Large-Scale Neural Tissue Simulations

A basket cell interneuron in simulated cerebral cortical tissue.
Figure 1. A basket cell interneuron in simulated cerebral cortical tissue from Figure 11 in “An ultrascalable solution to large-scale neural tissue simulation” (published September 19, 2011 in Frontiers in Neuroinformatics). Below the basket cell, traces of electrical activity from its dendrites are displayed. The article is being displayed in iPad using JournaLink version 1.4.

The author’s of the recent paper “An ultrascalable solution to large-scale neural tissue simulation” (published September 19, 2011 in Frontiers in Neuroinformatics) define neural tissue simulations as having the following characteristics:

  • multi-compartment Hodgkin-Huxley models of neurons derived from anatomical reconstruction of real neurons
  • support synaptic coupling between compartments that attempt to match synaptic distributions from real tissue
  • incorporate the three-dimensional coordinate system of neural tissue

The incorporation of structural constraints is a critical factor in neural tissue simulations. They guide the arrangement in synaptic connectivity of simulated compartments that make up the component neurons and, ultimately, the neural tissue. These are the same constraints placed on real neural tissue. Simulations that adhere to these real world constraints have the potential to provide insights into the functioning of real brain tissue.

The paper reports on large-scale simulations of cerebral cortex including 1 million neurons comprised of 1 billion compartments and connected through 10 billion conductance-based synapses and gap junctions. The neurons were derived from the morphology data of real neurons accessed at the public neuromorphology data repository

A significant feature of the reported neural tissue simulations is the use of complete compartment models of axons. The traversal of electrical current and action potentials are not only simulated across the dendrites and cell bodies but also along axon branches. This enables realistic modeling of action potential failures and conduction times.

Clearly simulations of this detail and scale need special machinery. The research team is from IBM’s T. J. Watson Research Center so, not surprisingly, they used the second generation Blue Gene supercomputer known as Blue Gene/P. Nevertheless, these simulations demanded computational ingenuity which takes up much of the discussion in the paper. The Neural Tissue Simulator utilizes what appear to be proprietary technologies known as Model Definition Language (MDL) and Model Graph Simulator (GSL).

Note: The authors provide MDL and GSL scripts and other files for creating and running the largest simulation reported in the paper. You may download them from the supplemental data link available at the paper’s website. The authors also say they’d like to share the Neural Tissue Simulator software and source code.

The current paper is the first report of simulations of more than one million neurons. The work demonstrates the computational feasibility of human brain scale neural tissue simulations within the next decade or so. To actually accomplish the feat, of course, more than computational capabilities will need to be met. For one thing, knowledge of connectivity in the human brain is far from complete. What will the relevant question or questions be when we do run these very large-scale brain simulations?