Multi-timescale learning signals in Drosophila dopaminergic neurons

Abstract

Learning often requires inference over hidden task structure, including features that predict outcomes. In mammals, prediction-error signaling by midbrain dopamine neurons is considered central to learning, but how these neurons reflect the progression and stability of learning remains unclear. Using Drosophila to monitor calcium activity at trial-by-trial resolution during aversive conditioning, we found that PPM3 dopaminergic neurons exhibit hallmark prediction-error responses in which activity shifted from the unconditioned stimulus (US) to the conditioned stimulus (CS), was suppressed when an expected US was omitted, and increased when the US exceeded expectations. Strikingly, the same neurons also exhibited slower state-like dynamics across trials, including tonic activity transitions that emerged with learning and tracked the acquisition of learned behavior, and perturbation-related dynamics that were briefly disrupted when expectations were violated. Increasing task demands by inserting a temporal gap between the CS and US (trace conditioning) delayed both response types to later trials, accompanied by corresponding delays in behavior acquisition. In addition, dopaminergic activity developed an anticipatory response that tracked the expected timing of the US during the gap. These findings reveal that a single dopaminergic neuron type integrates moment-to-moment prediction errors, expected outcome timing, and a longer-timescale signal reflecting learning stabilization, establishing Drosophila as a bona fide model for dissecting the neural mechanisms that shape learning under changing task demands.