FRONTOSTRIATAL CIRCUITS REFLECT OPPOSITE REINFORCEMENT LEARNING COMPUTATIONS

The dorsomedial frontal cortex (dmFC) is important for learning to associate stimuli and rewards. However, it remains unknown whether distinct populations of dmFC neurons, defined based on their projection pathways, play different roles in this process. To address this gap, we trained mice in a classical conditioning task in which different visual stimuli were associated with different probabilities of reward (0, 50, or 100%). Mice learned to discriminate between the stimuli, indicated by significantly different preparatory licking to visual stimuli. Once mice reached asymptotic behaviour, we used two-photon calcium imaging to record the activity of dmFC excitatory neurons – with a subset of them projecting to the dorsal striatum (dmFC→dStr), labelled using retrograde injections – during the task. In subsequent sessions, we imaged dmFC neural activity while stimulus-reward contingencies were reversed. dmFC & dmFC→dStr neurons showed similar encoding of task variables (i.e. stimulus value and motor correlates of reward anticipation) during established behaviour. However, these neuronal populations showed substantial differences during contingency reversals: non-projecting dmFC neurons primarily learned from rewards that after the reversal became unexpected (i.e. positive reward prediction errors) while dmFC→dStr neurons adapted their responses to unexpected reward omissions, learning from negative reward prediction errors. This suggests opposite learning computations are represented by different dmFC subpopulations, depending on error valence and underpinned by projection target. Further work will elucidate and formalize these observed differences, providing a principled computational account for neural bases of behavioural adaptability versus stability.