Brain regions have different risks of developing seizures depending on their biochemistry, electrophysiology, and circuitry. We propose that in addition to these known structural factors, seizure susceptibility can be related to the function of a neural network. Computational tasks that require information to flow readily across a continuum of network states may encourage excessive activity to spread into seizures. Tasks that require information to be confined to discrete, well-separated network states may instead disfavor the spread of pathologic activity. To substantiate these ideas, we construct spiking attractor networks that differ only in whether they possess a continuous attractor manifold or discrete attractor states. We also train spiking recurrent neural networks to perform either a continuous autoencoding task or a discrete classification task using the same architecture. When epileptogenic perturbations are applied to each network, the continuous networks respond more intensely compared to their discrete counterparts. To test our theoretical predictions, we perform in vivo electrophysiological recordings of mice during acute, pharmacologically induced limbic seizures. We find that entorhinal cortex (EC), whose grid cells encode positions in space as a continuous attractor, is more strongly implicated in seizure initiation compared to the hippocampal subfield CA3, which is believed to store memories as a discrete attractor. Moreover, epileptogenic perturbations cause the EC population activity pattern to be destabilized along continuous trajectories, in agreement with our simulations. These results relate seizure susceptibility to trade-offs between network responsiveness and stability and, more broadly, to connections between function and dysfunction.

Abstract to be announced.