The level of detail possible in network models developed with the current version of neuroConstruct is sufficient to reproduce most published conductance-based neuronal network models. However, there are a number of known biological phenomena which would be interesting to model, but which require extensions to the code to be fully supported. Many of these developments will go hand in hand with expansion of the scope of the NeuroML standards. We will outline what can currently be done with neuroConstruct and the approaches being taken to increase the functionality of the application.
We have shown that neuroConstruct can be used to construct models of various different brain regions, but it is not inconceivable that the specific anatomy of a particular brain region will bring up technical issues that require future software development. neuroConstruct currently provides the possibility to create spatial clusters of connections, local subcircuits (Song et al., 2005) and also 'small world' networks with a large number of local connections and sparse long-range connections (Watts and Strogatz, 1998). Anything that is currently beyond the capabilities of the GUI can be implemented by adding specialist NEURON or GENESIS code, or by importing networks that are represented by previously generated NetworkML files.
The scale of network models which can be built and visualized with the current version of neuroConstruct is limited by the physical memory of the machine on which it is running and the capabilities of the graphics card, respectively. It is currently possible to visualize and run networks of 50000 cells in a highly interconnected network on a machine with 256MB graphics memory and 8GB RAM. Larger networks can be simulated, but a network of this size is getting close to the maximum feasible on a single machine given the time it takes to execute the simulation.
We are currently investigating the feasibility of automatic generation of code for the parallel version of NEURON (Migliore et al., 2006), which is being used for network simulations in the Blue Brain Project (Markram, 2006). There are also a number of other initiatives to allow parallel execution of neuronal simulations, e.g. PGENESIS (Howell et al., 2000), GENESIS 3/MOOSE, Neurospaces (Cornelis and De Schutter, 2003) and any components added to support parallelization of NEURON (i.e. algorithms for partitioning the cells to computing nodes) could be reused for any supported parallel platform in the future. Support for specification of networks in a distributed environment will also be added to NetworkML, potentially allowing partitioned networks to be generated by other applications and imported into neuroConstruct.
Detailed signaling pathways neuroConstruct can be used to study the effect of pharmacological interventions on network behavior if the effect of a drug on single cell behavior is well characterized phenomenologically, such as the effect of benzodiazepines on GABAa receptor physiology. Such effects can be implemented in a network model by using a phenomenological approach, for example by representing the effect of benzodiazepines as slowed decay of GABAA receptor-mediated synaptic conductances, possibly with an increase in their peak amplitude. However, the mechanistic effect of drugs on the receptor kinetics is currently beyond the scope of neuroConstruct. Moreover, it is not yet possible to model the multitude of complex intracellular processes affecting cytoplasmic calcium dynamics, such as calcium induced calcium release or calcium and buffer diffusion in the presence of obstacles like endoplasmic reticulum (ER) cisternae or mitochondria. In the future, the integration of detailed biochemical models will be facilitated by the development of Level 4 of the NeuroML standards and by collaboration with initiatives in systems biology (Finkelstein et al., 2004; Kitano, 2002) such as SBML (Hucka et al., 2003) and CellML (Lloyd et al., 2004).
Representing neuronal network structure in 3D opens up the possibility to explore phenomena that are directly dependent on 3D spatial information. These phenomena include interactions between neurons that are mediated by diffusible substances, for example neurotransmitter spillover or communication through modulators like endocannabinoids or NO. Although it is already possible to create basic phenomenological models of diffusion of signaling molecules that affect synaptic plasticity in a distance dependent manner, more realistic diffusion models will require further development of neuroConstruct. The collaboration with systems biology initiatives should also open up the interaction with more detailed lower-level reaction-diffusion software packages like MCell (http://www.mcell.cnl.salk.edu) and VCell (http://www.nrcam.uchc.edu), although the difference in levels of description might require the development of separate but interacting models with different amounts of biological details.
The creation of a 3D network also opens up the possibility to generate different types of data that can be used to constrain and test the model. Extracellular electric fields could be simulated in 3D and these data could be compared with extracellular recordings in vivo and potentially used to link electroencephalogram (EEG) measurements to neuronal behavior. Vasculature could also be implemented and the modeled diffusion of metabolites and oxygen could be used for comparison with neuronal activity measurements resulting from functional magnetic resonance imaging (fMRI) studies. Moreover, simulating the 3D diffusion of oxygen and metabolites could form the basis of modeling studies into neurodegenerative states that are linked to an altered supply of these substances.
When constructing a network comprised of neuronal models with detailed morphologies, neuroConstruct currently assumes that all neurons of a given class have the same morphology. In reality, a biological network will contain a variety of morphologies and future developments of neuroConstruct will include the possibility to work with heterogeneous neuronal morphologies as have been used in applications like L-Neuron (Ascoli et al., 2001) and NeuGen (Eberhard et al., 2006). This will also allow the possibility to study the effect of variable morphologies on information processing in networks.
Was this article helpful?