Unsupervised inference of brain-wide functional motifs underlying behavioral state transitions

During prolonged periods of stress, animals switch from active to passive coping strategies to manage effort expenditure. These normally adaptive behavioral state transitions become maladaptive in disorders such as depression. Such behavioral state transitions occur as animals continually evaluate the fruitfulness or futility of their actions in response to stressors. This process is mediated by interactions across numerous brain regions over long timespans, integrating information about the stressor with the outcome of avoidance actions performed, and tracking the accrual of stress, ultimately driving transitions between different coping strategies. Here, we disentangled both the spatial and temporal dependencies of the neural mechanisms driving behavioral state transitions using computational models directly constrained by longitudinal, whole-brain, cellular-resolution neural recordings from larval zebrafish during active to passive coping in the face of persistent, inescapable stress [1]. We built and analyzed large-scale recurrent neural network (RNN) models that reproduced the long time-scale dynamics of over 10,000 simultaneously-recorded neurons. We combined this model’s outputs–connectivity [1] and inter-region currents [2]–with tensor decomposition [3] to infer, in an unsupervised manner, separate “functional motifs” capturing multi-region dynamics that describe the time-varying flow of source and target currents. We found three distinct functional motifs corresponding to key behavioral signals: shocks, tail movements, and stress accumulation. All three motifs included the habenula and raphe nucleus—regions previously implicated in passive coping [4]—as key targets of brain-wide networks corresponding to each behavioral signal. We show that these two regions integrate distinct sets of input currents from a number of other regions, including dorsal thalamus and telencephalon to drive the transition from active to passive coping. We provide an unbiased mechanistic framework to disentangle the simultaneous encoding of behaviorally-relevant signals across interacting regions brain-wide and demonstrate that behavioral state transitions require simultaneous integration of inputs from distinct networks over slow timescales.