New RNN model reveals how chronic stress rewires brain circuits

A new study from Mauricio A. Diaz and colleagues uses recurrent neural networks (RNNs) to uncover the computational mechanisms behind chronic stress-induced cognitive decline. The researchers trained RNNs on a working memory task and simulated chronic stress by systematically adjusting eight different synaptic or neuronal parameters. Only one manipulation—strengthening inhibitory-to-excitatory synapses—simultaneously reproduced all three experimental hallmarks of prefrontal dysfunction under stress: an overall inhibitory dominance, reduced excitatory activity, and impaired task accuracy. This pinpoints a single synaptic locus that could explain how prolonged stress degrades executive function.

When networks were trained under this stress mechanism (a proxy for resilience), they maintained task performance but paid a hidden cost: they generalized poorly to longer memory demands beyond their training regime. This trade-off between resilience and flexibility persisted across network sizes and stress intensities, offering a computational analogue of the shift from adaptive to rigid, habit-like behavior observed in chronically stressed animals. The findings provide a neural network framework for designing targeted interventions, such as synaptic modulation therapies, to restore cognitive function in stress-related psychiatric conditions.