Dynamic Reorganization of Neuronal Activity Patterns in Parietal Cortex

Abstract

Neuronal representations change as associations are learned between sensory stimuli and behavioral actions. However, it is poorly understood whether representations for learned associations stabilize in cortical association areas or continue to change following learning. We tracked the activity of posterior parietal cortex neurons for a month as mice stably performed a virtual-navigation task. The relationship between cells' activity and task features was mostly stable on single days but underwent major reorganization over weeks. The neurons informative about task features (trial type and maze locations) changed across days. Despite changes in individual cells, the population activity had statistically similar properties each day and stable information for over a week. As mice learned additional associations, new activity patterns emerged in the neurons used for existing representations without greatly affecting the rate of change of these representations. We propose that dynamic neuronal activity patterns could balance plasticity for learning and stability for memory.

Keywords: chronic imaging; in vivo calcium imaging; learning; mouse parietal cortex; navigation; optogenetics; population dynamics; virtual reality.

Figure 1. Chronic imaging during stable performance of a virtual-navigation decision task

(A) Schematic of the task. (B) Behavioral performance for five mice. (C) Example imaging plane with a subset of cells identified across days in color. See Figure S1 for PPC coordinates. (D) Example cells over weeks. Left columns: mean fluorescence image. Middle columns: deconvolved fluorescence signal on correct white cue-left turn and black cue-right turn trials. Right columns: mean activity on correct white cue-left turn (blue) and black cue-right turn (red) trials. See Figure S2 for cell identification protocol. (E) Sorted peak-normalized mean activity of neurons with a significant peak of activity, combined from one day each for 5 mice. (F) Left: Task performance on optogenetic inactivation and control trials for 4 mice, combined across days. Right: Task performance for inactivation and control trials within 7 day time bins. Error bars: mean ± sem. n = 3,3,4,2,1 mice for the bins, respectively. *** p < 0.001, control vs. inactivation trials, permutation test. Control mouse not expressing ChR2: p = 0.18.

Figure 2. Reorganization of activity within a trial across days

(A) Normalized mean activity of neurons identified on all three imaging days with a statistically significant peak in the sorted day. Sorting was the same for each day within a row and was different across rows. (B) For cells with a highly significant (99% confidence) peak of activity on a given day, the fraction of cells that had a significant (95% confidence) peak of activity at a similar location (< 70 cm shift) on a subsequent day (p < 10⁻⁸ vs. time, ANOVA). Shading: mean ± sem. n = 5 mice, except for large intervals (Figure S1A). The gray area indicates 95% confidence intervals of chance levels based on shuffling the cell IDs separately on each day. (C) Fraction of cells with a significant peak. p = 0.85 vs. time, ANOVA. Error bars: mean ± sem. n = 5,5,4 mice, respectively. (D) Left: For cells with a significant peak on day n and day n+x, the fraction of peaks that shifted by greater than 0.35 m, 0.5 m and 1 m. Fraction moved 0.35 m vs. time: p = 0.019, ANOVA. Center: For cells with a highly significant peak on day n, the fraction of cells that did not have a significant peak on day n+x. Fraction lost vs. time: p < 10⁻⁹, ANOVA. Right: For cells without a significant peak of activity on day n, the fraction of cells with a highly significant peak on day n+x. Fraction gained vs. time: p = 0.96, ANOVA.

Figure 3. Reorganization of information about trial-type across days

(A) Decoding accuracy for trial type based on the activity of individual neurons, sorted by Day 1. (B) For an example day, cells were sorted by their trial-type decoding accuracy. Decoding accuracy is shown the day used for sorting (black) and on subsequent days (gray). (C) For the cells that were identified on ≥ 15 days, the fraction of days in which a cell’s activity had above chance decoding accuracy, relative to the number of days in which the cell was identified. Chance: p < 0.05, permutation test. (D) Given high confidence for significant decoding accuracy on day n (99% confidence), the fraction of cells with greater than chance decoding accuracy (95% confidence) on day n+x. p < 10⁻¹¹ vs. time, ANOVA. Error bars: mean ± sem. n = 5 mice. Gray area: 95% confidence intervals of chance levels based on shuffling cell IDs separately on each day. (E) Decoding accuracies for cells with the top and bottom 20% decoding on day n over time. (F) Trial type preference for the cells in panel A sorted by trial type preference on day 1, based on decoding model weights averaged over spatial bins in the maze. (G) Example cell with dynamic trial-type information. Top: mean fluorescence image. Bottom: mean activity on correct white cue-left turn (blue) and black cue-right turn (red) trials. (H) Decoder weights for the cells with the 20% largest and 20% smallest decoding weights on day n over time.

Figure 4. Using a GLM to track changes in neuronal activity-behavior relationships across days

(A) For each neuron on each day, a GLM was fit to the activity of the neuron based on behavior features. Model coefficients for behavioral features were fit on one day and applied to behavioral data from another day to predict neuronal activity across days. See Figure S3–6. (B) For an example cell, mean activity for white cue-left turn (blue) and black cue-right turn (red) trials on imaging days 3 and 9, with model predictions (gray) for fitting and testing on the same or opposite days. (C) Left: Deviance explained by models fit on a one day and tested on the same or a different day, averaged across cells and then mice. See Figure S1C for n values. Right: Average deviance explained as a function of time between the fitting and testing days. Shading: mean ± sem. (D) Schematic for model comparisons binarized as significant and poor predictions; threshold of 0.2 deviance explained, chosen based on a bootstrap analysis (Methods). (E) Left: For cells without a significant model prediction on day n, the fraction of cells with a significant model prediction after a given interval. Right: For cells with a significant model prediction on day n, the fraction of cells with consistent (black), lost (medium gray), or switched (light gray) activity-behavior relationships after a given interval. Shaded area: mean ± sem. n = 5 mice, except for large intervals (Figure S2K). (F) Mean activity for example cells with varying consistency in activity-behavior relationships. (G) Left: For example cells from panel F, fitting and testing comparisons as in panel D. Right: Fraction of models with significant predictions as a function of time between the fitting and test days. Exponential fits are shown. (H) Histogram of the fraction of significant model predictions after 10–20 days between fitting and training days for cells identified on ≥ 15 days. n = 690 cells. (I) Contribution of different categories of behavior features to neuronal activity, estimated as the standard deviation of the linear part of the model (Methods). Comparisons of cells with the 20% most and 20% least consistent models: Position/cue, p = 0.026; treadmill velocity, p = 0.98; whether the mouse was in the inter-trial interval, p = 0.79; t-test. Error bars: mean ± sem. n = 5 mice. Gray lines: 95% confidence intervals for randomly selected cells. (J) For a given day, how many previous days the model of that day’s activity provided a good prediction of previous days’ activity. Models with significant predictions on ≥ 2 consecutive previous days were more likely to provide a good prediction in future days than models with significant predictions on only 1 previous day. * p < 0.05, t-test.

Figure 5. Stable statistical features of population activity

(A) For two example mice, mean population activity vs. maze position. (B) Ratio of activity in the cue period to the delay period. (C) For two example mice, population decoding accuracy of trial type vs. maze position. Separate decoders were trained at each spatial bin and on each day. (D) Population decoding accuracy of trial type. (E) For two example mice, distributions of cell-cell correlations of deconvolved calcium signals smoothed with a 2-second sliding window. (F) Summary of cell-cell correlation distributions. Boxes: 25th and 75th percentiles; white dots: mean; whiskers: 99% range. (G–L) Same as in panels E–F, except for correlations of population activity (cells × maze position) in trials of the same type (G–H), population activity event rates (I–J), and classification accuracy of trial type based on single-cell activity (K–L).

Figure 6. Decoding task information across days

(A) Population decoding accuracy of trial type on correct trials as a function of position in the maze. Separate decoders were trained at each spatial bin and on each day. Shading: mean ± sem. n = 5 mice. (B) Decoding accuracy of trial type on correct trials using 20 cells with the least (yellow) or most (purple) consistent activity-behavior relationships or 20 randomly selected cells (gray). Error bars: mean ± sem. n = 5 mice. * p < 0.05, t-test, between least and most consistent groups. See Figure S7 for inactivation experiments. (C) Decoders trained on a given session and tested on a later session using 20 cells. Error bars: mean ± sem. n = 5,5,4,4,2 mice for Δdays 0,2,5,10,20 respectively. * p < 0.05, t-test, between least and most consistent groups. (D) Decoders trained on a given session and tested on a later session using 30, 50 or 100 randomly selected cells. Decoding performance vs. number of neurons used for decoding: p < 10⁻³, ANOVA.

Figure 7. Neuronal activity during learning of a novel trial type

(A) Mouse perspective of virtual maze trial types. (B) Behavioral performance on 2 days preceding and 4 days following introduction of a novel trial type. Error bars: mean ± sem. n = 4 (2 mice with 2 novel trial types each). (C) Population decoding performance of different trial types during the cue, delay, and turn periods. Error bars: mean ± sem, n = 2 mice. (D) Example population activity during cue, delay, and turn periods in the same dimensionality reduced space (one mouse, four trial types). Dots: single trials. (E) Fraction of cells with consistent (black), lost (medium gray), switched (light gray) activity-behavior relationships within one session based on our GLM analysis on days with only familiar trials or with novel trial types. Dots: single sessions. First/Second Half, Consistent: p < 10⁻³, Lose: p < 10⁻³, t-test. (F) For models fit to activity-behavior relationships on one day, deviance explained of predictions of activity on other days. p = 0.91, ANOVA, comparing linear regression slopes for two-trial-type and novel-trial-type days. (G) Left: For individual cells, novel cue-related activity on day n (measured by > 0 contribution for novel cue onset/offset in the GLM) vs. familiar cue-related activity on day n-20. Right: For the same cells, novel cue-related activity on day n vs familiar cue-related activity on day n. (H) For cells with novel cue related activity on day n, the fraction of cells with activity related to either white or black cue onset/offset on the same day (day n) and on previous days. n = 4 circles (2 mice with 2 novel trial types each); bars: means; shading: 95% confidence interval from a random subset of neurons.