A computational perspective on Neuromodulation (in progress)
Neuromodulations can be characterized by their spatial and temporal characteristics, and in the computational framework chosen here, can also be characterized by its level of coupling with the specific neural computations under consideration.
1. Extrinsic and Intrinsic Neuromodulation
A first class of neuromodulatory signals may originate from an area extrinsic to the neural substrate whose computation is under study, so that lesioning the neuromodulatory center does not usually perturb the function itself, but only modifies its ‘quality’. The computational functions of such extrinsic neuromodulation are expected to be somewhat global, because they usually influence many functionally different sites, simultaneously. A second class consists in neuromodulations that either originate in the relevant substrate itself or originate in a distant site, but are controlled locally within the substrate. In such systems, neuromodulation is an integral part of the computation. Co-transmission (Brezina, Orekhova, & Weiss, 1996; Chan-Palay & Palay, 1984; Kupfermann, 1991; Marder, Christie, & Kilman, 1995), pre-synaptic receptions (Marder, 1996; Starke, Gothert, & Kilbinger, 1989), glial modulation (Hansson & Ronnback, 1994) or volume transmission (Fuxe & Agnati, 1991; Ridet, Rajaofetra, Teilhac, Geffard, & Privat, 1993) are examples of such phenomena. The functions of such intrinsic modulations are more specific to the substrate under consideration (Katz & Frost, 1996).
1.1 Extrinsic Neuromodulation:
In many models the origin of the modulation is known, but does not depend, in general, on the computation of the substrate being modulated. It rather depends on the parallel activity of functionally distinct systems, extrinsic to the substrate. Such is the case of most ‘neuromodulatory centers’ releasing specific neuroactive substances that modify the cellular and synaptic properties of their targets. Most of the actions of dopamine (Cooper, 1991) or norepinephrine (van Dongen, 1981) enter in this category. Here, we illustrate this point with a recent model of sequence learning in hippocampal region CA3 showing that computations may crucially depend on the extrinsic modulation by GABAergic and cholinergic inputs from the septum (Wallenstein & Hasselmo, 1997b). Other models considered extrinsic modulation as a signal influencing synaptic mechanisms. Such is the case of the reward signal entering the weight modification rule, between VTA and cortex (Montague, Dayan, & Sejnowski, 1996) or the direct change of synaptic efficacy triggered by an external center (Linster & Gervais, 1996; Linster & Masson, 1996; Raymond, Baxter, Buonomano, & Byrne, 1992).
1.2 Intrinsic Neuromodulation:
In some instances, it is not possible to isolate the neuromodulatory phenomenon from the system it modulates. In such cases, neuromodulation is intrinsic to the network whose computation is under study. Some experimental evidence for intrinsic neuromodulation are reviewed elsewhere (Katz & Frost 1996), we will briefly mention two examples below.
In the stomatogastric ganglion (STG) of the lobster, an afferent axon (SNAX1) has been characterized as both a participant in the rhythmicity of the gastric mill network, and as a conveyor of modulatory information (Nusbaum, Weimann, Golowasch, & Marder, 1992). SNAX1 receives (inhibitory) synaptic inputs from the STG and is capable of initiating action potentials (intrinsically, within the STG and not near the cell body, few centimeters away) which generates EPSPs on the STG elements, participating therefore in the generation of the rhythm. However, because SNAX1 is also electrically coupled with key neurons of the central pattern generator, its level of depolarization (whether or not action potentials are present) modulates the activity of the network. Similarly, in the tritonia, DSI (a serotonergic CPG neuron) is known to enhance synaptic transmission presynaptically at synapses made by a key CPG neuron (Katz & Frost, 1995b; Katz & Frost, 1996; Katz, Getting, & Frost, 1994). DSI elicits both a fast, neurotransmitter-like EPSP, and a slow neuromodulatory-like EPSP (Katz & Frost, 1995a), both pharmacologically separable. DSI, therefore, modulates the oscillatory pattern it is contributing to.
It is of course possible to envision dual extrinsic and intrinsic neuromodulations, whereby the former would express state or stimulus-dependency, and the later would be activity-dependent. In the computational framework of modeling studies, extrinsic neuromodulations can be easily implemented by choosing appropriate sets of parameters (tuning), whereas intrinsic neuromodulations require that the neuromodulatory mechanisms be regulated by the computations under consideration.