Neuro-musculoskeletal modeling reveals muscle-level neural dynamics of adaptive learning in sensorimotor cortex
The neural activity of the brain is intimately coupled to the dynamics of the body. Yet how our hierarchical sensorimotor system dynamically orchestrates the generation of movement while adapting to incoming sensory information remains unclear. In mice, the extent of encoding from posture to muscle-level features across the motor (M1) and primary sensory forelimb (S1) cortex and how these are shaped during learning are unknown. To address this, we built a large-scale model that captures hypothesized neural computations and use this to control a novel 50-muscle model of the adult forelimb amenable to studying motor control and learning in a physics simulation environment. We show that we can imitate 3D limb kinematics collected during a joystick task by solving inverse kinematics and deriving a sensorimotor control model that drives the same actions. Using the internal computations from our model, we find that populations of layer 2/3 M1 and S1 neurons encode high-level position, and lower-level muscle space and proprioceptive dynamics. During adaptive learning, these functionally distinct neurons map onto specific computational motifs. Strikingly, S1 neurons more prominently encode sensorimotor prediction errors, and M1 and S1 support optimal state estimation. Moreover, we find that neural latent dynamics differentially change in S1 vs. M1 during this within-session learning. Together, our results provide a new model of how neural dynamics in cortex enables adaptive learning.