Control of cortical population activity with patterned microstimulation

Abstract

Closed-loop control of cortical activity is a central goal in systems neuroscience and clinical neuromodulation, but most approaches either rely on detailed circuit models that are unattainable in vivo or on open-loop stimulation tuned by trial and error. Here we introduce REACHable manifold Control (REACH-Ctrl), a data-driven brain-computer interface that achieves real-time control of population spiking activity using patterned microstimulation and multi-electrode recordings. REACH-Ctrl learns a finite-horizon controllability map directly from short training epochs in which random multi-electrode pulse sequences are delivered through a subset of electrodes while recording evoked responses. From these input-output data, it identifies the "reachable manifold" of population states and computes low-current microstimulation sequences that steer activity toward designated targets, without explicit knowledge of the underlying connectivity or dynamics. We test REACH-Ctrl in macaque prefrontal cortex, demonstrating high control accuracy, robust across sessions and stimulation parameters. Geometric analyses showed that multi-pulse sequences traverse a well-defined reachable manifold with substantial, but incomplete, overlap with the intrinsic neural activity manifold, revealing both on- and off-manifold components of control. Encoding models further revealed that, in our weak-stimulation regime, population responses to multi-electrode sequences are well approximated by the linear sum of localized "stimulation fields" with modest history dependence, explaining the success of our linear control approach. These results demonstrate precise, sample-efficient control of cortical population activity with clinically relevant microstimulation hardware, and provide a general blueprint for designing perturbations for sparsely observed neural circuits.