NeuroGate: waveform translation between brain

Abstract

The brain processes information across distributed circuits, yet a typical experiment records only a few regions, leaving the rest unobserved. Connectivity and latent-embedding methods relate brain regions but do not return the waveform of an unrecorded one. Here we introduce NeuroGate, a framework for cross-regional neural signal translation that recovers an unrecorded region's waveform from a recorded one. Across 56 pathways spanning rodent LFP, human sEEG and ECoG, and scalp EEG, NeuroGate predicts waveforms more accurately than 13 deep-learning and linear baselines, with low across-session variance and high median accuracy. We tested the predictions against a circuit perturbation: silencing piriform-cortex output with tetanus toxin light chain collapsed piriform-to-bulb predictions in 8 of 8 folds while the reverse was preserved, matching the monosynaptic anatomy. We demonstrate two applications: quantifying how recording redundancy bounds the information from each added source, and a reusable perturbation protocol for verifying directional translation in any circuit.