Extended Data Fig. 6: The first-phase responses of midbrain DANs encode a distribution that is similar across cues and closer to the prior distribution of the rewards in the task than would be expected by chance.

a, Raster aligned to odour valve opening for two example neurons. The green shaded area depicts the window used to compute the first-phase responses (50–200 ms) and the blue to the second phase (300–450 ms). b, Second-phase responses vary more with the upcoming delay than first-phase responses. First-phase mean responses across the 1.5-s, 3-s and 6-s cues for different neurons as a function of the second-phase mean responses. The slopes are the fitted linear regression models for each neuron. Points and slopes are colour-coded by the estimated temporal discount factor. Inset: distribution of slopes across neurons. The variation in first-phase responses is smaller than in second-phase responses, because the mean absolute slope (vertical line) is smaller than one: two-tailed p-value=0.001, 95% CI =(0.0099,0.71) bootstrapped n=10,000 times. c, Decoded joint density of reward over magnitude and time, using the first (left) and second (right) phase population responses aligned to the different cues. d, Decoded density over reward time using the first-phase dopamine population responses for all cues. The grey lines depict the decoded density when the population temporal discount factors are shuffled. The light lines represent decoded densities using the responses of 70% of randomly selected trials and the thicker lines represent the mean decoded densities. e, 90% CI of the mean Kullback–Leibler (KL) divergence between the true prior distribution of reward times and magnitudes in the task and the distribution decoded from DAN activity. This comparison considers population tuning for reward time and magnitude, either preserved or shuffled, across 100 decoder runs. Data underlying the figure can be found in the Supplementary Data.