Hippocampal representations drift in stable multisensory environments

Abstract

Experiments that track hippocampal place cells in mice navigating the same real environment have found significant changes in neural representations over a period of days^1,2. However, whether such 'representational drift' serves an intrinsic function, such as distinguishing similar experiences that occur at different times^3,4, or is instead observed due to subtle differences in the sensory environment or behaviour^5-7, remains unresolved. Here we used the experimental control offered by a multisensory virtual reality system to determine that differences in sensory environment or behaviour do not detectably change drift rate. We also found that the excitability of individual place cells was most predictive of their representational drift over subsequent days, with more excitable cells exhibiting less drift. These findings establish that representational drift occurs in mice even with highly reproducible environments and behaviour and highlight neuronal excitability as a key factor of long-term representational stability.

Extended Data Figure 1:. Running behavior of individual mice during training and imaging days, and cell type examples and additional representational drift measurements

a) Left: The speed profile (m/s) of five mice performing the task in a familiar environment over five training and five imaging days. Each column shows a single-track traversal lap. The horizontal dash line indicates the water reward location, and the vertical white lines separate each day. Right: The speed correlation of laps 11–30 on each day to the mean lap from laps 11–30 on the reference day (cyan, n=20 laps per n=5 mice, error bars indicate SEM). Horizontal dashed line indicates the correlation threshold. b-c) The average speed profile of each mouse versus track position across days (b) and their mean (n=6 mice) (c). d-e) The average lick profile of each mouse versus track position across days (d) and their mean (n=6 mice) (e). f,g) Example cells showing stable (f) and unstable (g) spatial tunings across days. Scale bar=10 um. The confidence intervals are the lap-by-lap SEM. h) Quantification of total number of active neurons (left), fraction of active neurons that are place cells (middle), and spatial information of place cells (right), two-way ANOVA across animals and days. The thick black line shows the average. i) Tuning curve (TC) correlations for place cells within day. Odd versus even laps (black) are separated from 10,000 randomly chosen pairs of place cells (gray) at 0.4 threshold (dashed line). j) Representational drift measured by place field peak shift, Bayesian decoding error, Frobenius norm, representational similarity, and representational drift index, two-way ANOVA across animals and days. The grey lines indicate the random shuffle mean, and the confidence interval of the shuffle indicates standard deviation (std).

Extended Data Figure 2:. Running behavior and representational drift measurements for similar and dissimilar sets

a) The speed profile of mice across days; laps selected based on similar or dissimilar running sets. b-c) The average speed profile of the similar and dissimilar running laps of each mouse versus track position across days (b) and their mean (n=6 mice) (c). The confidence intervals indicate animal-by-animal SEM. d) Representational drift of similar (blue) and dissimilar (red) running sets measured by various measurements (three-way ANOVA across animal, day and set; lap set effect, n.s. p>0.05) The shaded region shows SEM.

Extended Data Figure 3:. Running behavior and representational drift measurements for flat, variable, and spatial odor tasks, and odor concentration measurements for variable odor task

a-f) The average speed and lick profiles of each mouse versus track position across days in flat odor (a), variable odor (c), and spatial odor (e) tasks and their mean (b, d, and f). g) Photoionization detector (PID) measurements acquired from the olfactometer nose cone for the five-day protocol of the variable odor task described in Fig. 3c. Day 1 of the variable odor protocol is the same as the flat odor protocol in Fig. 3b. The measurements were performed while a mouse was running on a treadmill in a virtual environment. In the inter-trial-interval, after the mouse completed one lap and before the start of the next, the odor concentration was set to 0. h) Representational drift of various odor mouse groups measured by various measurements. The shaded region shows SEM. The three groups have little or no statistical difference by non-directional multiple repeated measures ANOVA comparisons (see Supplementary Information Table 1 for all comparisons).

Extended Data Figure 4:. Licking behavior and representational drift measurements for most selective and non-selective pre-lick sets

a-d) The lick profile of mice across days for most selective (a-b) and non-selective (c-d) pre-lick laps. The tick black curve shows the average lick rate of all mice, and the red dashed line indicates reward location (b and d). e-f) Cross-validated heatmaps of 397 cells (6 mice) identified as place cells on Day 1, calculated based on most selective (e) and non-selective (f) pre-lick sets. g) Representational drift of most selective (blue) and non-selective (red) pre-lick sets measured by various measurements, three-way ANOVA across days, animals, and set (lap set effect). The shaded region shows SEM.

Extended Data Figure 5:. Representational drift measurements vs pre-lick index across all mice, and examples of hippocampal internal states and spatial tunings

a) Plots of representational drift measurements vs pre-lick index across all mice (n=30 mice) for five different tasks. The x-axes are the average prelick index across all days and the y-axes are the difference between the first and last day values for each drift measurement. Pearson correlation ρ and p-values from two-tailed Pearson correlation test are listed above each panel. b-d) Internal and spatial tuning curves of hippocampal neuronal activities for example mice from uncontrolled odor (b), flat odor (c), and variable odor (d) tasks. Top: Internal and spatial tuning of hippocampal population activity in PCA space. Bottom: Cross-validated heatmaps of cells calculated based on internal state and position.

Extended Data Figure 6:. Subtle visual sensory variability does not detectably affect hippocampal representational drift in mice.

a-b) In variable visual task, VR brightness varies lap-by-lap at different levels within and across days). c-d) Running and licking behavior of individual mice (c) used in the task and their averages (n=7 mice) (d). e-f) Within-day speed vector correlations for each of the five days (e, non-directional repeated-measures ANOVA, task by day interaction, F(4,44)=0.19. p=0.94) and across-day PCA space distance (f, two-sided Wilcoxon rank-sum test, p=0.95) (n=7 mice, n.s., p > 0.05). g) Cross-validated heatmaps of 344 cells (7 mice) identified as place cells on Day 1. h) Representational drift of variable visual (pink) and non-variable visual (black) mouse groups measured by various measurements (n=7 mice, n.s., p > 0.05. Non-directional repeated measures ANOVA task by day interaction. The shaded region shows SEM).

Extended Data Figure 7:. Examples of stable and unstable hippocampal place cells over days

Stable (a) and unstable (b) place cell examples. Top: normalized DF/F trace for the first 10 laps on the emerge day (black) and the two following days (dark and light gray, respectively). Each row is a lap. Middle: mean DF/F vs position (tuning curves) over the three days. Bottom: DF/F time series of each example neuron.

Extended Data Figure 8:. Properties of stable and unstable place cells within and across days

a) Fractions of total pooled neurons by type (n=8014 cells). Unclassified place cells were those that were not categorized into stable or unstable cell categories. b) Additional measures of representational drift for stable and unstable place cells across days. c) Additional excitability properties of stable and unstable neurons. d-f) Cumulative distributions of cell-by-cell excitability properties (d), spatial properties (e), and signal quality (f) for stable and unstable place cells. g-i) Excitability properties (g), spatial properties (h), and signal quality (i) for stable and unstable place cells across days. Two-sided Wilcoxon signed-rank p values (panels b-i) and Cohen’s d (panel c-e) are listed for each feature.

Extended Data Figure 9:. Multicollinearity among various hippocampal neuronal properties of place cells

Pairwise comparisons of each neuronal property with other neuronal properties used in the logistic regression model. Each data point is a single neuron. Pearson’s correlation values from two-tailed Pearson correlation test for each pair of properties are shown above the panels.

Extended Data Figure 10:. Higher neuronal excitability correlates with and is predictive of more stable representation in hippocampal CA1 across all active cells

a) Tuning curve (TC) correlations for all active cells within day. Odd versus even laps (black) are separated from 10,000 randomly chosen pairs of all active cells (gray) at 0.25 threshold (dashed line). This threshold is used for classification of cells as recurring and therefore, stable (same analysis as the Fig. 4 and Extended Data Figure 1i, but using all active cells instead of only place cells). b-c) Cross-validated heatmaps for all stable and unstable cells from the emerge day and two following days. d-f) Representational drift of stable and unstable cell populations across days (each point represents a mouse, n=30 mice) measured by recurrence probability (d), PV correlation (e), and TC correlation (f) (d-f, two-sided Wilcoxon signed-rank test). g-i) Stable vs. unstable cell excitability properties (g), spatial properties (h), and signal quality (i). Each point is the emerge day average for each mouse (n=30 mice, two-sided Wilcoxon signed-rank tests and Cohen’s d). J-k) Average dF/F (j) and baseline noise (k) from the emerge day and two following days for stable and unstable cell populations (n=30 mice, two-sided Wilcoxon signed-rank test). l) Logistic regression classification of stable vs. unstable cells. Confusion matrix of fitted regression model predicted whether an active cell becomes a stable or unstable cell from emerge day properties (n=50 repeats of logistic regression tests with different training and test set combinations; 100 stable, 100 unstable neurons; mean shown on matrix, stable cell true positive std=2.4%, stable cell false positive std =2.1%, unstable cell true positive std=2.1%, unstable cell false positive std =2.4%). m) Accuracy of logistic regression prediction when each group of properties left out (n=50 repeats of leave-out logistic regression tests with different training and test set combinations; middle values are medians, and vertical bars are 25th and 75th percentiles. None vs. excitability: two-sided Wilcoxon rank-sum test, p=6.8e-18. None vs. spatial: p=0.98. None vs. signal quality: p=0.41. * p-value <0.05).

Figure 1:. Hippocampal representational drift occurs in a familiar virtual environment

a) Experiment setup: head-fixed mouse running on cylindrical treadmill in front of VR display and receiving water rewards during 2P imaging. b) Mouse navigates 3-m one-dimensional virtual track. c) Imaging dorsal CA1 neurons using GCaMP8m through hippocampal window. d) Speed profile of trained exemplar mouse performing task in familiar environment over five training and imaging days. Each column shows single-track traversal lap. Horizontal dashed line and blue droplet indicate water reward location, and vertical white lines separate days. e) Speed correlation of laps 11–30 on each day to mean of laps 11–30 on reference day (cyan). Horizontal dashed line indicates correlation threshold (n=20 laps from 1 mouse, error bars indicate SEM). f-i) Across-day volumetric plane registration method. f) Z-stack images (15 slices, 2 μm apart) were obtained around Day 1 imaging plane. Maximum cross-correlation values for the 15 z-planes to Day 1 imaging plane were calculated and peak selected as the matching plane (red dot). g) Cross-correlation of spatial footprints for all ROI pairs aligned across days. h) Histogram of spatial footprint correlations for all ROI pairs aligned across days. Cross-cell shuffles shown in gray. i) Selected imaging planes in example mouse across five days. Cell ROIs identified using Suite2p, aligned across days using CellReg. Three example ROIs identified on all five days highlighted in cyan. j) Three example place cells from (i) showing heterogeneous place tuning curves across days. Mean DF/F versus track position shown for each cell across days. First place cell stable over days, second place cell field emerges day 3, third place cell field shifts location across days. Morphologies of the cells shown in insets, confidence intervals show lap-by-lap SEM. k) Sorted and cross-validated population heatmaps of 391 cells (6 mice) identified as place cells on Day 1. Each horizontal row shows mean DF/F of each neuron vs track position. Day 1: rows sorted based on place field locations on odd laps, even lap mean DF/F is plotted. Days 2–5: same 391 cells sorted according to day 1 order, plotting mean DF/F for even laps on each indicated day. l-n) Population representational drift measured by recurrence probability (l), population vector (PV) correlation (m), and tuning curve (TC) correlation (n). Thin line, each mouse; thick line, mean across mice; grey line, random shuffle.

Figure 2:. Hippocampal representational drift persists even with highly reproducible behavior

a-b) Speed profile of traversal laps of example mouse across days, selected based on “similar” (a) or “dissimilar” (b) lap sets. 20 laps for each imaging day selected for each subset. c-d) Principal component analysis (PCA) space representation of laps (position binned speed vectors) of example mouse in a-b for “similar” (c) and “dissimilar” (d) lap sets. Each dot represents single lap speed vector, and colors show different days. e-f) Within-day speed vector correlations for each of the five days (e, n=6 mice, three-way ANOVA across all laps, similar and dissimilar sets, animals, and days; set: F(2,78)=21.5, p=3.7e-8; Tukey-Kramer post-hoc test p = 0.61. Lines show average across mice and SEM, n.s. p>0.05) and across-day PCA space distance (f, n=6 mice, two-way ANOVA across all days, similar and dissimilar sets and animals; similar and dissimilar sets: F(2,10)=190.1, p=1.1e-8; Tukey-Kramer post-hoc tests, shuffle laps vs. similar laps, p=2.1e-6. Shuffle laps vs. dissimilar laps, p=1.6e-5. Similar laps vs. dissimilar laps, p=6.9e-9. *indicates p<0.05). g-h) Cross-validated heatmaps of 391 cells (6 mice) identified as place cells on Day 1, plotted on similar and dissimilar laps. Day 1: rows sorted based on place field locations across all laps, and only similar (g) or dissimilar (h) sets mean DF/F plotted. Days 2–5: same 391 cells sorted according to Day 1 order, plotting mean DF/F for similar (g) or dissimilar (h) lap sets on each day. i-k) Representational drift measured by recurrence probability (i), population vector (PV) correlation (j), and tuning curve (TC) correlation (k) (i-k, three-way ANOVA, sets effect, n.s. indicates p>0.05). Shaded region: SEM.

Figure 3:. Hippocampal representational drift persists even in a highly reproducible sensory environment

a) Visual-olfactory multisensory VR provides controlled odorant through nose chamber while mouse navigates virtual environment. b-d) Schematics of odor presentation protocols across days. b) Flat odor task, 50% α-pinene presented on each lap. c) Variable odor task, different concentrations and mixtures of α-pinene and Methyl valerate odorants presented each day. d) Spatial odor task, α-pinene presented with fixed gradient profile across all laps and days. e) Within-day speed vector correlations for each of the five days in different tasks (n=6 mice in uncontrolled odor, n=5 mice in flat odor, n=6 mice in variable odor, n=6 mice in spatial odor. Lines show average across mice and SEM, non-directional repeated-measures ANOVA used. n.s. p>0.05) .f) Across-day PCA space distance in different tasks (n=6 mice in uncontrolled odor, n=5 mice in flat odor, n=6 mice in variable odor, n=6 mice in spatial odor, non-directional repeated-measures ANOVA, n.s. p>0.05). g-h) Fraction of place cells (g, n=6 mice in uncontrolled odor, n=5 mice in flat odor, n=6 mice in variable odor, n=6 mice in spatial odor. Lines show average across mice, each dot is a mouse. Two-tailed chi-squared test used. n.s. p>0.05, * p<0.05) and their spatial information (h, n=391 cells from 6 mice in uncontrolled odor, n=300 cells from 5 mice in flat odor, n=600 cells from 6 mice in variable odor, n=504 cells from 6 mice in spatial odor. Two-sided Wilcoxon rank-sum test used) across different tasks. i-k) Cross-validated heatmaps of flat odor (i), variable odor (j), and spatial odor (k) tasks. l-n) Population representational drift in different odor tasks measured by recurrence probability (l) PV correlation (m), and TC correlation (n) (l-n, non-directional repeated measures ANOVA, task by day interaction, n.s. p>0.05. Shaded region: SEM).

Figure 4:. Higher neuronal excitability correlates with and is predictive of more stable representation in hippocampal CA1 place cells

a-c) Example stable and unstable place cells. Lap-by-lap DF/F vs position (a) and mean tuning curves (b) from “emerge day” to two days after and their corresponding DF/F vs time traces (c). d-e) Cross-validated heatmaps for all stable and unstable place cells from the emerge day and two following days. f-h) Representational drift of stable and unstable cell populations across days (each point represents a mouse, n=30 mice) measured by recurrence probability (f), PV correlation (g), and TC correlation (h) (f-h, two-sided Wilcoxon signed-rank test). i-k) Stable vs. unstable place cell excitability properties (i), spatial properties (j), and signal quality (k). Each point is emerge day average for each mouse (n=30 mice, two-sided Wilcoxon signed-rank tests and Cohen’s d). I-m) Average DF/F (l) and baseline noise (m) from emerge day and two following days for stable and unstable cell populations (n=30 mice) (l,m, two-sided Wilcoxon signed-rank test). n) Logistic regression classification of stable vs. unstable place cells. Confusion matrix of fitted regression model predicted whether a place cell becomes stable or unstable place cell from emerge day properties (n=50 repeats of logistic regression tests with different training and test set combinations; 100 stable, 100 unstable neurons; mean shown on matrix, stable cell true positive std=2.7%, stable cell false positive std =2.5%, unstable cell true positive std=2.5%, unstable cell false positive std =2.7%). o) Accuracy of logistic regression prediction when each group of properties left out (n=50 repeats of leave-out logistic regression tests with different training and test set combinations; middle values are medians, and vertical bars are 25th and 75th percentiles. None vs. excitability: two-sided Wilcoxon rank-sum test, p=4.1e-17. None vs. spatial: p=0.74. None vs. signal quality: p=0.67. * p-value <0.05).