Cue coding models

To assess cue coding schemes, we fit a new set of models focusing on a more restricted time window (−1–2.5 s from cue onset) using only cues and licks as predictors. Cue and lick neurons were identified as before, and subsequent cue characterization was performed on neurons with only a unique contribution of cues. To identify value coding among cue neurons, we fit a new kernel models with a single cue kernel that scaled according to the cue as well as six block constants (as above) with full rank. We inputted cue values as 1, 0.5, and 0 for each CS+, CS50, and CS−, respectively, ranked according to their reward probability. We fit 152 additional models with alternative configurations of cue value: all permutations of 1, 1, 0.5, 0.5, 0, 0 across the six cues, as well as all permutations of high responses (1) for 6 (we call this the ‘untuned’ model), 5, 4, 3, 2, or 1 cues, with other cues set to 0. Among the 153 total models, some were more similar to the ranked value model, which we quantified by correlating the six cue values of each alternative model with the ranked model. We termed all models with a correlation greater than 0.5 as ‘value-like.’ For each neuron, we found the model that best explained its activity. The models, their correlation with value, and the proportion of cue neurons best fit by each model are illustrated in Figure 3—figure supplement 1. To verify the robustness of value coding in the neurons best fit by the ranked value model, we fit each of those neurons with 1000 iterations of the cue value model with shuffled cue order to create a null distribution. The fits of the original value model exceeded the 98^th percentile of the null for all value neurons.