Mice can learn to lick for reward in response to a novel tactile whisker stimulus in a single session.

(A) Water-restricted mice were first trained to lick a water reward spout in response to a brief auditory pure tone (Days -2 and -1 correspond to the final two days of the auditory pre-training phase). On Day 0, the whisker stimulus (brief deflection of the C2 whisker) was introduced for the first time as an interleaved active (rewarded / unrewarded) trial type, and whisker training continued for two additional days (Day +1 and Day +2). Rewarded mice (R+, green) received a water reward if they licked in response to the whisker stimulus (whisker hit trials), whereas non-rewarded mice (R-, magenta) did not. Auditory hits (blue) were rewarded for both groups, and catch trials (black, no stimulus) were always unrewarded. At the end of each active behavioral session, a block of 50 consecutive passive whisker trials was presented. (B) Example probability of licking in blocks of 20 trials for one whisker rewarded (R+) and one whisker non-rewarded (R-) mouse across the five training days. Blue data points show licking probability on auditory trials; green (R+) and magenta (R-) data points show licking probability on whisker trials; and black data points show licking probability in no-stimulus trials. (C) Behavioral performance across days (n = 19 R+ mice and n = 16 R- mice, during functional imaging described later; dark blue, auditory lick probability in R+ mice; light blue, auditory lick probability in R-mice; green, whisker lick probability in R+ mice; magenta, whisker lick probability in R- mice; black, catch trial lick probability in R+ mice; grey, catch trial lick probability in R- mice). Thin lines indicate single-mouse trajectories and thick lines indicate averages across reward groups (left, mean ± 95% confidence interval). Bar plot quantifying whisker performance across days, comparing the two reward groups (right, mean ± 95% confidence interval; p-values from Mann-Whitney U test). (D) Whisker performance dynamics during Day 0. The top shaded bar indicates the p-value obtained from Mann-Whitney U tests comparing reward groups at each whisker trial, corrected for multiple testing with the Benjamini-Hochberg procedure. Data points show the mean Bayesian lick probability per trial ± 95% confidence interval. Divergence between the two groups first became significant after 22 whisker trials.

Pharmacological and optogenetic inactivation of barrel cortex prevent learning.

(A) Schematic illustrating muscimol inactivation of the barrel cortex (wS1) or the forepaw primary somatosensory cortex (fpS1). (B) Behavioral performance across days during wS1 muscimol inactivation (left; n = 13 R+ mice) and fpS1 muscimol inactivation (right; n = 8 R+ mice). Inactivation days, Days 0, +1, +2, are indicated in red; recovery on Days +3, +4, +5. (C) Statistical comparison of whisker performance during muscimol inactivation for wS1 and fpS1 (Mann-Whitney U test). (D) Schematic illustrating spatiotemporally-specific inactivation through optogenetic activation of local inhibitory neurons. (E) Behavioral performance across days during optogenetic inactivation of wS1 (left; n = 6 R+ mice) and fpS1 (right; n = 6 R+ mice) on Day 0 and recovery on Day +1. (F) Statistical comparison of whisker performance during optogenetic inactivation for wS1 and fpS1 (Mann-Whitney U test).

Learning induces a rapid reorganization of barrel cortex population activity.

(A) Two-photon calcium imaging of layer 2/3 neurons expressing GCaMP6f in the barrel cortex of Rasgrf2-Cre × Ai148 mice. Top: example field of view with individual cell ROIs color-coded by their learning modulation index (LMI: red, positive; blue, negative). Bottom: calcium traces in response to passive whisker trials for example cells. Orange vertical bars indicate whisker stimulus onset. (B) Heatmap of the trial- averaged whisker-evoked responses in the post-behavior passive epoch across the five imaging days (Day -2 to Day +2) for cells significantly modulated by learning (same example mouse as panel A) ordered by their LMI. (C) Population PSTHs over passive whisker trials across the five imaging days for R+ (green) and R- (magenta) mice. Thick lines show grand averages across mice (n = 19 R+ mice; n = 16 R- mice) and shading indicates 95% confidence interval. (D) Comparison of population whisker response amplitude (ΔF/F₀, averaged over 0-300 ms from stimulus onset) before (Days -2 and -1; pre) and after (Days +1 and +2; post) learning in R+ mice (n = 19 mice). Left: population PSTHs for pre- (grey) and post- (green) learning. Right: bar plot quantification of individual mouse averages over the same time window (Wilcoxon signed-rank test). (E) As in panel D, but for R- mice (n = 16 mice). (F) Distribution of LMI across all imaged neurons for R+ mice (green; n = 3,210 neurons) and R- mice (magenta; n = 2,846 neurons) (Kolmogorov–Smirnov test). (G) Proportion of neurons with a significant positive LMI (left) and significant negative LMI (right), compared between R+ and R- mice (Mann–Whitney U test). (H) Trial-by-trial population similarity matrices averaged across R+ (left) and R- (right) mice. Each entry represents the mean cosine similarity between neural population activity vectors from the indicated pair of days, computed using responses to passive whisker trials averaged over 0-300 ms from stimulus onset. (I) Average cosine similarity between trials from the same day compared between R+ and R- mice (Mann–Whitney U test). Thin lines indicate individual mice, thick lines average ± 95% confidence interval. (J) Network reorganization index quantifying the difference in cosine similarity between pre-learning (Days -2 and -1) and post-learning (Days +1 and +2) population representations for R+ and R- mice. Points indicate individual mice, bar graph shows mean ± 95% confidence interval. (K) Schematic of the linear decoding approach. A logistic regression classifier was trained to discriminate pre-learning (Days -2, -1) from post-learning (Days +1, +2) passive whisker trial population activity. (L) Pre- vs post-learning decoding accuracy for R+ and R- mice, with 10-fold stratified cross-validation. Dashed line indicates chance level (50%). Points indicate individual mice, bar graph shows mean ± 95% confidence interval (Mann–Whitney U test). (M) Pairwise day decoding accuracy matrices using the fixed pre- vs post-learning decoder, for R+ (left) and R- (right) mice. Each entry shows the mean decoding accuracy when the classifier is applied to classify trials from the indicated pair of days. (N) Decoding accuracy comparing pre-learning days (Days -2, -1) vs Day 0 (Pre) and post-learning days (Days +1, +2) vs Day 0 (Post) for R+ and R- mice. Dashed line indicates chance level (50%). Points indicate individual mice, bar graph shows mean ± 95% confidence interval (Wilcoxon signed-rank test).

Progressive realignment of barrel cortex representations and online reactivations.

(A) Schematic of the Day 0 decoding approach. The pre- vs post-learning classifier (trained on passive whisker trials from Days -2 and -1 vs Days +1 and +2) was applied to active whisker trials during Day 0 to compute a trial-by-trial projection on the learning axis. (B) Example Day 0 trajectories for one R+ mouse (left, green) and one R- mouse (right, magenta). Top: whisker hit probability, estimated by a Bayesian state-space model fitted to binary hit/miss outcomes. Bottom: projection on learning axis applied to the same whisker trials. (C) Average Day 0 trajectories across all mice for R+ (left, green) and R- (right, magenta) groups. Top: licking performance across whisker trials. Bottom: projection on the learning axis across whisker trials. Thick lines report averages across mice with shading indicating the 95% confidence interval. (D) Linear slope of the projection on the learning axis over Day 0 whisker trials, between R+ and R- mice. Small data points show individual mice, and big points show mean ± 95% confidence interval (Wilcoxon one-sample test). (E) Pearson correlation between the projection on the learning axis and whisker performance across Day 0, between R+ and R- mice. Small data points show individual mice, and big points show mean ± 95% confidence interval (Wilcoxon one-sample test). (F) Example reactivation heatmap for a R+ mouse on Day 0. Left strips: Learning modulation index (LMI); average response to passive whisker trials (reactivation template); and reactivation participation rate, for the top 20 cells ranked by participation rate. Right: neural activity heatmap of 180 seconds of concatenated catch trial data. Top: template correlation trace with detection threshold of 0.25 (horizontal dashed line); red vertical lines indicate detected reactivation events. (G) Example template correlation traces across all five training days for the same R+ mouse as in panel F. Red vertical lines indicate detected reactivation events and grey dotted lines indicate 0 correlation. (H) Reactivation frequency across training days for R+ (green) and R- (magenta) mice. Bar graph shows mean ± 95% confidence interval (Mann–Whitney U test). (I) Scatter plots of Day 0 reactivation participation rate versus LMI for individual neurons, shown separately for R+ (left, green) and R- (right, magenta) mice. Each point represents one neuron. Linear regression line is overlaid (Pearson correlation). (J) Participation rate across training days for LMI-positive (red) and LMI-negative (blue) neurons, shown separately for R+ (left) and R- (right) mice. Bar graph shows mean ± 95% confidence interval (Kruskal–Wallis test).

Behavioral quantification of rapid whisker detection learning.

(A) Trial number of the first whisker hit on Day 0 for R+ and R- mice. Points indicate individual mice. Bar graph shows mean ± 95% confidence interval (Mann-Whitney U test). (B) Whisker performance on Day 0. Points indicate individual mice. Bar graph shows mean ± 95% confidence interval (Mann-Whitney U test). (C) Stimulation particle off control performed on Day +3 for a subset of 14 R+ mice where the particle is removed from the C2 whisker to control for potential stimulus artefacts. First particle-on block (ON₁), particle-off block (OFF), and second particle-on block (ON₂). Left: whisker hit rate across the three epochs. Middle: catch trial false alarm rate across the three epochs. Right: comparison of whisker hit rate and false alarm rate during the OFF epoch alone. Grey lines show individual mice. Bar graph shows mean ± 95% confidence interval (Mann-Whitney U test). (D) Single-trial whisker binary hit/miss outcomes averaged across mice during Day 0 for R+ (green) and R-(magenta) mice. Top grey-scale bar indicates statistical significance comparing R+ and R- for each day (Mann-Whitney U test). (E) As in panel D with lick probability estimated by a Bayesian state-space model fitted to binary hit/miss outcomes and with whisker trials realigned to the first whisker hit. (F) Lick probability for whisker, auditory, and no-stimulus trial types across time on Day 0, interpolated onto a common time grid. (G) Mean reaction time per stimulus type (auditory, whisker, catch) across training days for R+ and R- mice, computed from lick trials. Points indicate individual mice. Bar graphs show mean ± 95% confidence interval (Mann-Whitney U test). (H) Mean reaction time ± 95% confidence interval across lick trials within Day 0 for each stimulus type, for R+ and R- mice.

Relationship between single-cell learning modulation and population decoding.

(A) Stability of the field of view across imaging days for an example mouse. Images are obtained by averaging the first 1,000 frames of the movement corrected data. (B) Responses to passive whisker stimulations of three example learning modulated cells during the five imaging days: a cell with a strong positive LMI that did not respond before learning (top), a cell with a strong positive LMI with an enhanced response (middle) and a cell with a strong negative LMI with an abolished response (bottom). Individual trials are shown in grey; the trial-averaged trace is shown in black. Orange line indicates whisker stimulus onset. (C) Scatter plot of logistic regression classifier weight versus LMI for all imaged neurons, color-coded by mouse. Black line indicates linear regression line. (D) Pre- vs post-learning decoding accuracy as a function of the proportion of most-modulated cells removed from the population. Cells are ranked by |LMI| in descending order and progressively excluded; at each step, decoding accuracy is computed by 10-fold stratified cross-validated logistic regression.

Pathway-specific reorganization of barrel cortex representations.

(A) Example two-photon calcium imaging field of view (top) and labelling of projecting neuron subtypes (bottom). Green: GCaMP6f-expressing neurons; yellow: CTB retrogradely labelled neurons projecting to the secondary somatosensory cortex (wS2p); blue: CTB retrogradely labelled neurons projecting to the primary motor cortex (wM1p). (B) Population PSTHs over passive whisker trials across the five imaging days for wS2p (top) and wM1p (bottom) neurons, shown separately for R+ mice (green) and R- mice (magenta) (wS2p: n = 17 R+ mice, n = 17 R- mice; wM1p: n = 12 R+ mice, n = 11 R- mice). Orange line indicates whisker stimulus onset. (C) Comparison of whisker-evoked response amplitude (ΔF/F₀, averaged over 0–300 ms from stimulus onset) before (Days -2 and -1; Pre) and after (Days +1 and +2; Post) learning in R+ mice, for wS2p neurons (n = 17 mice). Left: population PSTHs for pre- (grey) and post- (green) learning. Right: bar plot quantification of individual mouse averages over the same time window. Points indicate individual mice. Bar graphs show mean ± 95% confidence interval (Wilcoxon signed-rank test). (D) As in panel C, but for R- mice (n = 17 mice). (E) Distribution of the LMI across wS2p neurons for R+ mice (green; n = 328 neurons from 17 mice with 19 ± 10 neurons per mouse) and R- mice (magenta; n = 254 neurons from 17 mice with 21 ± 12 neurons per mouse). (F) Proportion of wS2p neurons with a significant positive LMI (left) and significant negative LMI (right), compared between R+ and R-mice. Points indicate individual mice. Bar graphs show mean ± 95% confidence interval (Mann-Whitney U test). (G) As in panel C, but for wM1p neurons (n = 12 mice). (H) As in panel D, but for wM1p neurons (n = 11 mice). (I) As in panel E, but for wM1p neurons (green, R+ mice: n = 304 neurons from 12 mice with 18 ± 10 neurons per mouse; magenta, R- mice: n = 182 neurons from 11 mice with 17 ± 12 neurons per mouse). (J) As in panel F, but for wM1p neurons. (K) Contribution of projection subtypes to LMI values. Cumulative distribution of LMI values for wS2p (orange) and wM1p (blue) neurons, shown separately for positive and negative LMI populations and for R+ and R- groups. (L) As in panel K, but for logistic regression classifier weights. (M) Pairwise Pearson correlation between wS2p neuron pairs (wS2p–wS2p) and wM1p neuron pairs (wM1p–wM1p) during the 2-second pre-stimulus quiet window, compared between pre- (Days -2, -1) and post- (Days +1, +2) learning for R+ and R-groups. Bar graphs show mean ± 95% confidence interval (Wilcoxon signed-rank test).

Spontaneous activity does not account for the LMI–reactivation relationship.

(A) Scatter plot of reactivation participation rate versus spontaneous calcium transient frequency on Day 0 for individual neurons, shown separately for R+ (left) and R- (right) mice. Each point represents one neuron, color-coded by LMI value. The black lines indicate the Pearson correlation. (B) Correlation between residuals of LMI and reactivation participation rate obtained from regressing out spontaneous frequency, shown separately for R+ mice (left) and R- mice (right). This analysis tests whether the LMI–participation relationship holds independently of baseline spontaneous activity. Dots indicate individual neurons. The black lines indicate the Pearson correlation. (C) Proportion of neurons classified as significantly participating in reactivation events across training days for LMI-positive (red) and LMI-negative (blue) neurons, shown separately for R+ mice (left) and R- mice (right). Each cell was classified as significantly participating if its observed participation rate exceeded the 95th percentile of a null distribution built from 1,000 circular shifts of the neuronal data. The effect of days on the proportion of reactivations was tested with a Kruskal–Wallis test. This analysis tests whether the LMI-participation relationship holds for a binary measure of significant reactivation independent of the magnitude of the participation rate and accounts for cells that would take part in reactivation by chance due to high spontaneous activity.