Rapid cortical reorganization tracks goal-directed sensorimotor learning in real time

Reviewer #1 (Public review):

Summary:

The paper submitted by Renard et al. seeks to capture the moment when learning occurs and to identify the associated changes in neuronal activity within cortical circuits. Specifically, the study aims to test whether sensory representations in the cortex reorganize on the same timescale over which behavioral changes first emerge.

To address this question, the authors developed a new behavioral paradigm in which mice were first trained on an auditory detection task and then introduced to whisker stimulation, which they learned to associate with reward. This design allowed mice to form a new whisker-reward association within a single behavioral session, enabling the authors to track learning-associated neuronal changes during the course of the experiment.

Using pharmacological and optogenetic interventions, the authors first show that learning depends on the whisker somatosensory cortex. They then combined the task with longitudinal two-photon calcium imaging to examine real-time changes in neuronal representations that accompany improvements in task performance over trials within a session and across days. By applying a range of analytical approaches, they show that learning induces a rapid reorganization of sensory cortical representations over tens of trials, on the timescale of minutes. They further propose that spontaneous reactivation of neurons during the task may contribute to these representational changes during learning.

Strengths:

(1) Overall, the experiments are thoughtfully designed, well controlled, and clearly presented. The conclusions are generally well supported by the data. The manuscript is clearly written, and the Discussion acknowledges potential caveats while outlining future directions.

(2) A major strength of the study is the design of a new learning paradigm in which head-fixed mice rapidly form a new sensory-motor association within a single session, on the timescale of minutes. This offers a unique opportunity to track real-time changes in neuronal dynamics associated with learning during a single recording experiment.

(3) Taking advantage of this behavioral design, the authors show that learning induces rapid reorganization of sensory cortical representations. They also report an increase in spontaneous reactivation of neurons that gained stimulus responsiveness during training, and propose that these reactivations may contribute to rapid representational reorganization. These findings provide important insights into the neural dynamics associated with learning.

Weaknesses:

(1) The authors propose that spontaneous reactivation mediates rapid reorganization of neuronal representations and thereby supports rapid task learning. However, as they also acknowledge in the Discussion, the present study does not directly test a causal role for these reactivations in facilitating representational changes or behavioral improvement.

(3) The authors show reorganization of neuronal representations even on the first day of training with the new whisker task. However, because there is no explicit control for natural representational drift, it remains unclear to what extent these changes reflect learning-related reorganization rather than spontaneous day-to-day drifts in neuronal responses.

Reviewer #2 (Public review):

Summary:

Renard, Foustoukos and colleagues present a study of rapid sensorimotor learning in the mouse barrel cortex. Head-fixed water-restricted mice already trained on an auditory detection task are introduced to a novel C2 whisker stimulus, and the authors show that reward-paired mice acquire the whisker-lick association within a single behavioral session, with the two groups (rewarded vs non-rewarded) diverging behaviorally within ~22 whisker trials and ~14 minutes. Both pharmacological inactivation of wS1 across Days 0/+1/+2 and optogenetic inactivation on Day 0 impair whisker-guided performance, while fpS1 manipulations do not, establishing that wS1 activity is required for whisker-guided behavior during the initial learning period. Longitudinal two-photon imaging of GCaMP6f-expressing L2/3 neurons across five days (-2 to +2 relative to whisker introduction) reveals a bidirectional, reward-dependent reorganization of population responses to passive whisker stimuli: rewarded mice show enhancement, non-rewarded mice show suppression. The authors use a logistic-regression decoder trained to discriminate pre- vs post-learning passive trials and then project Day 0 active whisker trials onto this learning axis; the projection rises monotonically across Day 0 in R+ mice and is significantly correlated with behavioral performance, with no such trajectory in R- mice. Finally, the authors detect reactivation events during catch trials by template-matching to the average passive whisker response, and show that on Day 0, the neurons most positively modulated by learning (LMI-positive) participate in these reactivations more than LMI-negative neurons in R+ but not R- mice. The authors interpret this as evidence that online, reward-gated reactivations may act as an upstream selection mechanism for which neurons undergo learning-related plasticity, operating on the minutes-timescale of within-session learning. There is much to like in this paper, with some moderate-to-major concerns that could largely be addressed with re-analysis or re-framing.

Strengths:

The single-session learning paradigm is a key aspect of this paper, given the rapid learning observed. Coupled with the R+ and R- design, there's a lot to like with the behavioral approach. The bidirectional response change across these R+ and R- groups (enhancement vs suppression) is also a nice finding.

The causal manipulations demonstrate that the imaged region is used during the task. By doing both pharmacological and optogenetic inactivation, each with a control in the spatially adjacent region (fpS1), the authors make a strong case that wS1 activity is necessary for whisker-guided behavior during the initial learning period (though see below about the limitations of the current approach).

The longitudinal two-photon imaging of the same L2/3 neurons across five days underlies essentially every neural analysis in the paper and enables the single-cell LMI and population-trajectory analyses.

The pathway-specific analysis in Figure 3 - figure supplement 2 is very interesting, but not much time is spent on it (lines 151-155). The dissociation between wS2-projecting neurons (which show learning-related enhancement in R+ and suppression in R-) and wM1-projecting neurons (which do not) is (in my opinion) a nice instance of projection specificity - it also aligns with the known routing of task-relevant whisker information through the wS1→wS2 pathway. I would encourage the authors to motivate this experiment in the main text rather than leaving it all to the discussion (lines 256-262).

The methods are generally well documented and easy to follow.

Weaknesses:

(1) Conflation of de novo association learning with generalization from auditory pre-training.

All mice have already learned a task structure with the auditory task - "detect the salient sensory cue → lick → reward". Under these conditions, the rapid emergence of licking to the whisker stimulus could reflect either de novo formation of a whisker-specific association or generalization of an instrumental policy to a novel salient cue. The manuscript frames the result as the former ("acquisition of a novel sensorimotor association"), but the experiment cannot distinguish between the two alternatives. This distinction between de novo learning and generalization may have a meaningful impact on the interpretation, though it doesn't impact the specific results. It would be helpful for the authors to discuss the two possibilities and generally consider the contribution of generalization from auditory pre-training to Day 0 performance.

Relatedly, the R- group is introduced (lines 69-74) and later used (lines 244-247) as a passive-exposure control that rules out representational drift. While R- group is an important control for repeated whisker stimulation and task context, it does not appear to be a pure passive-exposure control: Figure 1B shows that on Day 0 the mice lick more to the R- stimulus than with no stimulus and then extinguish that licking by Day 1. Thus, one possibility is that R- mice actively learn to suppress licking to an unrewarded stimulus (whisker) in a context where other stimuli (auditory) remain rewarded. This would be a different cognitive operation (response suppression) from a purely passive exposure condition. The manuscript therefore lacks a true passive-exposure baseline, and several claims that rely on R- as such a baseline (including that bidirectional changes are reward-driven rather than reflecting passive drift, lines 244-247) need to be reframed.

(2) The inactivation experiments establish that wS1 is necessary on Day 0, but they cannot separate detection, acquisition, and expression.

Both the muscimol manipulation (whole session, Days 0/+1/+2) and the optogenetic manipulation (0.1 s before stimulus onset through the 1 s reporting window) silence wS1 during the moments when the whisker stimulus must be detected for a successful trial. Under these conditions, impaired performance could reflect that the animal cannot detect the stimulus, cannot express the learned response on that trial, or cannot acquire the association. These are causally distinct processes, and the manuscript currently treats them as equivalent.

Specifically, on Day +1 of the opto experiment (light off), do mice learn at the same rate as a naive Day 0 cohort (e.g., the R+ imaging mice on Day 0), or is performance already higher than the naive group? If higher than the naïve group, this would suggest that there is learning occurring and would suggest that something that may have been acquired during Day 0 inactivation, even if it could not be expressed.

(3) The interpretation of the LMI-participation correlation is complicated by the peaked LMI distribution and neuron-level pooling.

Two related issues arise from the results shown in Figure 4I. First, the LMI distribution in Figure 3F (and visible in 4I) is sharply peaked near zero. The reported r = 0.24 in R+ mice is therefore difficult to interpret biologically because the distribution is dominated by near-zero-LMI neurons and the slope may be disproportionately influenced by neurons in the tails. The key claim is better tested by comparing significantly LMI-positive, LMI-negative, and non-modulated neurons. The authors do address this in Figure 4J - showing that participation rate rises across days for significantly LMI-positive R+ neurons (p = 5×10⁻⁴) but not for LMI-negative neurons (p = 0.05) - but this analysis is not the lead result. To my understanding, Figure 4J is more interpretable and should be the key piece of data supporting their claim.

Second, the p-value of p = 1×10^-41 in Figure 4I comes from treating thousands of neurons pooled across 19 mice as independent observations. Neurons within an animal are correlated through shared behavioral state, shared imaging session, and circuit-level interactions, so it would be helpful to consider a different statistical unit of comparison (FOV, animal, etc). For example, a linear mixed-effects model with mouse as a random effect could work.

(4) The reactivation-LMI relationship is partially circular, and the framing in the abstract could be more constrained.

The "reactivation template" is the trial-averaged passive whisker-evoked population vector from each session, and reactivations are detected as moments in catch-trial activity that correlate with this template above a shuffled threshold. This approach is reasonable, but it means that the reactivation-LMI relationship is not fully independent of template construction, and the framing in the abstract blurs that line. LMI-positive neurons are defined as neurons whose passive whisker-evoked responses increase from pre- to post-learning. Therefore, neurons with strong whisker responses, or neurons that become stronger components of the whisker-evoked template across learning, may be more likely to contribute to template-matching events by construction. Thus, the LMI-participation relationship could partly reflect template weighting or sensory-response amplitude, rather than showing that reactivation events selectively recruit neurons for future learning-related plasticity. It would be helpful and more reassuring if the authors could control for each neuron's whisker-template weight, baseline whisker responsiveness, and overall calcium event rate when relating LMI to reactivation participation.

A complementary unsupervised approach could also help: rather than starting from the whisker template, one can derive co-activity assemblies directly from spontaneous activity (e.g., via PCA or ICA on the catch-trial population activity), and then ask, separately, whether any of these assemblies overlap with the whisker ensemble. The interesting test is then whether whisker-like assemblies become more frequently expressed across Day 0 in R+ but not R- mice, and whether LMI-positive neurons are preferentially loaded onto these whisker-like assemblies. This logic inverts the current pipeline and can be complementary to the current analysis. By identifying structure in nominally spontaneous activity first and then comparing to the whisker response, this could help avoid the circularity in which the template both defines the events and contains the cells being tested. The Figure 4 - figure supplement 1B partial-correlation analysis is a step in this direction but addresses only spontaneous firing rate, not template coupling. Without such a complementary approach, the authors may want to clarify that the reactivation detection is anchored to a template defined in part by the same cells whose participation is being tested.

(5) The reactivation-as-selection-mechanism interpretation is not supported by the current data.

The Discussion (lines 278-281) acknowledges that the authors have not shown necessity, but the end of the intro and part of the discussion (Lines 275-277) frame reactivations as a "reward-gated selection mechanism" for plasticity. An equally plausible alternative is that neurons whose synaptic inputs or intrinsic excitability have been potentiated by reward-driven learning will simply co-fire more often during quiet periods - meaning reactivations would be a consequence of plasticity that has already occurred rather than a mechanism that selects which neurons to potentiate. The current data cannot distinguish these.

A separate concern is the use of the term "spontaneous." The authors' usage is defensible in one sense - catch trials are stimulus-free, so the activity is not externally driven. However, "spontaneous" in the reactivation literature typically connotes offline, internally generated activity during quiet wakefulness or sleep, which carries different implications for plasticity than activity during active task engagement. Catch trials in this paradigm occur within the behavioral session, with the animal still engaged in the task, potentially anticipating reward or licking. The authors should either acknowledge this distinction in the text or qualify the term - "within-session" or "inter-trial" reactivations would be more accurate and would avoid borrowing the conceptual weight of the offline-replay literature.

The authors should also clarify whether catch-trial activity around licks (false alarms, anticipatory licks) is excluded from the reactivation analysis, and whether reactivation rates depend on recent reward, recent whisker trial outcome, or behavioral state. Specificity controls - template-matching with shuffled templates and with auditory templates - would help establish that detected events reflect whisker-specific patterns rather than generic high-coactivity moments.

(6) Motor, lick, and behavioral-state confounds in the neural analyses are not fully addressed.

I have two specific concerns. First, for the Day 0 active-trial projection, mean whisker reaction times in Figure 1 - figure supplement 1G are around 350-500 ms, but the distributions extend into the 0-300 ms analysis window. The correlation between the projection trajectory and the behavioral learning curve (Figure 4E, lines 196-198) is the key piece of evidence that the neural shift tracks learning. However, on hit trials the lick may fall within or close to the analysis window, so a motor confound could in principle contribute to the rising projection. The authors could repeat the projection using an earlier/shorter window, exclude trials with early licks, or regress out lick timing. It would be helpful to better understand whether this effect is, in part, driven by licking activity.

Second, the central evidence for representational reorganization (Figure 3) rests on a post-session passive epoch in which 50 whisker stimulations are delivered after "task disengagement" (lines 131, 387-389). The concern is that the brain state during this epoch is unlikely to be matched across groups or across days. R+ mice receive additional water rewards on whisker trials, whereas R- mice receive rewards only on auditory trials. This could lead to systematic differences in satiety, arousal, and disengagement state during the passive block. Because cortical sensory responses are strongly modulated by arousal, some of the apparent learning-related enhancement (R+) or suppression (R-) of passive whisker responses across days could reflect systematic state differences during the passive epoch rather than plasticity. The disengagement criterion ("stopped licking in all trial types") is also qualitative - no consecutive-miss or time-window threshold is specified - so the epoch may begin at slightly different behavioral states across mice. To resolve this, the authors could (i) specify the disengagement criterion quantitatively and (ii) compare pupil diameter and whisker self-motion (if available) across R+ vs R- and across days during the passive epoch.

Reviewer #3 (Public review):

This is a methodologically sound manuscript and provides reasonably interpretable results. While being appropriate, they do not seem to bring entirely novel concepts; nevertheless, most of my comments concern the calibration of the interpretive claims rather than the quality of the data.

Strengths:

(1) Longitudinal within-subject imaging:
Tracking the same layer 2/3 neurons across learning allows the bidirectional effect (enhancement in R+, suppression in R-) to be measured within identified cells rather than inferred across cohorts.

(2) Appropriate behavioural controls:
The R+/R- design controls for repeated sensory exposure, and maintaining rewarded auditory trials in both groups controls for engagement and arousal, arguing against disengagement as the source of the R- effect.

(3) Convergent causal manipulations:
Muscimol and optogenetic inactivation both abolish acquisition and include an adjacent control region (fpS1); the temporally restricted optogenetic result partially addresses the concern (Hong et al., 2018) that sustained inactivation may destabilise downstream circuits.

(4) Convergent analyses:
Single-cell learning modulation indices, population similarity measures, and a trial-resolved decoder projection onto a naïve-to-expert axis provide consistent evidence that representational change is concurrent with behavioural acquisition.

(5) Projection-specific resolution:
Retrograde labelling shows learning-related changes in wS2-projecting, but not wM1-projecting neurons, consistent with preferential routing of task-relevant signals through the wS1 to wS2 pathway.

(6) Mechanistically motivated reactivation analysis:
Relating rapid, reward-dependent plasticity to spontaneous reactivations on a timescale of minutes is an original use of the single-session paradigm.

Weaknesses and points requiring clarification

(1) The stimulus is not strictly novel: Passive whisker stimulations were delivered on pre-training Days -2 and -1, so what changes on Day 0 is the stimulus-reward contingency rather than the stimulus itself. This resembles contingency reassignment with reversal-like properties (and possible habituation or latent inhibition) rather than de novo learning, and the licking response is already established during auditory training. The framing should be qualified accordingly.

(2) Barrel cortex dependence should be stated more narrowly: The data show that acute wS1 suppression prevents acquisition of this task, not that whisker detection in general requires barrel cortex; cortical dependence varies with task and manipulation (Hong et al., 2018; Ryan et al., 2022 vs Miyashita and Feldman, 2013). The near-threshold explanation would require psychometric or stimulus-intensity data.

(3) The passive block carries confounds: It is acquired after task disengagement, when satiety, arousal, and reward history differ across groups and days. The authors should report within-block response adaptation and the robustness of the main results to early versus late passive trials. Additionally, could passive presentation of the stimulus without reward delivery lead to devaluation of the stimulus, leading to additional behaviour and plasticity changes which are not addressed?

(4) Some statistics appear to be neuron-level rather than animal-level:
Very small p-values (e.g., the LMI-participation correlation r = 0.24, p on the order of 10^-41) suggest thousands of non-independent neurons treated as independent samples, risking pseudoreplication. Central claims should rest on hierarchical or animal-level statistics with effect sizes.

(5) The cosine-similarity decrease in R- animals needs clarification:
Because cosine similarity is scale-invariant, uniform suppression would leave it largely unchanged; the observed decrease therefore implies heterogeneous suppression, reduced signal-to-noise, or increased variability, and the favoured interpretation should be stated.

(6) The decoder requires cautious interpretation:
Training on passive trials and applying to active Day 0 trials could introduce a behavioural-state domain shift. The meaning of positive and negative values in Figure 4C should be defined, and near-zero early projections reflect the classifier boundary rather than a biological baseline.

(7) The reactivation analysis is the least conclusive and is susceptible to circularity:
The template and the LMI are both derived from the passive whisker response, predisposing responsive neurons to register as reactivation participants, and the Day 0 template is obtained after learning.

Leave-one-cell-out and pre-learning templates, cell-specific templates, and tests of whether reactivations predict subsequent trial responses would strengthen the claim; causal disruption would ultimately be required.

(8) Figure, sample-size, and specificity points:
The positive LMI shift in R+ animals is less visible than the R- shift in Figure 3F; the optogenetic cohort is small (n = 6 per group); and confirming that auditory detection was preserved during wS1 inactivation would establish whisker-specificity.

(9) The comparison to prior work is overly broad:
Banerjee et al. (2020) and Chéreau et al. (2020) are reversal learning and discrimination paradigms and may not be equated with simple whisker detection; the defensible novelty claim is the trial-resolved tracking within the first session and its concurrence with online reactivations.