Figure 6. Trial outcomes can be accurately decoded from neural activity in lesioned and non-lesioned mice.

(A) Average decoding accuracy of logistic regression models as a function of time against dummy models with a score of 0.5 meaning chance performance and a score of 1 being the maximum. Data shown depict the mean model accuracy across 37 (lesioned) and 38 (non-lesioned) sessions, respectively. Dots at the top indicate the time points (frames) where the model performance was significantly different between trained and dummy models for non-lesioned mice (teal) or lesioned mice (orange) (p<0.05, one-sided Wilcoxon signed-rank test or paired t-test with Bonferroni correction, depending on whether normality assumption was met), and between the trained models for non-lesioned vs lesioned mice (blue) (p<0.05, one-sided Mann-Whitney U test or t-test with Bonferroni correction, depending on whether normality assumption was met). (B) Same as (A) but the average model accuracy is plotted separately for mice with (near-)complete (22 sessions) and partial lesions (15 sessions). Dots at the top indicate the time points where the model performance was significantly different between partial vs (near-)complete mice (purple), (near-)complete vs non-lesioned mice (blue), and partial vs non-lesioned mice (red) (p<0.05, one-sided Mann-Whitney U test or t-test with Bonferroni correction, depending on whether normality assumption was met). Shaded areas represent 95% confidence intervals.

Figure 6—figure supplement 1. Trial outcome decoding is not meaningfully affected by differences in sound level distributions between hit and miss trials.

(A) Decoding results for one imaging session based on trials in which stimuli were presented at five (left), three (middle), or a single sound level (right). Thin colored lines show the results of each of the five cross-validation folds. Thick colored lines indicate averages across all five folds. Gray lines show results for the corresponding dummy models. (B) Superimposed averages from (A). (C) Hit and miss trial distributions for each of the five sound levels, as well as the mean sound level difference (Δ) between hit and miss trials for the three decoding conditions shown in (A) and (B). The mean difference was 3.08 dB, 1.01 dB, and 0 dB for the five, three, and one sound level conditions, respectively.

Figure 6—figure supplement 2. Greater number of recorded neurons was not associated with better decoding performance.

(A, B) Decoding performance (balanced accuracy) of the logistic regression models averaged over different 1 s time periods relative to stimulus onset as a function of the number of neurons recorded in a given session. A greater number of neurons obtained in a field of view was not associated with better decoding performance. Values above panels indicate Spearman’s rank correlation coefficient ρ. The only statistically significant relationship between the number of recorded neurons and decoding performance was found for late trial periods in non-lesioned mice (A), and indicated that for time periods >2 s after stimulus onset a smaller sample size was associated with better decoding performance. **p<0.01.

Figure 6—figure supplement 3. Similar fractions of task-modulated and sound-driven neurons in lesioned and non-lesioned mice.

(A) Fraction of neurons per session that exhibit a significant difference in response magnitude between hit and miss trials. (B) Fraction of neurons per session that exhibit a significant stimulus response in miss trials. *p<0.01, Mann Whitney U test.

Figure 6—figure supplement 4. Lick rates in peri-catch trial periods approximate next-trial-probability.

(A) Peri-catch trial lick raster for all catch trials that followed a hit trial for one example mouse. The peri-catch trial period was defined as the period from the reward delivery in the hit trial to the onset of the trial following the catch trial. (B) Lick rate averaged across the peri-catch trial periods shown in (A) and binned into 100ms wide bins. The thick blue line shows the smoothed (20-point running average) lick rate. The inset gives a magnified view of the average lick rate during the period indicated by the gray rectangle. The red line illustrates the distribution of ‘reward-to-next-trial-onset’ intervals experienced by the example mouse. Given that licks are plotted time-locked to reward delivery, we plotted the distribution of intervals between reward delivery and the onset of the next trial rather than the inter-trial interval (ITI) distribution. In practice, the difference between the two is roughly the latency between the stimulus and the first lick and thus barely distinguishable at this scale. As the distribution indicates the probability of the next trial presentation as a function of time since the preceding reward delivery we refer to it as ‘next-trial-probability’. (C) Same as the inset in (B) averaged across all mice. Next-trial-probability was smoothed with a 20-point running average. (D) Next-trial probability as a predictor of lick rate. The dotted lines indicate the 95% confidence bounds around the regression fit. Adjusted R²=0.59. Although the next-trial probability is a good predictor of changes in the average lick rate, the lick rate at the peak of the distribution is merely about a quarter higher than at its tails where next-trial probability approaches zero. Furthermore, to put the average lick rates into perspective, note that mice tend to lick in bouts, typically consisting of two to six licks in very quick succession (see lick raster in (A)), and that, consequently, the lick rate exceeds the underlying bout rate by a factor of about four. (E) Same as (C) but with peri-catch trials binned into four quarters before averaging in order to illustrate changes in lick behavior across different stages of the experiment. (F) Same as (E) for all peri-catch trials during the initial training with a single-level stimulus. While the peri-catch trial lick rate profile changed substantially over the course of the initial training (F) and started to approximate the stimulus probability distribution towards the end of training, it remained broadly stable throughout the main experiment (E). In order to increase the statistical power of this analysis, we included data from several additional mice used in other projects. These additional mice received the same training and performed the same task, but differed from those in the main dataset in that they had a different genetic background and/or had been fitted with a cranial implant for cortical rather than midbrain imaging. N for panels (C–F) = 34 mice.