Alzheimer's Disease as a Result of Stimulus Reduction in a GABA-A-Deficient Brain: A Neurocomputational Model

Abstract

Several research studies point to the fact that sensory and cognitive reductions like cataracts, deafness, macular degeneration, or even lack of activity after job retirement, precede the onset of Alzheimer's disease. To simulate Alzheimer's disease earlier stages, which manifest in sensory cortices, we used a computational model of the koniocortex that is the first cortical stage processing sensory information. The architecture and physiology of the modeled koniocortex resemble those of its cerebral counterpart being capable of continuous learning. This model allows one to analyze the initial phases of Alzheimer's disease by "aging" the artificial koniocortex through synaptic pruning, by the modification of acetylcholine and GABA-A signaling, and by reducing sensory stimuli, among other processes. The computational model shows that during aging, a GABA-A deficit followed by a reduction in sensory stimuli leads to a dysregulation of neural excitability, which in the biological brain is associated with hypermetabolism, one of the earliest symptoms of Alzheimer's disease.

Figure 1

Intrinsic plasticity. (a) The continuous line at the center represents the activation function at a hypothetical initial state. (b) When the net-input values are diminished as in A, B, and C, the sigmoidal activation function shifts leftwards. (c) When the net-input values grow like in the case of D, E, and F, the sigmoidal activation function shifts rightwards.

Figure 2

Cytoarchitecture of the koniocortex network in which acetylcholinergic neurons from the parabrachial nucleus project to thalamocortical neurons. The neurons that specifically belong to the koniocortex are the spiny stellate (SS) neurons, the inhibitory basket neurons (B), and the shunting basket (SB) neurons.

Figure 3

Alphanumeric input patterns presented as inputs to the koniocortex model.

Figure 4

(a) One way of emphasizing the distinctive features of a set of vectors a, b, c, d consists in subtracting their average vector (as in the case of vector b). (b) In this way, vectors become more separated (in terms of the angle between them).

Figure 5

Example of koniocortex continuous learning under physiological (not defective) conditions. Each of the six rows represents the learning status at a selected epoch (iteration), being 1,000 the total number of epochs. In each of the six rows, ten synaptic weight matrices are corresponding to each one of the ten spiny neurons in the simulation, being each spiny neuron identified by a number below each matrix. The relative size of the fifteen green tiles in each matrix corresponds to the relative value of the weights of the recurrent virtual connections from spiny to input neurons. When, at epoch 500, one of the training patterns, pattern one, was substituted by a new pattern, pattern X, there is a process of weight reorganization for deciding which spiny neuron will fire in front of pattern X. At the end (see rows corresponding to epochs 650 and 700), spiny neuron three fires in front of pattern X and spiny neuron ten fires when pattern two is presented to the koniocortex model. In the bottom graph, we depict two curves: the red curve is the average output and the blue curve the average shift along with iterations. Both curves exhibit stable and regular behaviors.

Figure 6

Normal aging. The two columns of this table show the koniocortex model behavior while performing a short-term memory (STM) task consisting of substituting one of the numerical patterns by pattern X. This substitution starts at 70% of epochs. In both experiments, the levels of GABA-A and ACh levels are normal, and pruning is also normal (starting at 75% of epochs). “Mem = no” means that in these experiments, we did not simulate the introduction of an NMDA blocker (memantine). At the bottom of each column, we show the evolution of the average output (in red) and average shift (in blue) of all neurons. Experiment a.1: the label “S_R = no” at the header of the first column means that there is no reduction of stimuli in this experiment. As seen in the block corresponding to 80% of training, the network successfully learns pattern X. Experiment a.2: the label “S_R = 60%” at the header of the second column indicates that there was a reduction of stimuli at 60% of repetitions. From this point, the average output and average shift experiment a sudden fall but stabilize rapidly. The presentation of a new pattern X at 70% of epochs does not alter the ongoing network dynamics in any significant way. Labels: S_R = sensory reduction; STM = short-term memory; Mem = memantine application; ACh = acetylcholine reduction in thalamocortical neurons.

Figure 7

This table shows two cases without GABA-A signaling: the column on the left without sensory reduction and the column on the right with sensory reduction. At the bottom of each column, we show the evolution of the average output (in red) and average shift (in blue) of all neurons. Labels: S_R = sensory reduction; STM = short-term memory; Mem = memantine; ACh = acetylcholine reduction in thalamocortical neurons. In experiment b.1 (left column), there is no reduction of stimuli, and we see that, although an episodic period of acute output oscillations occurred, homeostatic mechanisms are capable of driving the network to equilibrium again. By examining the blocks, we see that pattern recall was permanently impaired. However, acute oscillations related to hypermetabolism and disease progression were extinguished. In experiment b.2, the withdrawal of GABA-A at 50% of epochs was the precondition for the production of intense oscillations when the sensory reduction took place at 60% of epochs. At the same time, at 60% of epochs, neurons lost their pattern specificity so that several neurons processed the same pattern (number 1). Sustained oscillations are associated with hypermetabolism and the progression of AD.

Figure 8

This table shows the koniocortex model behavior when the reduction of sensory stimuli takes place after a reduction in GABA-A. A short-term memory (STM) learning task is shown in the first and second columns, consisting of learning an “X” pattern in substitution of one of the numerical patterns that constitute the learning set. In the second column, we evaluate the usage of NMDA blockers like memantine. At the bottom of each column, we show the evolution of all neurons' average output (in red) and average shift (in blue). Experiment b.3: the pattern “X” presented at 70% of repetitions is successfully learned. Despite this apparent success, the univocal correspondence between patterns and neurons that were compromised at 60% of epochs did not return to normal. Notice that episodes of oscillation and stabilization are intermingled and that learning a new pattern stabilizes the network, although this stable situation is usually transitory. In experiment b.4, the administration of an NMDA blocker like memantine took place at 55% of epochs. Numerical pattern memories are kept intact, but, when introducing pattern “X” at 70% of repetitions, the network was incapable of learning it. Stimulus reduction at 60% in a GABA-A-depleted network contributed to the initiation of persistent oscillations. Labels: S_R = sensory reduction; STM = short-term memory; Mem = memantine; ACh = acetylcholine reduction in thalamocortical neurons.

Figure 9

In these experiments, we reduce the effect of ACh at 50% of epochs. Experiment c.1, in which no other modification is involved, represents a control test for comparison with subsequent experiments. In this case, the bottom graph exhibits sparse oscillations of the average output. Experiment c.2, in which stimulus reduction takes place at 60% of epochs, produced smaller but sustained oscillations that are harmful to learning patterns. Labels: S_R = sensory reduction; STM = short-term memory; Mem = memantine; ACh = acetylcholine reduction in thalamocortical neurons.

Figure 10

This table shows the koniocortex model response when GABA-A levels are normal, and we reduce ACh at 50% of repetitions. Sensory reduction also takes place at 60% of total repetitions, and a memory test is performed at 70%. Experiment c.3: an NMDA blocker (like memantine) is not used. Oscillations are present initially when ACh is reduced and, especially afterward, during the process of learning the new pattern X. Experiment c.4: an NMDA blocker (like memantine) is applied at 55% of epochs. In this case, plasticity is eliminated; the network remains at its former stability level (without oscillatory activity) although it is not able to learn the testing pattern X. Labels: S_R = sensory reduction; STM = short-term memory; Mem = memantine; ACh = acetylcholine reduction in thalamocortical neurons.

Figure 11

This table presents a case in which GABA-A normalization is recovered despite the usual process of network aging. After the elimination of GABA-A normalization that occurred at 50% of epochs, pattern learning was impaired; that is to say, patterns were poorly recovered, and neurons lost their pattern specificity. With the reduction of sensory stimuli at 60% of epochs, the average shift and output started an intense oscillatory dynamic that suddenly disappeared when GABA-A normalization was reinstalled at 75% of repetitions. After this recovery, the network was capable of learning pattern X, while the recovery of the remaining numerical patterns was damaged. At this later stage, in which many of the connections were pruned, learning is a difficult process requiring much more repetitions than in an intact network. Labels: S_R = sensory reduction; STM = short-term memory; Mem = memantine; ACh = acetylcholine reduction in thalamocortical neurons.

Figure 12

(a) Shape of curves obtained with real neurons for different initial synaptic weights w_i. In this case, the experiment consisted of injecting current in the presynaptic neuron and measuring the postsynaptic voltage. The point where curves cross the horizontal axis is called the long-term potentiation threshold. (b) Family of curves obtained in the computer model through the presynaptic rule (adapted from Figure 2 in [105]). Both graphs exhibit metaplasticity: the rightward elongation of the curves along the horizontal axis for higher values of initial synaptic weights. For a more detailed explanation of how the curves were obtained, we suggest the reader to study References ([105] (Section 2), ([22] (Section3.2)).

Figure 13

When a time series f(x) is placed at the single input of a neuron with intrinsic plasticity property, the output of the neuron is the same time series f(x) but with its moving average removed.