Interpreting computational mechanism from multi-area neural recordings

Distilling computational mechanisms from neural recordings requires disambiguating intrinsic dynamics from external inputs. Here, we introduce Dynamically Interacting Systems (DynISys), a multi-compartment dynamical-systems model that infers per-area dynamics and structured inter-area communication from multi-area spiking activity. DynISys combines neural ordinary differential equation (ODE) latent dynamics with injective latent-to-neural mappings to achieve expressive and trustworthy dynamical inference. When applied to simultaneous dorsal premotor (PMd) and primary motor (M1) recordings from a continuous reaching task, DynISys accurately reconstructs neuronal responses without inventing spurious latent features. Analyzing the model's dynamics reveals putative mechanisms for motor planning and execution, with PMd-to-M1 inputs positioning a stable fixed point in M1 that encodes execution dynamics for the upcoming movement, subsequently evoked by external (non-PMd) inputs. Finally, multi-session DynISys achieves consistent latents when fit to neural recordings spanning multiple days. In summary, DynISys allows distillation of multi-area recordings into an interpretable account of distributed neural computation.