Computational demands shape seizure susceptibility in recurrent neural networks

Abstract

Brain areas differ in their inherent susceptibility to focal seizures, but the principles governing this risk remain unclear. While prior work has focused on anatomical and physiological factors, here we observed a fundamental contribution from the computations performed by the underlying neural network. Handcrafted and trained recurrent neural networks supporting continuous representations respond to seizure perturbations with higher activity and earlier performance decline relative to matched networks stabilizing discrete, well-separated states. Consistent with this prediction, in vivo recordings revealed that medial entorhinal cortex, whose grid cells exhibit continuous attractor dynamics, drives acute epileptiform discharges with stronger involvement and smoother state trajectories compared to CA3, a hippocampal subfield associated with discrete memory storage. Moreover, selective synaptic silencing demonstrated that this difference in seizure responses depends on intact entorhinal connectivity. Thus, the computations that enable neural networks to process information also influence their vulnerability to pathological transitions.