MODELING NEUROMODULATOR-DRIVEN CORTICAL DYNAMICS WITH LOW-RANK RECURRENT NEURAL NETWORKS

Neuromodulators play a central role in organizing brain function, flexibly reconfiguring neuronal circuits to support adaptive behavior. Yet, how neuromodulatory signals reshape population dynamics remains poorly
understood.
Recurrent neural networks (RNNs) provide a powerful framework for studying neural dynamics but typically lack explicit, biologically interpretable mechanisms for neuromodulation. To address this gap, we developed
a data-driven modeling framework that integrates neuromodulatory effects into RNNs, allowing modulation of neuronal gain, excitability, and connectivity.
We trained these models directly on large-scale electrophysiological recordings from mouse prefrontal cortex (PFC), reproducing the activity of individual neurons while incorporating simultaneously measured,
spontaneously fluctuating norepinephrine (NE) levels obtained via fiber photometry during behavior. We systematically evaluated four candidate neuromodulatory mechanisms: additive input, presynaptic gain,
postsynaptic gain, and low-rank connectivity modulation, by fitting each model directly to the data. Analysis of the learned dynamics revealed that intermediate NE levels induce a line attractor in PFC population
activity, which collapses into a single fixed point at low or high NE concentrations. This predicts maximal cortical responsiveness and variability at intermediate NE, a prediction confirmed in the neural recordings.
To uncover the underlying circuit mechanism, we derived an analytically tractable mean-field reduction of the trained networks, identifying two interacting neuronal populations differentially modulated by NE. Their gain
and firing-rate interactions governed the emergence and stability of the line attractor, a structure mirrored by clustering within the full RNN.
Together, this work establishes a principled framework linking neural data, RNNs, and theory to uncover mechanisms of neuromodulator-driven population dynamics.