The first paper that discussed ways to relate neural modeling to functional brain imaging data was by Horwitz and Sporns (1994). The first actual study that compared simulated data generated by a biologically-based neural model to hemodynamic-based functional neuroimaging data was by Arbib et al. (1995), who used a large-scale model of saccade generation (Dominey & Arbib 1992) and adapted it to generate simulated PET data. Their model included a number of brain structures, such as posterior parietal and visual cortex, superior colliculus, the frontal eye field, the mediodorsal and lateral geniculate nuclei of the thalamus, and the caudate nucleus and substantia nigra of the basal ganglia. Because some of the pathways involving the basal ganglia are inhibitory, this model was a good testing ground for examining the effects of inhibitory synaptic activity on simulated blood flow. The main hypothesis tested in the PET simulation was that regional cerebral blood flow, as measured by PET, correlates with local synaptic activity. Although at the time there was some, but not very much, experimental support for this hypothesis (Jueptner & Weiller 1995), since the publication of the Arbib et al. study, the evidence has grown much stronger that the hemodynamic methods are indicative of synaptic and postsynaptic activity (e.g., Lauritzen 2001; Logothetis et al. 2001). One consequence of this notion is that increases in excitatory and inhibitory synaptic activity can lead to increased blood flow and metabolic activity (Logothetis 2003). In the Arbib et al. (1995) study, PET activation in the model was computed by summing the absolute values of both excitatory and inhibitory synaptic weights times firing rates of presynaptic neurons and integrating these values over a time period that corresponded to the time scale of PET while the model performed a specific task.
Computed PET activity was calculated by Arbib et al. (1995) during two different tasks (generating simple saccades and memory-driven saccades) and the differences between the two conditions were evaluated in all regions of the model. Memory driven saccades are generated in the model by activation of a memory loop between the frontal eye field and mediodorsal nucleus (MD) of the thalamus, which is disinhibited by the substantia nigra, generating a saccade (via the superior colliculus) to a remembered target when there is no stimulus present. When compared to the simple saccade task, in which disinhibition of the superior colliculus allows a saccade to be generated to a target present in the field of view, spiking activity in the modeled MD region increased as a result of the disinhibition, while simulated PET activation decreased in MD. This result showed how modeling can illuminate a counterintuitive effect: during the simple saccade, synaptic activity from the tonic inhibition of the MD contributes more synaptic activity to the PET measure than the increase in excitation that results from disinhibition.
As shown by the above study, the interpretive difficulty associated with synaptic inhibition provided a good example of how large-scale modeling can help interpret the activations observed in functional neuroimaging studies, even when they lead to results that appear counterintuitive. Because different parts of the brain have different neural architectures, and because the composition of the excitatory and inhibitory elements will be different in these various architectures, a number of separate modeling efforts will be needed to understand fully the role of inhibition in PET/fMRI activation patterns. For neocortex the inhibition theme was explored by Tagamets and Horwitz (2001) using a large-scale model of the ventral cortical visual processing stream (the details of this model will be discussed below). They used the same assumption as did Arbib et al. (1995) - that PET regional cerebral blood flow is indexed by the absolute value of the total synaptic activity in a brain region. They identified three factors that may play a role in how neural inhibition affects imaging results: (1) local connectivity; (2) context (e.g., the type of task being performed by the network); and (3) type of inhibitory connection.
Simulation results showed how the interaction among these three factors can explain seemingly contradictory experimental results. Specifically, the modeling indicated that neuronal inhibition can raise brain imaging measures if there is either low local excitatory recurrence or if the region is not otherwise being driven by excitation. On the other hand, with high recurrence or actively driven excitation, inhibition can lower observed neuroimaging values.
To summarize this section, we have shown that an important use for large-scale modeling is to help interpret how task-specific mixtures of excitatory and inhibitory neural activities result in the complex patterns of activations often seen in functional neuroimaging studies.
Was this article helpful?