Learning how to experience the world: From circuits to cell types to genes

Perception depends on the brain's ability to transform high-dimensional sensory inputs into low-dimensional internal models that support adaptive behavior. Evidence supports two frameworks for sensory perception-representational processing, in which stimulus features are progressively integrated into complex perceptual objects across a cortical hierarchy, and predictive processing, in which internally generated predictions are continuously reconciled with incoming sensory signals. Yet how these frameworks are mechanistically implemented in neural circuits, and whether they can be unified, remains an open question. Here, we review recent studies in mouse primary sensory and higher-order association cortex demonstrating that cell-type-specific transcriptional programs may provide a critical mechanistic link between these frameworks and circuit functions. In primary sensory cortices, neurons that function as stable feature detectors or respond to sensory prediction errors correspond to distinct molecularly defined cell types. In higher-order association cortices, distinct inhibitory cell-type compositions and plasticity-related gene expression support both associative learning for representational processing and error learning for predictive processing. We discuss how cell-type-specific transcriptional programs may endow cell types and circuits with the capacity to support both representational and predictive processing modes in a behavioral state-dependent manner. This could potentially enable active sensation during behavioral engagement as well as memory consolidation and model updating during behavioral quiescence. Together, these studies suggest that examining how gene expression programs equip specific cell types with relevant computational properties is a promising approach that can integrate these frameworks and provide a new understanding of how sensory perception is implemented in the brain.