Touch inhibits feeding through a neural bottleneck in <em>C. elegans</em>

Neural bottlenecks are ubiquitous network motifs where multiple neurons project onto a smaller subset. This convergence suggests that the network compresses information encoded in incoming signals. However, measuring such compression remains challenging because it requires observing all the neurons involved, as well as the animal's behavior, to determine how much of the sensory information is used to guide behavior. RIP and I1 neurons in the nematode <em>C. elegans</em> are a compact implementation of a neural bottleneck where such measurements are possible. RIP receive massive converging sensory inputs and provide the only connections to the pharyngeal network controlling feeding via gap-junctions to the pharyngeal neurons I1. We developed a behavioral assay to deliver substrate vibrations as a controllable touch stimulus while automatically reading out the feeding behavior in freely moving <em>C. elegans</em>. We found that vibrations elicit both an escape response and a feeding inhibition. Both responses are abolished in mutants lacking the six touch receptor neurons (TRNs), whereas only the feeding response is abolished in animals where we genetically ablated RIP or I1 neurons, the escape circuit being upstream of the bottleneck. Furthermore, we showed that the escape response increases with the stimulus intensity, whereas the feeding response saturates at lower intensity, suggesting that the bottleneck may apply an intensity threshold to the touch signal. To determine how touch intensity information is encoded through the bottleneck, we adapted the calcium indicator GCaMP8f in <em>C. elegans</em> to enable sensitive measurements of neuronal activity in the input (TRNs) and bottleneck (RIP, I1) layers. Single neuron calcium imaging in immobilized animals showed that the input layer linearly encodes the behaviorally relevant stimulus intensity. However these experiments did not show activity in bottleneck neurons, indicating putative motor feedback. We therefore implemented simultaneous feeding and neuronal calcium imaging in behaving animals and showed that RIP activity is driven both by stimulus and feeding states. This provides a proof-of-concept for further measurements of intensity representation in the bottleneck layer. The touch-feeding system in <em>C. elegans</em> provides a neuroethological window to explore the role of neural bottleneck networks, potentially shedding light on generalizable principles of information compression.