Correlative light and electron microscopy identifies circuit motifs underlying processing of noisy visual motion in larval zebrafish

To integrate noisy visual motion, vertebrate circuits must transform transient sensory signals into persistent, lateralized activity. The synaptic architecture supporting this computation remains unclear. We combined calcium imaging and correlative electron microscopy (EM) in larval zebrafish with an independent whole-brain EM dataset, reconstructing 1,246 neurons in and connected to the anterior hindbrain, a region previously implicated in the temporal integration of visual evidence. We identified four broad functional and morphological classes interconnected through recurrent excitation, interhemispheric and ipsilateral inhibitory motifs, as well as long-range inputs from rostral visual regions. We then trained bilateral neural networks constrained by the measured neuron-to-neuron connectivity as a fixed anatomical mask, while optimizing unknown physiological weights. These models reproduced population dynamics and predicted responses to previously untested motion stimuli. Shuffling connectivity or selectively removing recurrent connections impaired model performance, whereas removing an equivalent number of random connections across the network had much smaller effects. Our results identify an anatomical network compatible with visual-motion integration and implicate specific recurrent connections in its computational dynamics. Together, our results link functional dynamics with synaptic architecture in a vertebrate sensorimotor circuit and provide a framework for interpreting connectomic datasets through combined functional, anatomical, and computational analyses.