Dynamics of directional motor tuning in the primate premotor and primary motor cortices during sensorimotor learning

Ebina, Teppei; Sasagawa, Akitaka; Hong, Dokyeong; Setsuie, Rieko; Obara, Keitaro; Masamizu, Yoshito; Kondo, Masashi; Terada, Shin-Ichiro; Ozawa, Katsuya; Uemura, Masato; Takaji, Masafumi; Watakabe, Akiya; Kobayashi, Kenta; Ohki, Kenichi; Yamamori, Tetsuo; Murayama, Masanori; Matsuzaki, Masanori

doi:10.1038/s41467-024-51425-3

Download PDF

Article
Open access
Published: 20 August 2024

Dynamics of directional motor tuning in the primate premotor and primary motor cortices during sensorimotor learning

Nature Communications volume 15, Article number: 7127 (2024) Cite this article

1650 Accesses
1 Citations
13 Altmetric
Metrics details

Subjects

Abstract

Sensorimotor learning requires reorganization of neuronal activity in the premotor cortex (PM) and primary motor cortex (M1). To reveal PM- and M1-specific reorganization in a primate, we conducted calcium imaging in common marmosets while they learned a two-target reaching (pull/push) task after mastering a one-target reaching (pull) task. Throughout learning of the two-target reaching task, the dorsorostral PM (PMdr) showed peak activity earlier than the dorsocaudal PM (PMdc) and M1. During learning, the reaction time in pull trials increased and correlated strongly with the peak timing of PMdr activity. PMdr showed decreasing representation of newly introduced (push) movement, whereas PMdc and M1 maintained high representation of pull and push movements. Many task-related neurons in PMdc and M1 exhibited a strong preference to either movement direction. PMdc neurons dynamically switched their preferred direction depending on their performance in push trials in the early learning stage, whereas M1 neurons stably retained their preferred direction and high similarity of preferred direction between neighbors. These results suggest that in primate sensorimotor learning, dynamic directional motor tuning in PMdc converts the sensorimotor association formed in PMdr to the stable and specific motor representation of M1.

Context-dependent limb movement encoding in neuronal populations of motor cortex

Article Open access 23 October 2019

Single-trial cross-area neural population dynamics during long-term skill learning

Article Open access 13 August 2020

Cortical preparatory activity indexes learned motor memories

Article 26 January 2022

Introduction

In tennis and baseball, it is necessary to hit the ball back in various ways in response to the sensory stimulus from the ball, which is coming towards you from one of a great variety of possible trajectories. In addition, when performing actions in general life, the relationship between the stimulus and response is often arbitrary, such as the operating of a screen cursor with a joystick or keyboard. To perform such actions, we learn new sensorimotor maps every day, adding them to our previous maps and improving the new movements. In particular, for both humans and non-human primates, the execution of a variety of forelimb movements is very important, regardless of the requirement for hand dexterity. The forelimb motor cortical area is included in the dorsal part of the premotor cortex (PMd) and primary motor cortex (M1)^1,2. In primates, the rostral part of the PMd (PMdr) connects densely with the prefrontal cortex (PFC) and is associated with cognitive processes, while the caudal part of PMd (PMdc) strongly connects with M1 and possesses corticospinal neurons that imply that it is more strongly related to motor processes than is PMdr^3,4,5,6,7. In forelimb movement, PMdr neurons show an earlier response to action-related sensory signals than do PMdc and M1 neurons, and PMdc activity follows PMdr activity after the action to be taken is determined^8,9. PFC and PMdr are thought to be important in the association of sensory cues with movement in sensorimotor learning^10,11,12. PM and M1 neurons show sensorimotor learning-related and motor skill learning-related changes in activity^{13,14,15,16,17,18}. However, the activities of neurons in PMdr and PMdc, or in PMdc and M1, are frequently collected as the same population when recording motor cortical activity. Thus, it is unclear how the PMdr, PMdc, and M1 motor representations differ during sensorimotor learning. PMdc and M1 have a hallmark property in that individual neurons possess a preferred direction for reaching movements and neighboring neurons show a similar preferred direction^19,20,21.

Studies using electrical recording and calcium imaging revealed that the preferred direction of forelimb reaching in the superficial and deep layer neurons in PMdc and M1 is relatively stable across days after motor learning^22,23,24. However, it is unclear how PMdc and M1 neurons tune directional preference when a new sensorimotor map is introduced into a previously learned map, and while the new movement is improved over dozens of days.

The neuroanatomical structures (including PMdr and PMdc) of the marmoset brain share a high degree of similarity with other primates^25,26,27, but the marmoset cerebral cortex is much smaller than that of the macaque, and has a flat and smooth surface that is suitable for calcium imaging^28,29,30,31. In the current study, we conducted wide-field one-photon calcium imaging of PMdr, PMdc, and M1, and two-photon calcium imaging of layer 2/3 (L2/3) neurons in PMdc and M1, while the marmoset learned a simple two-target reaching (pull/push) task over more than 2 weeks. Six marmosets were used in total (marmosets 1–3 for one-photon imaging and marmosets 4–6 for two-photon imaging), with these already being experts in a one-target reaching (pull) task²³. We revealed how the pull-related and push-related activity changed, and how the preferred direction (PD) (between pull and push) and PD spatial distribution changed over the learning period. We also conducted wide field-of-view simultaneous two-photon calcium imaging of L2/3 neurons in PMdc and M1 after more than 100 training sessions in one marmoset (marmoset 7) to show the PD distribution across PMdc and M1.

Results

Learning of the two-target reaching task

We trained six head-fixed marmosets (marmosets 1–6) under similar conditions. The animals performed a visual-cue-triggered one-target reaching (OTR) task for 31–85 days, and were then trained to perform a two-target reaching (TTR) task over 15–43 sessions²³ (Fig. 1a, b and Supplementary Figs. 1,2a). In the TTR training sessions, we conducted one-photon calcium imaging of marmosets 1–3 and two-photon calcium imaging of marmosets 4–6 (Supplementary Fig. 1 and Supplementary Table 1). In the OTR task, the animal needed to pull a pole that could move two-dimensionally in the horizontal plane with their left forelimb to move a cursor from a fixation area to a target area that appeared below the fixation area on the monitor²³. The two-dimensional pole movement was directly linked to the two-dimensional cursor movement. After the cursor was moved to the target area and held within it, a sucrose water reward was delivered. In the TTR task, the animal needed to pull or push the pole to move the cursor to a target area that appeared either below or above the fixation area, respectively (Fig. 1b). As the training sessions progressed, the movements in both OTR and TTR tasks were made more difficult by shortening the width of the target area, and sometimes lengthening the target distance (distance from the fixation area to the target area) and the cursor holding time within the target area (Fig. 1c and Supplementary 2b–d). Over the training sessions, the cursor trajectory in the rewarded trials became straight to fulfill the final threshold for the reward acquisition (Fig. 1c). We defined the rewarded trials that did not show any initial opposite movement as successful trials (see Methods for details). During the TTR training sessions, the rate of successful trials to total pull trials (SR_pull) remained high (~0.8), while the rate of successful push trials to total push trials (SR_push) increased to ~0.8, a similar level to SR_pull (Fig. 1d, e). In all of the following analyses of the neuronal activity, we use only data from successful trials in those sessions with at least 10 successful pull and push trials. From these sessions, and for each animal, two or three sessions were selected as early sessions and three sessions as late sessions (Supplementary Table 1). In five out of six animals, the variability of the trial-to-trial cursor trajectory compared with the averaged trajectory in the successful push trials (Var_push) was smaller in the late sessions than in the early sessions (Fig. 1f, g). The reaction time (the duration from the target onset to the time at which the cursor exited the fixation area) in the successful pull trials (RT_pull) increased from early to late sessions (Fig. 1h, i). By contrast, the animals did not show any common change in the reaction time in the successful push trials (RT_push).

**Fig. 1: Behavioral changes during learning of TTR task.**

This increase in RT_pull might be related to the decrease in the target size for pull trials (TS_pull) (Supplementary Fig. 2b). However, although TS_pull in the late sessions of the OTR task was smaller than that in the early sessions of the TTR task, RT_pull did not differ between these sessions (Supplementary Fig. 2e, f). To determine which task-related variables explained the change in RT_pull, we constructed three multivariable regression models. In the full model, the explanatory variables consisted of five task performance variables: SR_pull, SR_push, RT_push, the trial-to-trial variabilities of the cursor for pull trials (Var_pull), and Var_push; and six task difficulty-related variables: TS_pull, the target size for push (TS_push) trials, the target distance for pull (TD_pull) and push (TD_push) trials, and the holding time for pull (HT_pull) and push (HT_push) trials. The variables (including RT_pull) except for the success rate were normalized for each animal. In the task performance model, the explanatory variables were five task performance-related variables. In the task difficulty model, the explanatory variables were six task difficulty-related variables. We used 5-fold cross validation to estimate the accuracy of each model to predict RT_pull. The full model showed the highest prediction accuracy and the task difficulty model showed much lower prediction accuracy (Fig. 1j). In the full model, RT_push, SR_pull, SR_push, TD_push, and HT_pull had statistically significant weights (Fig. 1k). These results suggest that the change in RT_pull was more related to the changes in task performance than to the task difficulty.

One-photon calcium imaging of wide cortical areas including PMdr, PMdc, and M1 during learning of the TTR task

For calcium imaging, adeno-associated viruses (AAVs) carrying tetracycline-inducible tandem GCaMP6s^32,33,34 were injected into multiple sites on an intervening day during the OTR training sessions (Fig. 2a and Supplementary Figs. 1, 3a, b)²³. After the imaging experiments were finished, muscimol injection into the right PMdr and M1 inhibited the pull and push movements (Supplementary Fig. 3c, d), indicating that these areas were necessary for the TTR task performance. During calcium imaging, we also recorded body movements (forelimb, hindlimb, and orofacial movements) with two cameras and extracted the movements of several body parts^35,36 (Supplementary Fig. 4a–d).

**Fig. 2: Wide-field one-photon calcium imaging during TTR training sessions.**

One-photon imaging clearly showed the dynamics of the neuronal activity (relative fluorescence change, ΔF/F) in the TTR task over a wide area including PMdr, PMdc, and M1 (Fig. 2a–c). In the following analyses, the image size was down-sampled to 64 × 64 pixels (0.23 mm/pixel) to reduce both the noise affecting each pixel and the calculation time. The density of the AAV injection and the fluorescence intensity before the cue presentation did not differ largely between these areas or animals (Supplementary Fig. 5a, b). ΔF/F was not contaminated by a change in the intrinsic fluorescence reflecting hemodynamic activity^37,38 (Supplementary Fig. 5c–f). Thus, on the basis of ΔF/F, we compared the activity across PMdr, PMdc, and M1, and across sessions, in the following analyses.

The activity flow from PMdr → PMdc → M1 was stable during TTR training sessions

Although a variety of activity patterns were detected over the frontoparietal cortex, we focused on the three motor cortical areas (PMdr, PMdc, and M1) because our aim was to reveal the spatiotemporal dynamics of PM and M1 activity during the TTR task. The activities in these three areas increased after cue onset and reached a peak around 2 s after cue onset (Fig. 2d). A transient response to the target presentation was not apparent, and most activity appeared to be movement related (Fig. 2e). First, we estimated the peak timing of the trial-averaged activity (T_peak) for pull and push trials in the early and late sessions in marmosets 1–3. PMdr showed the earliest peak timing regardless of the twelve conditions (early and late sessions in successful pull and push trials in three animals) (Fig. 3a, b and Supplementary Fig. 6a). T_peak was earlier in PMdc than in M1 (Fig. 3b). The change in T_peak tended to be larger on the putative border between PMdr and PMdc than on the putative border between PMdc and M1 (Supplementary Fig. 6b, c). When we calculated the correlation between T_peak in each area and the reaction time across the training sessions, the T_peak of PMdr and the reaction time correlated in all six conditions (successful pull and push trials in three animals; Fig. 3c, d), although T_peak was slower than the reaction time. These results suggest that PMdr activity before the peak timing might be strongly related to the reaction time of the pull and push movements throughout the training sessions.

**Fig. 3: The order of activity in PMdr, PMdc, and M1 during TTR training sessions.**

To examine whether there were functional hierarchical relationships in task-related activity between PMdr, PMdc, and M1, we calculated transfer entropy, an algorithm for quantifying causal strength³⁹, in the time window from 2 s before to 1 s after the onset of the movement in the successful trials. PMdr was the critical area for predicting the activity of PMdc and M1 in both pull and push trials (Fig. 3e, f). PMdc also contributed to predicting the activity of M1 in both types of trials (Fig. 3e, f). Although M1 also contributed to predicting the activity of PMdc in both types of trials in early sessions, the transfer entropy was higher in the direction from PMdc to M1 than from M1 to PMdc in all four conditions (Fig. 3e, f). When the order of T_peak is also considered, this analysis suggests the stable flow of information from PMdr → PMdc → M1 during the period from the sensory input to the motor initiation in both early and late sessions.

PMdc and M1 retained stronger motor representation than PMdr

Next, we examined how information on the forelimb movement was possessed by each area throughout learning. In push trials, the cursor trajectory returned to the fixation area immediately after the initial push movement, which indicates that the push movement was frequently followed by pull movement (Supplementary Fig. 4c, d). To exclude the activity that might be related to the opposite-directed movement, we focused on the activity related to initial movements in pull and push directions. To assess the motor representation at this time, we predicted the cursor movement at each time point from 1 s before to 0.13 s after the movement onset that included only pull or push movement. For this, we used a linear decoding model with the neuronal activity in PMdr, PMdc, and M1 as explanatory variables (Fig. 4a; see Methods for details). For each pixel of the imaging field, we calculated the cross-validated coefficient of determination (cvR²) to represent the prediction accuracy^40,41 (Fig. 4a). The area-averaged cvR² was much higher in PMdc and M1 than in PMdr in both early and late sessions and both types of trials, although the cvR² values varied across the three animals (Fig. 4b, c and Supplementary Fig. 7a). The cvR² was higher in M1 than in PMdc (Fig. 4c and Supplementary Fig. 7a). As in the case of T_peak, the change in cvR² tended to be larger on the putative border between PMdr and PMdc than on the putative border between PMdc and M1 (Supplementary Fig. 7b, c). Even when the low-dimensional population activity (the first to tenth principal components extracted from all pixels) in each area was used to calculate cvR² (population cvR²), it was highest in M1 (Supplementary Fig. 7d, e). However, the increase in the population cvR² over the area-averaged cvR² was highest in PMdr, and higher in PMdc than in M1 (Supplementary Fig. 7f). This suggests that the motor representation in the local regions might be more heterogenous in PMdr than in PMdc and M1, so that the combination of multiple local regions of PMdr could contribute more to the populational representation than could a combination of PMdc and M1 local regions, although the motor representation in terms of individual local regions was weaker in PMdr than in PMdc and M1.

**Fig. 4: Changes in the prediction accuracy of initial movements in PMdr, PMdc, and M1 during TTR training sessions.**

PMdr decreased the area-averaged cvR² for push trials across training sessions, although this was not statistically significant in marmoset 1 (Fig. 3d). PMdc and M1 did not show a consistent trend of change in the area-averaged cvR² for pull or push movements across training sessions (Fig. 3d). Similar results were obtained for the population cvR² (Supplementary Fig. 7g). These results suggest that PMdc and M1 retained the representation of both pull and push forelimb movements, while the representation of the newly introduced push movement in PMdr was weak and tended to decrease throughout learning. To determine which task-related variables explained the decrease in the area-averaged cvR² for the push movement trajectory (cvR²_push) in PMdr in marmosets 2 and 3, we constructed three regression models similar to those used to explain RT_pull. Among these models, only the task performance model showed significant prediction accuracy (Supplementary Fig. 7h). In this model, the bias term, SR_push, and RT_pull, showed significant weights, with the weight of RT_pull being negative (Fig. 4e). These results suggest that the decrease in the push representation of PMdr during training sessions was partly related to the increase in RT_pull.

The activity of the imaged PM and M1 might reflect task-related orofacial movements rather than the contralateral forelimb movement, although orofacial movement should be more represented in the ventral PM and ventral M1⁴². When the pupil diameter, Z-axial eye movement, and lick rate were predicted from the neuronal activity in each area, the cvR² was smaller than the cvR² of the initial pole movement in most cases, and the prediction accuracy of any of the three orofacial movements did not consistently increase or decrease (Supplementary Fig. 8a, f). The Y-axial pole movement showed constant strong correlations with the Y-axial left hand and Y-axial left elbow movements, but did not strongly correlate with the Z-axial eye movement, pupil diameter, lick rate, right hand, right elbow, or bilateral knees throughout the training sessions (Supplementary Fig. 8g). Thus, the orofacial, right forelimb, and hindlimb movements did not have a large effect on the measured activity in PMdr, PMdc, and M1 although these movements should be involved in the neuronal activity to some extent.

The change in preferred direction was larger in PMdr and PMdc than in M1

We next examined the change in PD in PMdr, PMdc, and M1 pixels. Although we examined only two directions, we defined the preferred direction index (PDI) as (cvR² in pull trials – cvR² in push trials)/(cvR² in pull trials + cvR² in push trials). A PDI of 1 or –1 indicates that only a pull or push movement was represented, respectively. A PDI of 0 indicates that pull and push movements were represented equally. There was no apparent millimeter-scale segregation between the high PDI sub-areas and low PDI sub-areas that were common across the three animals in either the early or late sessions (Fig. 5a). PMdr showed a relatively broad distribution of PDI in both early and late sessions, whereas more than 80% of pixels in PMdc and M1 showed a PDI within the range of –0.2–0.2 in both early and late sessions and a cvR² of >0.3 for both pull and push movements (Fig. 5b). This indicates that many sub-areas in PMdc and M1 strongly showed both pull- and push-related activity in both early and late sessions. The degree of change in PDI in each pixel from the early to late sessions differed between these areas. Compared with M1, PMdr and PMdc showed larger changes in PDI from early to late sessions, while the sign of the area-averaged PDI was not consistent in early or late sessions across the three marmosets (Fig. 5a, c). The change in PDI in each pixel (ΔPDI_EL) in M1 from early to late sessions was distributed around zero in all three animals, whereas ΔPDI_EL in PMdr and PMdc deviated from zero. Although the direction of the deviation was the same between PMdr and PMdc within the same animal, it was not consistent across the animals (Fig. 5d). These results suggest that more sub-areas altered their PD in PMdr and PMdc than in M1. Thus, the manner of the directional motor tuning during the training sessions differed between these three areas.

**Fig. 5: Changes in the preferred direction of PMdr, PMdc, and M1.**

Wide field-of-view two-photon calcium imaging performed after learning revealed that M1 neighboring neurons showed a more similar preferred direction than PMdc neighboring neurons

The sub-areas that represented both pull and push movements might reflect the fact that individual neurons showed similar activity between pull and push trials, or that neurons with different PDs were intermingled within the sub-areas. In the following analyses and experiments, we focused on the PMdc and M1, which showed higher area-averaged cvR² than PMdr. To examine the PD distribution of individual neurons over PMdc and M1 after learning, we applied wide field-of-view two-photon microscopy with a 3 × 3 mm field of view (FOV) (which was first developed for mouse brain imaging⁴³) to a marmoset that was trained to perform the TTR task (marmoset 7; Supplementary Fig. 9a, b). We expanded the space between the large objective lens and the sample stage, and to make the optical axis perpendicular to the cranial window, we set the head-fixed marmoset sitting in the chair so that it was substantially tilted (Supplementary Fig. 9c).

The FOV was so large that parts of PMdc and M1 could be simultaneously imaged (Fig. 6a). We used a constrained non-negative matrix factorization (CNMF)⁴⁴ to extract active ROIs with denoised ΔF/F as active neurons. We detected 15,961 active neurons from six imaging sessions (in 4 days) in marmoset 7 (10,510 PMdc and 5451 M1 neurons; Fig. 6b–d). As for the pixels of the one-photon imaging data, the cvR² for pull and push movements and the PDI were calculated for each active neuron (Fig. 6e). When neurons had a cvR² of >0.02, they were defined as pull-related or push-related neurons (Fig. 6f and Supplementary Fig. 9d; see Methods for details). Their proportions ranged from 0.15–0.25, and were slightly higher in M1 than in PMdc (Fig. 6g). In both pull and push trials, the cvR² of the motor-related neurons was higher in M1 than in PMdc (Fig. 6h). The peak timing of the activity of movement-related neurons was earlier in PMdc than in M1 (Fig. 7a–c). This result is consistent with the results of the population activity obtained with one-photon imaging (Fig. 3b).

**Fig. 6: Wide field-of-view two-photon calcium imaging of PMdc and M1 after learning.**

**Fig. 7: Time course of activity and PDI in PMdc and M1 neurons after learning.**

Next, we examined the PDI of each neuron. Within the 3 × 3 mm FOV, there was no apparent segregation in the millimeter-scale areas of pull and push movements, but there were small scattered clusters for each direction (Fig. 7d). In both PMdc and M1, many neurons showed a high PD; ~40% of neurons in PMdc and M1 showed PDI of <–0.8 or >0.8 (Fig. 7e). The proportion of neurons with a PDI of –0.2 to 0.2 was only ~10% in both areas (Fig. 7e). As PDI increased, cvR² for the pull movement increased and cvR² for the push movement decreased in both areas. Thus, individual neurons with strong movement-related activity in PMdc and M1 showed clear directional preference. The distributions of PDI and cvR² of individual neurons differed substantially from the distributions of the PDI of individual pixels that were imaged by one-photon imaging (Fig. 5b). These results suggest that many neurons showed high directional preference and that neurons with different PDs are intermingled in PMdc and M1.

When we calculated the absolute value of the difference in PDI between a pair of neurons (|ΔPDI_pair|), the |ΔPDI_pair| at a cellular distance of <300 μm was significantly smaller than the |ΔPDI_pair| when the cellular distance was shuffled in both M1 and PMdc (Fig. 7f). |ΔPDI_pair| at a cellular distance of <300 μm was also significantly smaller in M1 than in PMdc (Fig. 7g). Thus, in M1, the neurons with similar PD tended to cluster more strongly in a local region of a few hundred micrometers than they did in PMdc.

Chronic two-photon calcium imaging during learning revealed that PMdc and M1 neurons show stable and flexible changes in their motor representation

From the results so far obtained, we assumed that any FOV of ~0.6 × 0.6 mm in PMdc and M1 included both pull-related and push-related neurons to some extent, and that some of these neurons might change their activity during the TTR training sessions. For the next part of the study, we conducted two-photon imaging of PMdc and M1 neurons with a microscope whose scanning head could be tilted and whose FOV was ~0.6 × 0.6 mm while marmosets 4–6 sat in the slightly tilted chair learning the TTR task²³. We analyzed three FOVs in PMdc L2/3 and three FOVs in M1 L2/3 for all three animals (Supplementary Figs. 3b,4d). We pooled the data for each area and were able to pursue 398 PMdc neurons and 336 M1 neurons in at least one session of each of the early and late sessions (Fig. 8a, b). Among these pursued neurons, the neurons that showed cvR² of >0.02 (Supplementary Fig. 10a, b) for pull or push movement were defined as movement-related neurons (pull-related neurons or push-related neurons, respectively). When we compared cvR² between early and late sessions for each pursued neuron (Fig. 8c), there were neurons in both PMdc and M1 that showed increased cvR², and those that showed decreased cvR². We defined increase or decrease neurons as those neurons that showed an increased or decreased cvR² of more than 0.02 from the early to late sessions, respectively. The proportion of decrease neurons was larger than the proportion of increase neurons in both areas (Fig. 8d). When cvR² was averaged over all pursued neurons except for those that were unrelated to either movement in both early and late sessions (black in Fig. 8c), the cvR² for push movement in all three FOVs in PMdc decreased from the early to late sessions, although the decrease was not significantly different in one FOV (Fig. 8e). By contrast, the change directions in the cvR² for pull movement in PMdc, or in the cvR² for pull and push movements in M1, were not consistent across FOVs (Fig. 8e, f). When the population activity of these neurons was used to predict the initial pole movement along the Y axis, the direction of the change in the prediction accuracy for either movement type from the early to late sessions was not consistent across the three animals (Fig. 8g, h). This is consistent with the result obtained from one-photon imaging (Fig. 4c), and suggests that while the populations of PMdc and M1 neurons did not consistently change their prediction accuracy for either pull or push movement during the training sessions, a subset of PMdc neurons did show decreased prediction accuracy for the push movement.

**Fig. 8: Changes in the prediction accuracy of the initial pole movement in individual PMdc and M1 neurons during TTR training sessions.**

Dynamic changes in the PD of PMdc might be related to the averaged PDI and success rate for push trials in the early sessions

Finally, we examined whether the PD of individual neurons changed more in PMdc than in M1, as predicted from the results of the one-photon imaging. The distributions of PDI in the early and late sessions were similar; the proportion of neurons with a PDI of <–0.8 or >0.8 was high in both PMd and M1, although the PDI distribution was biased to either positive or negative values within each FOV (Fig. 9a, b and Supplementary Fig. 11a, b). By contrast, the distribution of the difference in PDI between the early and late sessions (ΔPDI_EL) was broader in PMdc neurons than in M1 neurons (Fig. 9a, b), and |ΔPDI_EL| was larger in PMdc neurons than in M1 neurons (Fig. 9c). These results suggest that even at the single-neuron level, the directional preference changed more dynamically in PMdc than in M1 from the early to late sessions.

**Fig. 9: Changes in the PD of individual PMdc and M1 neurons during TTR training sessions.**

We then used one-photon and two-photon imaging data from six marmosets to examine whether the averaged ΔPDI_EL within the imaging field was related to the averaged PDI in the early or late sessions, even though the averaged PDI within a two-photon imaging field was not necessarily representative of the averaged PDI of the area to which it belonged. We found that the averaged PDI in the early sessions was relatively broad (–0.6 to 0.5) and negatively correlated with the averaged ΔPDI_EL, whereas the averaged PDI in the late sessions ranged from –0.3 to 0.3 and did not correlate with the averaged ΔPDI_EL (Fig. 9d). This suggests that irrespective of whether the PDI was positive or negative, the absolute amplitude of PDI decreased towards near zero from the early to late sessions, and the movements in both directions came to be represented to a similar extent. In the early sessions, the averaged PDI negatively correlated with SR_push, but not SR_pull (Fig. 9e and Supplementary Fig. 11c). The averaged ΔPDI_EL tended to correlate with SR_push (Supplementary Fig. 11d). Thus, the inconsistency of the sign of ΔPDI_EL across the animals might depend on the strength of PDI in the early sessions, which was related to SR_push in those sessions.

The |ΔPDI_pair| at any cellular distance of <300 μm in PMdc differed between the early and late sessions; |ΔPDI_pair| was ~0.6 in the early sessions and ~0.8 in the late sessions (Fig. 9f). This suggests that the heterogeneity of PD within the PMdc local circuit increased over the training sessions. By contrast, in M1, |ΔPDI_pair| did not differ between the early and late sessions (Fig. 9g). Consistent with the result of the wide field-of-view two-photon microscopy, |ΔPDI_pair| at a cellular distance of <100 μm in the late sessions was smaller in M1 than in PMdc (PMdc, 0.82 ± 0.03, n = 325 pairs, M1, 0.67 ± 0.02, n = 739 pairs, P < 0.01, Wilcoxon rank-sum test, two-sided). These results suggest that PMdc neurons changed PD more dynamically than M1 neurons, even at local-network levels. When we constructed three (full, task performance, and task difficulty) models to explain the |ΔPDI_pair| of pursued neuron pairs whose distances were <300 μm (|ΔPDI_pair|_<300 μm), the task performance model showed significant prediction accuracy, although this accuracy was low (~0.3) (Supplementary Fig. 11e). In this model, only Var_push had a significant weight, and this was negative (Fig. 9h). Thus, the increase in the heterogeneity of preferred directions in the local PMdc region from early to late sessions might be related to stabilization of the trajectory of the push movement.

Discussion

In the current study, the order of the peak timing of the activity and the transfer entropy analysis both suggest that the activity flow of PMdr → PMdc → M1 was stable throughout the training sessions. We also demonstrated that the order from highest to lowest motor representation was M1, PMdc, and PMdr throughout the sessions (Supplementary Fig. 12). These results are consistent with many previous sensorimotor control studies using electrophysiological recording in macaque^6,8,45,46.

Furthermore, we found some notable learning-related changes that have not been reported. First, we found that introducing a new sensorimotor map (top target-push movement) extended the reaction time for the already learned sensorimotor map (bottom target-pull movement). This RT_pull extension was related to the increase in SR_push. Second, the representation of the new (push) movement in PMdr weakened throughout the training sessions in two marmosets. This decrease in the cvR²_push of PMdr was related to the increase in RT_pull. Third, although the directional preferences for the new (push) and previously learned (pull) movements were similarly distributed in PMdc and M1, the change in directional preference during the training sessions was greater in PMdc. The direction and magnitude of this change depended on the preferred direction in the early stage of learning. The clustering of neurons with the similar preferred direction was stronger in M1 than in PMdc in the late stage of learning, and the increase in the heterogeneity of the preferred direction in PMdc during the training sessions might be related to stabilization of the push movement.

Stable activity flow from PMdr → PMdc → M1 during TTR training sessions

The current task, in which the animal did not need to retain the sensory memory before decision-making and movement onset, differs from the tasks in many previous studies, in which an instructed delay period was set^8,47. Nevertheless, activity flow from PMdr → PMdc → M1 was stably observed throughout the training sessions. In the macaque, PMdr is strongly associated with cognitive processes based on visual information, whereas PMdc is closely related to motor processes, and PMdr activity precedes PMdc activity^8,45,46. The caudal part of the posterior parietal cortex is associated with visual processing of movement and projects more strongly to PMdr than to PMdc^48,49. The activity flow from PMdr → PMdc → M1 would be consistent, regardless of the stage of learning and whether or not memory is retained during the delay period.

The increase in the reaction time for pull trials and the decrease in push representation in PMdr during TTR training sessions

In the OTR task, the target consistently appeared below the fixation area and there was only one movement to be performed, and therefore the pull movement would be prepared immediately. By contrast, in the TTR task, the animal needed to decide which direction to move in after the cue presentation. We speculate that the introduction of push movement to another target extended RT_pull, and that this extension was in parallel with the increase in SR_push. The increase in reaction time with the increase in the number of sensory-response alternatives is formulated as Hick’s law^50,51. Although we did not examine more than two alternatives, our result suggests that Hick’s law also applies to the learning process of the newly introduced forelimb movement.

The reaction time for multiple sensory-response alternatives is thought to be related to the response selection stage⁵¹. Considering that the T_peak in PMdr correlated most strongly with the reaction time, the decrease in cvR²_push in PMdr suggests that PMdr might serve to associate pushing with the top target in the early stages of learning, but that as learning progressed and SR_push increased, the attention for this association and push-related PMdr activity might decrease. Given the result of the regression analysis, we infer that the decrease in cvR²_push of PMdr occurred in conjunction with the increase in RT_pull. Thus, PMdr activity would also reflect the changes in cognitive processes during the sensorimotor learning.

Dynamic direction-associated motor tuning in PMdc during TTR training sessions

In PMdc, ΔPDI_EL negatively correlated with PDI in the early sessions, while PDI in the early sessions negatively correlated with SR_push in the early sessions. Therefore, the direction and magnitude of positive and negative changes in PMdc varied among animals, but animals that were good at the push movement in the early stage of learning tended to show a higher push preference in PMdc, whereas animals that were poor at it tended to show a lower push preference in PMdc. As learning progressed, the average preference became balanced (that is, a PDI of around zero) for animals that were either good or poor in the early-stage learning. We, therefore, conclude that the inconsistency in the change in the sign of PDI might be caused by the fact that the speed of learning the new sensorimotor association differed among the animals, but nevertheless, all animals learned this movement well later in learning, the bias in the preferred direction of the neuronal population decreased, and the proportion of neurons that strongly represent only pull and push, respectively, became similar.

By contrast, the PD of M1 neurons was more stable than that of PMdc neurons. Although we examined activity for only pull and push directions, our results do not contradict the previous finding that macaque M1 neurons did not substantially change their preferred direction within a session of a visuomotor association task with eight-way reaching⁵². Many movement-related neurons possessed a strong preference for the reaching direction, and the similarity of PD between pairs of neurons was statistically significant at a cellular distance of <300 μm in M1 and PMdc. In macaque, neighboring motor cortex neurons (at a cellular distance of <400 μm) show strong synaptic linkages^53,54,55. Thus, the clustering of neurons that are related to the same movement direction is a fundamental self-organizing property in non-human primates^20,21,54,56. However, the extent of such clustering was stronger in M1 than in PMdc in the late stage of learning and the increase in the heterogeneity of preferred directions in the PMdc local region from early to late sessions was weakly related to the stabilization of the trajectory of the push movement. Thus, the clustering of PMdc neurons with similar directional preferences might have some different functions to that of clustering of M1 neurons. However, we have not imaged the same area at different depths, nor the activity related to other movement directions. Three-dimensional mapping of reaching in eight directions is necessary to clarify whether the local clusters that we detected are parts of functional columns, and how much local size is represented in all directions^19,20.

These results suggest a principle of spatiotemporal patterns of the activity during sensorimotor learning as follows (Supplementary Fig. 12). PMdr forms a new sensory-response map that is incorporated into the previously learned map. This increases the reaction time for the previously learned response. The representation of the newly learned movement in PMdr then decreases as learning progresses. In PMdc, the directional preference is biased towards the newly learned movement according to the accuracy of the newly introduced movement in the early stage of learning, but this bias weakens as the animal becomes able to execute the newly learned movement as well as the previously learned movement. These flexible reorganizations in PMdr and PMdc process different sensorimotor associations, converting the sensory signal to the appropriate motor signal, and transferring it to clusters of M1 neurons that have specifically differentiated to accurately execute individual movements.

Possible reasons why motor representation in PMdc and M1 did not change consistently during TTR training sessions

A subset of L2/3 neurons in PMdc and M1 increased their motor representation while other neurons decreased theirs, with the prediction accuracy reflected by the ensemble activity showing no substantial change. These results are consistent with those of our previous study in mice. In that study, there was a subset of neurons that increased their motor representation and other neurons that decreased theirs in the forelimb M1 L2/3 during learning of the forelimb lever-pull task, but the prediction accuracy reflected by the population activity did not substantially change⁵⁷. Thus, in the superficial layer, a subset of neurons that flexibly change the amount of motor information they carry may contribute to motor learning in both primates and rodents.

However, the PM and M1 of macaques and humans frequently show expansion of the learned-movement-related area and an increase in the learned-movement-related activity^2,13,58,59. Although the reaching movement to the target should be novel for the marmosets, grasping the pole by the left hand and applying force (pull or push) to the pole should be familiar to them because their home cages had fences that could be grasped, and therefore a subset of neurons in PMdr, PMdc, and M1 might have already formed a high preference for the push direction. Since we only estimated the prediction accuracy of the initial phase of the movement, we might not have detected neuronal activity reflecting the fine improvement in the push movement that was newly introduced in the TTR task. We previously reported context-dependent reorganization with fine movement proficiency in mouse M1 L2/3⁴⁰. Another possibility was that large changes in push-related activity might have occurred in the earliest stage in which the number of successful trials was less than 10, a stage that we did not analyze. In macaque PM and M1, activities related to learning to associate new sensory cues to predetermined movements rapidly emerge within a session^13,14,18. Alternatively, substantial reorganization might occur in the deep layer. When electrical recording in macaque shows activity change during learning, the recorded neurons generally originate from the deep layer. In mice, the activity dynamics differ between M1 L2/3 and L5 neurons during learning of the lever-pull task⁵⁷. Endoscopy imaging and three-photon imaging could potentially be used to measure the activity of deep-layer neurons in the marmoset^22,60,61.

Differentiation of the premotor cortex during evolution

PMdr and PMdc differed substantially in the manner of their motor representation, as did PMdc and M1. The rodent M2, which is assumed to correspond to PM and includes the rostral forelimb area (RFA), is responsible for a wide range of information processing, including action planning, decision making, working memory, behavioral adaptation, and grasping^62,63,64,65. These functions are not spatially separated within M2, and hierarchical areas that correspond to PMdr and PMdc have not been found within M2. Thus, it is apparent that the motor cortex is more differentiated in the marmoset than in the rodent⁶². This difference is probably related to differences in the repertoires of forelimb movements between primates and rodents. Primate ancestors evolved the ability to move the forelimb independently of other body parts, and to differently move left and right forelimbs for grasping and leaping in arboreal life⁶⁶. In addition, primates need to flexibly associate many sensory targets, such as fruits and insects, with a variety of behavioral repertoires⁶⁷. This requirement would lead to differentiation of the M2 (more specifically, the RFA) seen in rodents to PMdr and PMdc (and also the pre-supplementary and supplementary motor cortices) in primates. The application of wide-field one-photon and wide field-of-view two-photon imaging methods to the neocortex of behaving marmosets offers promise for further understanding the cortical mechanisms of the complex sensorimotor transformations occurring in primates.

Limitations

In the present study, we did not image neuronal activity during learning of the OTR task. Thus, we were not able to detect the dynamics related to the emergence of the association between the target cue and forelimb movement, nor the process for improving the pole-pull movement. It is not clear how the learning order from the pull to push movements affected the neuronal dynamics during the imaging sessions. It is also unclear whether the pull and push movements are interchangeable, or whether learning the pull followed by the push would produce symmetrical results. We also used only a simple linear decoding model to estimate the motor representation, and therefore changes in motor representation that could not be detected by this model were not examined. The marmoset brain is lissencephalic, and it is therefore difficult to determine the borders of PMdr, PMdc, and M1 according to landmarks on the cortical surface such as the arcuate sulcus. Although we conducted ICMS in individual animals to account for the effects of individual differences, this was not sufficient to accurately identify the borders. Registration of the imaging field to a brain atlas with MRI would be useful to more accurately identify the areas⁶⁸.

Methods

Animals

All experiments were approved by the Animal Experimental Committee of the University of Tokyo and the Animal Care and Use Committees of the RIKEN Center for Brain Science. Seven laboratory-bred common marmosets (Callithrix jacchus; laboratory-bred and RIKEN Center for Brain Science, Research Resource Division) were used in the present study. Their age ranged from 20–68 months and their weight from ~250–350 g when the habituation started. Marmosets 1–3 were used for one-photon imaging throughout motor learning, marmosets 4–6 were used for two-photon imaging throughout the learning, and marmoset 7 was used for wide field-of-view two-photon imaging after more than 140 sessions of learning. Marmosets 1 and 6 were females and the others were males. The seven marmosets were randomly allocated to these experiments. All seven marmosets were kept under a 12:12-h light-dark cycle. None of them were used for any other experiments prior to the present study.

Methods details

Virus production

The AAV plasmid of human synapsin I promoter (hSyn)-tetracycline-controlled transactivator 2 (tTA2) was constructed by subcloning the DNA fragments containing hSyn and tTA2 into pAAV-MCS (Agilent Technologies, CA, USA). The AAV vector was produced as described previously^23,36,69. The generation of pAAV-TRE-GCaMP6s-P2A-GCaMP6s-WPRE (tandem GCaMP6s) is described in detail in ref. ³⁴. AAV plasmids were packaged into AAV serotype 9 using the AAV Helper-Free system and AAV-293 cells (Agilent Technologies).

Surgical procedures

All the surgeries and viral vector injections were carried out under aseptic conditions described previously^23,36. Each marmoset was placed in a stereotaxic instrument (SR-6C-HT; Narishige, Tokyo, Japan); 0.8–4.0% isoflurane anesthesia at a flow rate of 1 L/min was maintained; and the saturation of percutaneous oxygen (SpO₂), pulse rate, and rectal temperature were monitored throughout surgery. In the perioperative period, the following medicines were intramuscularly (i.m.) administered: ampicillin (16.7 mg per kg of body weight) as an antibiotic, carprofen (4.4 mg/kg) as an anti-inflammatory agent, and maropitant (1.0 mg/kg) as an antiemetic. In order to avoid dehydration, Ringer’s solution (10 mL) and riboflavin (vitamin B₂) were subcutaneously infused.

In the head plate implantation procedure, we depilated and sterilized the scalp, incised it with an external application of lidocaine, and removed the connective tissue to expose the skull. After six to seven small screws were anchored to the skull, a headplate was attached to the skull with universal primer (Tokuyama Dental, Tokyo, Japan), dual-cured adhesive resin cement (Bistite II or Estecem II; Tokuyama Dental), and dental resin cement (Superbond; Sun Medican, Siga, Japan). The task training under head fixation started more than a week after the head plate implantation.

ICMS

After sufficiently long OTR training sessions (Supplementary Fig. 1), craniotomy and durotomy were carried out under anesthesia and the medications described above were administered with additional intramuscular administration of dexametazone (0.5 mg/kg) and subcutaneous administration of D-mannitol (2 g/kg) to prevent cerebral edema. The exposed cortex was covered with silicone elastomer (Kwik-Sil; World Precision Instruments, FL, USA) and further covered with dental resin cement (Superbond; Sunmedical). To identify the border between the primary motor cortex (M1) and premotor cortex (PM), we conducted ICMS in a similar way to that described previously (Ebina et al.³⁶), 2–7 days after the craniotomy for marmosets 3–7, and after the imaging experiments were finished for marmosets 1–2 (Supplementary Fig. 1). During the ICMS, we anesthetized each marmoset with ketamine (initial dose 15 mg/kg; additional dose 5 mg/kg) and xylazine (initial dose 0.75 mg/kg; additional dose 0.25 mg/kg), and administered atropine (0.050 mg/kg) for sialoschesis to prevent respiratory obstruction. A silver reference electrode was immersed in the cerebrospinal fluid on the exposed cerebral surface, and then a tungsten microelectrode with an impedance of 0.5 MΩ and a diameter of 100 μm was inserted into the cerebral cortex to a depth of 1.5 or 1.8 mm. Twelve or fifteen 0.2-ms cathodal pulses of 333 Hz were applied. We increased the stimulation currents from 10 to 100 μA in steps of 10 μA until a body movement was detected.

The cortical map was determined as follows. First, we identified the cortical area that had a relatively low movement threshold in comparison with more posterior or anterior areas; we considered this area to be the main part of M1. Then, the border between M1 and PMdc was determined as the line at which the threshold current apparently increased along the posterior-to-anterior direction. The marmoset cortical map⁷⁰ was aligned with the dorsal cortex of the animal so that the midline and the border between M1 and PMdc matched between the marmoset cortical map and the animal. The putative border between M1 and PMdc was 12.6 ± 0.31 mm (mean ± standard deviation; n = 7 marmosets) from the inter-aural line along the anterior-posterior axis. The standard deviation was only 0.31 mm, so we therefore considered that the ICMS mapping was relatively consistent among the animals. However, even when the border location was correct, the shape of each cortical area might have still differed between the animals to some extent.

AAV injection

Mineral oil (Nacalai Tesque, Kyoto, Japan) was back-filled into a Hamilton syringe (25 μL) and a quartz pipette with an outer tip diameter of ~30 μm (Sutter Instruments, CA, USA). Then, the viral solution containing rAAV2/9-TRE3 promoter-tandem GCaMP6s (1.55 × 10¹³ or 1.36 × 10¹³ vector genomes [vg]/mL) and rAAV2/9-hSyn-tTA2 (3.9 × 10¹³ vg/mL) was front-loaded with a syringe pump (KDS310; KD Scientific, MA, USA). The viral solution was vertically injected into each site at a depth of 500 μm from the cortical surface at a rate of 0.10 μL/min to a total amount of 0.50 μL. Then, the pipette was maintained in place for 5 min before being slowly withdrawn. After injections into multiple sites, the exposed cortical surface was covered with a rectangular glass window of 15 × 8 mm for marmosets 1–3, 9 × 5 mm for marmosets 4–6, and a circular window with a diameter of 5.5 mm for marmoset 7.

Task behaviors

The marmosets were seated on the task apparatus while wearing a marmoset jacket (Ebina et al.²³). The task consisted of five steps, as described in ref. ²³: primary training without head fixation, pole-pull task without head fixation, pole-pull task with head fixation, OTR task, and TTR task (Supplementary Fig. 1). In the primary training, the animals were given water and food while seated in the chair to acclimate them. After the animals were sufficiently acclimated, we started training them in the pole-pulling task, in which a sucrose water reward was delivered when they pulled a pole with their left forelimb under a head-unfixed condition. After the animals were able to perform the pole-pull task for approximately 60 min, the pole-pull task with head fixation was started.

For the OTR and TTR tasks, a seven-inch liquid crystal display (LCD) monitor was placed 10–17 cm in front of the animal. Pulling and pushing of the pole moved the cursor downward and upward, respectively. Moving the pole to the right and left moved the cursor to the right and left, respectively. Each trial of the OTR task began with a holding period during which the pole was moved to the center position by a spring force of 0.25 N. That is, the cursor was moved to the fixation area. If the cursor stayed within the fixation area for 1400–2100 ms (randomly chosen for each trial), the fixation square and the spring force disappeared, and the green target rectangle appeared below the fixation area signaling the beginning of the reaching period. The target rectangle was presented during this period of 10–20 s. Animals were rewarded when they pulled the pole to move the cursor to the target within the reaching period and the cursor was held within the target rectangle for a set time (rewarded trials). In rewarded trials, the color of the target changed to white, and the white target was presented for 500 ms before it disappeared.

When the holding time was 400 ms or longer and the rate of the rewarded trials to the total trials was 70% or higher, the animals were considered to be experts and the TTR task started. In the TTR task, either of two green target rectangles appeared: a pull target below or a push target above the fixation square. The animals needed to move the cursor to the target within the allotted time and hold it within the target for the allotted time to receive 30–100 μL of sucrose water as a reward. The final parameters were 1000–1200 ms fixation period, 8 × 8 mm fixation area size, 16 × 16 mm target size, 14 mm distance between the centers of the fixation and target, an upper limit of 20 mm movement in the opposite direction, and 100 ms holding time within the target. When the TTR training session started, the target size (width) in push trials (TS_push) was set to 24–100 mm. Then, the target size narrowed throughout the sessions. In marmosets 1, 2, and 4, the target size in pull trials (TS_pull) was set to the same as TS_push. In marmosets 3, 5, and 6, TS_pull was fixed to 16 mm throughout the TTR training sessions and only TS_push was changed. The distance from the center of the fixation area to the center of target 1 (TD_pull) was set to 12–14 mm, and the distance from the center of the fixation area to the center of target 2 (TD_push) was set to 6–14 mm. The holding times in pull and push trials (HT_pull and HT_push, respectively) were the same in each session in marmosets 1, 2, and 4. In the other marmosets, HT_pull was 100 ms in all sessions and only HT_push was changed. These time courses are shown in Supplementary Fig. 2b–d. The minimum inter-trial interval between the cue-offset and fixation onset was 2.5 s for marmosets 1 and 2, 4.5 s for marmosets 3–6, and 2.0 s for marmoset 7.

In the TTR training sessions, we defined the early and late sessions as follows. Among the imaging sessions with at least 10 successful pull and push trials each, the early sessions were the first two or three imaging sessions before the 11th training session or those in which the rate of successful push trials was less than 0.5 (for two early sessions in marmoset 4), and the late sessions were the last three sessions in which both the rate of successful pull and push trials was more than 0.5 (Supplementary Table 1).

The trial-to-trial variability of the pole trajectory for each session was defined as the mean of the root mean square deviations (RMSDs) of the X and Y coordinates of individual trajectories from those of the trial-averaged trajectory. For each trial, the RMSD was calculated as $\sqrt{\frac{1}{n}{\sum }_{t}^{n}{\left({x}_{t}-\bar{{x}_{t}}\right)}^{2}+\frac{1}{n}{\sum }_{t}^{n}{\left({y}_{t}-\bar{{y}_{t}}\right)}^{2}}$, where n is the number of time points during the period from −500 to +200 ms of the pole movement onset, and ${x}_{t}$/${y}_{t}$ and $\bar{{x}_{t}}$/$\bar{{y}_{t}}$ are the X/Y coordinates of the trajectory in the trial and trial-averaged trajectory at time point t, respectively. The marmosets performed the task 1–5 days per week, over which their body weight was maintained at ~90% of their normal level by restriction of food and water. On the off-duty days, the food and water restrictions were weakened to allow them to return to their normal weight. The task events were controlled by LabVIEW software (National Instruments, TX, USA). The task data were sampled at 1 kHz.

To monitor body movements during the task performance, two CMOS cameras (DMK33UP1300, ImagingSource, Taipei, Taiwan) were placed at 35° and 90° angles from the front of marmosets 1 and 2, with single focal length lenses with f-numbers of 35 mm and 3 mm, respectively. Images of 320 × 240 pixels were acquired at a frame rate of 100 Hz. For marmosets 3–6, the cameras were placed at a 35° angle from the front of the marmoset with a single focal length lens (f = 3 mm) and a varifocal lens (f = 2.7–12.0 mm, Spacecom, Japan). For these marmosets, the pixel resolution and frame rate were changed to 480 × 480 pixels and 50 Hz to increase the prediction performance of the body movement tracking with DeepLabCut (https://github.com/DeepLabCut/DeepLabCut). For marmoset 7, the cameras were placed at 35° and 90° angles from the front of the marmoset with a single focal length lens (f = 35 mm and 3 mm, respectively). The pixel resolution and the frame rate of the images were set to 480 × 480 pixels and 30 Hz, respectively.

One-photon imaging

One-photon imaging in marmosets 1 and 2 was conducted with a variable zoom microscope (Axio Zoom.V16; Carl Zeiss, Jena, Germany) equipped with an air objective lens (Plan-NEOFLUAR Z 2.3×; numerical aperture 0.5; Carl Zeiss) and a FOV of 12.6 × 12.6 mm. The marmoset chair was rotated by 0°–6° in the anterior-posterior direction and 0°–12° in the lateral-medial direction. For marmosets 1 and 2, the microscope was not tilted. For marmoset 3, the microscope was tilted by 5°–15° in the anterior-posterior direction. These adjustments allowed the introduction of excitation light into the frontoparietal cortex perpendicular to the glass window²³. A fluorescence light source (HXP 200C; Carl Zeiss) and a filter set (38HE, Carl Zeiss; 470/40-nm excitation filter, 495-nm dichroic mirror, and 525/50-nm emission filter) were used for the imaging experiments. The intensity of the emitted excitation light was 5.0–6.5 mW. During the imaging, the animal’s head and the objective lens were covered with lightproof cloths to shut off possible stray light. A scientific CMOS camera (Sona; Andor Technology) with a resolution of 2048 × 2048 pixels was used as a photodetector, and the imaging was conducted at a frame rate of 30 Hz. Each series of imaging data consisted of 5400 frames (three minutes). Two-to-seventeen series of imaging data were acquired during a session. In marmoset 3, one-photon imaging was conducted with a custom-made zoom-variable microscope equipped with an air objective lens (Plan-NEOFLUAR Z 1.0×; numerical aperture 0.25; Carl Zeiss). The FOV size was set to 14.6 × 14.6 mm. A fluorescence light source (M470L5; Thorlabs) and a filter set were used for the imaging experiments. The excitation light intensity under the objective was 6.0 mW. A scientific CMOS camera (ORCA-Fusion; Hamamatsu Photonics) with a resolution of 2304 × 2304 pixels was used as a photodetector and the images were acquired at a frame rate of 30 Hz. Each series of imaging data consisted of 5400 or 10 800 frames (3 or 6 min). Two-to-seventeen series of imaging data were acquired during a session.

One-photon imaging to estimate the contamination of hemodynamic signals in the calcium imaging was also conducted in marmoset 3. An illumination light at 405 nm (M405L4, Thorlabs) was used to detect non-calcium-dependent fluorescence^38,71. The excitation light intensity under the objective was 9.0 mW for 470 nm and 3.0 mW for 405 nm. The images were acquired at a frame rate of 40 Hz. The excitation wavelength was switched from frame to frame, resulting in a 20-Hz frame rate for each excitation light.

Two-photon imaging during motor learning

For marmosets 4–6, we conducted two-photon imaging with a custom-built two-photon microscopy system (Olympus, Tokyo, Japan) that is described in ref. ²³. A femtosecond pulse laser (Femtolite FD/J-FD-500; pulse width, 191–194 fs; repetition rate, 51 MHz; wavelength, 920 nm; IMRA, MI, USA) was introduced to the microscope scanning head through a neodymium-based fiber so that X-Y scanning was possible without any tilt effect on the microscope body. The excitation beam was then passed through a dichromic mirror (transmission wavelength range, 800–1300 nm; reflection wavelength range, 400–755 nm) and a water immersion objective lens (XLPLN10XSVMP; numerical aperture, 0.6; working distance, 8 mm; Olympus). The intensity of the excitation beam under the objective lens was 35.0–65.0 mW. The fluorescence signal from the cortical tissue was reflected by the dichromic mirror and delivered to a cooled high-sensitivity photomultiplier tube through a liquid light guide with an infrared-cut filter (32BA750 RIF; wavelength range, 400–760 nm; Olympus).

The optical axis of the objective lens was inclined by an angle of 5.5°–15° and the chair was rotated horizontally to make the optical axis perpendicular to the cranial window. A bowl-shape made from aluminum foil was attached to the head plate on the animal with silicone elastomer (Kwik Cast, World Precision Instruments; Dentsilicone-V, Shofu, Japan) and the space above the animal’s head was covered with light-shielding cloths. Two-to-ten imaging series were acquired at a frame rate of 30 Hz for 3 min (5400 frames) using FV30S-SW software (Olympus). The resolution of the imaging field was 512 × 512 pixels, with a pixel length corresponding to 1–1.2 μm. Body movements were recorded in the same way as for one-photon imaging.

Wide field-of-view two-photon imaging

Imaging was conducted with a wide field-of-view microscopy system (Nikon, Tokyo, Japan)⁴³. This was equipped with a large objective lens (dry objective, 0.8 numerical aperture, 56-mm pupil diameter, Strehl ratio ~0.99 over the FOV, working distance of 4.5 mm, 35-mm focal length) and large-aperture (14-mm² aperture) gallium arsenide phosphide photomultipliers (GaAsP PMTs; R15248-40, Hamamatsu Photonics, Japan) with high current output (50 µA). A Ti:sapphire laser (Chameleon Vision-S Coherent Inc) tuned to 920 nm was introduced into a pre-chirper and then led to the resonant and galvanometric mirrors, and to the pupil of the objective lens through the scan tube lenses. The laser power under the objective lens was 60–90 mW. Emitted light was collected through 775-nm and 560-nm long-pass dichroic mirrors and 515–565-nm and 600–681-nm band-pass emission filters with GaAsP PMTs. A series of images were acquired at a frame rate of 7.5 Hz and a resolution of 2048 × 2048 pixels using Falcon software (Nikon). The total imaging duration was 10 min (4600 frames) for each imaging session.

The anterior-posterior (AP) axial-angle-adjustable stage was placed above the goniometer, and the lateral-medial-axial-angle and AP-angle adjustable marmoset chair was placed above the stage. The marmoset chair was used to restrain the body and to fixate the head of the marmoset²³. The marmoset chair and the stage were tilted to position the focal plane of the objective lens parallel to the glass window placed on the cortical surface. The angle of the stage was adjusted every session. The space between the objective lens and the animal’s head was covered with aluminum foil to shield the objective from sprayed light. Before the surgery for the virus injection and glass window placement, marmoset 7 was habituated to perform the TTR task in the tilted chair in the microscopy environment. Since the conditions of the microscopy environment were different for marmoset 7 in comparison with the others, marmoset 7 was not included in the comparison of behavioral performance among the animals (Fig. 1d–i and Supplementary Fig. 4c, d).

In the imaging sessions, the session-averaged rates of the successful pull and push trials were 0.781 ± 0.047 and 0.640 ± 0.03 (n = 4 days), respectively. The variability values of the trial-to-trial pole trajectory in the successful pull and push trials were 1.36 ± 0.06 mm and 1.65 ± 0.04 mm, respectively. The reaction times in the successful pull and push trials were 709.4 ± 35.5 ms and 917.9 ± 56.4 ms, respectively.

Pharmacological inactivation

To examine the effect of neuronal inactivation on task behavior, muscimol was injected into appropriate brain areas of marmosets 1 and 2 after all imaging experiments were finished. Before the first injection, the glass cranial window was replaced with a silicon-based window (thickness of 100 μm, 6-9085-12, AS-ONE Corporation, Japan), through which a microinjection needle was penetrated. Using a needle with an outer diameter of approximately 60 μm made from a quartz pipette and joined to a Hamilton syringe (25 μL), 0.25 μL of 5 μg/μL muscimol was injected at a rate of 0.10 μL/min at a depth of 500 μm in one of the following cortical areas: PM, M1, the somatosensory cortex, and the posterior parietal cortex (Supplementary Fig. 3c). As a control experiment, the same amount of saline was injected on other days. Two hours after the injection, the marmosets performed the TTR task that they had already learned. The interval between each injection experiment was more than or equal to 1 day. We did not inject muscimol into PMdc because we considered that if muscimol was injected into PMdc it might diffuse to M1 to some extent, and the origin of the effect on behavior might therefore be obscured. However, in our recent study³¹, the lateral spread of muscimol in the marmoset neocortex was estimated to be at most ~1–2 mm. Taken together with the fact that inactivation effects on the forelimb movement were not detected following muscimol injection into the somatosensory cortex in the current study, we conclude that muscimol spreading would be mainly limited to within the injected area.

Processing of one-photon imaging data

Tangential drifts in the imaging were removed with a finite Fourier transform algorithm⁷² and the data were down-sampled from 2048 × 2048 or 2304 × 2304 pixels to 256 × 256 pixels. For each pixel of the down-sampled image, ΔF/F(t), the relative change in fluorescence at a time point t was defined as (F(t) – F₀(t))/F₀(t), where F(t) is the fluorescence intensity at a time point t, and F₀(t) is the 8th percentile of F(t) across t ± 15 s. Data for the initial and end 15-s periods, and data that included time with missing video data, were excluded. The position of each cortical area across the experimental sessions was registered with the NoRMCorre program (available at https://github.com/flatironinstitute/NoRMCorre).

Imaging data were further down-sampled from 256 × 256 pixels to 64 × 64 pixels. For each pixel, the z-scored ΔF/F traces were calculated and normalized within each session. The z-score of each pixel was denoised by singular value decomposition (SVD), and the imaging data in the analyzed trials were concatenated in the time direction to create a data matrix. Next, SVD was performed on this matrix to decompose the pixel × time matrix data into multiple spatiotemporal components. Of these, the first 200 components explained more than 90% of the variance. Therefore, the original data were reconstructed from these 200 elements to remove the noise from the data⁷¹. The reconstructed data (denoised ΔF/F data) were used for the following analyses.

The timing of the peak activity (T_peak) was calculated from the successful trials in which the movement onset occurred more than 0.5 s and less than 4.0 s after the cue onset to exclude possible trials in which the animal quickly moved the pole by a random guess without looking at the target properly, or trials in which the animal might not have looked at the target. The z-scored trace was processed with Savitzky-Golay filtering with two orders and 15 frames (=0.5 s). For each type of pull and push trial, the timing of the maximum peak of the trial-averaged z-scored trace within 3 s after the cue onset was defined as T_peak for each pixel. If the pixel had no peaks within 3 s after the cue onset, it was removed from the analysis.

For the analysis of pairwise conditional transfer entropy (TE), pixel-averaged activity during the period from 2 s before to 1 s after the movement onset in the successful pull or push trials in each session was binned into 0.1-s intervals. For each trial, the time-averaged binned activity was subtracted from the binned activity to satisfy a stationarity criterion. The binned activity was then concatenated within early or late sessions and used to calculate TE with the DCode package (https://github.com/smsxiaomayi/DCcode/blob/main/DCcode.zip)³⁹. We confirmed that the concatenated data are stationary by KPSS test (P > 0.05; MATLAB kpsstest function). The TE was further averaged across three animals. To examine the statistical significance of this TE value, the activity was shuffled along the temporal and trial direction and the TE (shuffled TE) was calculated in the same manner. This was repeated 200 times. If the TE value was above the 99th percentile of the shuffled TEs, it was considered significant.

To estimate the motor representation of each pixel, we constructed a decoding model to predict the pole movement from the neuronal activities, referring to a previous study⁷³. We applied a multiple linear regression model to the ΔF/F trace to predict the z-scored Y-axial trajectory of the pole from 1 s before to 0.13 s after the movement onset. The model was fitted separately for pull and push trials. In this process, we first down-sampled pole trajectory data at 30 Hz by averaging the positions during the acquisition of each imaging frame. We then aligned the ΔF/F trace to the movement onset, defined as when the cursor was moved outside the fixation square, and set seven-time windows that were shifted by 100 ms (corresponding to three frames) in a 300-ms range before and after each time point. The predicted Y-axial trajectory at time t, $\hat{y}\left(t\right)$, was expressed using the following formula:

$$\hat{{{{\rm{y}}}}}\left({{{\rm{t}}}}\right)={{\sum}_{\varDelta t}}{\beta }_{\varDelta t}{{\Delta }}F/F\left(t+\Delta t\right)+{\beta }_{{bias}}$$

(1)

where Δt was set to 0, ±0.1, ±0.2, or ±0.3 s, β_Δt is the coefficient at Δt, and β_bias is the bias term. $\hat{y}\left(t\right)$ was calculated by fivefold cross-validation. To estimate the decoder performance, the cross-validated coefficient of determination (cvR²) was used as a measure of prediction accuracy. We calculated the coefficient of determination as the square of the correlation coefficients between the observed and predicted Y-axial pole trajectory.

In addition to the single-pixel decoder, another decoding model was constructed to predict pole movement from the activities of multiple pixels in individual motor cortical areas. To reduce overfitting by the decoder constructed from high-dimensional neuronal activity data, we extracted a low-dimensional neuronal subspace that captured most of the variance in the original neuronal activity space as follows. First, the neuronal activity data during –1.0 to +0.13 s from the movement onset were concatenated and principal component analysis was used to generate a matrix that transformed the original neural activity to 10-dimensional population activity, which explained more than 95.99% of variance in the neuronal activity (Supplementary Fig. 7d). The predicted Y-axial trajectory at time t, $\hat{y}\left(t\right)$, was expressed using the following formula:

$$\hat{{{{\rm{y}}}}}\left({{{\rm{t}}}}\right)={{\sum}_{D}}{{\sum}_{\varDelta t}}{\beta }_{D,\varDelta t}\,{X}_{{pca}}\left(t+\Delta t\right)+{\beta }_{{bias}}$$

(2)

where X_pca is the low-dimensional population activity, Δt was set to 0, ± 0.1, ± 0.2, or ± 0.3 s, β_D,_Δt) is the coefficient for the low-dimensional activity in the Dth dimension at Δt, and ${\beta }_{{bias}}$ is the bias term. $\hat{y}\left(t\right)$ was calculated by fivefold cross-validation.

The PDI for each pixel was defined as (cvR² in pull trials – cvR² in push trials)/(cvR² in pull trials + cvR² in push trials). A PDI of 1 or –1 indicates that only pull or push movement was represented, respectively. To ensure the denominator was not too small, we used only pixels that showed cvR² > 0.02 (the reason for using 0.02 is described in the next subsection) in at least one type of successful pull and push trial for the PDI calculation. However, all PMdr, PMdc, and M1 pixels in marmosets 1–3 showed cvR² > 0.02, and therefore all these pixels were used for the PDI calculation. When ΔF/F is used as a measure of PDI, we need to assume that the amplitude and velocity of the pole are the same for push and pull movements, and the activities of neurons that respond specifically to each direction have the same amplitude. Without this assumption, PDI will not be zero in a neuron that represents pull and push movements equally. It is also necessary that movements that are unrelated to the pole movement do not largely affect the neuronal activity. However, these assumptions did not apply to the current experiment because the cursor trajectory in successful pull and push trials was not symmetric, and orofacial and body movements were allowed. Therefore, we used only a relatively simple linear regression between the pole trajectory and ΔF/F to extract the push/pull representation-specific activity and calculate the PDI.

When the hemodynamic-signal contamination of the fluorescence signal from calcium imaging was estimated, the ΔF/F from the data obtained with the 405-nm light (violet light-excited ΔF/F, violet-excited ΔF/F) was z-scored and smoothed using a moving average filter with a time window of 400 ms. For each pixel, the z-scored ΔF/F from the data obtained with the 470-nm light (blue light-excited ΔF/F, blue-excited ΔF/F) was linearly fitted with the smoothed z-scored violet-excited ΔF/F. Then, the smoothed z-scored violet-excited ΔF/F multiplied by the weights used for fitting was subtracted from the z-scored blue-excited ΔF/F to calculate the hemodynamic corrected ΔF/F (Supplementary Fig. 5c–f). As shown in Supplementary Fig. 5d–f, the blue-excited ΔF/F and the corrected ΔF/F mostly overlapped, indicating that the hemodynamic contamination in the present one-photon imaging dataset was subtle. Therefore, we did not correct the one-photon calcium imaging dataset for hemodynamic contamination.

Processing of two-photon imaging data

Motion correction for two-photon imaging data was conducted with the same protocol as used for the one-photon imaging data. Then, active neuronal somata were extracted using a CNMF algorithm⁴⁴. We defined the active neurons as those ROIs whose automatically extracted activities reached the following criteria: a minimum spatial component size of 50 μm²; a minimum spatial component ellipticity of 0.5; a minimum signal-to-noise ratio of the estimated component of 2. For each active neuron, we calculated the detrended relative fluorescent change ΔF/F = (F–F₀)/F₀, where F₀ is the eight-percentile value over an interval of ±15 s around each time point. We corrected the X-Y shift between imaging datasets with the NoRMCorre algorithm and then identified the same ROI in the different sessions with the register_ROIs function in the CaImAn package (https://github.com/flatironinstitute/CaImAn-MATLAB). Neurons that were identified as the same neuron in at least one early session and at least one late session were defined as pursued neurons.

We also calculated the signal-to-noise ratio and decay time constant of the calcium transients of the pursued neurons in the two-photon imaging dataset. To calculate the signal-to-noise ratio, we first detected calcium events from the ΔF/F traces in early and late sessions according to Prsa et al.⁷⁴. We then calculated the “signal” of the calcium events as the mean of concatenated ΔF/F traces during 0.0 to 0.5 s from the event onset. “Noise” was also computed as the standard deviation of the ΔF/F traces during –0.5 to 0.0 s from the onset. The signal-to-noise ratio was calculated by dividing the signal by the noise. The decay time constant was calculated by fitting an exponential decay function to the event-averaged ΔF/F trace.

We constructed a decoding model to predict the pole movement from the activity of a single neuron detected in both early and late sessions with the same regression model used for the one-photon imaging data (see Eq. 1). To estimate the change in cvR² of the single-neuron decoding performance during learning, the cvR² of each neuron was averaged over early or late sessions and the averaged cvR² was compared between early and late sessions.

In addition to the single-neuron decoder, we also constructed a population decoder using the activities of multiple neurons in individual recording datasets. To reduce the possible influence of the difference in the number of trials and neurons between early and late sessions on the decoding performance, we randomly selected 10 pull/push trials and the same number of pursued neurons with early and/or late cvR² of >0.02 from early and late imaging sessions. The number of neurons used for the decoder construction was determined as the minimum number of pursued neurons in the early and late imaging sessions in the set. The numbers for individual datasets are shown in Fig. 8g, h. We excluded sessions with less than ten pursued neurons from this analysis (imaging session 3 in marmoset 5 and session 1 in marmoset 6). To reduce overfitting by the decoder constructed from high-dimensional neuronal activity data, we then searched for the neuronal subspace that captured most of the variance in the original neuronal activity space as follows. First, the neural activity data during –1.0 to +0.13 s from the movement onset were trial-averaged and principal component analysis was used to generate a matrix that transformed the original neural activity to low-dimensional population activity capturing more than 95% of the variance in the trial-averaged activity. The low-dimensional activity in the individual trials was calculated by applying the same transformation matrix to the neural activity in the trials.

The predicted Y-axial trajectory at time t, $\hat{y}\left(t\right)$, was expressed using Eq. 2. To assess the changes in the cvR² during learning, the cvR² was averaged over early sessions or late sessions, and a Δpopulation decoder for cvR² was calculated by subtracting the late sessions-averaged cvR² from the early sessions-averaged cvR². These processes started with a random selection of ten pull/push trials and pursued neurons, and were repeated 1000 times, with the difference being considered statistically significant when the 2.5th percentile of the 1000 Δpopulation decoder cvR² was above 0.0, or the 97.5th percentile was below 0.0.

To determine whether cvR² for each neuron was statistically significant, the neuronal activity was randomly shuffled along the temporal axis and cvR² was calculated. This was repeated 200 times. When the actual cvR² exceeded the 95th percentile of the distribution in the 200 instances, cvR² was considered significant. As shown in Supplementary Figs. 9d and 10a, b, the majority (87–100%) of cvR² > 0.02 were significant. Therefore, we defined these neurons as movement-related neurons and used them in the PDI calculation. The value of 0.02 was also used in the analysis of one-photon imaging data to ensure consistency throughout the analysis.

Raw image sequences acquired with wide field-of-view two-photon microscopy were motion-corrected with the NoRMCorre package. A time-averaged image was used as the target image for the motion correction. Then, to improve the signal-to-noise ratio of the imaging data, we performed shot noise reduction using the deep self-supervised learning-based denoising algorithm DeepCAD-RT⁷⁵. All datasets were denoised with the convolutional neural network (CNN), which was trained on all datasets (11 imaging sequences from six imaging sessions). The training process was terminated at the 10th iteration because the performance of the trained CNN was optimal, as mentioned in the original article. The CNMF algorithm was employed to extract neuronal activities from a time series of images (CaImAn package for MATLAB)⁴⁴. The factor of the autoregressive system was such that the inferred neuronal activities had one decay time constant in this step. We defined those ROIs whose automatically extracted activities reached the following criteria as the active neurons: a maximum spatial component size of 250; a minimum spatial component size of 10; a threshold for the spatial correlation between the estimated component and raw data of 0.6; a threshold for the temporal correlation between the estimated component and raw data of 0.5; a minimum signal-to-noise ratio of the estimated component of 2. For each active neuron, we calculated the detrended relative fluorescent change (ΔF/F) using the detrend_df_f function in the CNMF package with a time window of ±15 s for calculating background fluorescence. The single-neuron decoder was constructed with the same protocol as described above, but the pole trajectory was down-sampled to 7.5 Hz and Δt was set to 0, ±0.133, ±0.266, or ±0.399 s, because the frame rate of the wide field-of-view two-photon imaging was 7.5 Hz.

To determine whether the calculated PDI values of pixels or neurons significantly differed from zero, we randomly changed the combinations of trials of five segments for fivefold cross-validation for each pixel or neuron, and then calculated cvR² for successful pull and push trials and PDI. This procedure was repeated 200 times, and if more than 95% of these PDI values were above zero or below zero, we considered the PDI to be statistically significant.

Regression models

To determine which task-related variables were related to the changes in reaction time in successful pull trials (RT_pull), the cvR² of PMdr for push movement, and the |ΔPDI_pair| of pursued neuron pairs whose distances were <300 μm, we constructed three multivariable regression models for each. In the full model, the explanatory variables consisted of six task performance variables and six task difficulty-related variables. The six task performance variables were the success rates for pull trials (SR_pull) and push trials (SR_push), RT_pull (this was removed in the models to explain RT_pull), the reaction time for push trials (RT_push), and the trial-to-trial variabilities of the pole trajectory for pull trials (Var_pull) and push trials (Var_push). The six task difficulty-related variables were TS_pull, TS_push, TD_pull, TD_push, HT_pull, and HT_push. For each session, TS_pull and TS_push were defined as the largest target size within the session, and HT_pull and HT_push were defined as the longest holding time within the session. The variables (including RT_pull) except for the success rate were normalized to the range 0–1 for each animal. In the task performance model, the explanatory variables were six (or five) task performance-related variables. In the task difficulty model, the explanatory variables were six task difficulty-related variables. Then, we conducted fivefold cross-validation and calculated the prediction accuracy. In addition, the values for individual explanatory variables were randomly shuffled across sessions and the prediction accuracy was calculated for each model. This was repeated 1000 times. When the actual prediction accuracy exceeded the 95th percentile of the simulated values, the model was considered significant. We used the same results of the 1000× shuffling to consider whether the actual weight of each variable in the model was significant according to whether it was above the 97.5th percentile or below the 2.5th percentile of the weight values in the shuffled data.

Tracking of body movements

Movements of the upper limbs, tongue, and pupil in the videos were tracked with DeepLabCut³⁵. Fifty frames from video representing the dataset were automatically extracted on two experimental days. Since the tongue was not visible in many frames for marmosets 1 and 2, another 50-frame dataset that included the tongue movement was manually prepared from the same video. In the extracted dataset, the position of each part of the body was manually labeled, and on the basis of this labeled dataset, the position of each body part in the other frames was predicted with DeepLabCut. We adopted the data with a likelihood of >0.95 (for the marmosets 1 and 2) or >0.6 (for the marmosets 3–6), and the movement in the discarded frames was linearly interpolated using the data before and after these frames. The position of the pupil was determined as the midpoint of the predicted positions of the left and right edges of the pupil. Pupil diameter was calculated as the distance between these positions. The predicted position of each part of the body was smoothed with a moving average of five frames. The trajectories of the pole, left hand, right hand, left elbow, left shoulder, left knee, right knee position, pupil position, and pupil diameter were z-scored in the same way as ΔF/F. As described above, the images for marmosets 1 and 2 had a lower pixel number and a higher frame rate (shorter exposure time) than the images for marmosets 3–6. Thus, the quality of the former images was worse than that of the latter images, meaning that the tongue position when the mouth was closed was frequently assigned to the wrong place in marmosets 1 and 2. Therefore, the following criteria to detect licking were also set for marmosets 1 and 2: the predicted tongue position should be inside the ROI set near the mouth, and the intensity of another ROI over the mouth should be above a threshold value (for marmoset 1, the 30th percentile of the intensity values during the imaging; for marmoset 2, the 50th percentile of the values) because the tongue showed higher intensity than the closed mouth. No licking was assumed during periods when these criteria were not met.

Statistics and reproducibility

Statistical analyses were performed using MATLAB (2018–2020a, MathWorks). The Wilcoxon signed-rank test, Wilcoxon rank-sum test, Spearman’s rank correlation coefficient test, Pearson’s rank correlation test, and random permutation tests were used for statistical comparisons. No statistical tests were run to predetermine the sample size. However, sample sizes were estimated according to previous methodically comparable laboratory experiments and were similar to those generally employed in the field. Data are presented as mean ± SEM unless otherwise noted. Statistical values for all tests used in the figures are described in an Excel file (Statistics_Ebina et al.xlsx). Images similar to Fig. 6a were obtained in the total six sessions from one animal. Images similar to Fig. 8a were obtained in the total 54 sessions from three animals.

Reporting summary

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

Data availability

The imaging, behavioral, and task-event datasets are available at https://dataportal.brainminds.jp/caimaging-matsuzaki-2. All other data are available from the corresponding author. Source data are provided with this paper.

Code availability

All computer codes are available from the corresponding author.

References

Strick, P. L., Dum, R. P. & Rathelot, J. The cortical motor areas and the emergence of motor skills: a neuroanatomical perspective. Annu. Rev. Neurosci. 44, 425–447 (2021).
Article CAS PubMed Google Scholar
Nudo, R. J., Milliken, G. W., Jenkins, W. M. & Merzenich, M. M. Use-dependent primary motor alterations of movement representations cortex of adult squirrel monkeys. J. Neurosci. 16, 785–807 (1996).
Article CAS PubMed PubMed Central Google Scholar
Luppino, G., Rozzi, S., Calzavara, R. & Matelli, M. Prefrontal and agranular cingulate projections to the dorsal premotor areas F2 and F7 in the macaque monkey. Eur. J. Neurosci. 17, 559–578 (2003).
Article PubMed Google Scholar
Barbas, H. & Pandya, D. N. Architecture and frontal cortical connections of the premotor cortex (area 6) in the rhesus monkey. J. Comp. Neurol. 256, 211–228 (1987).
Article CAS PubMed Google Scholar
Wise, S. P., Boussaoud, D., Johnson, P. B. & Caminiti, R. Premotor and parietal cortex: corticocortical connectivity and combinatorial computations 1. Annu. Rev. Neurosci. 20, 25–42 (1997).
Article CAS PubMed Google Scholar
Picard, N. & Strick, P. L. Imaging the premotor areas. Curr. Opin. Neurobiol. 11, 663–672 (2001).
Article CAS PubMed Google Scholar
Abe, M. & Hanakawa, T. Functional coupling underlying motor and cognitive functions of the dorsal premotor cortex. Behav. Brain Res. 198, 13–23 (2009).
Cisek, P. & Kalaska, J. F. Neural correlates of reaching decisions in dorsal premotor cortex: specification of multiple direction choices and final selection of action. Neuron 45, 801–814 (2005).
Article CAS PubMed Google Scholar
Grafton, S. T., Fagg, A. H. & Arbib, M. A. Dorsal premotor cortex and conditional movement selection: a PET functional mapping study. J. Neurophysiol. 79, 1092–1097 (1998).
Article CAS PubMed Google Scholar
Hikosaka, O., Nakamura, K., Sakai, K. & Nakahara, H. Central mechanisms of motor skill learning. Curr. Opin. Neurobiol. 12, 217–222 (2002).
Article CAS PubMed Google Scholar
Deiber, M. et al. Frontal and parietal networks for conditional motor learning: a positron emission tomography study. J. Comp. Neurol. 522, 3683–3716 (2014).
Google Scholar
Shadmehr, R. & Holcomb, H. H. Neural correlates of motor memory consolidation. Science 277, 821–825 (1997).
Article CAS PubMed Google Scholar
Mitz, A. R., Godschalk, M. & Wise, S. P. Learning-dependent neuronal activity in the premotor cortex: activity during the acquisition of conditional motor associations. J. Neurosci. 11, 1855–1872 (1991).
Article CAS PubMed PubMed Central Google Scholar
Zach, N., Inbar, D., Grinvald, Y., Bergman, H. & Vaadia, E. Emergence of novel representations in primary motor cortex and premotor neurons during associative learning. J. Neurosci. 28, 9545–9556 (2008).
Article CAS PubMed PubMed Central Google Scholar
Germain, L. & Lamarre, Y. Neuronal activity in the motor and premotor cortices before and after learning the associations between auditory stimuli and motor responses. Brain Res. 611, 175–179 (1993).
Article CAS PubMed Google Scholar
Brasted, P. J. & Wise, S. P. Comparison of learning-related neuronal activity in the dorsal premotor cortex and striatum. Eur. J. Neurosci. 19, 721–740 (2004).
Article PubMed Google Scholar
Li, C. S. R., Padoa-Schioppa, C. & Bizzi, E. Neuronal correlates of motor performance and motor learning in the primary motor cortex of monkeys adapting to an external force field. Neuron 30, 593–607 (2001).
Article CAS PubMed Google Scholar
Paz, R. & Vaadia, E. Learning-induced improvement in encoding and decoding of specific movement directions by neurons in the primary motor cortex. PLoS Biol. 2, 264–274 (2004).
Article Google Scholar
Georgopoulos, A. P., Merchant, H., Naselaris, T. & Amirikian, B. Mapping of the preferred direction in the motor cortex. Proc. Natl. Acad. Sci. USA 104, 11068–11072 (2007).
Article ADS CAS PubMed PubMed Central Google Scholar
Amirikian, B. & Georgopoulos, A. P. Modular organization of directionally tuned cells in the motor cortex: Is there a short-range order? Proc. Natl. Acad. Sci. USA 100, 12474–12479 (2003).
Article ADS CAS PubMed PubMed Central Google Scholar
Ben-Shaul, Y. et al. Dynamical organization of directional tuning in the primate premotor and primary motor cortex. J. Neurophysiol. 89, 1136–1142 (2003).
Article PubMed Google Scholar
Bollimunta, A. et al. Head-mounted microendoscopic calcium imaging in dorsal premotor cortex of behaving rhesus macaque. Cell Rep. 35, 109239 (2021).
Article CAS PubMed PubMed Central Google Scholar
Ebina, T. et al. Two-photon imaging of neuronal activity in motor cortex of marmosets during upper-limb movement tasks. Nat. Commun. 9, 1879 (2018).
Article ADS PubMed PubMed Central Google Scholar
Chestek, C. A. et al. Single-neuron stability during repeated reaching in macaque premotor cortex. J. Neurosci. 27, 10742–10750 (2007).
Article CAS PubMed PubMed Central Google Scholar
Bakola, S., Burman, K. J. & Rosa, M. G. P. The cortical motor system of the marmoset monkey (Callithrix jacchus). Neurosci. Res. 93, 72–81 (2015).
Article PubMed Google Scholar
Burman, K. J., Palmer, S. M., Gamberini, M., Spitzer, M. W. & Rosa, M. G. P. Anatomical and physiological definition of the motor cortex of the marmoset monkey. J. Comp. Neurol. 506, 860–876 (2008).
Article PubMed Google Scholar
Krubitzer, L. A. & Kaas, J. H. The organization and connections of somatosensory cortex in marmomsets. J. Neurosci. 10, 952–974 (1990).
Article CAS PubMed PubMed Central Google Scholar
Walker, J., MacLean, J. & Hatsopoulos, N. G. The marmoset as a model system for studying voluntary motor control. Dev. Neurobiol. 77, 273–285 (2017).
Article PubMed Google Scholar
Matsuzaki, M. & Ebina, T. Common marmoset as a model primate for study of the motor control system. Curr. Opin. Neurobiol. 64, 103–110 (2020).
Article CAS PubMed Google Scholar
Kalaska, J. F. Emerging ideas and tools to study the emergent properties of the cortical neural circuits for voluntary motor control in non-human primates [version 1; peer review: 4 approved]. F1000Research 8, 1–13 (2019).
Article Google Scholar
Obara, K. et al. Change detection in the primate auditory cortex through feedback of prediction error signals. Nat. Commun. 14, 6981 (2023).
Sadakane, O. et al. Long-term two-photon calcium imaging of neuronal populations with subcellular resolution in adult non-human primates. Cell Rep. 13, 1989–1999 (2015).
Article CAS PubMed 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 Google Scholar
Matsui, T. et al. Orthogonalization of spontaneous and stimulus-driven activity by hierarchical neocortical areal network in primates. Preprint at https://doi.org/10.1101/2024.07.01.601463 (2024).
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 Google Scholar
Ebina, T. et al. Arm movements induced by noninvasive optogenetic stimulation of the motor cortex in the common marmoset. Proc. Natl. Acad. Sci. USA 116, 22844–22850 (2019).
Article ADS CAS PubMed PubMed Central Google Scholar
Pisauro, M. A., Dhruv, N. T., Carandini, M. & Benucci, A. Fast hemodynamic responses in the visual cortex of the awake mouse. J. Neurosci. 33, 18343–18351 (2013).
Article CAS PubMed PubMed Central Google Scholar
Allen, W. E. et al. Global representations of goal-directed behavior in distinct cell types of mouse neocortex. Neuron 94, 891–907.e6 (2017).
Article CAS PubMed PubMed Central Google Scholar
Shi, J., Chen, L. & Aihara, K. Embedding entropy: a nonlinear measure of dynamical causality. J. R. Soc. Interface 19, 20210766 (2022).
Terada, S., Kobayashi, K. & Matsuzaki, M. Transition of distinct context-dependent ensembles from secondary to primary motor cortex in skilled motor performance. Cell Rep. 41, 111494 (2022).
Article CAS PubMed Google Scholar
Kondo, M. & Matsuzaki, M. Neuronal representations of reward-predicting cues and outcome history with movement in the frontal cortex. Cell Rep. 34, 108704 (2021).
Article CAS PubMed Google Scholar
Burish, M. J., Stepniewska, I. & Kaas, J. H. Microstimulation and architectonics of frontoparietal cortex in common marmosets (Callithrix jacchus). J. Comp. Neurol. 507, 1151–1168 (2008).
Article PubMed Google Scholar
Ota, K. et al. Fast, cell-resolution, contiguous-wide two-photon imaging to reveal functional network architectures across multi-modal cortical areas. Neuron 109, 1810–1824.e9 (2021).
Article CAS PubMed Google Scholar
Pnevmatikakis, E. A. et al. Simultaneous denoising, deconvolution, and demixing of calcium imaging data. Neuron 89, 285–299 (2016).
Article CAS PubMed PubMed Central Google Scholar
di Pellegrino, G. & Wise, S. P. A neurophysiological comparison of three distinct regions of the primate. Brain 114, 951–978 (1991).
Article PubMed Google Scholar
Sayegh, P. F., Hawkins, K. M., Hoffman, K. L. & Sergio, L. E. Differences in spectral profiles between rostral and caudal premotor cortex when hand-eye actions are decoupled. J. Neurophysiol. 110, 952–963 (2013).
Article PubMed Google Scholar
Kurata, K. & Wise, S. P. Premotor cortex of rhesus monkeys: set-related activity during two conditional motor tasks. Exp. Brain Res. 69, 327–343 (1988).
Burman, K. J., Bakola, S., Richardson, K. E., Reser, D. H. & Rosa, M. G. P. Patterns of afferent input to the caudal and rostral areas of the dorsal premotor cortex (6DC and 6DR) in the marmoset monkey.J. Comp. Neurol. 522, 3682–3716 (2014).
Google Scholar
Andersen, R. A. & Buneo, C. A. Intentional maps in posterior parietal cortex. Annu. Rev. Neurosci. 25, 189–220 (2002).
Article CAS PubMed Google Scholar
Hick, W. E. On the rate of gain of information. Q. J. Exp. Psychol. 4, 11–26 (1952).
Article Google Scholar
Kveraga, K., Boucher, L. & Hughes, H. C. Saccades operate in violation of Hick’s law. Exp. Brain Res. 146, 307–314 (2002).
Article PubMed Google Scholar
Paz, R., Boraud, T., Natan, C., Bergman, H. & Vaadia, E. Preparatory activity in motor cortex reflects learning of local visuomotor skills. Nat. Neurosci. 6, 882–890 (2003).
Article CAS PubMed Google Scholar
Lee, D., Port, N. L., Kruse, W. & Georgopoulos, A. P. Variability and correlated noise in the discharge of neurons in motor and parietal areas of the primate cortex. J. Neurosci. 18, 1161–1170 (1998).
Article CAS PubMed PubMed Central Google Scholar
Stark, E., Globerson, A., Asher, I. & Abeles, M. Correlations between groups of premotor neurons carry information about prehension. J. Neurosci. 28, 10618–10630 (2008).
Article CAS PubMed PubMed Central Google Scholar
Smith, W. S. & Fetz, E. E. Synaptic interactions between forelimb-related motor cortex neurons in behaving primates. J. Neurophysiol. 102, 1026–1039 (2009).
Article CAS PubMed PubMed Central Google Scholar
Asanuma, H. & Rosén, I. Topographical organization of cortical efferent zones projecting to distal forelimb muscles in the monkey. Exp. Brain Res. 14, 243–256 (1972).
Article CAS PubMed 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 Google Scholar
Karni, A. et al. Functional MRI evidence for adult motor cortex plasticity during motor skill learning. Nature 377, 155–158 (1995).
Sanes, J. N. & Donoghue, J. P. Plasticity and primary motor cortex. Annu. Rev. Neurosci. 23, 393–415 (2000).
Article CAS PubMed Google Scholar
Kondo, T. et al. Calcium transient dynamics of neural ensembles in the primary motor cortex of naturally behaving monkeys. Cell Rep. 24, 2191–2195.e4 (2018).
Article CAS PubMed Google Scholar
Ouzounov, D. G. et al. In vivo three-photon imaging of activity of GCaMP6-labeled neurons deep in intact mouse brain. Nat. Methods 14, 388–390 (2017).
Article CAS PubMed PubMed Central Google Scholar
Barthas, F. & Kwan, A. C. Secondary motor cortex: Where ‘sensory’ meets ‘motor’ in the rodent frontal cortex. Trends Neurosci. 40, 181–193 (2017).
Article CAS PubMed Google Scholar
Svoboda, K. & Li, N. Neural mechanisms of movement planning: motor cortex and beyond. Curr. Opin. Neurobiol. 49, 33–41 (2018).
Article CAS PubMed Google Scholar
Ebbesen, C. L. & Brecht, M. Motor cortex—to act or not to act? Nat. Rev. Neurosci. 18, 694–705 (2017).
Article CAS PubMed Google Scholar
Chen, T. W., Li, N., Daie, K. & Svoboda, K. A map of anticipatory activity in mouse motor cortex. Neuron 94, 866–879.e4 (2017).
Article CAS PubMed Google Scholar
Nyakatura, J. A. Early primate evolution: insights into the functional significance of grasping from motion analyses of extant mammals. Biol. J. Linn. Soc. 127, 611–631 (2019).
Article Google Scholar
Schiel, N., Souto, A., Huber, L. & Bezerra, B. M. Hunting strategies in wild common marmosets are prey and age dependent. Am. J. Primatol. 72, 1039–1046 (2010).
Article PubMed Google Scholar
Ose, T. et al. Anatomical variability, multi-modal coordinate systems, and precision targeting in the marmoset brain. Neuroimage 250, 118965 (2022).
Article PubMed Google Scholar
Konno, A. & Hirai, H. Efficient whole brain transduction by systemic infusion of minimally purified AAV-PHP.eB. J. Neurosci. Methods 346, 108914 (2020).
Article CAS PubMed Google Scholar
Paxinos, G., Watson, C., Petrides, M., Rosa, M. & Tokuno, H. The Marmoset Brain in Stereotaxic Coordinates. (Academic Press, 2012).
Musall, S., Kaufman, M. T., Juavinett, A. L., Gluf, S. & Churchland, A. K. Single-trial neural dynamics are dominated by richly varied movements. Nat. Neurosci. 22, 1677–1686 (2019).
Article CAS PubMed PubMed Central Google Scholar
de Castro, E. & Morandi, C. Registration of translated and rotated images using finite fourier transforms. IEEE Trans. Pattern Anal. Mach. Intell. 9, 700–703 (1987).
Article PubMed Google Scholar
Tanaka, Y. H. et al. Thalamocortical axonal activity in motor cortex exhibits layer-specific dynamics during motor learning. Neuron 100, 244–258.e12 (2018).
Article CAS PubMed Google Scholar
Prsa, M., Galiñanes, G. L. & Huber, D. Rapid integration of artificial sensory feedback during operant conditioning of motor cortex neurons. Neuron 93, 929–939.e6 (2017).
Article CAS PubMed PubMed Central Google Scholar
Li, X., Zhang, G., Wu, J., Zhang, Y. & Zhao, Z. Reinforcing neuron extraction and spike inference in calcium imaging using deep self-supervised denoising. Nat. Methods 18, 1395–1400 (2021).
Article CAS PubMed Google Scholar

Download references

Acknowledgements

We thank Y. Hirayama, M. Hirokawa, Y. Takahashi, and A. Yamamoto for animal handling. We thank T. Ode for helping to build the wide field-of-view two-photon microscope system. This work was supported by Grants-in-Aid for Scientific Research on Innovative Areas (17H06309 to M. Matsuzaki, 19H05307 to T.E., and 21H00302 to T.E.), for Transformative Research Areas (A) (22H05160 to M. Matsuzaki and 23H04977 to T.E.), for Scientific Research (A) (19H01037 and 23H00388 to M. Matsuzaki), for Scientific Research (B) (20H03546 to T.E.), and for Young Scientists (A) (17H04982 to Y.M.) from the Ministry of Education, Culture, Sports, Science, and Technology, Japan; AMED (JP19dm0207069 to M. Matsuzaki; JP15dm0207001 to T.Y., M. Murayama, and M. Matsuzaki; JP18dm0207027 to M. Matsuzaki; JP19dm0107150 to M. Matsuzaki; JP19dm0207085 to T.E. and M.Matsuzaki; JP21dm0207014 to K. Ohki; JP23wm0625001 to K. Ohki and M.Matsuzaki); and the Tokyo Society of Medical Sciences (to T.E.). This work was also supported by the program for Brain Mapping by Integrated Neurotechnologies for Disease Studies (Brain/MINDS) from AMED under Grant number JP21dm0207111.

Author information

Authors and Affiliations

Department of Physiology, Graduate School of Medicine, The University of Tokyo, Tokyo, Japan
Teppei Ebina, Akitaka Sasagawa, Dokyeong Hong, Masashi Kondo, Shin-Ichiro Terada, Masato Uemura, Kenichi Ohki & Masanori Matsuzaki
Brain Functional Dynamics Collaboration Laboratory, RIKEN Center for Brain Science, Saitama, Japan
Rieko Setsuie, Keitaro Obara, Yoshito Masamizu, Katsuya Ozawa & Masanori Matsuzaki
Laboratory of Functional Brain Circuit Construction, Graduate School of Brain Science, Doshisha University, Kyoto, Japan
Yoshito Masamizu
Laboratory for Molecular Analysis of Higher Brain Function, RIKEN Center for Brain Science, Saitama, Japan
Masafumi Takaji, Akiya Watakabe & Tetsuo Yamamori
Laboratory for Haptic Perception and Cognitive Physiology, RIKEN Center for Brain Science, Saitama, Japan
Masafumi Takaji, Tetsuo Yamamori & Masanori Murayama
Institute of Innovative Research, Tokyo Institute of Technology, Kanagawa, Japan
Masafumi Takaji
Laboratory for Molecular Mechanisms of Brain Development, RIKEN Center for Brain Science, Saitama, Japan
Akiya Watakabe
Section of Viral Vector Development, National Institute for Physiological Sciences, Aichi, Japan
Kenta Kobayashi
International Research Center for Neurointelligence (WPI-IRCN), The University of Tokyo Institutes for Advanced Study, Tokyo, Japan
Kenichi Ohki & Masanori Matsuzaki
Institute for AI and Beyond, The University of Tokyo, Tokyo, Japan
Kenichi Ohki
Central Institute of Experimental Animals, Kanagawa, Japan
Tetsuo Yamamori
Department of Biological Sciences, Graduate School of Science, The University of Tokyo, Tokyo, Japan
Masanori Matsuzaki

Authors

Teppei Ebina
View author publications
You can also search for this author in PubMed Google Scholar
Akitaka Sasagawa
View author publications
You can also search for this author in PubMed Google Scholar
Dokyeong Hong
View author publications
You can also search for this author in PubMed Google Scholar
Rieko Setsuie
View author publications
You can also search for this author in PubMed Google Scholar
Keitaro Obara
View author publications
You can also search for this author in PubMed Google Scholar
Yoshito Masamizu
View author publications
You can also search for this author in PubMed Google Scholar
Masashi Kondo
View author publications
You can also search for this author in PubMed Google Scholar
Shin-Ichiro Terada
View author publications
You can also search for this author in PubMed Google Scholar
Katsuya Ozawa
View author publications
You can also search for this author in PubMed Google Scholar
Masato Uemura
View author publications
You can also search for this author in PubMed Google Scholar
Masafumi Takaji
View author publications
You can also search for this author in PubMed Google Scholar
Akiya Watakabe
View author publications
You can also search for this author in PubMed Google Scholar
Kenta Kobayashi
View author publications
You can also search for this author in PubMed Google Scholar
Kenichi Ohki
View author publications
You can also search for this author in PubMed Google Scholar
Tetsuo Yamamori
View author publications
You can also search for this author in PubMed Google Scholar
Masanori Murayama
View author publications
You can also search for this author in PubMed Google Scholar
Masanori Matsuzaki
View author publications
You can also search for this author in PubMed Google Scholar

Contributions

T.E. and M. Matasuzaki designed the experiments. T.E., A.S., R.S., K. Obara, and K. Ozawa. conducted the experiments. T.E., A.S., D.H., and M.K. analyzed the data. T.E., Y.M., and S.T. optimized the task apparatus for the experiments. Y.M., M.U., M.T., A.W., K.K., K.Ohki, and T.Y. designed and prepared AAVs. M. Murayama developed the wide field-of-view two-photon microscope system. T.E., A.S., and M. Matsuzaki wrote the paper, with comments from all authors.

Corresponding author

Correspondence to Masanori Matsuzaki.

Ethics declarations

Competing interests

The authors declare no competing interests.

Peer review

Peer review information

Nature Communications thanks Takashi Hanakawa and the other, anonymous, reviewer(s) 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

Reporting Summary

Source data

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

Ebina, T., Sasagawa, A., Hong, D. et al. Dynamics of directional motor tuning in the primate premotor and primary motor cortices during sensorimotor learning. Nat Commun 15, 7127 (2024). https://doi.org/10.1038/s41467-024-51425-3

Download citation

Received: 01 October 2023
Accepted: 05 August 2024
Published: 20 August 2024
DOI: https://doi.org/10.1038/s41467-024-51425-3