Determining the neural encoding of sensorimotor function in the mouse cortex using machine learning

Brain-machine interfaces (BMIs) aim to restore sensorimotor function to individuals suffering from neural injury and disease. Elucidating the neural encoding of sensorimotor function in the cortex and decoding sensory and motor commands from these brain areas is essential for the advancement of BMIs. In this dissertation, I present two studies for determining the neural encoding of sensorimotor function in the mouse cortex using machine learning on neural data recorded with two-photon (2p) calcium imaging. The first study applies deep learning to decode multi-limb movements of running mice from 2p calcium imaging data, demonstrating that an artificial neural network can accurately infer movements of all four limbs from neural activity in a single cortical hemisphere. Feature importance analysis reveals that decoding is driven by a small, sparsely distributed subset of neurons rather than correlations between movements of different limbs. This work provides insights into sensorimotor processing and paves the way for the development of optical BMIs. The second study investigates how somatosensation is encoded at the population level in the mouse sensorimotor cortex, using principal component analysis on three 2p calcium imaging datasets from anesthetized and awake mice with passive or spontaneous limb movements. The results reveal that somatosensory encoding is low-dimensional, with a few principal components capturing large variance. Additionally, we show that limb movement representations are conserved across animals, including the orthogonality between ipsilateral and contralateral limbs, highlighting similarities with motor processing. While individual neurons mainly encode intrinsic variables, population-level activity also represents extrinsic ones. Together, these analyses demonstrate that population-level encoding of somatosensory information in the mouse sensorimotor cortex is structured to facilitate sensorimotor integration across the brain and provides insights for the development of BMIs and neural prosthesis. Together, these studies leverage 2p calcium imaging and machine learning to deepen our understanding of how sensorimotor information is encoded in the mouse cortex. By demonstrating the application of deep learning for neural decoding and revealing the low-dimensional encoding of somatosensation at the population level, this work enhances the foundational knowledge required to develop advanced BMIs.