Alzheimer-like degradation: first results from a simulated cortical network

AbstractIn a 100,000-neuron simulation of a spiking cortical network, with Alzheimer-like damage imposed from outside, the only early signal we found is a local rise in the mean membrane voltage in the damaged zone, produced by slower recovery of the neurons after each spike. The coherence field, the voltage noise and the firing rate of the surviving neurons show nothing before neurons begin to die: the apparent "holes" come from the dead neurons. The signal appears with amyloid-driven damage and with tau-driven, amyloid-free damage, so the model does not distinguish the origin of the damage, and it appears only when the network is in the active regime. The rise scales with the strength of the change in recovery time (about 1 : 0.5 : 0.2 at strengths 4, 2 and 1) and, with a null that shares the noise, the zone found at strength 4 coincides with the imposed centre of the damage. In this version we consider five different working hypotheses. First, we place the network's epochs on a time axis by anchoring them to a separate population model of sporadic amyloid accumulation: depending on the anchor, the voltage signal precedes the first neuron death by about 0.7–3.3 years, and only one anchor can be used because the growth shapes of the two models do not coincide. Second, we close the loop in the network by making damage growth depend on neuronal activity, in both directions and at two strengths, at 20,000 and 100,000 neurons: progression changes by at most a few percentage points (deaths −3% and +6% for the two signs, A and T about 1%), within what a single run per condition cannot distinguish from noise. Third, we keep the network as a time-compressed simulation and compare ratios rather than rates per year (Section 4.4); its tau-like parameters are of the same order as those derived from tau-PET data, while the amyloid-like growth is 5–19 times faster. Fourth, an exploratory vesicle-like transport with long-range jumps (a hypothesis, not retained as the main mechanism) creates remote foci that start about 30 epochs after the central one and extend the damage; the voltage rises in their zone with the same delay as the damage, without anticipating the first signal (Section 5.4). Fifth, we combine the signal test and the propagation test in a single 100,000-neuron run, with a control that shares the noise: the dose-response of voltage to damage from the first test returns without recalibration (about 6% in the most damaged neurons versus 5%), and in the first run the voltage rises 11–15 epochs before the first death in the first two zones; in three repetitions with other seeds and other starting points, the lead of the inhibitory neurons at the strongest strength is positive (8–13 epochs) in all five of the eight cases (four seeds, two zones) in which the network is active between the first death and the detection, while in the other three the network is mostly silent and the comparison cannot be made; with all neurons together and at weaker strengths it is less regular (Section 5.5). This is the result of a simulation that starts from literature data and from data extrapolated from statistical-probabilistic simulations: the disease is imposed from outside, the network is not calibrated on data, the population model is