Neuron simulation environment
|Ted Carnevale (2007), Scholarpedia, 2(6):1378.||doi:10.4249/scholarpedia.1378||revision #154131 [link to/cite this article]|
The NEURON Simulation Environment (aka NEURON --see http://www.neuron.yale.edu/) is designed for modeling individual neurons and networks of neurons, and is widely used by experimental and theoretical neuroscientists. It provides tools for conveniently building, managing, and using models that are numerically sound and computationally efficient. NEURON is particularly well-suited to problems that are closely linked to experimental data, especially those that involve cells with complex anatomical and biophysical properties.
NEURON began in the laboratory of John W. Moore at Duke University, where he and Michael Hines started their collaboration to develop simulation software for neuroscience research. It has benefited from judicious revision and selective enhancement, guided by feedback from the growing number of neuroscientists who have used it to incorporate empirically-based modeling into their research strategies. Most papers that report work done with NEURON have addressed the operation and functional consequences of mechanistic models of biological neurons and networks. Readers who wish to see specific examples are encouraged to peruse the online bibliography. Working code for many published NEURON models can be downloaded from ModelDB.
Advantages over general-purpose simulators
NEURON's computational engine employs special algorithms that achieve high efficiency by taking advantage of the structure of the equations that describe neuronal properties (for examples see (Hines, 1984), discussion of staggered Crank-Nicholson method in chapter 4 of The NEURON Book (Carnevale and Hines, 2006)). It has functions that are tailored for conveniently controlling simulations, and presenting the results of real neurophysiological problems graphically in ways that are quickly and intuitively grasped. Instead of forcing users to reformulate their conceptual models to fit the requirements of a general purpose simulator, NEURON is designed to let them deal directly with concepts that are familiar at the neuroscience level. Consequently, users can think in terms of the biophysical properties of membrane and cytoplasm, the branched architecture of neurons, and the effects of synaptic communication between cells.
Separates biology from purely computational concerns
One of NEURON's design goals has been to help modelers focus on high-level neuroscience research questions without being distracted by low-level mathematical or computational issues. Here are some of the strategies that help it approach this goal.
Working with familiar idioms
The most important strategy is to offer users what might be called "natural syntax", which allows them to specify model properties in familiar idioms, instead of laboriously recasting kinetic schemes or differential equations into C or some other generic programming language. Perhaps the most noticeable example of natural syntax in NEURON is the notion of a section, which is a continuous unbranched cable (directly analogous to an unbranched neurite). Sections can be connected together to form branched trees, and endowed with properties that can vary continuously with position along their lengths. Sections let investigators represent neuronal anatomy without having to wrestle with the cable equation. Sections also easily lend themselves to manipulation by graphical tools for building and managing models, such as the CellBuilder. Other examples of natural syntax are seen in the NMODL programming language, and in the Linear Circuit Builder and Channel Builder (see User-extendable library of biophysical mechanisms below).
Integrator-independent model specification
NEURON offers several different numerical integration methods that users can select according to their needs.
- The default integration method is implicit Euler, which provides robust stability and first order accuracy in time (sufficient for most applications).
- There is also a Crank-Nicholson method that provides second order accuracy at little additional computational cost. However, this is prone to numerical oscillations if dt is too long, voltage clamps are present, or system states are described by algebraic equations.
- Adaptive integration may achieve increased accuracy, faster run times, and sometimes both. It does this by adjusting integration order and time step as necessary to satisfy a local error criterion.
NEURON automatically selects one of two adaptive integration methods, depending on whether or not a model involves states that are described by algebraic equations: IDA (Hindmarsh and Taylor, 1999) or CVODEs (Hindmarsh and Serban, 2002).
Users can switch between these integration methods without having to rewrite the model specification because NEURON avoids computation-specific representations of biological properties. This convenience is essential because deciding which method is best in any particular situation is often an empirical question. Further details about numeric integration in NEURON are provided in chapter 4 of The NEURON Book (Carnevale and Hines, 2006).
Efficient and painless spatial and temporal discretization
Another important strategy is to allow users to defer or avoid the explicit specification of spatial and temporal discretization. NEURON's approach to spatial discretization, which sets it apart from other modeling tools, eliminates preoccupation with compartments. This approach has two components:
- Each section has its own discretization parameter nseg, which governs the number of internal points at which the discretized form of the cable equation is integrated, and can be changed after the anatomical and biophysical properties of the model have been specified.
- Parameters and variables that are spatially nonuniform are described in terms of normalized distance from one end of the section, e.g. soma.v(0.5) means the membrane potential midway between the 0 and 1 ends of the soma section.
In addition, NEURON provides tools that can automatically take care of discretization. Numerical simulation of models of biological neurons can pose a special challenge because such models tend to include mechanisms that span a wide range of spatial and temporal scales, with state variables that involve a wide range of magnitudes. For spatial discretization, the CellBuilder offers users the d_lambda rule ((Hines and Carnevale 2001), p. 122 et seq. in (Carnevale and Hines 2006)), which selects values of nseg based on the model's anatomical and biophysical properties. The d_lambda rule can also be applied via hoc code. Similarly, the error tolerances that govern time step size (and integration order) during adaptive integration can be adjusted automatically by the GUI's Absolute Tolerance Scale Tool, which is part of the VariableTimeStep tool.
Convenient user interface
Most NEURON programming has been done with hoc, an interpreted language with C-like syntax (Kernighan and Pike, 1984) that has been extended to include functions specific to the domain of modeling neurons, implementing a graphical interface, and object-oriented programming. This provides a tremendous degree of flexibility for setting up the anatomical and biophysical properties of the model, defining the appearance of the graphical interface, controlling simulations, and plotting results.
Beginning in 2006, Python was added as an alternative interpreter. This feature is now part of the standard distribution, so that programs can be developed entirely in Python, or in a combination of Python and hoc in which hoc variables, objects, etc. can be accessed from Python or vice versa (Hines et al., 2009).
With either hoc or Python, one can take advantage of multiprocessors or workstation clusters in order to accelerate embarassingly parallel problems, such as optimization and parameter space exploration. Furthermore, network models, and complex models of a single neuron, can be distributed over multiple processors to achieve nearly linear speedup (Migliore et al., 2006).
The default graphical user interface can be used to create and exercise models that have a wide range of complexity. By using the GUI, it is possible to generate publication-quality results without having to write any program code at all. However, the most flexible and powerful use of NEURON is to combine hoc programming and the GUI in order to exploit the strengths of both approaches.
Many GUI tools are designed to be analogous to biology and lab instrumentation, and are very convenient for initial exploratory simulations, setting parameters, common control of voltage and current stimuli, and graphing variables as functions of time and position. Perhaps the most noteworthy of the model specification tools are
- The Channel Builder, which is discussed below (see User-extendable library of biophysical mechanisms).
- The CellBuilder, which can be used to create new models of cells from scratch and modify existing models without writing any code. Biophysical properties may be constant over the length of each section, or governed by user-specified functions of position according to radial distance from a point, path distance from a reference location in the cell, or distance along an axis in the xy plane.
- The Import3D tool, which can convert morphometric data from Eutectic, Neurolucida, and SWC formats into model cells. It fixes many common errors automatically, and helps pinpoint complex problems that require judgment.
- The Linear Circuit Builder, a tool for setting up models that involve gap junctions, ephaptic interactions, dual-electrode voltage clamps, dynamic clamps, or other combinations of neurons and electrical circuit elements.
- The Network Builder, for prototyping small networks that can be mined later for reusable code in order to create large-scale networks.
There are also powerful tools for model analysis and optimization, e.g.
- The ModelView tool automatically discovers model properties and presents them in a browsable textual and graphical form. This can be very helpful for code development and maintenance. It is growing increasingly important as code sharing becomes more common, not only because it helps users understand each others' models, but also because it can emit and import model specifications in XML, thereby facilitating interoperability with other simulators.
- The Impedance tools perform electrotonic analyses of a model cell, including input and transfer impedance, voltage transfer ratios, and the electrotonic transformation.
- The Multiple Run Fitter facilitates setting up and managing optimization protocols for automated tuning of model parameters that are constrained by data from one or more experimental protocols.
User-extendable library of biophysical mechanisms
User-defined mechanisms such as voltage- and ligand-gated ion channels, diffusion, buffering, active transport, etc., can be added by writing model descriptions in NMODL, a high-level programming language that has a simple syntax for expressing kinetic schemes and sets of simultaneous algebraic and/or differential equations. NMODL can also be used to write model descriptions for new classes of artificial spiking cells. These model descriptions are compiled so that membrane voltage and states can be computed efficiently using integration methods that have been optimized for branched structures. A large number of mechanisms written in NMODL have been made available on the WWW by the authors of published models; many of these have been entered into ModelDB which makes it easy for users to find and retrieve model source code according to search criteria such as author, type of model (e.g. cell or network), ionic currents, etc..
As mentioned above, NEURON also has a GUI tool called the ChannelBuilder that makes it easy to specify voltage- and ligand-gated ion channels in terms of ODEs (HH-style, including Borg-Graham formulation) and/or kinetic schemes. ChannelBuilder channels actually execute faster than identical mechanisms specified with NMODL. Their states and total conductance can be simulated as deterministic (continuous in time), or stochastic (countably many channels with independent state transitions, producing random, abrupt conductance changes).
Enhancements for modeling networks
Although NEURON began in the domain of single-cell models, since the early 1990s it has been applied to network models that contain large numbers of cells and connections. Its suitability for network models has been enhanced by the introduction of features that include
- an event delivery system that manages delivery of spike events to synaptic targets, and has enabled the implementation of artificial spiking cells that are several orders of magnitude faster than biophysical model cells and can be used in network models that may also include biophysical model cells
- a Network Connection object class, which has greatly simplified the construction and management of networks
- generalized synapses which facilitate computationally efficient implementation of synaptic convergence, while allowing stream-specific synaptic plasticity governed by user-specified rules (e.g. use-dependent and/or spike-timing dependent plasticity)
- global and local variable-order, variable-stepsize integration methods, which can substantially shorten runtimes for some models
- ability to implement network models that are distributed across processors in a parallel computational architecture (Migliore et al., 2006). Properly written code will work on a single processor standalone PC or parallel hardware without modification. Networks can be load balanced by splitting cells into two or more subtrees at any point on each cell, and placing these pieces on different hosts; this does not degrade stability and accuracy, or increase computational complexity.
The process of setting up a large network with NEURON typically involves writing hoc or Python scripts with multiple "for" loops to position cells in space, establish the desired connections, and to set synaptic weights and propagation delays. At present, NEURON does not have single commands for setting up connections, weights, and delays for an assembly of cells in a network without using loops.
Currently the recommended practice is to use hoc to define the cell classes, and to use Python for everything else, e.g. setting up the network, specifying instrumentation (stimuli, exogenous afferent spike trains, recording of spike times and other key variables), implementing simulation flow control (e.g. "experimental protocols"). This may be somewhat simplified by using the Channel Builder and Cell Builder to create and save template cells to hoc files, and the Network Builder to manually create a small prototype network, which is also saved to a hoc file. Code (especially the cell class definitions) may then be extracted from the file for the prototype network model and modified to create the larger network.
Customizable initialization and simulation flow control
Simulation control in NEURON employs a standard run system that is designed for convenient customization (see chapters 7 and 8 of (Carnevale and Hines, 2006)). This is very helpful for dealing with complex models, which often require custom initialization and/or simulation flow control. For example, custom initialization can be used to ensure that resting potential is uniform in a model cell with inhomogeneous membrane properties, or to launch simulations of a spontaneously active model from a particular point on a dynamic trajectory. Custom simulation flow control has many specialized uses, e.g. to mimic an experimental protocol by perturbing one or more parameters in the course of a simulation, or to interrupt simulation execution periodically or conditionally for one purpose or another.
Large user base
As of mid 2008, more than 750 scientific articles and books have reported work that was done with NEURON (see online bibliography). The NEURON mailing list has more than 600 subscribers, and The NEURON Forum, which was started in May 2005 and is gradually supplanting the old mailing list, has more than 600 registered users and 5000 posted messages.
Development, support, and documentation
NEURON is actively developed and supported, with new standard releases appearing about twice a year, supplemented by bug fixes as needed. Alpha versions that contain "the very latest features" are made available at shorter intervals.
Support is available by email, telephone, and consultation. Users can also post questions and share information with other members of the NEURON community through a mailing list and The NEURON Forum. In addition to being a browsable and searchable venue for discussions on specific questions, the Forum contains a growing list of tips, how-tos, and hacks of interest to users at all levels of expertise.
Extensive documentation is available via links at http://www.neuron.yale.edu/neuron/docs/docs.html. This includes an on-line programmer's reference, FAQ list, tutorials, and preprints of key articles about NEURON. The authoritative book on NEURON is The NEURON Book (Carnevale and Hines, 2006). Four other books have made extensive use of NEURON (Destexhe and Sejnowski, 2001; Johnston and Wu, 1995; Lytton, 2002; Moore and Stuart, 2000), and several of these include source code or have made it available online.
Since 1997, one day courses on NEURON have been presented as a satellite to the annual meetings of the Society for Neuroscience, and intensive five day hands-on summer courses have been held at the University of California in San Diego and other locations; future courses are announced at http://www.neuron.yale.edu/neuron/courses. Special seminars and tutorials have been presented at The NEURON Simulator Meeting, an occasional gathering of neuroscience investigators and educators who are interested in using computational modeling in their work. Instruction and consultation on NEURON have also been provided at the European Advanced Course in Computational Neuroscience.
Availability and system requirements
NEURON is distributed free of charge from http://www.neuron.yale.edu/. It runs on all popular hardware platforms under MSWin (98 or later), UNIX, Linux, and OS X, and on parallel hardware including Beowulf clusters, the IBM Blue Gene, and the Cray XT3.
- Carnevale, N.T. and Hines, M.L. The NEURON Book. Cambridge, UK: Cambridge University Press, 2006.
- Destexhe, A. and Sejnowski, T.J. Thalamocortical Assemblies. New York: Oxford University Press, 2001.
- Hindmarsh, A.C. and Serban, R. User documentation for CVODES, an ODE solver with sensitivity analysis capabilities. Lawrence Livermore National Laboratory, 2002.
- Hindmarsh, A.C. and Taylor, A.G. User documentation for IDA, a differential-algebraic solver for sequential and parallel computers. Lawrence Livermore National Laboratory, 1999.
- Hines, M. Efficient computation of branched nerve equations. International Journal of Bio-Medical Computing 15:69-76, 1984.
- Hines, M.L. and Carnevale, N.T. NEURON: a tool for neuroscientists. The Neuroscientist 7:123-135, 2001.
- Hines, M.L., Davison, A.P. and Muller, E. NEURON and Python. Frontiers in Neuroinformatics 3:1. doi:10.3389/neuro.11.001.2009. http://frontiersin.org/neuroinformatics/paper/10.3389/neuro.11/001.2009/
- Johnston, D. and Wu, S.M.-S. Foundations of Cellular Neurophysiology. Cambridge,MA: MIT Press, 1995.
- Kernighan, B.W. and Pike, R. Appendix 2: Hoc manual. In: The Unix Programming Environment. Englewood Cliffs, NJ: Prentice-Hall, 1984.
- Lytton, W.W. From Computer to Brain. New York: Springer-Verlag, 2002.
- Migliore, M., Cannia, C., Lytton, W.W., Markram, H. and Hines, M.L. Parallel network simulations with NEURON. Journal of Computational Neuroscience 21:110-119, 2006.
- Moore, J.W. and Stuart, A.E. Neurons in Action Version 2: Tutorials and Simulations Using NEURON. Sunderland, MA: Sinauer Associates, 2007. http://neuronsinaction.com/
NEURON source code, installers, documentation, bibliography, course announcements, and The NEURON Forum are available via links at http://www.neuron.yale.edu/
- Olaf Sporns (2007) Complexity. Scholarpedia, 2(10):1623.
- James Meiss (2007) Dynamical systems. Scholarpedia, 2(2):1629.
- James M. Bower and David Beeman (2007) GENESIS. Scholarpedia, 2(3):1383.
- Mark Aronoff (2007) Language. Scholarpedia, 2(5):3175.
- Marc-Oliver Gewaltig and Markus Diesmann (2007) NEST (NEural Simulation Tool). Scholarpedia, 2(4):1430.
- Jeff Moehlis, Kresimir Josic, Eric T. Shea-Brown (2006) Periodic orbit. Scholarpedia, 1(7):1358.
- Philip Holmes and Eric T. Shea-Brown (2006) Stability. Scholarpedia, 1(10):1838.
- Alan C. Hindmarsh and Radu Serban (2007) Sundials equation solvers. Scholarpedia, 2(3):2860.
- John W. Moore (2007) Voltage clamp. Scholarpedia, 2(9):3060.
- Bard Ermentrout (2007) XPPAUT. Scholarpedia, 2(1):1399.