Supervised deep machine learning models predict forelimb movement from excitatory neuronal ensembles and suggest distinct pattern of activity in CFA and RFA networks

Abstract

Neuronal networks in the motor cortex are crucial for driving complex movements. Yet it remains unclear whether distinct neuronal populations in motor cortical subregions encode complex movements. Using in vivo two-photon calcium imaging (2P) on head- fixed grid-walking animals, we tracked the activity of excitatory neuronal networks in layer 2/3 of caudal forelimb area (CFA) and rostral forelimb area (RFA) in motor cortex. Employing supervised deep machine learning models, a support vector machine (SVM) and feed forward deep neural networks (FFDNN), we were able to decode the complex grid-walking movement at the level of excitatory neuronal ensembles. This study indicates significant differences between RFA and CFA decoding accuracy in both models. Our data demonstrate distinct temporal-delay decoding patterns for movements in CFA and RFA, as well as a selective ensemble of movement responsive neurons with higher distribution in CFA, suggesting specific patterns of activity-induced movement in these two networks.