A voxel-wise encoding model for VR-navigation maps view-direction tuning at 7T-fMRI

The brain derives cognitive maps from visual inputs, posing the important question how information is transformed and communicated across the neural systems involved. Inspired by the prior success of encoding models at characterizing the tuning and topography of responses in the visual cortex, we developed voxel-wise and multivariate encoding models to examine networks supporting spatial orienting during active navigation. We used 7T-fMRI to monitor brain activity at submillimeter resolution while participants freely navigated in a virtual environment. We combined their virtual view direction (vVD) with a variety of circular-gaussian vVD-basis functions that differed in kernel spacing and width. For each parameter set, we then estimated model weights using an iterative training procedure that maximized predictability to find the optimal parameters explaining each voxel’s time course. Using the resulting model weights, we examined fMRI-responses in held-out data to show that vVD predicts activity in early visual, medioparietal and parahippocampal cortices involved in self-motion- and scene processing, as well as in mediotemporal regions known to support navigation, like the entorhinal cortex. The activity in each region was best predicted with distinct vVD-tuning widths that increased anteriorally from medioparietal to entorhinal cortices. Inverting the encoding model reconstructed vVD from fMRI responses also in regions with high-level mnemonic function, like the hippocampus, and revealed a vVD-tuning topology within the entorhinal cortex, akin to the topology of head direction representations previously observed in rodents. Our approach demonstrates the feasibility of using encoding models to study visual tuning during naturalistic behaviors, and sheds new light on how the currently viewed scene is encoded in the brain during active navigation.