Data-Constrained Recurrent Network Neural Model Uncovers the Circuit Mechanism of Olfactory OFF Responses

Abstract

Sensory neural circuits must encode both the presence and termination of a stimulus. Following odor offset, projection neurons (PNs) in the insect antennal lobe (AL) exhibit transient increases in firing rate, termed OFF responses, yet the circuit mechanisms that generate them in recurrent excitatory-inhibitory networks remain poorly understood. Here, we constructed a biologically-constrained firing rate-based recurrent neural network (RNN) model of the locust AL and trained it on electrophysiological recordings from 110 in vivo PNs to reconstruct their odor-evoked temporal dynamics across five odorants. The trained model faithfully reproduced the firing rates of constrained neurons, while unconstrained PNs and LNs developed biologically plausible temporal response patterns and response-type diversity. Using targeted input and connectivity perturbations, we found that OFF responses arise through two mechanistically distinct pathways. A feedforward pathway transmits offset-type olfactory receptor neuron (ORN) input directly to downstream PNs, while a recurrent pathway generates post-stimulus excitation independently of offset input. Selective perturbation of individual recurrent connections identified LN-LN mutual inhibition as the dominant recurrent pathway, and decomposition of excitatory and inhibitory inputs revealed that it produces net excitation through the transient release of inhibition rather than through increased drive. The two pathways recruit largely non-overlapping PN populations, indicating that OFF response identity is an emergent property of the network state rather than a cell-intrinsic feature. These findings provide a circuit-level account of OFF response generation and demonstrate how data-constrained RNNs can dissect circuit mechanisms directly from in vivo recordings.