Understanding how neuronal populations interact to process information and generate behaviour is a central goal of neuroscience. This can be challenging due to the high dimensionality of the neuronal populations, the dense interactions between them and the unobserved factors that influence the dynamics. The neuronal manifold hypothesis tackles high dimensionality by positing that the relevant dynamics occurs on a lower-dimensional manifold. While the neuronal manifold may be useful in visualising global dynamics, it tells us very little about how various subsystems interact to produce these dynamics. The embedding dimensions are often highly entangled, representing arbitrary sets of neurons that may overlap with other dimensions, thereby obscuring the contributions of individual subsystems. Here, we introduce a BunDLe-Net-based architecture that embeds separate neuronal populations into distinct latent dimensions. By leveraging BunDLe-Net’s Markovian embedding, we ensure that every point in the latent space retains as much predictive information as the raw neuronal data about future behavioural dynamics. We apply our method to C. elegans behavioural and neuronal data grouped into sensory, motor and interneurons categories. Our method not only reveals recurring behavioural motifs but also shows how different populations interact to orchestrate these motifs. From the manifold, we can read off which populations encode information and drive the dynamics in each behavioural state. Thus, we present a powerful visual tool that reveals how information is processed and relayed across populations.