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. In order to predict and adapt to the sensorimotor consequences of our actions, compelling behavioral studies in humans, non-human primates, and in rodents have shown the existence of internal models - predictive models of our body in the environment. These internal models are theoretically used to compute an updated state estimate with prediction errors. Here, we directly test whether such errors are encoded in primary somatosensory (S1) or motor (M1) cortex during a motor adaptation task in mice. Using control theory-derived features that include prediction errors, we find that functionally distinct neurons are mapped onto specific computational motifs. We find that layer 2/3 population dynamics encode command-like signals and sensorimotor prediction errors (SPEs). S1 neurons encode SPEs more prominently than M1, and the neural latent dynamics change in S1 more than in M1 during this within-session learning. Then we asked, in which coordinate frameworks are such errors computed? To do this, we developed a novel 50-muscle model of the adult mouse forelimb that is capable of studying motor control and learning in a physics simulator. We identify both high-level 3D position and muscle spaces as coordinate frameworks for SPEs. Together, our results provide a new model of how neural dynamics in S1 enables adaptive learning.