DOES THE SPREAD OF ALZHEIMER'S DISEASE NEUROPATHOLOGY
INVOLVE THE MECHANISMS OF CONSOLIDATION?
MICHAEL E. HASSELMO and BRADLEY P. WYBLE
Dept. of Psychology and Program in Neuroscience
Harvard University, 33 Kirkland St., Cambridge, MA 02138
E-mail: hasselmo@katla.harvard.edu
A model of the hippocampus allows analysis of the role of network dynamics in the initiation and progression of neuropathology in Alzheimers disease. The model is neutral with respect to etiology, focusing on a final common breakdown in function termed runaway synaptic modification. This phenomenon could account for evidence showing that the neurofibrillary tangles associated with Alzheimers disease first appear and attain their highest concentration in subregions of the hippocampal formation, then successively spread into temporal lobe cortex and the cortex of the frontal and parietal lobes. The model demonstrates how the spread of neuropathology from the hippocampus into neocortical structures could result from the mechanisms of consolidation. Initial sensitivity of the hippocampus and entorhinal cortex to the development of neurofibrillary tangles is proposed to result from an imbalance of parameters regulating the influence of synaptic transmission on synaptic modification. Degeneration of cortical cholinergic innervation is proposed to result from exponentially increased demands on the feedback regulation of cholinergic modulation. Increased levels of amyloid are attributed to exponential increases in the modification and maintenance of synaptic connections, while development of paired helical filaments is ascribed to exponentially increased demands on the mechanisms of axonal transport or remodeling. Memory deficits are described as due to increased interference effects in recent memory caused by runaway synaptic modification which ultimately progresses to cause impairments of remote memory and semantic memory.
In the popular consciousness, Alzheimer's disease is identified as a disorder of memory function. While research on Alzheimer's disease has produced a range of etiological theories, ranging from the improper splicing of the amyloid precursor protein (Selkoe 1993) to epidemiological factors such as aluminum exposure (Crapper McLachlan and Van Berkum 1986) or prions (Goudsmit and Van der Waals 1986), these etiological theories have not explicitly accounted for why this disorder should show its earliest symptoms as a disorder of memory function (Jolles 1986; Morris and Kopelman 1986; Albert et al. 1991), and should so severely affect those structures associated with memory function (Arnold et al. 1991; Arriagada et al. 1992; Hyman et al. 1984; 1990). As it stands, the early effect on memory function is attributed to the unexplained specificity of Alzheimer's disease for hippocampal region CA1, the subiculum, and layers II and IV of entorhinal cortex (Hyman et al. 1984, 1990; Arnold et al. 1991; Braak and Braak 1991). Here it will be proposed that the causality is in fact reversed. The selective cortical neuropathology associated with the progression of this disorder may be rooted in the breakdown of the essential mechanisms of cortical memory function.
This paper presents a computational theory of the initiation and progression of Alzheimer's disease which attempts to account for evidence on the progression of Alzheimer's disease not in terms of a specific etiological factor, but in terms of the processing characteristics of cortical structures, and the stability of the learning mechanisms within these structures. This theory was inspired by the phenomenon of runaway synaptic modification, as demonstrated in models of cortical associative memory function (Hasselmo et al. 1992; Hasselmo 1993a, 1994; Hasselmo and Bower 1993; Barkai et al. 1994; Hasselmo and Barkai, 1995). In these models, interference during learning can lead to the exponential growth of a large number of synaptic connections within the network. Runaway synaptic modification of this sort may underlie the neuropathological characteristics of Alzheimer's disease. The theory provides a framework showing why this neuropathology should appear initially in particular cortical regions associated with memory function (Braak and Braak 1991; Arriagada et al. 1992), and why it should appear to progress into adjacent regions of association cortex along the observed anatomical connections (Pearson et al. 1985; Arnold et al. 1991; Arriagada et al. 1992). Finally, the theory suggests that the apparent degeneration of cortical cholinergic innervation in this disease (Davies and Maloney 1976; Perry et al. 1977; Whitehouse et al. 1982; Coyle et al. 1983; Saper et al. 1985) may result from feedback mechanisms placing too great a demand on the cholinergic modulation of learning processes.
In terms of specific causative factors, this theory is neutral in many respects. The phenomenon of runaway synaptic modification could be initiated by a variety of changes in the parameters of cortical function, due to either a genetic predisposition or environmental influences. The initial appearance of runaway synaptic modification is attributed to an imbalance of cortical parameters, which could result from insufficient normalization of synaptic strength within each neuron, from a lowering of the threshold of synaptic modification, from insufficient feedback inhibition, from an imbalance of the sensitivity of presynaptic and post-synaptic cholinergic receptors, or from direct overload of the information capacity of cortical structures.
2.1. Selective distribution and progression of neuropathology
A vast range of research on Alzheimer's disease has provided a body of empirical evidence which must be explained by any theory of the progression of Alzheimer's disease. Alzheimer's disease is diagnosed post-mortem on the basis the density of the neuropathological characteristics including neuritic plaques and neurofibrillary tangles (for review, see Price 1986; Katzman 1986; Hyman et al. 1990; Selkoe 1993). Neuritic plaques tend to be broadly distributed, appearing throughout the cortex, with a greater density in regions of frontal, parietal and temporal association cortex distant from the primary sensory and motor cortices (Pearson et al. 1985; Arnold et al. 1991). In addition, neuritic plaques appear in subcortical regions receiving projections from the cortex (Pearson et al. 1985). The distribution and component features of neuritic plaques have led to the suggestion that they reflect the degeneration of axonal processes from the same set of neurons which develop neurofibrillary tangles (Hyman et al. 1984). Here it is proposed that neuritic plaques result from a breakdown in the normal mechanisms for modification of synaptic strength.
Neurofibrillary tangles show a more localized initial distribution than plaques and spread in a characteristic sequence (Hyman et al. 1990; Arnold et al. 1991; Braak and Braak 1991). Tangles appear initially and attain their highest concentration in layer II of entorhinal cortex, region CA1 of the hippocampus, and the portions of the subiculum adjacent to region CA1 (Ball 1972; Hyman et al. 1984; Hyman et al. 1990; Arriagada et al. 1992; Braak & Braak 1991). As the severity of the disease progresses, tangles appear in regions receiving projections from these areas, initially in portions of the temporal lobe adjacent to the entorhinal cortex, and later in regions of parietal lobe and frontal cortex which are closely linked to entorhinal cortex (Brun & Gustafson 1976; Pearson et al. 1985; Arnold et al. 1991; Arriagada et al. 1992; Braak & Braak 1991). Tangles appear to be distributed in almost columnar fashion, with tangles in layer 2 and 3 appearing in register with tangles in layers 5 and 6 (Pearson et al., 1985). The primary sensory and motor cortices typically show the lowest density of neurofibrillary tangles, suggesting they are the least sensitive to the mechanisms underlying this disorder (Brun & Gustafson 1976; Pearson et al. 1985; Esiri et al. 1986). These patterns of distribution suggest that the disease progresses from the hippocampus along back-projections into cortical regions. In addition, it suggests that susceptibility to degeneration is somehow correlated with the level of involvement in higher order cognitive processes and associations between modalities -- processes which involve ongoing remodeling of cortical representations.
2.2. Loss of neuromodulatory agents
Studies have shown decreases in the levels of various cortical neuromodulatory agents in Alzheimer's disease, with a primary focus on enzymatic markers for acetylcholine. Acetylcholinesterase (AChE) and cholineacetyltransferase (ChAT) show marked decreases in concentration in the cortex of patients diagnosed with Alzheimer's disease (Davies & Maloney 1976; Perry et al. 1977; Coyle et al. 1983). In addition, the regions of the basal forebrain which provide the cholinergic innervation of cortical structures show decreases in the number of neurons, suggesting a degeneration in Alzheimer's disease (Whitehouse et al. 1982; Rapp & Amaral 1992). This evidence, coupled with research showing that cholinergic muscarinic antagonists such as scopolamine impair memory function in normal subjects (for review, see Hasselmo, 1995), led to the suggestion that the memory deficits of Alzheimer's disease might somehow be specifically related to the loss of cortical cholinergic innervation. Administration of acetylcholinesterase blockers cause a small but significant alleviation of the memory impairment associated with Alzheimer's disease.
2.3. Theories of the progession of Alzheimer's disease
Research on potential causes of Alzheimer's disease at the molecular level do not attempt to systematically describe the spread of pathology between different regions. Many forms of familial Alzheimer's disease have been linked to specific inherited differences in protein structure (Schellenberg et al. 1992), though the fact that monozygotic twins can show different susceptibility for the disease suggests the disease is not entirely genetic (Breitner et al. 1993). This class of theories depends upon the assumption that one of the protein components of the Alzheimer neuropathology, such as amyloid or tau, is a causative agent in the disease (see e.g. Hardy and Higgins 1992). The characteristic distribution of neuropathology places extra demands on this theory, suggesting a particular susceptibility of the hippocampus and association cortex but not the primary motor and sensory cortices. Theories based on environmental factors must also account for this selective sensitivity.
The theory presented here focuses on relating initiation and progression of the disease to functional characteristics of cortical regions. In this framework, the initial imbalance of cortical parameters which results in pathology could be due to defects of the genetic code, the spread of toxic factors, an infection by a prion, or a combination of different factors. But this theory proposes that the final effect of this imbalance is the initiation of runaway synaptic modification within cortical regions with a strong capacity for synaptic modification, such as the hippocampal formation. In this context, the progression of the disease depends upon the functional interaction of cortical regions. Rather than depending on the transmission of some substance from the axon terminal of an affected cell to an as yet unaffected post-synaptic neuron, this theory depends only on the normal mechanisms of synaptic transmission and synaptic modification at these connections. The basis of this theory is that the pattern of activity propagated from affected to unaffected regions may be pathological in itself. That is, runaway synaptic modification may cause a breakdown of function in one region, and the patterns of activity elicited can induce runaway synaptic modification in connected regions.
The focus of this model differs from other models of Alzheimer's disease (Horn et al., 1993; Herrmann et al., 1993), which do not attempt to address the dynamics of spread of cortical neuropathology. These previous models start with the assumption of loss of neurons or synaptic connections within models of cortex, and then analyze how the effects of this loss on memory function may be affected by synaptic compensation mechanisms.
3. Runaway synaptic modification in models of cortex.
This theory of Alzheimers disease focuses on the phenomenon of runaway synaptic modification as the final common pathway by which different etiological influences can result in Alzheimer neuropathology (Hasselmo, 1994). Runaway synaptic modification can occur in any system in which synaptic modification depends upon synaptic transmission. Thus, most neural network models of cortical function have the potential to undergo runaway synaptic modification.
Models of cortical associative memory function (Anderson 1972; Grossberg 1970; Hopfield 1982; Kohonen 1984; McClelland & Rumelhart 1988; Amit et al. 1990) focus on the anatomical evidence for widely distributed excitatory intrinsic and associational connections linking pyramidal cells within cortical structures, including neocortex and hippocampus. While they differ in detail, the function of all these models depends upon the synaptic modification of excitatory connections using some modification of the Hebb rule (Hebb 1949; Wigstrom et al., 1986). The basic feature of the Hebb rule is a change in synaptic strength proportional to pre-synaptic and post-synaptic activity during learning. These modified excitatory synapses can then form the basis for recalling associations between different patterns of activity. A simple example of this associative memory function is shown in Figure 1.
Neurophysiological data suggests that Hebbian synaptic modification depends upon combining post-synaptic depolarization with synaptic transmission to activate NMDA receptors at the synapse being modified (Wigstrom et al. 1986). However, if a modifiable synapse can influence post-synaptic activity during learning, strengthening this synapse will increase post-synaptic activity, and thereby increase subsequent strengthening of the synapse. This positive feedback effect can very rapidly lead to exponential growth of undesired synapses within the network, i.e. runaway synaptic modification.
The mechanism for runaway synaptic modification is illustrated in Figure 2. This figure shows that if synaptic transmission at modifiable synapses is allowed during learning, the spread of activity across previously modified connections causes the new synaptic modification to contain elements of proactive interference from previously learned memories (Hasselmo et al. 1992; Hasselmo & Bower 1993; Hasselmo 1993; 1994; Hasselmo et al., 1995). Without the proper balance of parameters of cortical function, this interference during learning can have disastrous effects in models of cortical memory function . Though the effect of interference during learning in any particular stage of learning may be small, this phenomenon can severely affect the function of the network over time, because the effects are compounded by subsequent learning. The progressive buildup of interference from previous retrieval leads to a malignant nostalgia resulting in runaway synaptic modification throughout the whole network. In this case, severe proactive and retroactive interference results in a complete breakdown of normal memory function. This runaway interference during learning has been described previously in detail using mathematical analysis (Hasselmo 1994) and computational models (Hasselmo et al. 1992; Hasselmo 1993; Barkai et al., 1994).
Because of the problems caused by synaptic transmission during learning, most associative memory models ignore the effects of synaptic transmission at modifiable synapses during learning (Kohonen 1972; Anderson 1972; Hopfield 1982; McClelland & Rumelhart 1988; Amit et al. 1990), allowing synaptic transmission only during recall. In computational models, this suppression of synaptic transmission at intrinsic and association fiber synapses during learning can prevent runaway synaptic modification (Hasselmo et al. 1992; Hasselmo 1993; 1994; Hasselmo & Bower 1993; Barkai et al., 1994; Hasselmo and Barkai, 1995). Though this suppression of synaptic transmission during learning has been used for decades in neural network models, researchers did not provide a neurophysiological mechanism for this effect until recently. Recently, it has been shown that acetylcholine has the capacity to selectively suppress intrinsic and association fiber synaptic transmission, while leaving afferent fiber synaptic transmission unaffected (Hasselmo & Bower 1992; Hasselmo and Schnell, 1994; Hasselmo et al., 1995). In addition, acetylcholine enhances the excitability of cortical neurons to the afferent synaptic input (Cole & Nicoll 1984; Barkai & Hasselmo 1994). In computational models of cortical associative memory function, application of this selective suppression of intrinsic fiber synaptic transmission during learning prevents interference from previously learned memories (Hasselmo et al. 1992; Hasselmo 1993, 1994; Hasselmo and Bower 1993; Hasselmo and Barkai, 1995). While the examples shown here are highly simplified, the prevention of runaway synaptic modification has also been explored in detailed biophysical simulations of cortical associative memory function (Hasselmo and Barkai, 1995; Barkai et al., 1994). Prevention of runaway synaptic modification by cholinergic suppression of synaptic transmission is illustrated in simplified form in Figure 4.
In this new framework for learning, synaptic modification must be maximal during the cholinergic suppression of synaptic transmission. But how does this allow activation of post-synaptic NMDA receptors? The activation of NMDA receptors and the mechanisms of synaptic modification are still possible because the cholinergic suppression of synaptic transmission is not complete. In neurophysiological experiments, the cholinergic suppression of synaptic transmission is usually less than 70% (Hasselmo & Bower 1992). Analysis of associative memory models incorporating feedback inhibition, a threshold for synaptic modification and gated decay of synaptic strength shows that this level of suppression is sufficient to prevent interference during learning, while allowing sufficient synaptic transmission for the modification of synapses (Hasselmo 1993; 1994). In models of cortical associative memory function, interference during learning can be prevented by the proper balance of cortical physiological parameters including 1.) pre-synaptic cholinergic modulation of synaptic transmission, 2.) Regulated decay of synaptic connectivity strength, 3.) post-synaptic cholinergic modulation of cellular excitability, 4.) the level of inhibition within the network, 5.) the threshold for synaptic modification, and 6.) the nature of the patterns being stored within the network. In addition, cholinergic agonists have been shown to enhance synaptic modification in cortical structures (Hasselmo and Barkai, 1995; Huerta and Lisman, 1994).
Associative memory models with fixed-point attractor dynamics have been utilized in other modeling work focused on Alzheimer's disease (Horn et al., 1993; Herrmann et al., 1993). In that work, the accuracy of memory recall is analyzed as different numbers of synaptic connections or processing units are deleted within the network. This causes a gradual impairment in memory function which can be offset by postulating various processes for synaptic compensation, in which strengthening of existing synapses offsets the loss of other synapses. Thus, this research focuses on compensatory mechanisms for decreasing the impaired memory function associated with Alzheimer's disease, rather than addressing the spread of pathology directly.
Synaptic transmission at synapses undergoing Hebbian synaptic modification is a regular feature of a different class of models, focused on self-organization of feature detectors in cortical structures (Linsker, 1988; Miller et al., 1990). These models avoid the exponential growth of the full population of synapses by using techniques such as normalization. However, even slight imbalances in the mechanism of normalization of synaptic strength causes a breakdown of function and runaway synaptic modification in these networks. Here it is proposed that neuronal degeneration in Alzheimer's disease may result from runaway synaptic modification due to an imbalance of synaptic normalization, which could include overproduction or flaws in the function of the amyloid precursor protein, or improper phosphorylation of the tau protein.
4. Runaway synaptic modification and the progression of Alzheimer's disease.
The phenomenon of runaway synaptic modification would greatly increase the metabolic and structural demands on the mechanisms of synaptic modification within cortical regions. As shown in Figures 2 and 3, runaway synaptic modification results in the strengthening of many additional connections -- in proportion to the number of associations stored in the network. Thus, runaway synaptic modification in a network storing 100 different associations would result in synaptic modification at a 100-fold greater number of synapses. While cortical structures do not show the complete connectivity used in these models, the axon collaterals of each pyramidal cell within a cortical region makes from 1000 to 10,000 excitatory synapses on other pyramidal cells. An exponential increase in the demands on synaptic modification might explain neuronal degeneration of the sort seen in Alzheimer's disease. This theory does not stand in opposition to any specific theory of the etiology of the disease, since it suggests that the progression of Alzheimer's disease could start from any one of a number of imbalances in cortical function. In fact, this theory could even allow for multiple different initiating factors -- including the gradual overload of the capacity of cortical networks.
4.1. Time course of onset of neuropathology.
Evidence for specific genetic markers of Alzheimer's disease raises the question: Why does the disease appear late in life? This question might be answered by analysis of how an imbalance of cortical parameters can result in the initiation of runaway synaptic modification. This analysis shows the relative importance of different parameters in preventing runaway synaptic modification, and how the magnitude of the imbalance relates to the speed of progression of runaway synaptic modification. As noted above, the imbalance could result from changes in the strength of any of a number of factors. As a specific example, a simplified analysis of an associative memory model suggests that runaway synaptic modification will appear if cholinergic suppression of synaptic transmission is not sufficient to bring the level of post-synaptic activity for undesired connections below the level of inhibition and the threshold of synaptic modification (Hasselmo 1993; 1994). Another important factor in determining whether runaway synaptic modification appears within the network is the normalization of synaptic weights. If the mechanisms of normalization of synaptic weights are somehow impaired, then runaway synaptic modification can be initiated.
In cortical function, the relative strength of any of these parameters could change due to toxic factors, or these parameters could be improperly balanced due to genetic predisposition. For low cholinergic suppression, low inhibition, and a low threshold of synaptic modification, runaway synaptic modification will occur after a smaller number of patterns are stored. For strong cholinergic suppression, strong inhibition, and a high threshold of synaptic modification, runaway synaptic modification will take longer to appear, or may be prevented entirely. It is the relative balance of these different parameters which determines how well the network resists the initiation of runaway synaptic modification.
The breakdown of function can take one of two forms: 1) The network can start with a given balance of parameters, and when the capacity of the network is exceeded, runaway synaptic modification will cause a breakdown of function, or 2) The network could start with a given balance of parameters, and suffer some change in this balance which causes the breakdown of function. In either case, the larger the imbalance, the earlier the breakdown appears, and the more rapidly it will progress. A simplified analysis of a linear associative memory model shows that the undesired connections strengthened due to interference during learning will grow exponentially. If cholinergic suppression is very low relative to other cortical parameters, interference will appear sooner within the network, and the spread of interference throughout the network will be more rapid. If there is only a slight imbalance, interference will take much longer to appear, and once it appears, it will progress more slowly (Hasselmo, 1994).
This phenomenon could underlie the evidence for differences in the time course of progression of presenile dementia of the Alzheimer type as compared with senile dementia of the Alzheimer type. It has been shown that quantitative measures of anatomical markers of Alzheimer's disease suggest greater severity in presenile dementia as compared with senile dementia of the Alzheimer type (Hansen et al. 1988). In addition, some studies show a correlation between age and cognitive performance on memory tasks, with younger Alzheimer's patients performing more poorly (Kopelman 1985), and clinical evidence suggests that the rate of progression from onset of symptoms to death may be more rapid in presenile dementia (Selzter & Sherwin 1983). However, other evidence suggests that the progression of neuropsychological deterioration is actually slower in presenile dementia (Huff et al. 1987).
This model of the initiation of Alzheimer's disease pathology could also be used to address the epidemiology of this disease. Alzheimer's disease has been characterized as a disease of old age primarily because the frequency of the disease increases rapidly with increasing age. The exponential progression of runaway synaptic modification in these computational models could be used as a model of the epidemiology of Alzheimer's disease. Since the risk factors for onset of runaway synaptic modification include many components of normal cortical function, all humans may have some probability of induction of these pathological effects. Depending upon individual variation in certain parameters, the induction will take place at different time points, but once initiated, it will progress exponentially. This means that the percentage of cases of senile dementia of the Alzheimer type should increase exponentially with increasing age (though an epidemiological prediction of this sort must take into account the reduction of the population at each age level due to death from Alzheimer's disease and death from other factors). There is certainly a rapid increase in the number of cases of Alzheimer's disease in later life (Katzman, 1986), and neuropathological data from normal elderly subjects suggest a rapid increase in the density of tangles with increasing age, even in tissue from subjects who have not been diagnosed with Alzheimer's disease (Arriagada et al., 1992). A more radical prediction of the model is that runaway synaptic modification will appear more rapidly in proportion to the greater overlap between stored patterns. This suggests that the amount of correlation in the environment could influence the propensity for development of Alzheimer's disease.
4.2. Selective distribution of neuropathology.
If the neuropathology associated with Alzheimer's disease results from runaway synaptic modification, this suggests that the apparent early and severe involvement of layers II and IV of the lateral entorhinal cortex, region CA1 of the hippocampus and the adjacent regions of the subiculum (Hyman et al. 1984; 1990; Arnold et al. 1991; Braak & Braak 1991; Arriagada et al. 1992) results from a particular sensitivity of these regions to runaway synaptic modification. For example, it is possible that sensitivity to runaway synaptic modification might be associated with 2 features: 1.) a strong capacity for Hebbian synaptic modification, and 2.) absence of the cholinergic suppression of synaptic transmission during learning. Considerable research has focused on the how subregions of the hippocampus resemble the structures of associative memory models (Marr 1971; McNaugh-ton & Morris 1987). As noted above, this region shows robust Hebbian synaptic modification (Wigstrom et al. 1986) and has been implicated in memory function in extensive research (Scoville & Milner 1957; Squire & Zola-Morgan 1991). The greater propensity for synaptic modification could underlie the early sensitivity of the hippocampal formation. The early sensitivity of lateral entorhinal cortex could be linked to the absence of cholinergic suppression at synapses arising from this region and terminating in the outer molecular layer of the dentate gyrus (Kahle and Cotman, 1990), while the relative sparing of region CA3 could result from the robust cholinergic suppression of synaptic transmission at synapses arising from CA3 pyramidal cells (Hasselmo and Schnell, 1994; Hasselmo et al., 1995).
4.3. Spread of degeneration between cortical regions.
The neuropathology of Alzheimer's disease appears to spread from hippocampus into the neocortex along well-established anatomical connections, the back projections from the subiculum and entorhinal cortex to association cortex (Pearson et al. 1985; Hyman et al. 1990; Arnold et al. 1991). In later stages of the disease, neurofibrillary tangles appear in regions of the temporal, parietal and frontal neocortex (Hirano and Zimmerman 1962; Pearson et al. 1985; Arnold et al. 1991). Neurofibrillary tangles primarily spread into cortical regions, though the plaques associated with terminal degeneration appear in subcortical regions as well (Pearson et al. 1985). This spread of degeneration is here proposed to result from the spread of runaway synaptic modification between cortical regions. This would use the same mechanisms important for transferring information stored in the hippocampus back into the neocortex -- the process of consolidation (Wilson and McNaughton, 1994; McClelland et al., 1995). A network simulation of the hippocampus (Hass-elmo, 1995b; Hasselmo et al., 1996) has been used to model the process of consolidation, as shown in Figure 5.
Runaway synaptic modification could spread from hippocampus back into neocortical structures using the same mechanisms as consolidation. It has been shown in simulations that if runaway synaptic modification occurs during the initial formation of representations of new memories in the hippocampus, then subsequent retrieval of these representations will result in a spread of runaway synaptic modification. For example, as shown in Figure 6, a decrease in the mechanisms of synaptic decay of the input from entorhinal cortex layer II to dentate gyrus results in the initiation of runaway synaptic modification in this pathway. Even if the parameters of other connections have not been altered, the initiation of runaway synaptic modification at these perforant path synapses results in the spread of runaway synaptic modification into back projections from region CA1 to neocortex. In simulations with multiple interacting layers, the initiation of runaway synaptic modification in one layer results in the progressive spread to other layers, even if those layers would not undergo runaway synaptic modification independently.
Modeling suggests that runaway synaptic modification will spread according to functional boundaries. That is, after occurring in neurons encoding a particular category of information, it will more rapidly influence similar or strongly associated information before influencing unrelated information. This might explain the apparent heterogeneous distribution of tangles in Alzheimer's disease and the apparent specificity for specific modalities in some cases. In particular, the distribution of neurofibrillary tangles may be on the order of the size of cortical columns, with tangles in layers 2 and 3 in register with tangles in layers 5 and 6 (Pearson et al. 1985).
The rate of spread of runaway synaptic modification depends upon the ongoing capacity for Hebbian synaptic modification and the amount of excitatory associative connectivity. Primary sensory and motor cortices have more restricted and specific connectivity of excitatory intrinsic and associational connections (Lund 1988), and are further removed from the highly plastic structures of the hippocampal formation. This may explain why the neuropathology in Alzheimer's disease is far less pronounced in the primary sensory cortices (Hirano and Zimmerman 1962; Brun and Gustafson 1976; Pearson et al. 1985; Esiri et al. 1986).
Based on the distribution of neuropathology in cortical regions and impairments in olfactory identification (Koss et al. 1988), it has been suggested that a pathogen for Alzheimer's disease enters via the olfactory epithelium (Pearson et al. 1985; Talamo et al. 1989). The olfactory epithelium does show changes which may be associated with Alzheimer's disease (Talamo et al. 1989), but the relative lack of pathology within the mitral cell layer of the olfactory bulb does not support the hypothesis that the disease spreads through these neurons (Hyman et al 1991). Even if an environmental factor arriving along this pathway is responsible, its effects appear highly specific to regions with modifiable excitatory intrinsic connections. For example, high concentrations of tangles have been found in regions with intrinsic excitatory synapses which may be modifiable such as the anterior olfactory nucleus (Esiri and Wilcock 1984; Hyman et al. 1991) and the piriform cortex (Reyes et al. 1987). In contrast, much lower concentrations of tangles are found in the olfactory bulb, which contains primarily inhibitory feedback connections (Kauer 1991).
4.4. Breakdown of cholinergic modulation.
A range of evidence suggests the degeneration of cortical cholinergic innervation in Alzheimer's disease, including decreased levels of AChE and ChAT in cortical structures (Davies and Maloney 1976; Perry et al. 1977) and decreased cell counts in basal forebrain nuclei (Whitehouse et al. 1982 ). However, other evidence shows increased sprouting of the cholinergic innervation of the dentate gyrus (Geddes et al. 1985) and increases in ChAT staining levels in the dentate gyrus and stratum moleculare of the subiculum (Hyman et al. 1986). In addition, memory impairment has been correlated with hypertrophy of basal forebrain neurons in aged non-human primates (Rapp and Amaral 1992). The latter effects have been suggested to reflect the reafferentation of hippocampal neurons which have lost their innervation from the entorhinal cortex. However, the effects of acetylcholine on dentate gyrus granule cells is considerably different from the effects of the glutamate released by the terminals of the perforant path. It is unclear why sprouting would occur to replace such innervation. The model presented here provides a different perpsective on this data.
As shown above, cholinergic modulation can help prevent runaway synaptic modification in models of cortical associative memory. While deficiencies of cholinergic modulation could underlie the initiation of runaway synaptic modification, runaway synaptic modification could also be due to changes in other parameters, without any significant change in the level of cholinergic modulation. However, once runaway synaptic modification begins to appear, it is likely that the normal feedback mechanisms which regulate the level of cholinergic modulation would be significantly increased as a mechanism for preventing the breakdown of function. This might place increasingly greater demands on cortical cholinergic modulation, initially resulting in a strong enhancement of cholinergic modulation. This effect could underlie the increased sprouting of cholinergic axons in the dentate gyrus (Geddes et al. 1985) and the increases in AChE staining in the dentate gyrus and subiculum (Hyman et al. 1986). The increases in AChE staining appear in exactly those regions where insufficient cholinergic modulation during learning might underlie the initiation of runaway synaptic modification.
While a large increase in cholinergic modulation might slow the progression of runaway synaptic modification, once the breakdown has been initiated it is very difficult to forestall. Even at very high concentrations of acetylcholine some synaptic transmission remains (Hasselmo and Bower 1992), allowing continued runaway synaptic modification. This would result in continually increasing demands which could lead to an initial hypertrophy and eventual degeneration of cells of the basal forebrain nuclei from which this innervation arises (Whitehouse et al. 1982). Thus, the model suggests two phases: an initial increase of innervation followed by degeneration.
4.5. Molecular components of the neuropathology of Alzheimer's disease.
The phenomenon of runaway synaptic modification in models suggests how the molecular components of the neuropathology of Alzheimer's disease might underlie the initiation of runaway synaptic modification, and how runaway synaptic modification could result in large scale deposits of these particular substances.
4.5.1. Amyloid precursor protein and runaway synaptic modification.
Neuritic plaques have as their main component feature the accumulation of the b-amyloid protein (BAP). BAP apparently arises from a normal component protein of the brain referred to as amyloid precursor protein (APP). The APP has a transmembrane domain, with the bulk of the protein being extracellular (Selkoe 1993; Marotta et al. 1992). Some forms of the protein contain a segment which can act as a protease inhibitor in the extracellular portion (Wagner et al. 1992; Hyman et al. 1992). Unfortunately, the functional role of the amyloid precursor protein has not yet been determined. However, the known properties of the APP are compatible with some role in the growth and maintenance of synaptic connections (Milward et al. 1992; Leblanc et al. 1992; Saitoh et al. 1989). For example, APP may prevent proteolysis of existing or newly strengthened synapses by extracellular proteases.
In this framework, if post-synaptic activity is not correlated with the activity of the pre-synaptic terminal, the APP will be cleaved at its usual site just outside the cell membrane (at amino acid 16 of the BAP), and will allow extracellular proteases to break down the components of the synaptic contact. However, if all synaptic connections are showing increased correlation of pre and post-synaptic activity, as would occur with the feedback mechanisms of interference during learning, then the APP will not be removed, but will continue to accumulate on the pre-synaptic terminals. This will allow runaway growth of the synaptic connection, protected from proteolysis by the protease inhibitors of the membranous APP, and could ultimately lead to high concentrations of APP throughout cortical regions with excessive synaptic modification. The pathways concerned with the breakdown of APP would come under increasing demands, possibly leading to a build-up of BAP in the region of the synapse. Alternately, the cell giving rise to the synapse might be unable to cope with the excessive metabolic demands of thousands of growing synapses. Its death would lead to the degeneration of a synaptic terminal with high concentrations of APP which might be broken down into BAP.
In this framework, the build-up of BAP is seen as a by-product of the runaway synaptic modification due to interference during learning, but this framework could also account for a causal role in Alzheimer's disease not of BAP, but of the overproduction of the precursor of BAP, APP. Over-production of APP due to genetic mutations, or the additional copy of chromosome 21 associated with Downs syndrome, would lead to excessive APP on the pre-synaptic terminal. This would slow the normal breakdown or weakening of synaptic connections not associated with correlation of pre and post-synaptic activity, and could ultimately trigger the initiation of runaway synaptic modification. This would account for the apparent linkage of some forms of familiar Alzheimer's disease with genetic defects within the APP or in other portions of chromosome 21 as well as the apparent development of Alzheimer's type symptoms in people with Downs syndrome. Perhaps the mental retardation found in Downs syndrome might also result from the insufficient breakdown of existing synaptic connections. Any influence of APP on the mechanisms of synaptic modification could result in runaway synaptic modification if that influence is somehow altered by mutations in the APP gene.
4.5.2. Tau protein and runaway synaptic modification
Neurofibrillary tangles appear to contain considerable levels of paired helical filaments, of which a primary component is an abnormally phosphorylated form of the tau protein (Grundkeiqbal et al. 1986) . The tau protein may play a role in the assembly of microtubules. Blockade of the expression of tau protein impairs the development of axons in culture (Kosik and Caceres 1991), suggesting possible involvement in the growth and remodeling of axonal connectivity, or in axonal transport mechanisms. Thus, the tau protein may be involved in formation and regulation of synaptic connections. A breakdown in the normal function of the tau protein, possibly due to improper phosphorylation, might lead to a slowing of the normal mechanisms of normalization of synaptic strength arising from a single neuron. A slowdown in this redistribution of synaptic strength could allow the initiation of runaway synaptic modification in a self-organizing network.
Once runaway synaptic modification begins to occur, it could place increasing demands on this aspect of cellular function, since it would require increased transport of substances produced in the cell body to the thousands of growing synapses. Ultimately, this could overload the capacity of the axonal transport system, leading to excessive production of structural elements of cellular transport, and ultimately an increased accumulation of the byproducts of this system, the paired helical filaments. Ultimately, the increased demands on each individual neuron might cause a complete breakdown of function, and the ultimate the death of neurons, leaving behind neurofibrillary tangles.
4.6. Neuropsychology of Alzheimer's disease
In its initial stages, a major characteristic of Alzheimer's disease is the impairment of memory function. The most common early symptoms of this impairment appear to primarily involve episodic or declarative recent memory. In later stages, the symptoms progress to impairments of semantic memory and other factors such as emotional disturbances.
In neuropsychological tests, a clear deficit can be shown in a range of memory tasks. The overall characteristics of memory loss in Alzheimer's disease have been reviewed elsewhere (Morris and Kopelman 1986). Impairments appear to be particularly severe on memory for recent events. The computational modeling presented here suggest that runaway synaptic modification could cause increased interference between stored representations, causing impairments in short-term memory tasks requiring free recall (Corkin 1982; Morris 1986) and increasing the number of intrusions reported in other tasks (Fuld et al. 1986; Troster et al. 1989; Jacobs et al. 1990; Delis et al. 1991). Continued interference effects during learning could eventually lead to the spread of runaway synaptic modification into neocortex via the mechanisms of consolidation. This would lead to impairments of remote memory (Wilson et al. 1981; Corkin et al. 1984) and semantic memory (Huff et al. 1986; Ober et al. 1986). It is important to discriminate the neuropsychological effects at different stages of the disease. In the early stages of the disease, the interference effects predicted by the model should be evident. However, in later stages of the disease, the degeneration of cortical structures could have memory effects which are less directly related to interference during learning.
The model predicts that interference between stored memories should impair memory function in early stages of the syndrome. Thus, storage and retrieval of information will still be possible, but the encoding of information should be impaired, particularly for overlapping memories. The effect of interference may not only result in erroneous answers due to intrusions from previously learned information. In some situations, interference could lead to the activation of many conflicting associations, so that the subject is unable to recall any particular association. In any case, this model predicts that in the early stages of Alzheimer's disease, interference effects should become more severe.
The notion of interference between memories is compatible with the presenting characteristics of Alzheimer's disease. The most common complaints refer to loss of memory for the location of simple household objects, loss of topographical memory, loss of orientation in time, and loss of recognition memory for acquaintances. All of these forms of memory require accurate memory for complex associations with a high degree of similarity. Development of higher order spurious associations between these items would severely interfere with day to day memory. A similar effect might underlie associations for days of the week or the various streets in a neighborhood. As summarized in previous work (Hasselmo, 1994), the model can account for the following neuropsychological data on Alzheimer's disease: 1. Increased intrusion errors. 2. Interference effects in short-term retention, 3. Sparing of implicit memory. 4. Homogeneous impairment of remote memory. 5. Impairments of semantic memory. However, this theory is far from complete. The predictions of the model must be tested experimentally using techniques ranging from brain slice physiology to behavioral memory tasks. In addition, more detailed biophysical simulations of the cortical structures affected by the disorder must be analyzed. However, computational modeling of the breakdown of function in cortical structures provides a unique means for linking evidence on this disorder across a range of different experimental techniques. Use of these types of modeling techniques will be vital to bringing together the disparate disciplines of neuroscience research in the understanding of Alzheimer's disease.
Albert, M., Smith, L.A., Scherr, P.A., Taylor, J.O., Evans, D.A. and Funkenstein, H.H. (1991) International Journal of Neuroscience 57: 167-178.
Amit, D.J., Evans, M.R. and Abeles, M. (1990) Network 1: 381-405.
Anderson, J.A. (1972) Mathematical Biosciences 14: 197-220.
Arnold, S.E., Hyman, B.T., Flory, J., Damasio, A.R., and Van Hoesen, G.W. (1991) T Cerebral Cortex 1: 103-116.
Arriagada, P.V., Marzloff, B.A. and Hyman, B.T. (1992) Neurology 42: 1681-1688.
Ball, M.J. (1972) Neuropathology and Applied Neurobiology 2: 395-410.
Barkai, E. and Hasselmo M.E. (1994) Journal of Neurophysiology 72: 644-658.
Barkai, E., Horwitz, G., Bergman, R.E. and Hasselmo, M.E. (1994) Journal of Neurophysiology. 72: 659-677.
Braak, J. and Braak, E. (1991) Acta Neuropathologica. 82: 239-259.
Breitner, J.C.S., Gatz, M., Bergem, A.L.M., Christian, J.C., Mortimer, J.A., McClearn, G.E., Heston, L.L., Welsh, K.A., Anthony, J.C., Folstein, M.F. and Radebaugh, T.S. (1993) Neurology 43: 261-267.
Brun, A., Gustafson, L. (1976) Archiv fur Psychiatrie und Nervenkrankheiten 223: 15-33, 1976.
Cole, A.E. and Nicoll, R.A. (1984) Journal of Physiology 352: 173-188.
Corkin, S. (1982) In: S. Corkin, K.L. Davis, J.H. Growdown, E. Usdin and R.J. Wurtman (eds.), Alzheimer's disease: A report of research in progress. New York: Raven Press.
Corkin, S., Growdown, J.H., Nissen, M.J., Huff, F.J., Freed, D.M. and Sagar, H.J. (1984) In: R.J. Wurtman, S. Corkin and J.H. Growdown (eds.) Alzheimer's disease: Advances in basic research and therapies. Center for Brain Sciences: Cambridge, MA.
Coyle J.T., Price, D.L. and DeLong, M.R. (1983) Science 219: 1184-1190.
Crapper McLachlan, D.R. and Van Berkum, M.F.A. (1986) In: D.F. Swaab, E. Fliers, M. Mirmiran, W.A. Van Gool and F. Van Haaren (eds) Progress in Brain Research 70: 399-409.
Davies, P. and Maloney, A.J.F. (1976) Lancet 2: 1403.
Delis, D.C., Massman, P.J., Butters, N., Salmon, D.P., Cermak, L.S. and Kramer, J.H. (1991) Psychological Assessment 3: 19-26.
Esiri, M.M. and Wilcock, G.K. (1984) Journal of Neurology Neurosurgery and Psychiatry 47: 56-60.
Esiri, M.M., Pearson, R.C.A., Powell, T.P.S. (1986) Brain Research 366: 385-387.
Fuld, P.A., Katzman, R., Davies, P. and Terry, R.D. (1982) Annals of Neurology 11: 155-159.
Geddes, J.W., Monaghan, D.T., Cotman, C.W., Loh, I.T., Kim, R.C., Chui, H.C. (1985) Science 230: 1179-1181.
Goudsmit, J. and Van der Waals, F.W. (1986) In: D.F. Swaab, E. Fliers, M. Mirmiran, W.A. Van Gool and F. Van Haaren (eds) Progress in Brain Research 70: 399-409.
Grossberg, S. (1970) Studies in Applied Mathematics 49: 135-166.
Grundkeiqbal, I., Iqbal, K., Tung, Y.C., Quinlan, M., Wisneiwski, H.M. and Binder, L.I. (1986) Proceedings of the National Academy of Sciences, U.S.A. 83: 4913-4917.
Hansen, L.A., DeTeresa, R., Davies, P. and Terry, R.D. (1988) Neurology 38: 48-54.
Hardy, J.A. and Higgins, G.A. (1992) Science 256: 184-185.
Hasselmo, M.E. (1993) Neural Computation. 5(1): 32-44.
Hasselmo, M.E. (1994) Neural Networks. 7(1): 13-40.
Hasselmo, M.E. (1995a) Behav. Brain Res. 67: 1-27.
Hasselmo, M.E. (1995b) In L. Niklasson, M.B. Boden (eds.) Current Trends in Connectionism. Lawrence Erlbaum Assoc.; Hillsdale, N.J. pp.15-32..
Hasselmo, M.E. and Barkai, E. (1995) J. Neurosci. 15(10); 6592-6604.
Hasselmo M.E., Anderson B.P. and Bower J.M. (1992) Journal of Neurophysiology 67: 1230-1246.
Hasselmo M.E. and Bower J.M. (1992) Journal of Neurophysiology 67: 1222-1229.
Hasselmo M.E. and Bower, J.M. (1993) Trends in Neurosciences 16: 218-222.
Hasselmo, M.E. and Schnell, E. (1994) J. Neurosci. 14(6): 3898-3914.
Hasselmo, M.E., Schnell, E., Barkai, E. (1995) J. Neurosci. 15(7): 5249-5262.
Hasselmo, M.E., Schnell, E., Berke, J. and Barkai, E. (1995) In G. Tesauro, D. Touretzky, T. Leen (eds.) Advances in Neural Information Processing Systems, Vol. 7. MIT Press: Cambridge, MA.
Hebb, D.O. (1949) The organization of behavior. New York.: Wiley.
Herrmann, M., Ruppin, E., Usher, M. (1993) Biol. Cybern. 68: 455-463.
Hirano, A. and Zimmerman, H.M. (1962) Archives of Neurology 7: 73-88.
Hopfield, J. J. (1982) Proceedings of the National Academy of Sciences USA 79: 2554-2559.
Horn, D., Ruppin, E., Usher, M. and Herrmann, M. (1993) Neural Comp. 5: 736-749.
Huerta, P.T., Lisman, J.E. (1994) Nature 364: 723-725.
Huff, F.J., Corkin, S., and Growdon, J.H. (1986) Brain and Language 34: 269-278.
Huff, F.J., Growdown, J.H., Corkin, S. and Rosen, T.J. (1987) Journal of the American Geriatrics Society 35: 27-30.
Hyman, B.T., Damasio, A.R., Van Hoesen, G.W., and Barnes, C.L. (1984) Science 225: 1168-1170.
Hyman, B.T., Van Hoesen, G.W., Kromer, L.J. and Damasio, A.R. (1986) Annals of Neurology 20:472-481.
Hyman, B.T., Kromer, L.J., and Van Hoesen, G.W. (1987) Annals of Neurology. 21: 250-267 1987.
Hyman, B.T., Van Hoesen, G.W., and Damasio, A.R. (1990) Neurology 40: 1721-1730.
Hyman, B.T., Arriagada, P.V. and Van Hoesen, G.W. (1991) Annals of the New York Academy of Sciences. 640: 14-19.
Jacobs, D., Salmon, D.P., Troster, A.I. and Butters, N. (1990) Archives of Clinical Neuropsychology 5: 49-57.
Jolles, J. (1986) In: D.F. Swaab, E. Fliers, M. Mirmiran, W.A. Van Gool and F. Van Haaren (eds) Progress in Brain Research 70: 399-409.
Kahle, J.S. and Cotman, C.W. (1989) Brain Research 482: 159-163.
Katzman, R., (1986) New England Journal of Medicine 314: 964-973.
Kauer, J.S. (1991) Trends in Neuroscience 14: 79-85.
Keane, M.M., Gabrieli, J.D.E., Fennema, A.C., Growdon, J.H. and Corkin, S. (1991) Behavioral Neuroscience 105: 326-342.
Keppel, G. and Underwood, B.J. (1962) Journal of Verbal Learning and Verbal Behavior 1:153-161.
Kopelman, M.D. (1985) Neuropsychologia 23: 623-638 1985.
Kopelman, M.D. (1986) Quarterly Journal of Experimental Psychology 38: 535-573.
Kosik, K.S. and Caceres, A. (1991) Journal of Cell Science S15: 69-74.
Koss, E., Weiffenbach, J.M., Haxby, J.V. and Friedland, R.P. (1988) Neurology 38: 1228-1232.
Leblanc, A.C., Kovacs, D.M., Chen, H.Y., Villare, F., Tykocinski, M., Autiliogambetti, L. and Gambetti, P. (1992) Journal of Neuroscience Research 31: 635-645.
Levy, W.B. and Steward, O. (1979) Brain Research 175: 233-245.
Levy, W.B. and Colbert, C.M. (1992) Soc. Neurosci. Abstr. 18: 628.14.
Liljenstrom H. and Hasselmo M.E. (1995) J. Neurophysiol. 74: 288-297.
Linsker, R. (1988) Computer 21: 105-117.
Lund, J.S. (1988) Annual Review of Neuroscience 11: 253-288.
Marotta, C.A., Majocha, R.E. and Tate, B. (1992) Journal of Molecular Neuroscience 3: 111-125.
Marr, D. (1971) Philosophical Transactions of the Royal Society B 262: 23-81.
McClelland, J.L. and Rumelhart, D.E. (1988) Explorations in Parallel Distributed Processing. Cambridge, MA: MIT Press.
McClelland, J.L., McNaughton, B.L. and O'Reilly, R. (1995) Psych. Rev. 102: 419-457.
McNaughton, B.L. and Morris, R.G.M. (1987) Trends in Neurosciences 10: 408-415.
Miller, K.D., Keller, J.B. and Stryker, M.P. (1989) Science 245: 605-615.
Milward, E.A., Papadopolous, R., Fuller, S.J., Moir, R.D. and Small, D. (1992) Neuron 9: 129-137.
Morris, R.G. (1986) Cognitive Neuropsychology 3: 77-97.
Morris, R.G. and Kopelman, M.D. (1986) Quarterly Journal of Experimental Psychology 38A: 575-602.
Pearson, R.C.A., Esiri, M.M., Hiorns, R.W., Wilcock, G.K. and Powell, T.P.S. (1985) Proceedings of the National Academy of Sciences USA 82: 4531-4534.
Perry, E.K., Gibson, P.H., Blessed, G., Perry R.H. and Tomlinson, B.E. (1977) Journal of Neurological Science 34: 247-265.
Postman, L. and Underwood, B.J. (1973) Memory and Cognition 1:19-40.
Rapp, PR. and Amaral, D.G. (1992) Trends in Neurosciences. 15: 340-345.
Reyes, P.F., Golden, G.T., Fagel, P.L., Fariello, R.G., Katz, L. and Carner, E. (1987) Archives of Neurology 44: 644-645.
Saitoh, T., Sundsmo, M., Roch, J.M., Kimura, N., Cole, G. (1989) Cell 58: 615-622.
Saper, C.B., German, D.C., and White, C.L. (1985) Neurology 35: 1089-1095.
Schellenberg, G.D., Bird, T.D., Wijsman, E.M., Orr, H.T., Anderson, L., Nemens, E., White, J.A., Bonnycastle, L., Weber, J.L., Alonso, M.E., Potter, H., Heston, L.L., Martin, G.M. (1992) Science 258: 668-671.
Scoville, W.B. and Milner, B. (1957) Journal of Neurology, Neurosurgery and Psychiatry 20: 11-21.
Selkoe, D.J. (1993) Trends in Neurosciences 16: 403-409.
Seltzer, B. and Sherwin, I. (1983) Archives of Neurology 40: 143-146.
Squire, L.R. and Zola-Morgan, S. (1991) Science 253: 1380-1386.
Talamo, B.R., Rudel, R., Kosik, K.S., Lee, V.M.Y., Neff, S., Adelman, L. and Kauer, J.S. (1989) Nature 337: 736-739.
Troster, A.I., Jacobs, D., Butters, N., Cullum, C.M., Salmon, D.P. (1989) Clinics in Geriatric Medicine 5: 611-632.
Wagner, S.L., Siegel, R.S., Vedvick, T.S., Raschke, W.C., Vannostrand, W.E. (1992) Biochemical and Biophysical Research Communications 186: 1138-1145.
Whitehouse, P.J., Price, D.L., Struble, R.G., Clark, A.W., Coyle J.T. and DeLong, M.R. (1982) Science 215: 1237-1239.
Wigstrom, H., Gustafsson, B., Huang, Y.-Y. and Abraham, W.C. (1986) Acta Physiologica Scandinavica 126: 317-319.
Wilson, R.S., Kaszniak, A.W. and Fox, J.H. (1981) Cortex 17: 41-48.
Wilson, M.A. and McNaughton, B.L. (1994) Science 265: 676-679.