P259 Local multi-gridding for detailed morphology, spines and synapses
Cecilia Romaro*1, William W. Lytton2,3,4,5, Robert A. McDougal1,6,7,,8
1Department of Biostatistics, Yale School of Public Health, New Haven, CT, United States
2Department of Physiology and Pharmacology, SUNY Downstate Health Sciences University, Brooklyn, New York, United States
3Department of Neurology, SUNY Downstate Health Sciences University, Brooklyn, New York, United States
4Department of Neurology, Kings County Hospital Center, Brooklyn, New York, United States
5The Robert F. Furchgott Center for Neural and Behavioral Science, Brooklyn, New York, United States
6Department of Biomedical Informatics and Data Science, Yale School of Medicine, New Haven, CT, United States
7Program in Computational Biology and Bioinformatics, Yale University, New Haven, CT, United States
8Wu Tsai Institute, Yale University, New Haven, CT, United States
*Email: cecilia.romaro@yale.edu
Introduction
The NEURON simulator (https://nrn.readthedocs.io) is one of the most widely-used tools for simulating biophysically detailed neurons and networks [1]. In addition to electrophysiology simulations, NEURON has long supported multi-scale models incorporating intra- and extracellular chemical reaction-diffusion [2], in both 1D and 3D. To accurately simulate whole cells in 3D requires capturing large regions like somas and small regions like spines. We demonstrate an algorithm in NEURON for achieving high-quality results with reasonable computational cost through local multi-gridding.
Methods
We extended NEURON's reaction-diffusion Region specification to support per-Section grid size specification. Sections with different grid sizes are independently discretized using NEURON's standard voxelization algorithm [3]. Small voxels are removed and/or added to produce a join with minimal voxel overlap. Neighboring voxels of different sizes are connected to allow molecules to diffuse between the grids. For ease of use, the model specification is in Python; for performance, coupling between grids and all simulation is done in C++.
Results
Multigrid-voxelization overhead due to the editing and alignment of the grids is small but measurable. Mass is conserved when diffusing across the grid-size boundary, however subtle differences may arise in numerical results due to the changes in volume and surface-area voxel-size-dependent estimates; implications for assessing convergence are discussed. Accuracy and performance are assessed for both simplified morphologies and detailed cell morphologies from NeuroMorpho.Org; initialization and simulation are necessarily slower than for the coarse grid, (but not for the finest grid) however the time cost and accuracy improvements are highly dependent on the problem.
Discussion
Using multiple grid sizes for 3D reaction-diffusion simulation allows increased accuracy in small parts of the morphology or in regions of interest with moderate compute overhead. This approach preserves the regular sampling and easy convergence testing of NEURON's finite-volume integration. This numerical simulation method pairs naturally with ongoing work to import high-resolution neuron spine morphologies into NEURON models, with the spine and the dendrites simulated using different grids. Carefully chosen grid sizes have the potential to enable high-fidelity simulations combining chemical, electrical, and network activity with modest compute resources.
Acknowledgements
This research was funded by the National Institute of Mental Health, National Institutes of Health, grant number R01 MH086638. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.
References
1.https://doi.org/10.3389/fninf.2022.884046
2.https://doi.org/10.3389/fninf.2022.847108
3.https://doi.org/10.1016/j.jneumeth.2013.09.011
