overhead.