Reward expectation signals in the superior colliculus during associative learning

Abstract

Predicting future rewards allows neutral sensory cues to acquire motivational significance and guide goal-directed actions that maximise reward acquisition. These reward predictions are updated when outcomes differ from expectations resulting in an error in prediction. Midbrain dopamine neurons provide a canonical neural correlate of reward prediction error signal that are broadcast to multiple brain areas. These learned reward predictions ultimately influence neural processes that enable animals to select, prepare and execute actions towards expected rewards. Reward expectation signals have been observed across sensory and motor circuits where they can modulate sensory processing and action selection and vigor. This raises the question of how integrative sensorimotor circuits represent information about reward expectation alongside sensory and motor variables that drive reward-seeking actions. The superior colliculus (SC), a sensorimotor area in the midbrain traditionally associated with orienting and action selection, is strongly influenced by reward context. Expected reward modulates visual and premotor activity, learned object values shape sensory responses, and recent reward alters visual processing in the mouse SC. The SC also provides input to midbrain dopamine neurons and can relay or acquire conditioned sensory responses that influence dopaminergic activity. However, SC projections can also drive orienting movements without supporting reinforcement, raising the possibility that cue-related activity reflects salience or motor preparation rather than reward prediction itself. Here, using an associative-learning task in which auditory cues predicted distinct reward probabilities, together with longitudinal monitoring of intermediate SC population activity across learning, we asked whether the SC acquires signals related to reward probability, whether these signals can be dissociated from anticipatory licking, and whether outcome responses exhibit the hallmarks of reward-prediction errors. We found that cue-evoked SC activity progressively differentiated reward probabilities across learning, and that cue-associated reward probability explained neural activity beyond anticipatory licking. Thus, the intermediate SC carries a learned reward-expectation signal that is multiplexed with, but not reducible to, motor preparation. By contrast, outcome responses did not exhibit a consistent population-level signature of reward-prediction errors. Together, these findings identify the SC as a site where predicted reward is integrated with anticipatory action during associative learning.