A biophysical model of two interacting cortical areas

We present a large-scale, data-driven, biophysically-detailed computational model of two interacting cortical areas, based on data from rodent somatosensory cortex.

This model is derived from a previous model (described in two manuscripts: anatomy, physiology), but it consists of a reduced setting tailored to the study of inter-areal interactions in cortical sensory processing. Details of the model and initial results can be found here.

Description

The model describes a system of two otherwise isolated cortical areas (X and Y), where area X is a primary sensory and area Y is the first higher-order area in a cortical processing hierarchy. Each area consists of about 200K morphologically-detailed conductance-based neurons, distributed across six cortical layers and of 60 different morphological types and 212 morpho-electrical types.

The model incorporates the following connectivity:

• Local touch-based connectivity within each area.
• Thalamocortical innervation from both VPM (core-type) and POm (matrix-type) nuclei to area X.
• Long-range data-driven projections between both areas with characteristic laminar termination profiles.

Additionally, each area receives nonspecific background noise to exhibit spontaneous activity comparable to experimental recordings of per-layer mean firing rates.

Setup

The model is provided in the SONATA format and can be run using Neurodamus, a simulator frontend for NEURON. Synaptic and ion channel mechanisms specific for neocortical neurons are also required to run this model (build instructions here).

To setup the model, all files must be placed in the same directory and all the Lzip-compressed TAR archives must be extracted. The total uncompressed size is 219 GB.

$ mkdir model_root
# download all files into model_root
$ cd model_root
$ for file in *.tar.lz; do tar -xf $file; done

Simulation

We provide some example configuration files (under example_simulation_configs) for simulations of spontaneous and evoked activity, as well as some network manipulations (layer-wise pathway blocks and TTX application).

In order to run a simulation, copy simulation_config.json into a new directory and set the network key to the path of the directory containing the extracted model (optionally, set the output key as well). Instructions for running a simulation can be found here and documentation for the simulation configuration file can be found here.

Analysis

Analysis of model composition and connectivity, as well as of simulation outputs, can be performed using Blue Brain SNAP or by directly accessing the HDF5 files with libsonata. Documentation on the SONATA format for all files making up the model can be found here.

Computational resources

Approximate scaling of computational resources is as follows (based on simulations of 5 s biological time running on a cluster with 40 cores @ 2.5 GHz and 376 GB of RAM per node, one MPI process per core, using CoreNEURON):

• Memory per process = 1106 GB / N ** 0.87
• Simulation time = 4278 h / N ** 0.93

For example, running with N = 1000 processes (25 nodes) results in 107 GB memory usage per node and 7h16m simulation time for 5 s biological time. Longer simulations scale approximately linearly in time, taking 13h13m for 10 s biological time and 21h48m for 15 s biological time.

Changelog

v1.0.1
Fixed (unused) key "node_sets_file" in example simulation configuration files.