VTA projections to M1 are essential for reorganization of layer 2-3 network dynamics underlying motor learning

Ghanayim, Amir; Benisty, Hadas; Cohen Rimon, Avigail; Schwartz, Sivan; Dabdoob, Sally; Lifshitz, Shira; Talmon, Ronen; Schiller, Jackie

doi:10.1038/s41467-024-55317-4

Download PDF

Article
Open access
Published: 02 January 2025

VTA projections to M1 are essential for reorganization of layer 2-3 network dynamics underlying motor learning

Nature Communications volume 16, Article number: 200 (2025) Cite this article

7047 Accesses
39 Altmetric
Metrics details

Subjects

Abstract

The primary motor cortex (M1) is crucial for motor skill learning. Previous studies demonstrated that skill acquisition requires dopaminergic VTA (ventral-tegmental area) signaling in M1, however little is known regarding the effect of these inputs at the neuronal and network levels. Using dexterity task, calcium imaging, chemogenetic inhibiting, and geometric data analysis, we demonstrate VTA-dependent reorganization of M1 layer 2-3 during motor learning. While average activity and average functional connectivity of layer 2-3 network remain stable during learning, activity kinetics, correlational configuration of functional connectivity, and average connectivity strength of layer 2-3 neurons gradually transform towards an expert configuration. Additionally, sensory tone representation gradually shifts to success-failure outcome signaling. Inhibiting VTA dopaminergic inputs to M1 during learning, prevents all these changes. Our findings demonstrate dopaminergic VTA-dependent formation of outcome signaling and new connectivity configuration of the layer 2-3 network, supporting reorganization of the M1 network for storing new motor skills.

Motor learning refines thalamic influence on motor cortex

Article 07 May 2025

Corticothalamic neurons in motor cortex have a permissive role in motor execution

Article Open access 21 May 2025

Task-specific modulation of corticospinal neuron activity during motor learning in mice

Article Open access 11 May 2023

Introduction

Motor learning is an essential process by which the organism acquires skilled movements and the ability to associate between new sensory information and actions, thus adapting to the ever-changing environmental demands of the world^1,2,3,4.

Learning, planning, and execution of movements, are carried out by a highly complex and distributed system, with the primary motor cortex (M1) being one of its main hubs, generating the output cortical motor commands to the downstream brainstem and spinal cord execution centers⁵. Surmounting evidence supports M1 to be crucial for motor skill learning^6,7,8,9,10. Previous studies have shown that during motor learning M1 undergoes major plasticity changes, manifesting at multiple levels ranging from motor representations, dendritic computations, branch spike properties, and spine turnover and clustering^{6,8,11,12,13,14}.

The functional role of M1 in motor control and learning is expected to be cell-type dependent^13,15,16,17. Here we concentrate on layer 2-3 pyramidal neurons (PNs), which were shown to represent movement, error estimation, and outcome-related activity^{8,13,16,18,19}. In addition, layer 2-3 PNs of M1 were shown to undergo plasticity changes during motor learning both at the activity and structural spine levels^16,19,20.

Central to motor learning is a reward-based learning process¹, where reward can in principle inform about the consequences of actions and drive the motor learning process via plasticity mechanisms²¹. The neuronal substrate carrying reward in motor learning and adaptation is the dopaminergic projection systems to both the basal ganglia and cortex²². The ventral tegmental area (VTA) was shown to be the main source of dopaminergic projections to M1, innervating dendrites of both superficial and deep layers^23,24 with forelimb regions being preferentially innervated²⁵.

Our working hypothesis is that the dopaminergic VTA projections to M1 directly drive synaptic plasticity mechanisms and the ensuing network connectivity changes in M1, which underlies motor learning. Previous studies have demonstrated the crucial role of dopaminergic signals in M1 for motor task acquisition as well as synaptic plasticity. Destroying the direct dopaminergic projections to M1 eliminated acquisition of a new skilled reaching motor behavior²⁴. Dopaminergic system was demonstrated to mediate plasticity changes in both the structural and cellular synaptic plasticity levels of M1 including long term potentiation (LTP) of layer 2-3 horizontal connections, immediate early gene expression, and regulation of spine turnover^21,26,27. In contrast to the effects of dopamine at the synaptic level, little is known regarding the effects of the dopaminergic inputs to M1 in altering the activity and functional connectivity of the M1 network during motor learning.

Here, we set out to investigate first, if and how functional activity and connectivity within layer 2-3 network of M1 are altered during learning, second, how the VTA projections to M1 participate in motor learning at the behavioral, network activity and functional connectivity levels, and third, whether VTA projections are essential for the development of outcome signaling in M1. Toward this end, we recorded the activity of layer 2-3 PNs in M1 using two-photon calcium imaging with the genetically encoded calcium indicator GCaMP6s²⁸ during learning of a head-fixed hand reach for pellet motor task^10,13 with and without inhibiting M1 dopaminergic projections from VTA using designer receptors exclusively activated by designer drugs DREADDs²⁹. We developed a novel non-linear analysis approach to explore the correlational structure of the network throughout training. We used graph theory and Riemannian geometry^{30,31,32,33,34} which allow us to compare the functional connectivity of the network at different stages of training. We find that motor learning is associated with gradual and monotonic reorganization of functional connectivity of the M1 layer 2-3 network. In contrast to previous reports^19,35, we find the majority of neuronal activity as well as the population connectivity gradually transforms towards a new “expert” configuration. Blockade of dopaminergic neurotransmission locally in M1 prevented motor learning at the behavioral level, and concomitantly halted plasticity changes at both the individual neuron activity level as well as at the network activity and functional connectivity levels.

Results

Dopaminergic VTA projections to M1 are essential for hand reach motor learning

To study the reorganization of network activity and functional connectivity of layer 2-3 PNs and their dependence on dopaminergic VTA activity in M1 during motor learning, we trained mice to perform a head-fixed version of the forelimb grasping task, where mice learn to reach, grab, and eat a food pellet^10,13 while performing longitudinal two-photon calcium imaging of GCaMP6s²⁸ (Fig. 1a–c). To explore the effects of the direct dopaminergic VTA projections to M1 on motor learning progression and the consequent changes in neuronal activity, we divided the animals into two groups: a manipulated group for which we inhibited the VTA dopaminergic projections to M1 for three consecutive training sessions after an initial shaping session, and a second group trained with no manipulations, serving as controls. Dopaminergic projections to M1 were inhibited by DREADDs (hM4D), which were expressed in dopaminergic neurons of the VTA in DAT-Ires-Cre mice. Inhibition was performed by locally injecting Clozapine-N-Oxide (CNO) to M1 via an access port (CNO sessions) (Fig. 1a–e). It should be noted that all mice expressed the DREADDs virus and later assigned to control or manipulated groups.

**Fig. 1: Experimental and behavioral set up for investigating the effect of direct ventral tegmental area (VTA) dopaminergic projections to M1 on motor learning.**

The control animal group demonstrated a steady and gradual improvement in motor execution of the task until reaching a stable level. We quantified task performance by the success rate and by evaluating the sequence of motor events during learning. To calculate the success rate, we determined trials successful if animals succeeded in grabbing and consuming the pellet. We observed a gradual increase in success rates at the first four training sessions (R² = 0.55, p = 2·10^–9), after which success rate did not change significantly (p = 0.7, F = 0.37, for one-way ANOVA comparing success rate in sessions 5 through 7, n = 12 mice) where all animals maintained high proficiency (0.55 ± 0.03 at 7^th session, n = 12 mice, Fig. 1f; Supplementary Movies 1–3). Learning as evaluated by success rate was impaired and remained low for the manipulated group in sessions where CNO was injected and dopaminergic afferents in M1 were inhibited (Fig. 1f; Supplementary Movies 4, 5; p = 3.6·10^-8 unpaired two-tailed t-test comparing success rate in the 3^rd session of manipulated group n = 10 mice to control group, n = 12). In consecutive sessions, once CNO was lifted, learning progressed gradually reaching comparable success values to expert controls only on the 8^th training day (p = 0.05 and p = 0.36 t-test for comparison of the 7^th and 8^th training days of the CNO group to the 7^th training days of control group, respectively; Supplementary Movie 6.

We performed several experiments to control for multiple parameters related to the CNO, the viral injection, the origin of dopaminergic afferents, and the spatial cortical location of the observed effect in the manipulated mice: 1. We expressed hM4D DREADDs in the VTA, but instead of CNO, we injected Ringer to M1. We did not observe a significant difference in the progression of the success rate of learning compared to the control group (Fig. 1g; p > 0.19 unpaired t-test for all training days comparing control and Ringer injections; n = 4). 2. We expressed a sham virus that contains only the fluorophore (pAAV-flex-tdTomato) to the VTA, and CNO to M1. Here as well we did not observe a significant difference in the progression of the success rate of learning compared to the control group (Fig. 1g; p > 0.15 unpaired t-test for all training days comparing control and sham viral injection; n = 3). 3. We expressed hM4D DREADDs in the VTA, but applied CNO in a seemingly unrelated cortical region; the primary visual cortex (V1). In this case as well we did not observe a significant difference in the progression of the success rate during learning compared to the control group (Fig. 1h; p > 0.16 unpaired t-test for all training days comparing CNO injections in M1 and CNO in V1; n = 3). 4. Although the majority of dopaminergic axons to the motor cortex originate from the VTA still some axons may originate from the substantia nigra pars compacta (SNc)²⁴. To address the question of the possible contribution of the SNc projections to M1, we expressed hM4D DREADDs in the SNc (Supplementary Fig. 1, and Fig. 1h; n = 3) and applied CNO locally in M1 during the learning process. In this case, we did not observe a significant delay in the learning process and the learning did not differ significantly from the control group (p > 0.41 unpaired t-test for all training days comparing control and SNc injected group). 5. Finally, to address the question of whether inhibiting VTA dopaminergic inputs disrupts motor performance we introduced CNO locally in M1 in mice expressing hM4D DREADDs in the VTA, after the mice reached expert level (Fig. 1i). We found that inhibiting VTA dopaminergic afferents to M1 using hM4D DREADDs at the expert level did not change significantly the success rate of motor performance (p = 0.2, n = 8, comparing expert sessions with and without CNO injections to M1). Thus, inhibiting dopamine afferents to M1 did not hamper motor performance but rather disrupted the learning process itself.

Taken together these results indicate that the effect of CNO injection on motor learning progression is mediated via the hM4D DREADDs inhibition of VTA dopaminergic axons projecting to M1 and is not due to the viral injection nor to the injection of CNO per se to M1 or general inhibition of movement.

Re-organization of layer 2-3 neurons’ activity during motor learning and the role of direct M1 dopaminergic projections in the process

We next explored changes in the activity of layer 2-3 neurons during the training sessions (Fig. 1e) using consecutive two-photon calcium imaging from the same neurons. When considering the overall mean activity (cells across trials, per training session) we found it did not change significantly with training for both control and manipulated groups^16,36 (Fig. 2a, p = 0.97, F = 0.2, n = 12 control group; p = 0.3, F = 1.2, n = 10 manipulated group, one way ANOVA test).

**Fig. 2: Reorganization of layer 2-3 PN activity in M1 during motor learning at the single neuron and the population levels.**

While these results might indicate that M1 layer 2-3 network activity remains unchanged with training, a closer look at the dynamics of the activity profile of the cells throughout the training sessions reveals that the network undergoes a gradual reorganization, with an emergence of a tri-phasic average activity pattern¹³. Examples of the mean activity of cells (across trials), per session, sorted by their variance at the 7^th session, extracted from a control animal and for a manipulated animal are shown in Fig. 2b and their corresponding average activity traces across cells and trials per training sessions (Fig. 2c). To quantify this transformation of activity dynamics during training, we calculated the correlation coefficient between the average activity traces of each cell in the different training sessions with its trace of average activity on the 7^th session (see Methods). For the control group, we observed a shift towards higher values with training, indicating that a growing number of cells adopt a dynamic pattern, similar to their individual dynamics as experts (i.e., 7^th session in the control group, Fig. 2d). For the manipulated group, this process emerged only at the 5^th session as training with no CNO is resumed (Fig. 2d). We quantified this process by counting the fraction of cells with a correlation coefficient higher than 0.5 (Fig. 2d, dashed line) in each training session (Fig. 2e). We found a gradual increase in the control group (R² = 0.45, p = 1.64∙10^-10, for linear fit) while for the manipulated group a gradual increase was observed only after lifting the CNO (Fig. 2e; R² = 0.3, p = 0.0001, for linear fit).

To further quantify the reorganization of the layer 2-3 PN population towards a specific activity profile throughout the training process, we trained a linear classifier to separate between the activity profile of the 1^st session and the 7^th session. The classifier was based on the ensemble activity profile averaged across a short sliding time window throughout the trials (1 second time window, with 0.5 seconds window hop, see Methods). We then applied this classifier to sessions 2–6 and counted the fraction of trials in each session that were classified as related to the 7^th session (Fig. 2f for a time window centered at go cue+1 sec and Fig. 2g for the entire trial time, window length = 1 sec, hop = 0.5 sec, see Methods). For the control group, we observed a monotonic increase of trials classified as related to the 7^th session, indicating a gradual and steady shift of population activity towards an expert profile (days in training 1-7, R² = 0.74, p = 5.9∙10⁻²⁶, linear fit for training days 1–6, n = 12). Interestingly, the gradual progression in similarity to the expert activity is apparent already at the pre-tone period, suggesting a reorganization in network state even before the go-cue. Repeating this analysis for the manipulated group revealed that CNO injection leads to an unrecognizable activity profile, as trials were not consistently classified as either the 1^st or the 7^th day in training. Once CNO was lifted at the 5^th session, we observed a gradual transformation of the neuronal population activity toward its configuration on the 7^th day of training (Fig. 2f; R² = 0.3, p = 0.002, linear fit for training days 4–6, n = 10).

It should be noted that for the two experimental groups controlling for the DREADDs manipulation (sham virus expression and the Ringer injection), the mean activity and the reorganization of the population activity of the layer 2-3 PN towards the expert profile on the 7^th day behaved similarly to the control mice (Supplementary Fig. 2a, b). This further emphasizes the specificity of the inhibition of dopaminergic axons in M1 by hM4D DREADDs.

Together, these results highlight the crucial role the dopaminergic innervation in M1 plays in the gradual development of the learning related dynamics at the population of layer 2-3 neurons.

The correlative changes of the behavioral events and activity during motor learning

Next, we investigated the behavioral changes during the learning process in a more detailed manner, examining their correlation with the neuronal activity changes. To do so we identified specific behavioral events during the task (lift, reach, grab, supinate, and at mouth, see Methods) using a semi-supervised software^37,38 as exemplified in Fig. 3a, b and Supplementary Fig. 3a, b. We segmented the events into sequences where each sequence is defined as a series of consecutive events ending with the animal bringing the food pellet to the mouth (an “at mouth” event) or with the animal returning its forelimb to the perch. We then evaluated the number of transitions between behavioral events within each sequence per trial at each training session (Fig. 3c). For the control group we observe a gradual increase in the first three training sessions (R² = 0.3, p = 0.04, n = 5 mice for linear fit), after which the number of transitions slightly decreased and then stabilized (no significant change during sessions 4 through 7, p = 0.98, F = 0.07 for one way ANOVA, n = 5 mice). For the manipulated group, the number of transitions remain low in the first 3 sessions (p = 0.9, F = 0.07 for one-way ANOVA test comparing training days 1-3, n = 5 mice) and was significantly different compared to the control group (p = 0.001, unpaired one tail t-test comparing number of transitions in session 2–4 for the manipulated group vs. the control group). In addition to the number of transitions, we evaluated the time delay between the sensory cue (tone) and the first lift event during the training sessions (Fig. 3d). We found that while in control mice this time delay decreases during the first four training sessions and then levels off, for the manipulated group the time to lift remains high during the CNO training sessions, and decreases only later in training sessions where CNO is lifted (p = 0.01, for unpaired one tail t-test asking if delay time in sessions 2–4 is longer than delay time in the 7^th session).

**Fig. 3: Behavioral analysis and its correlations with activity.**

We next used a generalized linear model (GLM) to model the activity of layer 2-3 PNs as a function of sensory, motor, and outcome task components during learning of the task¹³. Similar to our previous work, we modeled the calcium transients of layer 2–3 PNs based on 4 types of predictors: time series of 3D spatial location of the hand³⁹, time-varying orofacial features extracted from recorded videos⁴⁰, time-varying discrete behavioral events (tone, lift, reach, grab, supinate, at mouth and back to perch)^37,38, and outcome labeled as constant binary variables of success or failure (1 or 0, respectively) throughout the trial. We found that during learning, the full GLM (i.e., including all predictors; see Methods) captures the peri-movement activity (1 sec before until 2 sec after the tone) of an increasing fraction of layer 2–3 PNs. The fraction of neurons that were successfully modeled by the full GLM model increased from 4.5% on the first day of training, through 9.5% on the 3^rd day of training, to 13% in expert mice in the control group. These results indicate the development of a sparse representation of task variables in the layer 2-3 PNs. For the manipulated group, similar fractions were observed for the 1^st and expert mice sessions (3% and 10%, respectively) but remained low on the CNO injected days (2.5%). We further investigated the contribution of task components to modeling the calcium dynamics. For the control group, the relative contribution shifted from the sensory cue of the tone (51 ± 15% on the first day of training) to the outcome component of the task, which explained 58 ± 5% of the activity in expert mice (Fig. 3e and Supplementary Fig. 4). This shift from sensory to outcome was not observed in the manipulated group while CNO was present, as the contribution of the tone component was still high (43 ± 12%). After resuming training with no CNO, mice in the manipulated group exhibited the same trend as the control group where the outcome component was the most prominent (Fig. 3e and Supplementary Fig. 4; 42 ± 4%). Thus, during learning layer 2-3 PNs develop a high order outcome encoding of the task, which depends on activation of VTA dopaminergic inputs to M1.

Development of outcome-related signaling in layer 2-3 network of M1 during motor learning and the role of direct M1 dopaminergic projections in the process

Our previous study showed a specialized sub-population of layer 2-3 neurons, which reported the outcome of the hand reach trial. These outcome-related neurons developed during the learning process¹³ (Fig. 3e).

Here, we studied the involvement of M1 VTA projections in the development of the outcome signal. We calculated the fraction of indicative neurons that reliably report the outcome of the trial¹³. We defined “indicative neurons” as neurons that consistently predict success or failure trials with 99% confidence at the different training days in both the control and manipulated groups (Fig. 4a). Similar to our previous results¹³ we observed a gradual increase in the fraction of indicative neurons along the training sessions. In contrast, in the manipulated group we did not observe the emergence of indicative neurons during sessions where dopaminergic VTA axons in M1 were inhibited (Fig. 4a and supplementary Fig. 5). An increase was observed only after lifting the CNO in subsequent training sessions (Fig. 4a). However, both control and manipulated groups reached peak levels and stabilized the fraction of indicative neurons after three training days (control group by day 3 of training and CNO group by day 6 of training reaching a similar level, p = 0.7, unpaired t-test for comparison of day 3 in control and day 6 in manipulated group).

**Fig. 4: Development of success/failure signal in M1 network during motor learning in control and manipulated mice.**

We next evaluated the outcome signals at the population level using a dynamical system analysis as we described previously^13,41,42. We extracted principal components (PC) of the ensemble activity to capture 95% of variance, and then averaged the PCs across trials according to outcome to obtain two trajectories per session: one representing the average ensemble dynamics in successful trials and the other representing the average dynamics in failed trials (Fig. 4b). We quantified the distance between the success and failure trajectories as a function of time and used permutation testing to determine whether these differences are significant (see Methods). We counted the fraction of animals where success-failure trajectory distances were significant (p < 0.05, see Methods; Fig. 4c). In the first training session the average dynamics did not differentiate by outcome for both control and manipulated animals (Fig. 4b, c). With training, the trajectories of the control group became increasingly separated by outcome mostly after the movement time window. Significant differences emerged in sessions 2-3 and onward, at the post-movement time segment (5–6 seconds after the go cue) (Fig. 4b, c and Supplementary Fig. 6a), which indicates that the population activity adopted a different dynamic for success compared to failure trials. In contrast, for the manipulated group, the separation between success and failure trajectories did not emerge during training sessions with CNO injections (2^nd–4^th sessions). Instead, it began to emerge in training session 5 and onward, although it remained lower and did not reach the levels of separation seen in the control group by the 7^th training session.

We further quantified the extent to which outcome can be decoded from the entire population activity of layer 2-3 PNs by training a set of binary classifiers (see Methods). Sliding the analysis window across trial time (1 sec time window and 0.5 sec window hop) and training a classifier per time window produced a curve of accuracy vs. time window indicating how well the outcome was decoded at the population level above chance ($\Delta$ accuracy) in a specific session (Fig. 4d and Supplementary Fig. 6b). For the control group, the population activity gradually became predictive from the second session (p = 0.004 for a t-test comparing the 2^nd training day to the 3^rd, p = 0.6 for a t-test comparing the 3^rd training day to the 4^th, n = 12) and stabilized after three training sessions from the 4^th day of training at the post-movement time window (approximately 5-6 seconds after the go cue, p = 0.9, F = 0.19 for one-way ANOVA comparing training days 4-7, n = 12). For the manipulated group, the outcome became gradually predictive from the 5^th training session after CNO was lifted (p = 0.01, F = 4.95 one-way ANOVA comparing training days 5-7) reaching comparable value to control on day 7 (Fig. 4d; p = 0.47 for a t-test comparing the manipulated to control groups on day 7).

We next addressed the question of whether the population encoding of outcome gradually transforms towards a fixed configuration, or rather, the population representing outcome varies across training sessions being specific per training session. Toward this end, we used the model trained on the 7^th day of training at the last analysis window (go cue +7.5 sec) and applied it to predict outcome signals from the neuronal activity of previous training sessions. Delta accuracy (Fig. 4e) as a function of the training session showed that the 7^th day model was unable to decode the outcome from ensemble activity of beginner animals (1^st training session). However, with training, the accuracy of the 7^th day model in decoding outcome increased and became significantly above chance from the 4^th session and onward (p = 8.5 ∙ 10⁻⁴, p = 1.0 ∙ 10⁻⁴, p = 6.3 ∙ 10⁻⁵, p = 4.0 ∙ 10⁻⁹ for sessions 4–7 respectively, one tail t-test, n = 12). This monotonic increase in accuracy confirmed our hypothesis that encoding of outcome by population activity progresses towards a specific profile with training. For the manipulated group the accuracy of the 7^th training day model began to rise only when CNO is lifted but did not reach the same level as the control group (Fig. 4e; p = 0.01, unpaired t-test, comparing the 6^th session of the manipulated group, n = 10, to the control group, n = 12).

For the two experimental groups controlling for the DREADDs manipulation (sham virus expression and the Ringer injection), we observed a monotonic increase in the decoding outcome accuracy based on 7^th model day similar to the control mice (Supplementary Fig. 2c) further stressing the specificity of the inhibition of dopaminergic axons in M1 by hM4D DREADDs.

Collectively these results show that outcome representation by layer 2-3 PNs network is not inherent (as it does not exist in beginner animals) but acquired and gradually transforms towards a population activation profile through training, and that dopaminergic VTA projections to M1 play a fundamental role in facilitating this outcome network learning process.

Functional network connectivity reorganization of layer 2-3 PN during motor learning

In addition to the changes in activity dynamics, we set out to investigate the changes in the functional network connectivity configuration that evolves during the motor learning process. To get insight into the changes in the functional connectivity within the layer 2-3 network during the training sessions, we extracted a correlation matrix per trial for every training session based on pairwise Pearson’s correlation coefficients between all possible pairs of ROIs (Fig. 5a). We next calculated the centroid of all trial-based correlation matrices of a given session, creating a single matrix for the different trials of a session. As correlation matrices are Semi-Positive Definite, they are located on a curved manifold and do not obey Euclidean geometry. Therefore, instead of using a Euclidean average, we used Riemannian geometry to extract their centroid on the curved manifold^43,44 (see Methods). Overall, this analysis produced a single matrix per training session, the Riemannian centroid, representing the average correlational structure of the network in a given training session (Fig. 5a).

**Fig. 5: Transformation of the functional connectivity of layer 2-3 network during motor learning and the role of ventral tegmental area (VTA) dopaminergic projections to M1 in the process.**

For illustration purposes, we first computed a low-dimensional representation for the trial-based correlation matrices and their centroids for a control animal and a manipulated animal using diffusion maps^45,46 combined with Riemannian geometry (Fig. 5b, see Methods), ultimately obtaining a visual representation of each session, where each trial was represented by a colored dot and the whole session by its Riemannian centroid (black)^30,32. In the control animal, the embedded correlations were mainly grouped by the training session, that is the trials within each training session had similar connectivity matrices (up to some intrinsic trial variability), whereas trials from different sessions had different connectivity matrices. Furthermore, with training, we observed a typical trajectory of transformation of the connectivity matrices indicating an evolution of the layer 2-3 neurons’ functional connectivity that occurs along the training process. For the manipulated group, the embedded correlation matrices did not differentiate well between the trials of the training sessions with CNO (2^nd–4^th) which were grouped together with the 1^st session, and only in the following (5^th–7^th) sessions a differentiation developed. These results indicate the essential role of M1 dopaminergic projections in changing the functional connectivity profile of the network during learning.

We quantified the changes of the functional network connectivity with training, at three levels of analysis (Fig. 5c). First, the detailed micro level of individual functional connections, where we took into account the specific identity of pairwise neuronal connectivity. Second, the mezzo level of individual neurons, where we computed the degree of connectivity of each neuron with other neurons in the network regardless of the identity of the connection, thereby quantifying the strength each neuron is connected to the network. Third, the macro level of the entire network, where we evaluated the average functional connectivity of all neurons in the network. For these three levels of analyses, we evaluated the functional connectivity of the network in each session as the Riemannian centroid matrix of all related trial-based correlation matrices (Fig. 5a, see Methods).

To capture the transformation in functional network connectivity at the micro individual connection level, and to account for pairwise relationships between cells, we evaluated the Riemannian similarity between the Riemannian centroid matrices (Fig. 5a) in different training sessions (Fig. 5d, e, see Methods). In the control group, the Riemannian centroid connectivity matrices were most similar to subsequent training sessions, where the similarity to the 7^th session was gradually rising (Fig. 5d; R² = 0.55, p = 1.2·10⁻¹³ linear fit, n = 12), while for the manipulated group, this process was observed only after CNO was lifted during the training sessions (Fig. 5e; a rising trend from day 4–7, R² = 0.45, p = 7.4∙10⁻⁵ linear fit, n = 10).

The analysis at the micro level demonstrated that the functional connectivity reorganizes during the training and that the specific pairwise connectivity footprint each neuron makes in the network is important and evolves with training. Moreover, our finding shows that this reorganization process required dopaminergic activation in M1.

Next, we investigated whether the reorganization of the individual functional connections also impacts the average functional strength each neuron is connected in the network. It is possible that individual connections will change but the average connectivity per neuron will remain constant due to a balance between strengthening and weakening of connections of the neuron. To test this, we applied the mezzo-level analysis and computed the degree connectivity of each neuron at a given training session. The degree (overall connectivity) of a neuron was calculated by summing over the rows of each Riemannian centroid connectivity matrix, resulting in a vector expressing the extent each cell is correlated with all other cells within the network in a given training session (Fig. 5c) and compared the values of degree across training sessions (Fig. 5f, g; see Methods). We found that the degree profile of the network transformed toward the configuration on 7^th training session in the control group (Fig. 5f; R² = 0.46, p = 6.2∙10⁻¹¹, linear fit, n = 12). No such transformation was observed for the manipulated group in training sessions 2–4 where CNO was applied, however, once CNO was lifted during the consequent training sessions the degree similarity transformed fast (Fig. 5g; R² = 0.6, p = 1.6∙10⁻⁷, linear fit, n = 10) reaching similar level as the control group. Repeating this analysis for the pre-tone time segment leads to similar trends yet noisier possibly due to smaller number of time points used for computing the correlation matrices (See Supplementary Fig. 7a, b), suggesting a network organization that is not solely ascribed to the motor execution itself.

Our findings indicate that the reorganization of the network during learning also involves changes in the overall strength of connectivity of individual neurons to the network and verified that dopaminergic inputs at M1 were essential for this transformation.

It should be noted that the Riemannian and degree similarity developed faster after lifting the CNO compared to the control group indicating some network reorganization occurred despite the presence of CNO.

For the two experimental groups controlling for the DREADDs manipulation (sham virus expression and the saline injection), we observed a monotonic increase in the Riemannian similarity and Degree similarity to the 7^th day of training, similar to the control mice (Supplementary Fig. 2d, e). This further emphasizes the specificity of the inhibition of dopaminergic axons in M1 by hM4D DREADDs in the reorganization of network functional connectivity.

Finally, at the macro level, we tested whether the global connectivity of the network changes during the training sessions. We compared the average degree across all cells per training session and found that the mean degree of the network did not exhibit a significant change with training (p = 1, F = 0.01, n = 12, for the control group, p = 1, F = 0.15, n = 10 for the manipulated group, one way ANOVA test). This result indicates that despite the reorganization of functional connectivity at the micro and mezzo levels, the overall connectivity level of the network remained unchanged. This result may be explained by the fact that some cells increased, and some decreased their connectivity such that the overall connectivity in the network did not change with training.

Taken together, our results indicate that layer 2-3 network undergoes a reorganization towards an expert functional connectivity configuration with training, which critically depends on activation of the VTA dopaminergic projections to M1. The expert configuration depends on both the specific connectivity profile of each neuron and on the strength of connectivity of each neuron within the network.

Discussion

In this study, we examined the plasticity changes in the layer 2-3 PNs network of M1 during the learning process of a dexterous motor task, and the role that direct dopaminergic inputs to M1 play in this process. We found that while the average activity of layer 2-3 PNs did not change significantly during the training process, the activity kinetics along the trial changed and became more similar to the “expert” configuration of the 7^th day of training. Moreover, while we did not observe changes in the average functional connectivity of layer 2-3 PNs, we observed a gradual reorganization of the specific functional connectivity profile of neurons and the strength each neuron is connected in the network, which became more similar to the final “expert” functional connectivity configuration. The fact the average connectivity of the network remains stable probably reflects homeostatic mechanisms that maintain the overall excitability of the network. The monotonic transformation of activity kinetics as well as network functional connectivity during learning were critically dependent on the VTA dopaminergic inputs to M1, as inhibiting these projections during the learning sessions blocked both motor learning and plasticity changes in the reorganization of activity and functional connectivity of layer 2-3 PNs network. Importantly, inhibiting dopaminergic inputs to M1 in expert mice did not affect performance, underscoring that blocking these inputs did not impair motor performance but is essential for the learning process itself. In addition, we found dopaminergic VTA projections to M1 are crucial in encoding outcome of the learnt task at the single cell and population of layer 2-3 neurons. This VTA dependent reorganization of specific functional connections during learning can serve to generate specific subnetworks (ensembles) that underlie memory.

We report that the total averaged layer 2-3 network activity did not change significantly during learning of the hand reach task. This is consistent with a previous study, that showed stability in the fraction of active layer 2-3 neurons during a lever press task¹⁶. However, despite this seemingly stable network activity a deeper look revealed learning was associated with major reorganization of both the neuronal activity dynamics and the network connectivity. During learning, the activity dynamics of neurons along the trial changed and transformed to the expert activity configuration adopting a typical three phasic kinetics¹³.

In addition to activity dynamics, we examined the effect of learning on functional connectivity. During most experimental paradigms, the access to real connectivity levels of neurons in networks is extremely limited. Thus, a common practice is to evaluate the functional connectivity by measurements of pairwise correlations between neuronal signals for various signals such as functional magnetic resonance imaging (fMRI)^47,48, calcium imaging or single unit recordings⁴⁹ where configurational structure of the network is compared across time or across subjects through centrality measures such as degree⁵⁰.

In this work, we use a similar practice, where we evaluate the functional connectivity between individual neurons using the correlation matrix, calculated per trial. Here we developed a geometric approach pipeline for tracking changes in correlational structure throughout the training process. In this new approach, we calculate the Riemannian centroid across trials, to express the correlational structure of all neuronal pairs, in a given session. We found that while there were no significant changes in the overall connectivity of the network (globally averaged degree centrality), the learning process induced a significant change in the functional connectivity of the individual neurons and their total connections in the network thus keeping a homeostatic level of connectivity while introducing specific task related connectivity changes. The correlational structure expressed by all pairwise relations gradually transform and converges towards an “expert” configuration indicating that the learning influences specific connections in the network.

Taken together our findings support four processes that occur during learning. First, formation of new specific connections between neurons and thereby probably forming specific ensembles for storing the new task^51,52. Second, to maintain the overall excitability of the network, weakening of other unrelated functional connections occurs. Third, shaping of a new tri-phasic dynamic of the neuronal response¹³. This is probably mediated both by the formation of new connections between excitatory neurons, and probably by connectivity changes with inhibitory interneurons⁵³. The activity dynamics and connectivity of interneurons were not examined in this study, yet it is important to stress that interneurons in M1 are being innervated by the VTA⁵⁴ possibly contributing to the tri-phasic dynamics of the layer 2-3 activity. Finally, learning was associated with the development of a specialized subpopulation of neurons that encoded task outcome. We hypothesize that these outcome signals are crucial for reinforcement motor learning¹³, providing feed-back for future adaptations.

Our results are in line with the increased temporal correlations of pairs of neurons reported for anterolateral (ALM) and posterior medial motor cortexes during learning of lick no-lick odor task, though correlations were generally low, and the network correlative configuration was not studied³⁶. In addition, our results are also consistent with the previously reported increased stability of the temporal activity pattern of neurons and the relationship between activity and movement in pairs of trials that became more consistent with learning of a lever press task^16,19.

Our results, however, are inconsistent with two previously described findings with regard to layer 2-3 neurons during motor learning. First, a previous study that reported unchanging total predictive decoding accuracy of the motor behavior by both the neuronal population or single layer 2-3 neurons of M1 during lever pull motor learning³⁵, a property that was attributed specifically to layer 5a but not 2-3³⁵. Second, our results are also inconsistent with the notion that the single neurons are unstable in their activity during trials, tuning or predictive decoding of behavior during the learning sessions^16,19,35,36. We show that increasing fraction of neurons gradually change and converge toward the “expert” activity configuration where each neuron adopted gradually its specific activity pattern during the learning process, which can potentially explain the gradual transformation towards the expert configuration seen at the network functional connectivity level.

Previous work demonstrated that the main source of dopaminergic projections to M1 originate from the meso-cortical dopaminergic system, the VTA. Further, this pathway was shown to be necessary for hand reach motor skill learning²⁴ as eliminating the dopaminergic terminals with 6-OHDA in M1 or blocking D1 and D2 receptors affected the motor skill learning⁵⁵. Consistent with previous works, behaviorally, we found that dopaminergic inputs to M1 are essential for motor learning as inhibiting these inputs locally in M1 prevented learning progression^24,27,56. However, in contrast to a previous study that showed VTA lesioning did not affect learning within session rather only between sessions²⁴, we show that inhibiting VTA terminals with DREADDs in M1 halted both within and between sessions motor learning, preventing mice from increasing the success rate in the CNO sessions. In addition, our analysis of the motor behavior during the hand reach task revealed that inhibition of the direct dopaminergic projection also decreased the number of task related movements and delay time from go cue to first lift, indicating the importance of direct dopaminergic projections on learning not only through success rates but also on fine behavioral parameters. Furthermore, our GLM analysis indicate that layer 2-3 neurons represent task variables in a small but significant portion of the neurons ( ~ 13%) suggesting a sparse representation of movement in this layer. In contrast, layer 5 pyramidal tract (PT) neurons densely encode movement^13,57. Moreover, the GLM analysis shows a shift of the modeled activity from sensory to outcome representation during the learning process, a shift that was also dependent on the activation of VTA dopaminergic projections to M1 during learning. These findings are in line with recent work showing a role of dopamine in coordinating fine reaching skilled movements⁵⁸ though in this case midbrain dopaminergic neurons in the SNc and not VTA dopaminergic neurons were manipulated.

It should be noted, that in our experiments, although learning in the manipulated group was gradual after the release of CNO, the rate of learning was faster compared to the control group in the functional connectivity measures. This may indicate that some motor learning occurred in the presence of CNO, potentially due to either an incomplete blockade of the dopaminergic signals to M1, allowing some plasticity in M1 during the CNO days, or the involvement of additional mechanisms beyond dopaminergic inputs to M1. For example, dopaminergic signals in the striatum, conveyed to the cortex via thalamocortical inputs, or other plasticity mechanisms in M1 independent of the dopaminergic VTA axons, may also contribute to motor learning.

Interestingly, previous studies have reported the importance of dopaminergic SNc afferents to the striatum in motor learning^59,60,61,62. Worth noting that we did not observe an effect on the progression of learning when inhibiting SNc projection to M1, which is consistent with the sparser direct innervation from the SNc to M1 compared to that from the VTA²⁴. Thus, motor learning probably requires concerted activation of both direct dopaminergic signaling to M1 via the VTA and dopaminergic signaling to the striatum. It is yet unclear whether the two different pathways carry different information and/or are responsible for different aspects of learning or whether learning requires a double gate dopaminergic mechanism. The fact M1 receives both a general reward prediction error signals^1,22 and develops a specific task related outcome signal as part of the learning process may be related to this question. Further studies are required to unravel this interesting and important question.

While the role of direct dopaminergic VTA afferents to M1 on learning was described previously at the behavioral level, information regarding the role of dopaminergic VTA projections to functional connectivity and activity dynamics in M1 during learning was lacking. In this study we attempted to close this gap and show that for both single neuron dynamics and network functional connectivity reorganization to emerge, and for the concomitant motor learning to proceed, activation of direct VTA dopaminergic inputs to M1 is essential.

It is interesting to note that the dopaminergic VTA axons project to a wide range of areas in the brain⁶³, with the majority of projections innervating subcortical regions, primarily the nucleus accumbens (medial and lateral) and amygdala⁶⁴. VTA is also known to innervate other cortical regions, especially the prefrontal cortex⁶⁵, however, other cortical regions including M1 are innervated to a lesser degree. Here we show that despite this scarcer innervation, dopaminergic VTA projections to M1 are vital for motor learning and for the reorganization of the M1 network during the training process.

Dopamine is thought to play a major role in reinforcement learning providing a reward prediction error signal, comparing between expected and experienced outcome which than leads to strengthening of the appropriate synapses via plasticity mechanisms⁶⁴. In this study we did not address the cellular mechanisms by which dopamine participates in reorganization of the M1 network during learning. Previous studies showed that the dopaminergic innervation to M1 from the VTA is important for inducing LTP in M1 layer 2-3 synapses²⁷ in line with the findings that Horizontal connections of layer 2-3 motor cortex undergo LTP during forelimb motor skill learning^21,56. At the structural plasticity level task related spine appearance and elimination in layer 2-3 neurons during motor learning was shown^16,20. Moreover, in layer 5 neurons spine turnover was shown to depend on direct mesocortical projections to M1⁶⁶. It is likely that one or more of these mechanisms are responsible for the effects we describe in this study. Thus, our findings together support the notion that the reorganization at the single cell and network levels of M1 are supported by dopamine released during the motor learning process from the VTA (see summary Fig. 6).

**Fig. 6: Schematic highlighting the essential role of dopaminergic inputs from the VTA to M1 in motor learning.**

Methods

Experimental model and subject details

All animal procedures were performed in accordance with guidelines established by the NIH on the care and use of animals in research, as confirmed by the Technion Research Campus Institutional Animal Care and Use Committee. Adult male Slc6a3 (DAT-Cre) mice were used in this study. Mice (Jackson Laboratory; strain number #020080 from a mixed genetic background of C57BL/6 J and C57BL/6 N) were housed in a 12:12 hours reverse light: dark cycle, kept at controlled humidity (50%) and temperature (22⁰C). For behavioral training and experiments, food intake was limited to 2.65–3 g/day with ad libitum water. The typical age at which animals began experiments was 8 weeks.

Experimental design

To investigate the activity and network dynamics in M1 during motor learning, we trained mice to perform a head-fixed version of the hand reach grasping motor task. After the first successful hand reach to pellet attempt, all subsequent training sessions (1–7 sessions) were performed with simultaneous two-photon calcium imaging (control group). Chronic calcium imaging was performed with the genetically encoded calcium indicator GCaMP6s, which was introduced to the cells via a viral vector injection. The resulting permanent expression of the indicator allowed us to monitor the activity for many weeks. We selectively labeled layer 2-3 PN using a specific promotor (CaMKII). We implanted a chronic window over the M1 forelimb area that allowed us to image the activity for long time periods. For inhibiting the activity of VTA dopaminergic projections to M1 we used chemogenetic inhibition using DREADDs which was introduced to the cells via a viral vector injection. We selectively labeled dopaminergic neurons in the VTA using Cre-dependent viral expression in DAT-Cre mice, the virus that encodes for the inhibitory DREADDs (hM4D) was injected in the VTA of both control and manipulated groups and assignment to the two experimental groups was done later. In the manipulated group, CNO was injected for three consecutive training sessions (2, 3 and 4). Injections were performed locally to M1 via an access port⁶⁷ which allowed specific inhibition of the dopaminergic projections from VTA to M1. In some cases (as indicated in the text), we also applied CNO in expert sessions to examine the effect of the inhibiting dopaminergic projections to M1 on movement.

In addition to the control and manipulated groups, we had four additional groups, which controlled for the CNO, the viral injection, the origin of dopaminergic afferents, and the spatial cortical location:

1. The Ringer group. In this group, we used mice that underwent the same surgery, with injections of GCaMP6s in M1 and with DREADDs in the VTA. However, in the manipulated sessions (second, third, and fourth) instead of injecting CNO locally to M1, we injected Ringer solution. 2. The sham group. Mice were injected with GCaMP6s in M1, however, instead of injecting DREADDs to the VTA, we injected a sham virus (pAAV-flex-tdTomato). During the manipulated sessions, this group was injected locally with CNO to M1. 3. The V1 group. This group was injected with DREADDs in the VTA, however, during the manipulated sessions instead of injecting CNO locally to M1, CNO was injected locally in V1 (coordinates 2.7 mm posterior and 2.5 mm lateral to Bregma), which created a spatial negative control. 4. The SNc group. In the SNc group, instead of injecting hM4D DREADDs to the VTA, DREADDs were injected to the SNc bilaterally (3.4 mm posterior, 1.5 mm lateral to Bregma; depth of 4000 µm, volume of 120 nl for each side). This group was injected with CNO locally to M1 during the manipulated sessions.

Viral injections and cranial window surgery

The procedures were performed on 2-3 months old mice, under isofloren anesthesia (4% for induction and 1.5–2% for maintenance during the surgery). Mice were secured on a stereotaxic apparatus (KOPF), and a heat pad was used to maintain body temperature at 36–37 °C. Pre-operation medications were administrated subcutaneously for analgesia, Ketoprofen (5 mg/Kg) buprenorphine (0.1 mg/Kg). The scalp was shaved and cleaned with ethanol. The skull surface was exposed after a subcutaneous injection of 2% Lidocaine. Viral injection was performed using a hydraulic manipulator (M0-10 Narishige) via thinned skull at two points: 1. M1 forelimb area¹⁰ (0.6 mm anterior and 1.6 mm lateral to Bregma) at three depths (150 µm, 250 µm, 350 µm, with 70 nl each depth). For calcium imaging at M1, we injected AAV.CamKII.Gcamp6s. 2. VTA bilaterally (3.1 mm posterior, 0.3 mm lateral to Bregma; depth of 4200 µm, volume of 120 nl for each side). For VTA inhibition we used DREADDs AAV2-hSyn-DIO-hM4D(Gi)-mCherry. After injections we performed a circular craniotomy over the M1 forelimb area (0.6 mm anterior and 1.6 mm lateral to Bregma) and implanted a 3 mm diameter laser-cut glass optical window with a 1 mm access hole which we sealed with biocompatible silicone for future local injections. The window was sealed with superglue and dental cement. Next, a custom-made 3D printed plastic headpost¹³ was affixed to the skull with dental cement. Mice were injected with ketoprophen (5 mg/kg) and buprenorphine (0.1 mg/kg) administered subcutaneously for 2-days post operation and were left to recover from the surgery for two weeks with ad libitum food and water.

Local CNO/ ringer injections

CNO or Ringer were injected unilaterally in M1 and contralateral to the hand performing the task immediately before starting the manipulated sessions. In the Ringer group, Ringer was injected locally to M1, while in the V1 group, CNO was injected locally to V1. The injections were done via an access port in the cranial window which was sealed with biocompatible silicone⁶⁷ using hydraulic micromanipulator with a glass pipette that was inserted into the cortex with an angle to reach the specific imaged area in M1 (typically 250–300 µm from the dura). CNO was dissolved in saline to reach a concentration of 100 µM. 500 nl of CNO (100 µM) or of Ringer solution was injected according to the experimental group. The experiments started 5 min after the injection of the CNO and lasted for 30–40 min.

Behavioral training

After recovering from the surgery mice were restrained to 2.65–3 gr/day and ad libitum water. Training started when mice reached 80–90% of their initial body weight. Mice were habituated to head fixation in a custom-built apparatus¹³ in dark and quiet conditions, monitored by a webcam. We trained to reach for a food pellet (20 mg; Test Diet; St Louis, MO) from a rotating table placed in front of the animal and driven by either a NI USB DAQ device or a Teensy microcontroller driven by custom-made LabVIEW software. An auditory tone (200 ms, 1 kHz) was used as a cue during plate rotation. Initially, the rotating table was placed directly below the mouth allowing the animal to tongue reach the pellet upon the tone. We gradually placed the table in further position until the first successful hand reach. Thereafter, training was combined with two-photon calcium imaging in 7 subsequent sessions. The first training day with imaging was called “session 1”. Each training session consisted of ~50 trials. Each trial lasted 12 seconds; the food pellet comes to position 4 sec after trail start along with a tone. The pellet is available to the mouse for 8 sec.

Two-photon calcium imaging

Images (512 × 512 pixels) were acquired using a two-photon microscope (Bruker) equipped an 8 kHz bidirectional resonant galvo scanner and a Nikon 16X CFI Plan Fluorite objective (NA 0.8, 3 mm working distance), controlled by the software package PrairieView 5.3. GCaMP6s was excited at 940 nm using a femtosecond pulsed laser (InSight X3, Spectraphysics). Emission light was detected by a GaAsP photomultiplier tube (Hamamatsu).

Each trial was 12 sec with an auditory cue that was presented in the 4^th second, with an imaging frame rate of 30 Hz. The same field of view was imaged over all sessions in the same animal. Behavioral performance was monitored at 200 Hz using two cameras (side and front view; Flea3 FL3-U3-13Y3M, PointGrey). The timing of two-photon calcium imaging, behavioral task, and video recording were coordinated via a National Instruments board (PCI-6110), using custom-made software written in MATLAB.

Histology

At the end of experiments, mice were deeply anesthetized and transcardially perfusion with PBS 0.1% followed by paraformaldehyde 4%. The brains were removed and put in fixative solution (paraformaldehyde 4%) for 48 hours, and later stored in PBS 1% solution. Coronal sections (100 µm thickness) were made. The sections were mounted on slides embedded with Fluoroshield containing DAPI (Sigma). Slices were scanned using a spinning disk confocal system (Olympus IXplore-Spin), equipped with a X20 objective (UPLXAPO 20x NA 0.8 WD 0.6 mm Olympus). The presented histology images (Fig. 1d; Supplementary Fig. 1) were acquired as z-stacks (at a definition of 1024 × 1024 pixels with pixel size of 0.646 µm and Z stack interval 0.67 µm) and presented as max-projection images. Injection sites were validated via histology for all manipulated animals.

Behavioral data analysis

We labeled the behavioral events (lift, reach, grab, supinate, at mouth and Back to perch) using a semi-supervised software^37,38 as follows: lift – a vertical movement of a forelimb from the perch or the table, lasting till peak height is reached; reach – a horizontal movement towards the pellet; grab – starts when animals spread their fingers, ending when fingers are wrapped into a fist. Supinate – the forepaw is rotated outward while being brought towards the mouth; at mouth – fist is close to the mouth; back to perch - forepaw returns to perch.

To characterize behavior throughout training we segmented the data into sequences of behavioral events a sequence is defined as a series of consecutive events ending with an “at mouth” event or a “back to perch” event. Behavioral transitions were counted as the number of transitions between events within each sequence per trial at each training session.

Two-photon data analysis

The fluorescence data acquired by the two-photon microscope was first registered to correct for brain motion artifacts. Our registration method was based on⁶⁸, using Fourier transform based correlation between two successive images. The maximal value position in the correlation image specifies the relative shift between the two images; we designate them ${u}_{t}$ and ${v}_{t}$. This method requires a template specification and matching against an image stack. The template image ${I}_{{temp}}(x,y)$ was defined as the average of all images in the selected trial over time.

$${I}_{{temp}}\left(x,y\right)=\frac{1}{N}\sum _{t=1..N}I\left(x-{u}_{t},y-{v}_{t},t\right)$$

(1)

The set $\left\{{u}_{t},{v}_{t}\right\},t=1..N$ is an image shifted in the XY plane after alignment. We initially start with ${u}_{t}=0,{v}_{t}=0$ and then update their values according to the registration maxima. This procedure is repeated several times, when each time we compute the new template ${I}_{{temp}}(x,y)$ using previously computed $\left\{{u}_{t},{v}_{t}\right\}$ for each image. Typically, this procedure converges after several iterations, in our case three iterations.

To align the imaging data over many trials, we used a similar technique, utilizing the previously computed averaged templates for each trial. For each trial k, we performed a single trial registration using the template algorithm for three iterations. To align the image data over many trials, we treated the final templates ${I}_{{temp}}^{k}(x,y)$ from each trial k as unaligned image data and repeated the same registration procedure to find offsets $\left\{{u}_{k},{v}_{k}\right\}$ for each trial. These offsets, along with previously found offsets $\left\{{u}_{t},{v}_{t}\right\}$, account for the final image shift in the XY plane.

Regions of interest (ROIs) were detected manually using average fluorescence images and $\varDelta F/F$ projection images, which highlighted active neurons. The pixels within each ROI were averaged for every frame. The ROI “mask” was used to detect the same neurons on multiple imaging sessions on different days.

$\varDelta F/F$ was computed using the following formula:

$$\Delta F/{F}_{n}[t]=\frac{{F}_{n}\left[t\right]-{Min}10({F}_{n}\left[t\right])}{{Min}10\left({F}_{n}\left[t\right]\right)+{Bias}}$$

(2)

${Min}10\left({F}_{n}\left[t\right]\right)$ is a mean value of the lowest consecutive 10% values of the fluorescence signal ${F}_{n}\left[t\right]$. This minimal fluorescence value was calculated both per trial and across the entire experimental session, with no significant differences in the results with these two calculation variants. A small bias factor in the denominator prevented zeros when the cell was completely silent.

Statistical analysis

In each experiment, the mouse repeated the hand reach task several dozen times. The data were analyzed using custom Matlab code, except for training and testing the SVM classifiers^41,69, for which we used the standard LIBSVM software⁷⁰.

The imaging data collected at each session $l$ are stored in a 3-dimensional matrix (tensor) ${X}^{l}$ of size ${N}_{r}\times t\times {T}^{l}$, where ${N}_{r}$ is the number of neurons, $t$ is the number of time samples per trial, and ${T}^{l}$ is the number of trials. ${X}^{l}\left(i,j,k\right)$ is the neuronal activity of the $i$-th neuron at the $j$-th time sample in the $k$-th trial related to the $l$-th session, where $l=1,\ldots,7$.

Activity of cells through training

The overall average activity of the entire population of cells per training session was obtain as:

$${\bar{x}}^{l}=\frac{1}{{T}^{l}{N}_{r}}\mathop{\sum }\limits_{i=1}^{{N}_{r}}\mathop{\sum }\limits_{j=1}^{t}\mathop{\sum }\limits_{k=1}^{{T}^{l}}{X}^{l}\left(i,j,k\right),l=1,\ldots,7$$

(3)

The average dynamics of cells in each session was obtained by averaging over trials as follows,

$${\bar{X}}^{l}\left(i,j\right)=\frac{1}{{T}^{l}}\mathop{\sum }\limits_{k=1}^{{T}^{l}}{X}^{l}\left(i,j,k\right),j=1,\ldots,t;i=1,\ldots,{N}_{r};l=1,\ldots,7$$

(4)

We obtained the correlation coefficient between the average dynamics of each cell in the 7^th session with its own dynamics in all other sessions:

$${\rho }^{l}\left(i\right)=\frac{\sum _{j}\left({\bar{X}}^{l}\left(i,j\right)-{\bar{x}}^{l}\left(i\right)\right)\left({\bar{X}}^{7}\left(i,j\right)-{\bar{x}}^{7}\left(i\right)\right)}{\sqrt{\sum _{j}{\left({\bar{X}}^{l}\left(i,j\right)-{\bar{x}}^{l}\left(i\right)\right)}^{2}}\cdot \sqrt{\sum _{j}{\left({\bar{X}}^{7}\left(i,j\right)-{\bar{x}}^{7}\left(i\right)\right)}^{2}}}$$

(5)

where ${\bar{x}}^{l}\left(i\right)$ is the activity of the $i$-th neuron in the $l$ session, averaged across time and trials.

SVM classification of ensemble activity

To further quantify the activity profile of the cellular ensemble throughout training, we trained a linear SVM classifier to separate neuronal activity in trials related to the first session from the activity in trials related to the 7^th session. We used a sliding window of 1 sec with 0.5 sec hop. In every time window, we evaluated the average activity of each neuron at each trial,

$${\bar{x}}_{b}^{l}(i,k)=\frac{1}{B}\sum _{n\in {\rm{time}}\,{\rm{bin}}\,b}{X}^{l}(i,n,k)$$

(6)

where $b$ is the index of time window and $B$ is the window length measured in samples. The activity of the network in each time window for each trial was represented by the following ${N}_{r}\times 1$ vector:

$${\bar{x}}_{b,k}^{l}=\left(\begin{array}{c}{\bar{x}}_{b}^{l}\left(1,k\right)\\ {\bar{x}}_{b}^{l}\left(2,k\right)\\ \vdots \\ {\bar{x}}_{b}^{l}\left({N}_{r},k\right)\end{array}\right),k=1,2,...,T$$

(7)

For each time bin, we extracted the set ${\left\{{\bar{x}}_{b,k}\right\}}_{k=1}^{{T}^{l}}$, representing the averaged activity of the ensemble in all trials related to all 7 training sessions. We paired the activity of each trial with a label indicating whether it is associated with expert activity (7^th session) or with beginner activity (1^st, session):

$${y}_{k}=\left\{\begin{array}{cc}1 & {\text{trial}}\,{{k}}\,{\rm{is}}\; {\rm{related}}\; {\rm{to}}\; {\rm{expert}}\\ 0 & {\text{trial}}\,{k}\,{\rm{is}}\; {\rm{related}}\; {\rm{to}}\; {\rm{beginner}}\end{array}\right.$$

(8)

We trained a linear classifier (SVM) to separate expert from beginner, per time bin, using 10-fold cross validation. We then applied the trained classifiers to estimate trials from sessions 2–6. For the 1^st and 7^th sessions, we used the output of the cross-validation procedure. For each session we evaluated the fraction of trials identified as expert, per time window $b$,

$${a}_{l}(b)=\frac{1}{\#{\rm{trials\,{on}}}\,{\rm{session}}\,{l}}\sum _{k\in \,{\rm{session}}\,s}{\hat{y}}_{b,k}$$

(9)

where ${\hat{y}}_{b,k}$ is the estimated label. Figure 2f presents mean ± SEM values of ${a}_{l}\left(b\right)$, averaged across animals, for a time window centered at tone+1 sec. Figure 2g presents the average values as a function of the time bin $b$.

Generalized linear model (GLM)

We used a GLM to measure the relative contribution of behavioral-sensory signals to modeling the dynamics of the Calcium transients. To this end, we used a similar strategy to the previously published one^13,71 where a model is trained to predict the calcium signal of each neuron on a trial-by-trial basis, based on behavioral-sensory data. The predictor signals were of four types:

Hand trajectories: Time series of hand trajectories and 4 fingers were extracted using DeepLabCut software³⁹ from videos taken with side and front view cameras. We extracted x and y locations, altogether 20 predictors.

Orofacial features: Time series of orofacial motion variability features were extracted using FaceMap software⁴⁰ from recorded videos. A single ROI was placed to include the face area, where the first 20 principal components were taken as predictors.

Time-varying and binary events: tone, lift, reach, grab, supinate, at mouth, and back to perch). Movement events were extracted using the modified JAABA software^37,38.

Whole trial binary event: success/failure trial status (outcome). We convolved the time-varying binary events to model the time course of single-neuron calcium signals with a set of 7-degree-of-freedom regression splines. We used a set of 4 durations: 0.25 sec, 0.5 sec, 1 s, and 2 sec, all generated using the ‘bSpline’ package in R. We had 32 spline functions in total, which resulted in 32 × 7 = 192 convolved signals used as predictors. Altogether, we used 265 predictors (224 convolved predictors + 20 hand trajectories + 20 orofacial features + 1 whole binary success/failure outcome). We performed this analysis in the peri-movement time window (from one sec before until 2 s after the tone). The data for this analysis included 850 cells for the control group (pooled from 5 animals) and 606 cells for the manipulated group (also pooled from 5 animals). We trained a GLM per cell, per training session and accounted for cells for which the explained variance was at least 5%. We then evaluated the relative contribution of each component we trained a series of 9 partial models, each excluding predictors related to one category. We computed the contribution of each category as:

$$\alpha \left(i,c\right)=1-\frac{{R}_{p}^{2}\left(i,c\right)}{{R}_{f}^{2}\left(i\right)},c=1,\ldots,9$$

(10)

where ${R}_{f}^{2}\left(i\right)$ is the explained variance of the $i$ -th cell using the full model (i.e., including all predictors) and ${R}_{p}^{2}\left(i,c\right)$ is the explained variance of the $i$ -th cell based on the partial model (excluding the components related to the category c. All explained variance values were obtained using a 10-fold cross-validation process, where per fold 10% of trials were set aside as a test set. Figure 3e presents the relative contribution of each category evaluated as $\alpha \left(i,c\right)/{\sum }_{c=1}^{9}\alpha \left(i,c\right)$.

Indicative neurons

We examined whether single neurons can reliably report success or failure by using a similar analysis for the detection of “indicative neurons” as we described in ref. ¹³.

We trained an SVM-classifier for the prediction of trial outcome based on the average activity of each neuron in a given time bin (window length 1 sec, window hop 0.5 sec). We then evaluated the accuracy rate using a 10-fold cross-validation process. We obtained the mean and STD values of the accuracy rate, per neuron and at each time window, by averaging over folds. To determine whether a neuron is indicative, we compared its respective mean accuracy rate to prior probability, which is the maximum between the fraction of successful trials and the fraction of failure trials. In each time window, we marked a neuron as indicative if its mean accuracy rate was higher than the prior probability, with a 99%.

PCA trajectories of network dynamics

We explore the evolution of the network in encoding outcome using Principal Component Analysis (PCA). We reshaped the imaging data tensor of a given session, ${X}^{l}$, into a 2-dimensional matrix of ${N}_{r}\times [t{T}^{l}]$, where each row consists of the neuronal activity of a specific cell across all time samples and trials:

$${\widetilde{X}}^{l}=\left(\begin{array}{ccc}{X}^{l}\left(1,\cdot,1\right) & \ldots & {X}^{l}\left(1,\cdot,{T}^{l}\right)\\ \vdots & \vdots & \vdots \\ {X}^{l}\left({N}_{r},\cdot,1\right) & \ldots & {X}^{l}\left({N}_{r},\cdot,{T}^{l}\right)\end{array}\right)$$

(11)

We computed the sample covariance of ${\widetilde{X}}^{l}$ and extracted its eigenvalue decomposition to obtain $d$ principal components explaining 95% of the variance of ${\widetilde{X}}^{l}.$ Finally, we reshaped the new representation into a $d\times t\times {T}^{l}$ tensor of the principal components of ${X}^{l}$. Figure 4b presents the first 3 principal components, averaged according to outcome, computed per session for two animals.

Significance in separation

Let ${\boldsymbol{S}}_{{suc}}$ be a $d\times t\times {T}^{{suc}}$ tensor of all trajectories related to successful trials in each session, and ${\boldsymbol{S}}_{{fail}}$, a $d\times t\times {T}^{{fail}}$ be the corresponding tensor for failed trials. First, we computed the average trajectories for success and for failure, ${\boldsymbol{\mu }}_{{suc}}(n)$ and ${\boldsymbol{\mu }}_{{fail}}(n)$, and the corresponding sample covariance matrices, ${\Sigma }_{{suc}}(n),{\Sigma }_{{fail}}(n)$:

$${\boldsymbol{\mu }}_{i}\left(n\right)=\frac{1}{{T}^{i}}\mathop{\sum }\limits_{k=1}^{{T}^{i}}{S}_{i}( \cdot,n,k)$$

(12)

$${\Sigma }_{i}\left(n\right)=\frac{1}{{T}^{i}}\mathop{\sum }\limits_{k=1}^{{T}^{i}}\left({S}_{i}\left(\cdot,n,k\right)-{\mu }_{i}\left(n\right)\right){\left({S}_{i}\left(\cdot,n,k\right)-{\mu }_{i}\left(n\right)\right)}^{T}i=\left\{{suc},{fail}\right\}$$

(13)

Second, we calculated the separation between success and failure dynamics using:

$$ {d}_{{suc},{fail}}(n)=\\ {\left({\mu }_{{suc}}(n)-{\mu }_{{fail}}(n)\right)}^{T}{\left(\left({\Sigma }_{{suc}}(n)+{\Sigma }_{{fail}}(n)\right)/2\right)}^{-1}\left({\mu }_{{suc}}(n)-{\mu }_{{fail}}(n)\right)$$

(14)

Third, we shuffled the trial outcome labels and then re-computed the distances. We repeated this computation 1000 times per session and obtained the empirical probability for the randomized distance to be larger than the actual distance, denoted by $\hat{p}\left(n\right)$. Finally, for each training session $l$, and time point $n$, we counted how many animals yielded a significant difference, i.e., $\hat{p}\left(n\right)$ < 0.05, as presented in Fig. 4c.

Classification of outcome per training session

We further quantified outcome separation emerging with training based on the predictive power of the imaging data tensor ${{\boldsymbol{X}}}^{l}$ using a linear classifier. The classification accuracy was evaluated for each training session, using a sliding window of 1 second with 0.5-second hop. In every time window $b$, we evaluated the average activity of each neuron at each trial, ${\bar{x}}_{b,k}^{l}$, as defined above. We then trained a binary SVM classifier (linear) to separate successful trials from failed trials, per time window. Training and testing were performed using a 10-fold cross-validation procedure. As success rate varies with training and across animals, we evaluated the decoding power of the trained classifiers as the difference between accuracy and chance level, calculated per training session.

Classification of outcome using expert model

We applied the SVM models trained to separate success from failure based on the activity of expert animals, at the last time window (go cue+7 seconds) to data extracted from previous sessions (last time window as well).

Riemannian centroid of correlation matrices

We computed the pairwise Pearson’s correlation coefficients based on the activity of cells at each trial:

$$\begin{array}{c}{C}_{i,j,k}^{l}=\frac{\sum _{n}\left({X}^{l}\left(i,n,k\right)-{\bar{X}}^{l}\left(i,k\right)\right)\left({X}^{l}\left(j,n,k\right)-{\bar{X}}^{l}\left(j,k\right)\right)}{\sqrt{\sum _{n}{\left({X}^{l}\left(i,n,k\right)-{\bar{X}}^{l}\left(i,k\right)\right)}^{2}}\cdot\sqrt{\sum _{n}{\left({X}^{l}\left(j,n,k\right)-{\bar{X}}^{l}\left(j,k\right)\right)}^{2}}}\\ k=1,2,...,{T}^{l}\;\;\;\;\;\;\;\;i,j=1,2,\ldots,{N}_{r}\;\;\;\;\;\;\;\;l=1,\ldots,7\end{array}$$

(15)

where ${{\rm{C}}}_{k}^{l}$ is the Pearson correlation matrix of the $k-$ th trial in the $l$-th training session. The resulting Pearson correlation matrices ${\left\{{{\rm{C}}}_{k}^{l}\right\}}_{k=1}^{{T}^{l}}$ are Symmetric Positive Semi Definite (SPSD) matrices and therefore do not obey Euclidean geometry. To correctly evaluate the overall connectivity in a given session and compare connectivity across sessions, we utilized the particular geometry proposed in ref. ³⁰ and ref. ³² as follows. Consider two Pearson correlation matrices ${C}_{k}$ and ${{\rm{C}}}_{l}$ of rank r. We decomposed each of these matrices such that:

$${{\mathbf{C}}}_{{\rm{k}}}={{\mathbf{G}}}_{{\rm{k}}}{{\mathbf{P}}}_{{\rm{k}}}{{\mathbf{G}}}_{{\rm{k}}}^{T}$$

(16)

$${{\mathbf{C}}}_{{\rm{l}}}={{\mathbf{G}}}_{{\rm{l}}}{{\mathbf{P}}}_{{\rm{l}}}{{\mathbf{G}}}_{{\rm{l}}}^{T}$$

(17)

where ${{\rm{G}}}_{{\rm{k}}}$ and ${G}_{{\rm{l}}}$ are ${N}_{r}\times r$ matrices with orthogonal columns that lie on Grassmann manifold⁷² and ${{\rm{P}}}_{{\rm{k}}}$ and ${{\rm{P}}}_{{\rm{l}}}$ are $r\times r$ Symmetric Positive Definite (SPD) matrices that lie on the SPD cone manifold³¹. The SPSD distance proposed in³⁰ is given by:

$${d}_{{\mathcal{S}}}\left({{\rm{C}}}_{{\rm{k}}},{{\rm{C}}}_{{\rm{l}}}\right)={d}_{{\mathcal{G}}}\left({{\rm{G}}}_{{\rm{k}}},{{\rm{G}}}_{{\rm{l}}}\right)+\alpha {d}_{{\mathcal{P}}}({{\rm{P}}}_{{\rm{k}}},{{\rm{P}}}_{{\rm{l}}})$$

(18)

where $\alpha \, > \, 0$ is a tunable hyperparameter, ${d}_{{\mathcal{G}}}$ is the Riemannian distance on the Grassmann manifold⁷², and ${d}_{{\mathcal{P}}}$ is the Riemannian distance on the SPD manifold³¹. The centroid of the Pearson correlation matrices in each training session was calculated as their Frechet mean³²:

$${\bar{C}}^{l}={{\arg }}\mathop{\min }\limits_{{\rm{C}}}{\sum }_{k}{d}_{{\mathcal{S}}}^{2}\left({\rm{C}},{{\rm{C}}}_{k}^{l}\right)$$

(19)

Diffusion embedding of SPSD matrices

We made use of this Riemannian distance to illustrate the development of the correlations of cells along the training sessions, by incorporating it into Diffusion Maps⁴⁶, a non-linear dimensionality-reduction algorithm. Using Diffusion Maps, the collection of Pearson correlation matrices ${\left\{{{\boldsymbol{C}}}_{k}\right\}}_{k=1}^{T},T=\sum _{l}{T}^{l}$ corresponding to trials from all experiments were embedded into a Euclidean space. First, we constructed a $T\times T$ affinity matrix, whose elements are given by

$${\boldsymbol{W}}({\boldsymbol{k }},{\boldsymbol{l}})=\exp \left\{-\frac{{\left({d}_{{\mathcal{S}}}\left({C}_{k},{{\rm{C}}}_{l}\right)\right)}^{2}}{{\sigma }^{2}}\right\}$$

(20)

where $\sigma$ is a scaling tunable parameter. This affinity matrix could be viewed as the weight matrix of the edges of graph, whose nodes are the trials. By normalizing the rows of ${\boldsymbol{W}}$, we obtained a row stochastic matrix ${\boldsymbol{P}}$,

$$\begin{array}{cc}{\boldsymbol{P}}={\boldsymbol{D}}^{-1}{\boldsymbol{W}},{\boldsymbol{D}}(k,k)=\mathop{\sum }\limits_{l=1}^{T}W\end{array}(k,l)$$

(21)

Then, we applied eigenvalue decomposition to ${\boldsymbol{P}}$,

$$\begin{array}{cc}{\boldsymbol{P}{\psi }}_{\ell }={\lambda }_{\ell }{\boldsymbol{\psi }}_{\ell },& \ell=0,...,T-1\end{array}$$

(22)

where ${{\mathbf{\psi }}}_{\ell }$ and ${\lambda }_{\ell }$ are the right eigenvectors and eigenvalues of ${\boldsymbol{P}}$. Finally, we defined ${\Psi }_{k}$ as the diffusion map of the $k$ th trial to a Euclidean space ${R}^{d}$ by

$${{\mathbf{\Psi }}}_{k}=\left(\begin{array}{c}{\lambda }_{1}{\psi }_{1}\left(k\right)\\ {\lambda }_{2}{\psi }_{2}\left(k\right)\\ \vdots \\ {\lambda }_{d}{\psi }_{d}\left(k\right)\end{array}\right)\in {{\mathbb{R}}}^{d}$$

(23)

where $d\le T-1$. The Euclidean distance between trials in the embedded space constitutes the diffusion distance, which is related to a distance between the respective transition probabilities on the graph.

Riemannian similarity

We quantified the distance between centroids related to different training sessions using the SPSD distance, ${d}_{{\mathcal{S}}}({\bar{{\bf{C}}}}^{l},{\bar{{\boldsymbol{C}}}}^{m}),l,m=1,\ldots,7$, described above. As the number of cells greatly affect the dynamic range of these distances, we z-scored each matrix (per animal) to produce ${\widetilde{d}}_{{\mathcal{S}}}({\bar{{\bf{C}}}}^{l},{\bar{{\boldsymbol{C}}}}^{m})$. Finally, the similarity between centroids was obtained as:

$${{\rm{R}}}_{{sim}}\left(l,m\right)=\exp \left\{-{\widetilde{d}}_{{\mathcal{S}}}\left({\bar{{\bf{C}}}}^{l},{\bar{{\bf{C}}}}^{m}\right)\right\}$$

(24)

Degree similarity

We obtained the degree of each neuron per training session $l$:

$${d}_{i}^{l}=\sum _{j\ne i}{\bar{C}}_{i,j}^{l}$$

(25)

We calculated the distances between degree vectors ${{\boldsymbol{d}}}^{l}$ related to different days as:

$${\rm{D}}(l,m)={\left\Vert {\boldsymbol{d}}^{l}-{\boldsymbol{d}}^{m}\right\Vert }_{2}^{2},\;l,m=1,\ldots,7$$

(26)

We then z-scored each matrix ${\rm{D}}$ per animal to obtain $\widetilde{{\rm{D}}}$ and extracted the degree similarity presented in Fig. 5f, g as:

$${{\rm{Degree}}}_{{sim}}\left(l,m\right)=\exp \left\{-\widetilde{{\rm{D}}}\left(l,m\right)\right\}$$

(27)

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Data availability

Source data are provided with this paper. Further information that supports the findings of this study is available from the corresponding authors upon request. Source data are provided with this paper.

Code availability

The analysis code is available as a capsule on the Code Ocean platform at the following link: https://codeocean.com/capsule/5135512/tree/v1.

References

Wolpert, D. M., Diedrichsen, J. & Flanagan, J. R. Principles of sensorimotor learning. Nat. Rev. Neurosci. 12, 739–751 (2011).
Article CAS PubMed Google Scholar
Asanuma, H. & Keller, A. Neuronal mechanisms of motor learning in mammals. Neuroreport 2, 217–224 (1991).
Article CAS PubMed MATH Google Scholar
Krakauer, J. W., Hadjiosif, A. M., Xu, J., Wong, A. L. & Haith, A. M. Motor learning. Compr. Physiol. 9, 613–663 (2019).
Article PubMed MATH Google Scholar
Sanes, J. N. Neocortical mechanisms in motor learning. Curr. Opin. Neurobiol. 13, 225–231 (2003).
Article CAS PubMed MATH Google Scholar
Lemon, R. N. Descending pathways in motor control. Annu. Rev. Neurosci. 31, 195–218 (2008).
Article CAS PubMed MATH Google Scholar
Makino, H., Hwang, E. J., Hedrick, N. G. & Komiyama, T. Circuit mechanisms of sensorimotor learning. Neuron 92, 705–721 (2016).
Article CAS PubMed PubMed Central MATH Google Scholar
Kawai, R. et al. Motor cortex is required for learning but not for executing a motor skill. Neuron 86, 800–812 (2015).
Article CAS PubMed PubMed Central MATH Google Scholar
Heindorf, M., Arber, S. & Keller, G. B. Mouse motor cortex coordinates the behavioral response to unpredicted sensory feedback. Neuron 99, 1040–1054.e5 (2018).
Article CAS PubMed PubMed Central MATH Google Scholar
Peters, A. J., Liu, H. & Komiyama, T. Learning in the rodent motor cortex. Annu. Rev. Neurosci. 40, 77–97 (2017).
Article CAS PubMed PubMed Central MATH Google Scholar
Guo, J.-Z. et al. Cortex commands the performance of skilled movement. eLife 4, e10774 (2015).
Article PubMed PubMed Central Google Scholar
Fu, M., Yu, X., Lu, J. & Zuo, Y. Repetitive motor learning induces coordinated formation of clustered dendritic spines in vivo. Nature 483, 92–95 (2012).
Article ADS CAS PubMed PubMed Central MATH Google Scholar
Cichon, J. & Gan, W.-B. Branch-specific dendritic Ca2+ spikes cause persistent synaptic plasticity. Nature 520, 180–185 (2015).
Article ADS CAS PubMed PubMed Central Google Scholar
Levy, S. et al. Cell-type-specific outcome representation in the primary motor cortex. Neuron 107, 954–971.e9 (2020).
Article CAS PubMed MATH Google Scholar
Papale, A. E. & Hooks, B. M. Circuit changes in motor cortex during motor skill learning. Neuroscience 368, 283–297 (2018).
Article CAS PubMed MATH Google Scholar
Chen, J. L., Carta, S., Soldado-Magraner, J., Schneider, B. L. & Helmchen, F. Behaviour-dependent recruitment of long-range projection neurons in somatosensory cortex. Nature 499, 336–340 (2013).
Article ADS CAS PubMed Google Scholar
Peters, A. J., Chen, S. X. & Komiyama, T. Emergence of reproducible spatiotemporal activity during motor learning. Nature 510, 263–267 (2014).
Article ADS CAS PubMed MATH Google Scholar
Otsuka, T. & Kawaguchi, Y. Pyramidal cell subtype-dependent cortical oscillatory activity regulates motor learning. Commun. Biol. 4, 1–13 (2021).
Article Google Scholar
Tsubo, Y., Isomura, Y. & Fukai, T. Neural dynamics and information representation in microcircuits of motor cortex. Front. Neural Circuits 7, 85 (2013).
Huber, D. et al. Multiple dynamic representations in the motor cortex during sensorimotor learning. Nature 484, 473–478 (2012).
Article ADS CAS PubMed PubMed Central MATH Google Scholar
Lee, C., Kim, Y. & Kaang, B.-K. The primary motor cortex: the hub of motor learning in rodents. Neuroscience 485, 163–170 (2022).
Article CAS PubMed MATH Google Scholar
Rioult-Pedotti, M.-S., Friedman, D. & Donoghue, J. P. Learning-induced LTP in neocortex. Science 290, 533–536 (2000).
Article ADS CAS PubMed Google Scholar
Schultz, W. Multiple reward signals in the brain. Nat. Rev. Neurosci. 1, 199–207 (2000).
Article CAS PubMed MATH Google Scholar
Descarries, L., Lemay, B., Doucet, G. & Berger, B. Regional and laminar density of the dopamine innervation in adult rat cerebral cortex. Neuroscience 21, 807–824 (1987).
Article CAS PubMed Google Scholar
Hosp, J. A., Pekanovic, A., Rioult-Pedotti, M. S. & Luft, A. R. Dopaminergic projections from midbrain to primary motor cortex mediate motor skill learning. J. Neurosci. 31, 2481–2487 (2011).
Article CAS PubMed PubMed Central Google Scholar
Hosp, J. A., Nolan, H. E. & Luft, A. R. Topography and collateralization of dopaminergic projections to primary motor cortex in rats. Exp. Brain Res. 233, 1365–1375 (2015).
Article CAS PubMed MATH Google Scholar
Luft, A. R. & Schwarz, S. Dopaminergic signals in primary motor cortex. Int. J. Dev. Neurosci. 27, 415–421 (2009).
Article CAS PubMed MATH Google Scholar
Molina-Luna, K. et al. Dopamine in motor cortex is necessary for skill learning and synaptic plasticity. PLoS One 4, e7082 (2009).
Article ADS PubMed PubMed Central Google Scholar
Chen, T.-W. et al. Ultrasensitive fluorescent proteins for imaging neuronal activity. Nature 499, 295–300 (2013).
Article ADS CAS PubMed PubMed Central MATH Google Scholar
Krashes, M. J. et al. Rapid, reversible activation of AgRP neurons drives feeding behavior in mice. J. Clin. Invest. 121, 1424–1428 (2011).
Article CAS PubMed PubMed Central MATH Google Scholar
Bonnabel, S. & Sepulchre, R. Riemannian metric and geometric mean for positive semidefinite matrices of fixed rank. SIAM J. Matrix Anal. Appl. 31, 1055–1070 (2010).
Article MathSciNet MATH Google Scholar
Bhatia, R. Positive definite matrices. in Positive Definite Matrices (Princeton University Press, 2009). https://doi.org/10.1515/9781400827787.
Pennec, X., Fillard, P. & Ayache, N. A riemannian framework for tensor computing. Int J. Comput Vis. 66, 41–66 (2006).
Article MATH Google Scholar
Chung, F. R. K. Spectral Graph Theory (American Mathematical Soc., 1997).
Benisty, H. et al. Rapid fluctuations in functional connectivity of cortical networks encode spontaneous behavior. Nat. Neurosci. 27, 148–158 (2024).
Article CAS PubMed MATH Google Scholar
Masamizu, Y. et al. Two distinct layer-specific dynamics of cortical ensembles during learning of a motor task. Nat. Neurosci. 17, 987–994 (2014).
Article CAS PubMed MATH Google Scholar
Komiyama, T. et al. Learning-related fine-scale specificity imaged in motor cortex circuits of behaving mice. Nature 464, 1182–1186 (2010).
Article ADS CAS PubMed MATH Google Scholar
Kabra, M., Robie, A. A., Rivera-Alba, M., Branson, S. & Branson, K. JAABA: interactive machine learning for automatic annotation of animal behavior. Nat. Methods 10, 64–67 (2013).
Article CAS PubMed MATH Google Scholar
Bohnslav, J. P. et al. DeepEthogram, a machine learning pipeline for supervised behavior classification from raw pixels. eLife 10, e63377 (2021).
Article CAS PubMed PubMed Central Google Scholar
Mathis, A. et al. DeepLabCut: markerless pose estimation of user-defined body parts with deep learning. Nat. Neurosci. 21, 1281–1289 (2018).
Article CAS PubMed MATH Google Scholar
Stringer, C. et al. Spontaneous behaviors drive multidimensional, brainwide activity. Science 364, eaav7893 (2019).
Article CAS Google Scholar
Boser, B. E., Guyon, I. M. & Vapnik, V. N. A training algorithm for optimal margin classifiers. in Proceedings of the Fifth Annual Workshop on Computational Learning Theory 144–152 (Association for Computing Machinery, 1992). https://doi.org/10.1145/130385.130401.
Sulam, J., Romano, Y. & Talmon, R. Dynamical system classification with diffusion embedding for ECG-based person identification. Signal Process. 130, 403–411 (2017).
Article ADS MATH Google Scholar
Diamond, S. & Boyd, S. CVXPY: a python-embedded modeling language for convex optimization. J. Mach. Learn Res. 17, 83 (2016).
MathSciNet PubMed PubMed Central MATH Google Scholar
Venkatesh, M., Jaja, J. & Pessoa, L. Comparing functional connectivity matrices: a geometry-aware approach applied to participant identification. NeuroImage 207, 116398 (2020).
Article PubMed MATH Google Scholar
Coifman, R. R. et al. Geometric diffusions as a tool for harmonic analysis and structure definition of data: Diffusion maps. Proc. Natl. Acad. Sci. USA 102, 7426–7431 (2005).
Article ADS CAS PubMed PubMed Central MATH Google Scholar
Coifman, R. R. & Lafon, S. Diffusion maps. Appl. Comput. Harmonic Anal. 21, 5–30 (2006).
Article MathSciNet MATH Google Scholar
Fong, A. H. C. et al. Dynamic functional connectivity during task performance and rest predicts individual differences in attention across studies. NeuroImage 188, 14–25 (2019).
Article PubMed MATH Google Scholar
Zhao, W. et al. Task fMRI paradigms may capture more behaviorally relevant information than resting-state functional connectivity. NeuroImage 270, 119946 (2023).
Article PubMed Google Scholar
Cramer, S. W. et al. Wide-field calcium imaging reveals widespread changes in cortical functional connectivity following mild traumatic brain injury in the mouse. Neurobiol. Dis. 176, 105943 (2023).
Article CAS PubMed MATH Google Scholar
Feldt, S., Bonifazi, P. & Cossart, R. Dissecting functional connectivity of neuronal microcircuits: experimental and theoretical insights. Trends Neurosci. 34, 225–236 (2011).
Article CAS PubMed MATH Google Scholar
Buzsáki, G. Neural syntax: cell assemblies, synapsembles, and readers. Neuron 68, 362–385 (2010).
Article PubMed PubMed Central Google Scholar
Lansner, A. Associative memory models: from the cell-assembly theory to biophysically detailed cortex simulations. Trends Neurosci. 32, 178–186 (2009).
Article CAS PubMed MATH Google Scholar
Sreenivasan, V. et al. Movement initiation signals in mouse whisker motor cortex. Neuron 92, 1368–1382 (2016).
Article CAS PubMed PubMed Central MATH Google Scholar
Duan, Z., Li, A., Gong, H. & Li, X. A whole-brain map of long-range inputs to GABAergic interneurons in the mouse caudal forelimb area. Neurosci. Bull. 36, 493–505 (2020).
Article CAS PubMed PubMed Central MATH Google Scholar
Hosp, J. A., Molina-Luna, K., Hertler, B., Atiemo, C. O. & Luft, A. R. Dopaminergic modulation of motor maps in rat motor cortex: an in vivo study. Neuroscience 159, 692–700 (2009).
Article CAS PubMed Google Scholar
Rioult-Pedotti, M.-S., Pekanovic, A., Atiemo, C. O., Marshall, J. & Luft, A. R. Dopamine promotes motor cortex plasticity and motor skill learning via PLC activation. PLoS One 10, e0124986 (2015).
Article PubMed PubMed Central Google Scholar
Otor, Y. et al. Dynamic compartmental computations in tuft dendrites of layer 5 neurons during motor behavior. Science 376, 267–275 (2022).
Article ADS CAS PubMed MATH Google Scholar
Bova, A. et al. Precisely timed dopamine signals establish distinct kinematic representations of skilled movements. eLife 9, e61591 (2020).
Article CAS PubMed PubMed Central MATH Google Scholar
Beeler, J. A. et al. A role for dopamine-mediated learning in the pathophysiology and treatment of Parkinson’s disease. Cell Rep. 2, 1747–1761 (2012).
Article CAS PubMed PubMed Central MATH Google Scholar
Fee, M. S. & Goldberg, J. H. A hypothesis for basal ganglia-dependent reinforcement learning in the songbird. Neuroscience 198, 152–170 (2011).
Article CAS PubMed MATH Google Scholar
Shen, W., Flajolet, M., Greengard, P. & Surmeier, D. J. Dichotomous dopaminergic control of striatal synaptic plasticity. Science 321, 848–851 (2008).
Article ADS CAS PubMed PubMed Central Google Scholar
Wood, A. N. New roles for dopamine in motor skill acquisition: lessons from primates, rodents, and songbirds. J. Neurophysiol. 125, 2361–2374 (2021).
Article CAS PubMed PubMed Central MATH Google Scholar
Fuxe, K. et al. The discovery of central monoamine neurons gave volume transmission to the wired brain. Prog. Neurobiol. 90, 82–100 (2010).
Article CAS PubMed MATH Google Scholar
Schultz, W. Dopamine reward prediction-error signalling: a two-component response. Nat. Rev. Neurosci. 17, 183–195 (2016).
Article CAS PubMed PubMed Central MATH Google Scholar
Beier, K. T. et al. Circuit architecture of VTA dopamine neurons revealed by systematic input-output mapping. Cell 162, 622–634 (2015).
Article CAS PubMed PubMed Central MATH Google Scholar
Guo, L. et al. Dynamic rewiring of neural circuits in the motor cortex in mouse models of Parkinson’s disease. Nat. Neurosci. 18, 1299–1309 (2015).
Article CAS PubMed PubMed Central MATH Google Scholar
Roome, C. J. & Kuhn, B. Chronic cranial window with access port for repeated cellular manipulations, drug application, and electrophysiology. Front. Cell. Neurosci. 8, 379 (2014).
Fienup, J. R. & Kowalczyk, A. M. Phase retrieval for a complex-valued object by using a low-resolution image. J. Opt. Soc. Am. A 7, 450–458 (1990).
Article ADS MATH Google Scholar
Joachims, T. Training linear SVMs in linear time. in Proceedings of the 12th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining 217–226 (Association for Computing Machinery, 2006). https://doi.org/10.1145/1150402.1150429.
Chang, C.-C. & Lin, C.-J. LIBSVM: a library for support vector machines. ACM Trans. Intell. Syst. Technol. 2, 1–27 (2011).
Article MATH Google Scholar
Engelhard, B. et al. Specialized coding of sensory, motor and cognitive variables in VTA dopamine neurons. Nature 570, 509–513 (2019).
Article ADS CAS PubMed PubMed Central MATH Google Scholar
Edelman, A., Arias, T. A. & Smith, S. T. The geometry of algorithms with orthogonality constraints. SIAM J. Matrix Anal. Appl. 20, 303–353 (1998).
Article MathSciNet MATH Google Scholar

Download references

Acknowledgements

We thank Y. Schiller for helpful discussion and for comments on the manuscript. We also thank B. Engelhard, Y. Gutfreund and O. Barak for helpful comments on the manuscript. Funding: This study was supported by the Israeli Science Foundation (J.S. grant #2694/23 and H.B. grant #2418/24), Prince Funds (J.S.) and Rappaport Foundation (J.S.).

Author information

These authors contributed equally: Amir Ghanayim, Hadas Benisty.

Authors and Affiliations

Department of Neuroscience, Technion Medical School, Bat-Galim, Haifa, Israel
Amir Ghanayim, Hadas Benisty, Sivan Schwartz, Sally Dabdoob & Jackie Schiller
The Taub Faculty of Computer Science, Technion, Haifa, Israel
Avigail Cohen Rimon
Viterbi Faculty of Electrical and Computer Engineering, Technion, Haifa, Israel
Shira Lifshitz & Ronen Talmon

Authors

Amir Ghanayim
View author publications
You can also search for this author inPubMed Google Scholar
Hadas Benisty
View author publications
You can also search for this author inPubMed Google Scholar
Avigail Cohen Rimon
View author publications
You can also search for this author inPubMed Google Scholar
Sivan Schwartz
View author publications
You can also search for this author inPubMed Google Scholar
Sally Dabdoob
View author publications
You can also search for this author inPubMed Google Scholar
Shira Lifshitz
View author publications
You can also search for this author inPubMed Google Scholar
Ronen Talmon
View author publications
You can also search for this author inPubMed Google Scholar
Jackie Schiller
View author publications
You can also search for this author inPubMed Google Scholar

Contributions

Conceptualization, J.S.; Methodology, J.S., A.G., S.D., H.B., and R.T.; Software, H.B. and A.CR.; Formal Analysis, A.G., J.S., H.B., A.CR., R.T., S.S., S.D., and S.L.; Data Curation, H.B. and J.S.; Writing, J.S., A.G., H.B., and R.T.; Visualization, H.B., A.G., and A.CR.; Funding Acquisition, J.S. and H.B.

Corresponding authors

Correspondence to Hadas Benisty or Jackie Schiller.

Ethics declarations

Competing interests

The authors declare no competing interests.

Peer review

Peer review information

Nature Communications thanks Carl Petersen and the other, anonymous, reviewers for their contribution to the peer review of this work. A peer review file is available.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary information

Supplementary Information

Peer review file

Description of Additional Supplementary Files

Supplementary Movie 1

Supplementary Movie 2

Supplementary Movie 3

Supplementary Movie 4

Supplementary Movie 5

Supplementary Movie 6

Reporting Summary

Source data

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.

Reprints and permissions

About this article

Cite this article

Ghanayim, A., Benisty, H., Cohen Rimon, A. et al. VTA projections to M1 are essential for reorganization of layer 2-3 network dynamics underlying motor learning. Nat Commun 16, 200 (2025). https://doi.org/10.1038/s41467-024-55317-4

Download citation

Received: 29 November 2023
Accepted: 06 December 2024
Published: 02 January 2025
DOI: https://doi.org/10.1038/s41467-024-55317-4