Introduction

More than forty years following the initial report of the patient H. M. (Scoville and Milner, 1957), the role of the mammalian hippocampus in memory function remains the subject of debate (Eichenbaum, 1998). At the cellular level, investigation of the hippocampus has revealed a number of powerful computational elements including a specialized anatomical architecture (Amaral, 1993), regular spiking versus bursting pyramidal cells, diverse interneuron classes with distinct inputs and arborizations (Freund and Buzsáki, 1996), theta-band (4-10 Hz) oscillations (Vanderwolf, 1969), gamma-band (20­100 Hz) oscillations (Bragin et al., 1995), sharp waves, (Buzsáki et al., 1992), short-term synaptic plasticity (Zucker, 1989), long-term synaptic plasticity (Nicoll and Malenka, 1995; Bear and Abraham, 1996), and an assortment of neuromodulatory agents of intrinsic and subcortical origin. While the most recent data point to a critical role for the hippocampus in episodic as opposed to semantic memory (Vargha-Khadem et al., 1997), a central mystery remains: how do these cellular-level elements of the hippocampus interact to endow it with memory function? Even more importantly given the primary role played by the hippocampus in Alzheimerąs disease (AD), epilepsy and traumatic brain injury, how do specific perturbations at the cellular or subcellular level lead to the cognitive-level, clinical manifestations of disease?

With respect to both cortical function and dysfunction, the role of neuromodulators such as acetylcholine (ACh), norepinephrine (NE), serotonin (5-HT, 5-hydroxytryptamine), and dopamine (DA) are of particular concern since they exert subtle but significant effects on both single-cell physiology and network function by altering ionic conductances, membrane properties, and synaptic transmission (see Hasselmo, 1995 for a recent review). As a result, neuromodulators can endow subcortical nuclei with the ability to regulate information processing, behavioral states, and the sleep-wake cycle. Perturbations of these neuromodulatory systems, as occurs in AD and other diseases, is likely to have serious consequences for cognitive function but may also represent a promising avenue for therapeutic intervention.

Here we revisit the "cholinergic hypothesis" of Alzheimerąs disease via a computational investigation of the neuromodulation of single cells and networks by acetylcholine in the CA3 region of the hippocampus. We believe that the results provide insight into the regulatory role of acetylcholine in learning and recall and suggest novel mechanisms for the decline in memory function that accompanies AD.

The hippocampal formation is one of the first and most drastically damaged brain regions in Alzheimerąs disease as judged by the appearance of neuritic plaques and neurofibrillary tangles (Price, 1986; Katzman, 1986; Hyman et al., 1990; Selkoe, 1993). While these pathological markers are used to make the diagnosis of AD, an early and more insidious manifestation of the disease is the neuromodulatory denervation that accompanies the death and/or dysfunction of specific subcortical neurons. Often the early memory impairment in AD is attributed to cell death from the presumed neurotoxicity of plaques and/or tangles, particularly those occurring in hippocampal region CA1, the adjacent subiculum, and entorhinal cortex (Hyman et al., 1990; Arnold et al., 1991; Braak and Braak, 1991). This view, however, runs counter to theoretical considerations and computational studies which suggest that a significant number of neurons must die before memory function is seriously impaired (Ruppin and Reggia, 1995), far more than the 10% seen in the latest stages of AD (Katzman, 1986). In contrast to cell death or synaptic perturbations, the removal of neuromodulatory control of hippocampal neurons is likely to have a major impact on memory function.

The "cholinergic hypothesis" of AD initially arose from the observed AD-related decreases in the activity of choline acetyltransferase (ChAT), the enzyme responsible for synthesizing acetylcholine. This observation was made in post-mortem tissue (Bowen et al., 1976; Bowen et al., 1977b; Bowen et al., 1977a; Davies and Maloney, 1976; Perry et al., 1977) but has since been confirmed in ante-mortem biopsies (Palmer et al., 1987c). This theory, tying low levels of acetylcholine to cognitive impairment, gained support from the observed decreases in choline uptake (Rylett et al., 1983), decreases in ACh release (Nilsson et al., 1986), the death of subcortical cholinergic neurons (Whitehouse et al., 1982; Arendt et al., 1983) and the disappearance of cholinergic varicosities in the early stages of AD (Bowen et al., 1982). This hypothesis led to the recent clinical trials of a variety of acetylcholinesterase (AChE) inhibitors, currently the first-line pharmacological approach to treating AD.

The limited efficacy of these agents, however, in addition to the intense research efforts directed at the amyloid b-peptide, the amyloid precursor protein (APP), the tau cytoskeletal protein, the presenilin genes, and apolipoprotein E4, has caused the cholinergic hypothesis to fall somewhat out of favor. This has occurred even though little is yet known as to how acetylcholine and other neuromodulators regulate cortical cells and networks. In fact, the AD-related deaths of noradrenergic neurons in the locus coeruleus (Mann et al., 1984), serotonergic neurons in the dorsal raphé nucleus, the presence of neurofibrillary tangles in the surviving serotonergic cells (Curcio and Kemper, 1984; Mann et al., 1984), and decreased levels of norepinephrine and serotonin in the early stages of AD with a sparing of dopaminergic systems (Palmer et al., 1987a; Palmer et al., 1987c; Palmer et al., 1987b) all point to a complex perturbation of neuromodulatory centers. Such an intricate regulatory system is unlikely to be repaired by the global pharmacological approach that is currently the best available therapy. A primary goal of our studies has been to examine neuromodulation in detailed cellular-level models of hippocampal cells and networks in the hope of elucidating the regulatory roles of neuromodulators in memory function, how their depletion in disease leads to clinical manifestations, and hopefully novel approaches to therapy.

Cholinergic input to the mammalian hippocampus arises subcortically from the medial septal nuclei and the vertical nucleus of the diagonal band of Broca (Mesulam et al., 1983; Amaral and Cowan, 1980). This input is strongly correlated with the sleep-wake cycle such that hippocampal levels of ACh are significantly higher during waking and the paradoxical stage of sleep than during slow-wave sleep (Kametani and Kawamura, 1990). The neuromodulatory effects of ACh are attributed to a variety of muscarinic receptor subtypes (m1 through m5) acting upon potassium, calcium, and mixed cation channels some of which remain incompletely characterized. The classical effect of ACh is the inhibition of the "muscarinic" non-inactivating K⁺ current I_M (Brown and Adams, 1980; Halliwell and Adams, 1982), however this is among the least sensitive of its effects. It is now known that muscarinic activation also decreases the slow calcium-activated K⁺ afterhyperpolarizing current I_AHP (Cole and Nicoll, 1983; Cole and Nicoll, 1984a; Cole and Nicoll, 1984b; Madison et al., 1987),, decreases the fast transient repolarizing K⁺ current I_A (Nakajima et al., 1986), and decreases a resting K⁺ current (Madison et al., 1987; Benson et al., 1988; Brown et al., 1988). Several studies have described muscarinic inhibition of calcium currents (Gahwiler and Brown, 1987) where ACh appears to inhibit high-threshold channels and excite low-threshold channels (Toselli and Lux, 1989; Fisher and Johnston, 1990). In the case of high-threshold channels, it appears that both the L- and N-type high-voltage activated (HVA) channels demonstrate muscarinic inhibition (Toselli and Taglietti, 1994; Fisher and Johnston, 1990) while the low-voltage activated (LVA) T-type channel demonstrates an increased probability of opening (Fisher and Johnston, 1990). In addition to these ion-selective currents, the hyperpolarization-activated mixed cationic current I_h (also referred to as I_Q by some investigators) has been reported to be potentiated by muscarinic agonists (Colino and Halliwell, 1993). Although these results have all been obtained in hippocampal pyramidal cells, it has also been reported that ACh has similar effects on the array of ionic conductances in interneurons (Freund and Buzsáki, 1996).

The overall effect of the above muscarinic actions on individual ionic currents forms the basis for the generally "excitatory" physiological effects attributed to ACh. In pyramidal cells, ACh causes a slow depolarization due to the suppression of the above-mentioned tonically active K⁺ current (Madison et al., 1987). ACh also suppresses spike frequency adaptation in pyramidal cells due to decreases in calcium-dependent K⁺ currents and I_M (Madison et al., 1987). In hippocampal interneurons, ACh causes a rapid excitation of the cell via modulation of a K⁺ conductance (Reece and Schwartzkroin, 1991).

In addition to cellular-level effects, several studies have examined the synaptic effects of cholinergic neuromodulation as well. ACh has been reported to cause the presynaptic suppression of excitatory synaptic transmission (Hounsgaard, 1978; Valentino and Dingeldine, 1981) with a stronger effect in the stratum radiatum (s. r.) of the hippocampus than in the stratum lacunosum-moleculare (s. l.-m.) (Hasselmo and Schnell, 1994). This effect has led Hasselmo and colleagues to propose a role for ACh in the switching between learning and recall in hippocampal (Hasselmo et al., 1995) and olfactory networks (Barkai et al., 1994; Hasselmo and Barkai, 1995; Liljenstrom and Hasselmo, 1995). While these results suggest a suppressive synaptic effect for ACh, interestingly, muscarinic activation has been reported to enhance the NMDA-mediated component of EPSPs in CA1 pyramidal cells (Markram and Segal, 1990). Finally, muscarinic activation has also been reported to yield decreased mIPSCs most likely via a high density of presynaptic m2 receptors on terminals of those interneurons innervating the somata of pyramidal cells (Freund and Buzsáki, 1996).

In essence, there is a wealth of in vitro data regarding cholinergic modulation of single cells; what is required is a functional interpretation of the role of these processes in cellular and memory function in vivo. While this is a difficult task from a purely analytical standpoint, computational modeling of cells and networks is a valuable tool in using biological data to test the validity of various hypotheses.

To begin to understand how the cellular components of the hippocampus endow it with memory function, what control systems regulate this ability, and how specific deficits can lead to disease, we have begun investigating detailed biophysical models of single cells as well an anatomically-inspired network model of the CA3 region of the hippocampus comprised of these cellular models. While important insights arise from simpler, analytically tractable network models (Horn et al., 1993; Horn et al., 1996; Ruppin et al., 1996a; Ruppin et al., 1996b; Ruppin and Reggia, 1995), we have several motivations for taking the biophysical approach. First, in order to understand or predict the functional consequences of pathology occurring at the cellular or subcellular level, detailed models are required. Not all pathology leads to cell death nor to synaptic perturbations, those pathologies most amenable to neural network analysis. The level of detail provided by biophysical models is particularly important for the investigation of neuromodulators and disease processes where the effects are at the level of ion channels, G protein-coupled receptors, membrane properties and intracellular biochemistry. Second, it is most likely that therapeutic innovations will require the targeting of the cellular and subcellular foundations of the disease; an understanding of these processes and how they can be manipulated is critical. Third, biophysical-level network simulations are the best way of linking more abstract neural network theories to experimental data.

Our simulations of both single cell models and networks were constructed using the GENESIS development package (Bower and Beeman, 1994) and PGENESIS, its recent implementation for parallel platforms (Goddard and Hood, 1998). Simulations were performed on a four-processor Silicon Graphics Origin2000. Differential equations were solved using Crank-Nicholson implicit integration with a step size of 25us. This value was determined to be the largest step size that allowed for accurate convergence.

Full Text:
 PostScript
 Acrobat