Recurrent latent dynamics in spinal circuits enable a digital twin of autonomic state transitions

Physiological signals underlying homeostasis, such as blood pressure, are dynamic and nonlinear, shaped by interactions among diverse neurons. In the spinal intermediolateral nucleus (IML), neurons show heterogeneous, monotonic firing changes during blood pressure transitions, intrinsic to control yet difficult to decode. We introduce a manifold-based model reconstructing internal state dynamics from IML activity, beyond prior descriptive frameworks. Brief spinal stimulation lowering blood pressure showed that single-neuron responses could not explain the nonlinear trajectory, so we derived a low-dimensional circular manifold capturing its cyclical evolution. Embedding the manifold as a dynamic reservoir in a recurrent neural network reconstructed trial-to-trial variability and nonlinear blood pressure dynamics generated by the manifold, without a readout. Our results establish the neural manifold as a generative computational substrate for autonomic physiology, a capability that remains largely unexplored in spinal circuits, and a step toward closed-loop spinal neuromodulation.