Excitatory-inhibitory balance modulates the formation and dynamics of neuronal assemblies in cortical networks

Abstract

Repetitive activation of subpopulation of neurons in cortical networks leads to the formation of neuronal assemblies, which can guide learning and behavior. Recent technological advances have made the artificial induction of such assemblies feasible, yet how various patterns of activation can shape their emergence in different operating regimes is not clear. Here we studied this question in large-scale cortical networks composed of excitatory (E) and inhibitory (I) neurons. We found that the dynamics of the network in which neuronal assemblies are embedded is important for their induction. In networks with strong E-E coupling at the border of E-I balance, increasing the number of perturbed neurons enhanced the potentiation of ensembles. This was, however, accompanied by off-target potentiation of connections from unperturbed neurons. When strong E-E connectivity was combined with dominant E-I interactions, formation of ensembles became specific. Counter-intuitively, increasing the number of perturbed neurons in this regime decreased the average potentiation of individual synapses, leading to an optimal assembly formation at intermediate sizes. This was due to potent lateral inhibition in this regime, which also slowed down the formation of neuronal assemblies, resulting in a speed-accuracy trade-off in the performance of the networks in pattern completion and behavioral discrimination. Our results therefore suggest that the two regimes might be suited for different cognitive tasks, with fast regimes enabling crude detections and slow but specific regimes favoring finer discriminations. Functional connectivity inferred from recent experiments in mouse cortical networks seems to be consistent with the latter regime, but we show that recurrent and top-down mechanisms can dynamically modulate the networks to switch between different states. Our work provides a framework to study how neuronal perturbations can lead to network-wide plasticity under biologically realistic conditions, and sheds light on the design of future experiments to optimally induce behaviorally relevant neuronal assemblies.