Deep Representation Learning on Whole-Brain Population Dynamics Uncovers Geometrically Separable Neural Codes

Learning interpretable low-dimensional representations of whole-brain neuronal dynamics remains a major computational challenge in systems neuroscience. We present a wiring-agnostic deep-learning framework that couples a convolutional encoder with a temporal transformer to learn compact representations directly from volumetric calcium imaging of the entire Drosophila melanogaster brain. Trained to classify 16 experimental conditions that factorially combine metabolic state (fed, starved), sensory modality (olfaction, gustation, or combined), and stimulus valence (appetitive, aversive, or conflicting), the model organizes pan-neuronal whole-brain population activity into geometrically distinct, condition-specific clusters. Analysis of the models latent space reveals that state, modality, and valence are encoded along three near-orthogonal axes: a separable structure that emerges from the classification objective without explicit disentanglement constraints. Spatial attribution and regional importance analyses link modality decoding to distinct anatomical circuits, whereas metabolic state and valence related information show weaker regional specificity and broader distribution across the brain. Our approach does not require anatomical annotation, neuronal identification, or connectivity information, and thus provides a scalable foundation for comparative whole-brain imaging and representation learning of brain wide dynamics.