Coordinated geometric representations of learned knowledge in hippocampus and frontal cortex

Interactions between frontal cortex and hippocampus (HPC) play a key role in decision-making behaviors. Here, we test how these brain areas coordinate their representations of behavioral and cognitive variables in a task that requires both. We use a virtual reality task in which mice navigate in a T-maze and accumulate transiently visible and stochastic left/right cues to infer the correct reward location. Neural data is recorded acutely with up to 10 Neuropixel shanks simultaneously in bilateral HPC and medial prefrontal cortex (mPFC). To compare neural activity patterns, we analyze this data with state space models, in which each neuron is represented by an activity axis, and population activity constitutes a trajectory in this high-dimensional space. We demonstrate that mPFC neural activity across trials does not fill this space, but lies on a structured, low-dimensional, and non-linear subspace. Population activity can therefore be parameterized by a relatively small number of latent variables that span this neural activity manifold. Similar to previous observations in the HPC, we find that behavioral and cognitive variables, like position in the maze and accumulated visual evidence, are represented as smooth gradients on the mPFC manifold. HPC and mPFC also encoded position and evidence with similar precision and geometry. We demonstrate this similarity between regions by training a decoder on the HPC manifold and decoding behavioral variables from mPFC and vice versa. Despite similar encoding of position and evidence between regions, a communication subspace analysis suggests a differential representation of position and evidence in both areas. The results reported here, together with a growing literature of related findings across brain regions and tasks, contributes to a better understanding of how task-relevant variables might be integrated through mPFC-HPC interactions in decision making.