Extended Data Fig. 2: Low-order statistics do not explain differences in Tg + TAT-Cntrl mice from other groups.

(A) The number of neurons detected per animal does not differ significantly between groups. 2-way ANOVA, no significant main effects of Genotype or Treatment, no significant Genotype ⨯ Treatment interaction. (B) Peri-event plot of GCaMP activity of all neurons in seconds before and during freezing in the Test session (all freezing bouts aligned to Time 0, yellow dotted line) appears similar across groups. n = number of neurons. (C) Differences in number of recorded cells does not account for classifier performance across groups. All classifiers were trained on data from a randomly selected, fixed number of cells (10, 25, 100, or all cells; see Methods; Subsampling cells and frames). One-way analyses of variance on each experimental condition revealed a significant main effect of Classifier in all conditions except Tg TAT-Cntrl for all cell subsample conditions. (D) Differences in amount of freezing does not account for classifier performance across groups. All classifiers were trained on data from a randomly selected 300 or 1000 freezing frames and an equivalent number of non-freezing frames (see Methods; Subsampling cells and frames). One-way analyses of variance on each experimental condition revealed a significant main effect of Classifier in all conditions except Tg+-TAT-Cntrl for all frame subsample conditions. WT+TAT-Cntrl n = 10, WT+TAT-GluA2_3Y n = 6, Tg+TAT-Cntrl n = 4, Tg+TAT-GluA2_3Y = 7. (E) Differences in amount of freezing and number of cells together do not account for classifier performance across groups. Combining approaches from C and D, all classifiers were trained on data from a randomly selected 25 cells, comprising 1000 freezing frames and an equivalent number of non-freezing frames (see Methods; Subsampling cells and frames). One- way analyses of variance on each experimental condition revealed a significant main effect of Classifier in all conditions except Tg+TAT-Cntrl for all frame subsample conditions. WT+TAT-Cntrl n = 10, WT+ TAT-GluA2_3Y n = 6, Tg+TAT-Cntrl n = 4, Tg+TAT-GluA2_3Y = 7. (F) Test subsampling approach does not account for classifier performance across groups. To account for imbalances between classes in the test dataset, we compared a balanced accuracy approach to a balanced subsampling approach (see Methods; Balanced subsampling procedure). There were no significant differences between sampling types in any pairing. (G) Different linear classifiers have equivalent performance decoding freezing behavior. Linear discriminant analysis (LDA), logistic regression (LR) and linear support vector machine (SVM) classifiers showed no difference in balanced accuracy across groups. A 3- way ANOVA revealed no significant main effect of Classifier, and no significant interactions between Classifier and Genotype and/or Treatment. (H) Normalizing the high overall cell activity in Tg+TAT-Cntrl mice to WT+TAT-Cntrl levels does not increase accuracy of high-capacity NN classifier in decoding freezing behavior in Tg mice. (Left) Average cell activity in WT+TAT-Cntrl mice (dark gray bars), Tg+TAT-Cntrl mice (orange bars) and in Tg+TAT-Cntrl mice in which highly active cells were removed (light gray bar) such that average cell activity does not differ from that observed in WT +TAT-Cntrl mice. Boxplot presented as 1.5 × the interquartile range (whiskers), 25^th and 75^th percentile (box) and median (center line). (Right) NN classifier accuracy in decoding freezing behavior is higher in WT mice (dark gray bar) than Tg mice with all cells considered (orange bar). Removing cells with high activity from Tg mice such that overall cell activity does not differ between Tg and WT mice does not increase NN classifier accuracy (light gray bar). Therefore, normalizing cell activity in Tg+TAT-Cntrl does not improve the ability of the NN classifier to decode freezing behavior, suggesting that high neuronal activity alone cannot account for the relatively poor accuracy of the NN classifier in Tg mice relative to WT mice. n = number of neurons. (I) No difference in levels of pairwise cell correlation during test session across mouse groups (average pairwise correlation, no significant main effects of Genotype or Treatment, no significant Genotype ⨯ Treatment interaction.). (J) Eliminating coordinated population information impairs high-capacity classifier performance. Training classifiers on single-cell data (see Methods; Single-cell classifier analysis) reduces classification accuracy compared to classifiers trained on population data. (K) Differences in numbers of cells or amount of freezing between groups does not explain performance of single-cell classifiers. Training data was limited to a randomly selected 25 cells and 1000 frames of freezing and non-freezing data in all animals (as in E, see Methods; Subsampling cells and frames). Single-cell classifiers (see Methods; Single-cell classifier analysis) showed significantly worse performance than full classifiers when trained on this restricted dataset. † = p = 0.054. All data presented as mean ± SEM. Details of statistical analysis are listed in the Methods. *p < .05, **p < 0.01, *** p < 0.001, n.s. = not statistically different. In panels A, C-G & I-K, n = number of mice. In panels A,C,F,I-K, group sizes: WT+TAT-Cntrl n = 10, W +TAT-GluA2_3Y n = 7, Tg+TAT-Cntrl n = 6, Tg+TAT-GluA2_3Y = 7.