Sequential Finger-tapping Learning Mediated by the Primary Motor Cortex and Fronto-parietal Network: A Combined MRI-MRS Study

Shuki Maruyama National Institute for Physiological Sciences Masaki Fukunaga National Institute for Physiological Sciences Sho K. Sugawara National Institute for Physiological Sciences Yuki H. Hamano National Institute for Physiological Sciences Tetsuya Yamamoto National Institute for Physiological Sciences Norihiro Sadato (  sadato@nips.ac.jp ) Department of System Neuroscience, Division of Cerebral Integration, National Institute for Physiological Sciences (NIPS), 38 Nishigonaka, Myodaiji, Okazaki, Aichi 444-8585, Japan


Introduction
Motor learning refers to the acquisition of new spatiotemporal muscle activation patterns 1 . Practice is a critical factor for motor learning, which is characterized by a goal-seeking process with feedback, leading to a con gurational change in movement in terms of speed and accuracy 2 , representing the performer's attempt to reach a goal 3,4 . Thus, practice requires externally directed attention toward a goal and feedback, and internally directed attention focused on motor control. The difference between a goal and feedback is referred to as a challenge, which is crucial for motor skill learning and retention 4,5 . Challenge makes trainees exert an effort into the relevant training; for example, the sequential nger-tapping learning paradigm is frequently utilized in the neuroimaging eld [6][7][8][9][10][11] wherein participants are usually instructed to practice a given sequence "as fast and as accurately as possible." In these situations, participants must rst retrieve the sequence to conduct the practice. Speed pressure enhances the learning process because the instruction of "as fast as possible" maintains the difference between the goal and performed output; thus, the task remains challenging. In addition to motor control, this type of practice for motor learning requires exible cognitive control 12 of externally and internally directed attention.
The primary motor cortex (M1) plays an essential role in motor skills learning 1,[13][14][15] . M1 consists of the intrinsic horizontal connection network necessary to support learning-induced reorganization 1 dependent on the precise balance of excitatory and inhibitory signaling within the system. Control at the local inhibitory level is critical to enable the functional restructuring of intracortical connections, leading to a change in the M1 output map 1 . The modulation of inhibitory GABA levels in M1 by anodal transcranial direct current stimulation, as measured by magnetic resonance spectroscopy (MRS), enhances motor learning 16 . Kolasinski et al. 17 reported a dynamic reduction of GABA levels within M1 during motor skill learning using a serial reaction time task and MRS. Combining MRS and resting-state functional magnetic resonance imaging (fMRI) has permitted exploration of the relationship between learningrelated changes in the resting-state network and the formation of motor engrams in M1. A learningrelated reduction in GABA levels in M1 correlated with functional connectivity strength changes in the resting-state sensorimotor network (SMN) in long-term motor learning 18 . Baseline GABA levels in M1 are positively correlated with motor learning-related changes in resting-state functional connectivity between the bilateral M1s and between the right M1 and left superior parietal cortex 19 . These studies evaluated the relationship between GABA levels of the M1 and network changes in motor task-relevant regions.
Meanwhile, as a goal-seeking behavior, practice for motor learning requires executive control, suggesting the involvement of the fronto-parietal execution network (FPN) 20 . Herein, we hypothesized that learningrelated information during goal-seeking practice is provided by the FPN, in addition to the SMN. We measured the levels of glutamate (Glu) and GABA in the M1, as these neurotransmitters play an essential role in the neural circuits underpinning learning and memory 21,22 . We combined MRS, task fMRI, and resting-state fMRI to depict the network level changes during sequential nger-tapping learning with the non-dominant left hand, under speed-pressure, using a 7T MR machine (see Materials and Methods for details; Fig. 1). This study aimed to depict the M1-centered network for motor learning through goalseeking practice. To control the learning effect non-speci c to the sequence, another group of 13 participants underwent the identical procedure, except that they tapped 120 different sequences. Figure 2A represents an example of MR spectra within M1 obtained using the 7T MR system. The 2×2×2 cm 3 volume of interest was centered over the hand knob area of the right M1 identi ed using fMRI during a nger opposition task (red), and was superimposed on T1w MPRAGE images. To assess whether the changes in metabolite concentrations were due to uctuations in spectral quality, we evaluated the Cramer-Rao lower bounds (CRLB), linewidth, and SNR. MRS spectra provided reliable estimates of multiple metabolites, with a CRLB < 15%. A repeated-measures ANOVA revealed no signi cant main effect of time (pre-task vs. during-task vs. post-task) on CRLB, linewidth, and SNR (Table 1). Table 1 Magnetic resonance spectroscopy spectra quality of pre-task, during-task, and post-task periods Data are presented as mean ± standard deviation (SD) for n = 38.

MRS spectra
CRLB, Cramer-Rao lower bounds; Glu, glutamate; tCr, total creatine Figure 2B represents the distribution of the concentrations of GABA/tCr and Glu/tCr in the pre-, during-, and post-task periods. The variation in neurotransmitter concentration was analyzed using repeatedmeasures ANOVA with time as a factor (pre-task vs. during-task vs. post-task . Post-hoc one-sample t-tests revealed that the transition time did not signi cantly differ between blocks 4 and 5 (p = 0.389 with Bonferroni correction), indicating that performance plateaued. The relationship between the change in Glu/GABA ratio within M1 and performance improvement was evaluated using linear regression analysis. A positive correlation was observed between the change in the Glu/GABA ratio and performance improvement (r (25) = 0.42, p = 0.038) (Fig. 3B).

Linear increments in execution-related and preparation-related activity
We observed linear increments in execution-related activity in the right M1, S1, and the inferior occipital lobe with a lenient threshold (uncorrected p < 0.001 at voxel-level and FWE-corrected p < 0.05 at the cluster level; Fig. 5A). In contrast, linear increments in preparation-related activity were observed in the right M1 and S1, and the SMA. A linear increase in preparatory activity was also found in fronto-parietal regions, including the bilateral inferior parietal lobule (IPL), the MFG, STG, the thalamus, the CB lobules, the anterior cingulate cortex, and the middle cingulate cortex (FWE-corrected p < 0.05 at the cluster level; Fig. 5B).
Resting-state functional connectivity before and after motor learning The learning-related network, depicted as linear increments in preparation-related activity using task fMRI, overlapped with the FPN and SMN templates provided by the CONN toolbox ( and Glu/tCr concentration (F (2,24) = 3.014; p = 0.068) in the right M1 (Fig. 8).
As demonstrated in Figure 9A, the transition times in non-speci c learning were 466.960 ± 97.  In non-speci c learning, no signi cant effect was observed in the linear increment in preparation-related activity in right M1 (T-score = 1.91, un-corrected p = 0.080).

Discussion
In this study, MRS, task fMRI, and resting-state fMRI methods were combined to assess network-level changes during motor sequence learning using a 7T MR machine. This study replicated and extended previous ndings regarding the crucial role of M1 in motor sequence learning. To the best of our knowledge, this is the rst report to demonstrate that the local excitatory-inhibitory balance within M1 regulates M1 connectivity with the FPN.
GABA and Glu measurements were of high quality and reproducibility (Table 1). Although the neural excitatory-inhibitory balance is crucial for learning and memory, the main focus of the prior studies on motor learning has been limited to the evaluation of GABA 16,18,19,23,24 due to technical limitations. The subtraction of two independent spectra to remove the overlap of signals is required in 3T MRS.
Conversely, 7T MRS is able to concurrently resolve GABA, Glu, and glutamine (Gln), as sensitivity and chemical shift dispersion increase with increasing magnetic eld strength 25 . A higher SNR that increases linearly with the magnetic eld strength enables more accurate detection of weak signals from neurotransmitters in smaller voxels and with shorter measurement times 26 . Neurotransmitters were measured within M1 using a voxel size of 8 cm 3 (2×2×2) at 7T in this study; however, a voxel size of 27 cm 3 (3×3×3) was selected at 3T 27,28 . The 3T MRS is relatively insensitive to subtle changes in neurotransmitters underscoring cognitive functions due to large MRS voxel sizes 29 . Thus, a 7T MRS is advantageous over a 3T MRS in terms of observing neurotransmitter function in a speci cally localized brain region related to changes in cognitive and behavioral task performance.
We also measured GABA and Glu levels within M1 at pre-training and post-training resting-state conditions, and during task execution using a 7T MR system. We found signi cant reductions in Glu (p < 0.05, repeated-measures ANOVA and t-test with Bonferroni correction) between pre-task and post-task periods and between during-task and post-task periods (Fig. 2B). The decrease in Glu probably re ects the decrease in synaptic Glu or glutamatergic cycling as a part of energy metabolism in the tricarboxylic acid cycle 30 . These ndings point to a learning-related decrease in Glu within M1 in motor sequence learning. A previous study using 7T machine showed no change in Glu within M1 during motor sequence learning using a serial reaction time task 17 . It is important to point out that, instead of implicit learning which mainly involved in the M1 31 , we adopted the explicit motor sequence learning, which is known to recruit global brain network 9,32 . The motor engram was shown to be generated in the parietal regions distant from M1 during the explicit learning with the instruction of "tap the sequence as fast and correct as possible" (maximum mode), whereas generated in the M1 and dorsal premotor cortex during the implicit learning through visually guided constant speed execution (constant mode) 9 . Glu is known to exhibit a global effect on the BOLD response via glutamatergic projections to other cortical regions rather than modulating the BOLD response within the acquired MRS voxel 33,34 . Based on these ndings, the decrease in Glu is assumed to likely be related to the sequence learning-speci c recruitment of the global brain network.
With regard to GABA, although previous studies showed GABA reduction in M1 during motor learning 17,23 , no signi cant difference in GABA (Fig. 2B). Our results are in line with the recent study using similar motor sequential tasks 19 . The GABA measured using the MRS thought to re ect bulk GABA from a large volume of interest, and is thought to re ect cellular, rather than synaptic GABA levels predominantly 24,35 .
No correlation was observed between GABA with MRS and phasic GABA signaling using TMS 36,37 .
Although a signi cant correlation between GABA with MRS and tonic GABA was observed in one study 36 , no correlation was observed in a recent study 37 . The main factor of this difference in the two studies could be measurement methods of MRS: scan sequence (SPECIAL vs. STEAM) and magnetic eld strength of the MR system (3T vs. 7T). That is, there seems to be no consensus on the relationship between the GABA with MRS and tonic GABA.
We found that GABA changes were signi cantly correlated with the performance improvement (p = 0.018), consistent with the previous nding 19 . The disinhibition is to enhance Glu-related excitatory processes resulting in the decline of Glu concentration. The M1 comprises the intrinsic horizontal connection network necessary to support circuit reorganization during learning dependent on the precise balance of excitatory and inhibitory signaling within the M1 networks 1 . We adopted the Glu/GABA ratio to account for behavioral performance changes, as the stability of cortical areas during learning depends on the balance between cortical excitation and inhibition 38 . Further, considering that Glu is the precursor of GABA, their concentrations are likely to be reciprocally dependent. These factors indicate that the Glu/GABA ratio corresponds to cortical excitability 37 and is a more sensitive proxy for plasticity than Glu or GABA alone. We observed a positive correlation between changes in the Glu/GABA ratio and task performance improvement (Fig. 3B). This nding suggests that between-participant variation in the balance of GABA and Glu re ects improvements in motor sequence learning performance.
We observed that preparation-related activity increased linearly in fronto-parietal regions, especially in the right M1 (Fig. 5B). This result is consistent with that of our previous study 39 . In explicit motor sequence learning, participants needed to retrieve whole-sequence information at the preparation phases internally.
Electrophysiological studies in nonhuman primates demonstrated an increase in neuronal responses re ecting preparatory activity for movement in M1 as learning progressed 40 . Thus, the increase in preparation-related activity represents motor learning as an ecphoric process without being confounded by motor execution effects dependent on speed [41][42][43] and force 44 . The motor learning-related information of the speci c sequence was accumulated in M1 because no such signi cant effect was observed in the M1 of the control group which conducted the sequential nger tapping with 120 different sequences. In addition, an increment of the preparatory activity was highly present in regions included SMN and FPN, thereby suggesting that the learning-related information is distributed in networks associated with both motor and executive controls.
We also assessed resting-state M1 seed-based functional connectivity change elicited by motor sequence learning. As shown in Figure 6C, a positive correlation was observed between changes in the Glu/GABA ratio within M1 and M1 seed-based resting-state functional connectivity changes in FPN. In contrast, no correlation was found in the SMN. These results re ect the learning effect during motor sequence learning because no such correlation was observed in the control group (Fig. 10). The FPN controls coordinated behavior in a rapid, accurate, and exible goal-driven manner 12 . Therefore, this nding indicates that motor learning driven by cognitive control is associated with local changes in excitatoryinhibitory balance in the M1. As described above, these ndings re ect individual differences in skills, effort, and concentration of self-paced movement because participants were required to execute the task as quickly as possible during learning.
To further probe the relationship between M1 and FPN, we assessed the correlations of connectivity changes in the bilateral PFC and PPC with changes in the Glu/GABA ratio of the M1. These correlations were more prominent in parietal regions than in frontal regions, thereby suggesting that the Glu/GABA ratio of the M1 is more likely to affect the connectivity with the PPC in FPN (Fig. 7). This nding concurs with the notion that the PPC is necessary for early and late learning phases, whereas the PFC is primarily involved in early learning phases 45 . The PFC processes sensory inputs, motor outputs, and working memory [46][47][48] . The PPC, encompassing the IPL and SPL, processes spatial-sequential components 31,49 . Both the M1 and PPC are critical hubs for the late motor sequence learning phase because these areas contribute to the delayed recall of learned motor sequences 50,51 . In other words, in the later phase of learning, PPC and M1 are involved in retrieving the learned sequences acquired during the early learning phase. Our results, combined with our previous data, suggest that M1 integrates the accumulated information processed by the PPC in motor sequence learning.
These ndings are consistent with those of Sami et al. 52 , who investigated the consolidation effects on the resting state network using dual regression Independent Component Analysis (ICA) analysis following implicit and explicit learning, with serial reaction time task. The authors had demonstrated the role of FPN in the explicit learning group, six hours following the initial acquisition, and have interpreted this nding as bringing the learned sequence back to declarative awareness. Furthermore, they directly compared explicit and implicit groups at this late state, thereby identifying bilateral activation in both the parietal and premotor regions. The authors also speculated that this network might represent an engram of the extra procedural learning skill that had developed in the explicit acquisition group 52 . Therefore, we conclude that the M1 centered network with FPN represents the formation of declarative procedural skills.
As represented in Figure 4, we observed similar spatial patterns of activity in the execution and preparation phases. These areas represent the large-scale functional motor network, necessary for performing sequential motor tasks. The selection of a particular motor sequence is based on inputs from the prefrontal cortex and parietal-temporal regions to the ventral premotor cortex (PMv) 53,54 . The dorsal part of the IPL (dIPL) is a multimodal sensory association region involved in the initial acquisition and learning of a motor task. The anterior parts of the IPL, PMv, and M1 consist of the ne motor control network [54][55][56][57] . The dorsal premotor cortex (PMd) is involved in movement selection 58 . In addition, preparation-related activity was most prominently associated with enhanced activity in the putamen (Fig.   4B), suggesting that this preparatory activity represents preceding self-initiated movements 59 . Our ndings are consistent with previous results demonstrating preparatory activity in the motor, somatosensory, parietal, and prefrontal cortical regions, basal ganglia, and cerebellum in sequential nger movements 60 .
The participants recruited in this study were predominantly women, with body weights of 60 kg or less. This limitation contributed to technical challenges in MRS measurements using a single-transmit 7T MR system. First, the B1 transmit eld inhomogeneity was enhanced. The suppression of water signals for the measurement of metabolites may have been insu cient depending on the head size, and it was challenging to obtain good spectral quality. Second, adjustments of MRS sequence parameters may have been necessary, involving a lengthening of measurement time to solve the local speci c absorption rate limitations partly de ned using body weight. Gender differences are known to affect visuo-motor adaptation learning of throwing 61 ; given that the participants in this study were primarily women, the generalizability of the results remain limited, and further studies are warranted in which the number of men is high or at least equivalent to that of women.
In conclusion, our ndings indicate that motor learning driven by cognitive control is associated with local variation in the excitatory-inhibitory balance in M1 that regulates remote connectivity with the FPN, constituting the M1-centered motor learning network.

Materials And Methods
Participants A total of 43 healthy, right-handed adult volunteers participated in the study (7 male and 36 female: mean age (± SD) was 22.9 ± 4.4 years). Handedness was assessed using the Edinburgh Handedness Inventory 62 . None of the participants had a history of neurological or psychiatric diseases. All participants provided written informed consent for participation in the experiment. The study was conducted according to the Declaration of Helsinki and was approved by the Ethical Committee of the National Institute for Physiological Sciences, Japan.

Experimental design
We carried out MRS-fMRI experiments using a 7T MRI scanner (MAGNETOM 7T, Siemens Healthineers, Erlangen, Germany) with a 32-channel receiving head coil and a single-channel transmitting coil (Nova Medical Inc., MA, USA). All participants underwent resting-state fMRI and MRS scans before and after the motor sequence learning tasks, as well as one MRS and four fMRI scans during motor sequence learning tasks in the task session (Fig. 1A). Dielectric pads (CaTiO3) 63 were placed around each participant's head while scanning at 7T to improve the B1 transmit eld inhomogeneity. All scans were performed within the Speci c Absorption Rate (SAR) limit of the normal operation mode.

Motor sequence learning task
Thirty participants were asked to perform pre-determined ve-digit sequences "4-1-3-2-4" (n = 17) or "2-3-1-4-2" (n = 13) as quickly and accurately as possible in the MRI scanner (Fig. 1B) [9][10][11] . Additionally, 13 participants were asked to perform 120 different sequences to assess the non-speci c learning as the control condition. The sequence "4-1-3-2-4" corresponds to "index-little-middle-ring-index." The motor sequence task consisted of six 30-s tapping epochs followed by 30-s rest epochs that were repeated ve times (Fig. 1B). The visual feedback signals were displayed using a projector (Optoma EH503; Optoma Inc., Fremont, CA, USA) with a lens (APO 50-500 mm F4.5-6.3 DG OS HSM; SIGMA, Kanagawa, Japan) on a screen viewed by the participants via a mirror mounted to the receiving head coil. Response time was measured using Presentation software version 16.4 (Neurobehavioral Systems, NY, USA; RRID: SCR_002521). The rest epoch started with the appearance of the instruction "Rest" on the screen for 500 ms, followed by a 500-ms presentation of four blue circles aligned within an equally spaced horizontal array. The instruction "Task" appeared for 2 s at the end of the rest epoch as a signal to the participants to retrieve motor sequences and prepare for their execution (Fig. 1B) fMRI data acquisition fMRI images were acquired before, during, and after the motor sequence learning tasks using a multiband gradient-echo echo-planar imaging sequence 66 . The scan parameters were set as per the human connectome project (HCP) 7T protocol 67 (TR/TE = 1,000/22.2 ms; eld of view = 208×208 mm 2 ; matrix size = 130×130; slice thickness = 1.6 mm; 85 slices; multi-band/GRAPPA acceleration factor = 5/2; bandwidth = 1,924 Hz/Px; ip angle = 45°). The spin echo eld map was acquired 68 (TR/TE = 3,000/60 ms; eld of view = 208×208 mm 2 ; matrix size = 130×130; slice thickness = 1.6 mm; 85 slices; multiband/GRAPPA acceleration factor = 5/2; bandwidth = 1,924 Hz/Px; ip angle = 180°; acquisition time = 1 min 26 s). A B1 transmit eld map in the center of the brain, around the slice of the M1 hand knob area, was acquired for each participant to optimize the input power for accurately producing a 90° pulse for all fMRI scans. In particular, participants were instructed to keep their eyes open while viewing a xation cross and to avoid having any speci c thoughts or falling asleep during resting-state fMRI scans.

MRS data acquisition
A 2×2×2 cm 3 volume of interest was centered over the right M1 hand knob area ( Fig. 2A), without dura, on T1w MPRAGE images. The hand knob area was identi ed using fMRI during a sequential nger opposition task with the left hand (TR/TE = 1,000/24 ms; eld of view = 192×192 mm 2 ; matrix = 96×96; slice thickness = 2 mm; 20 slices; GRAPPA acceleration factor = 2; bandwidth = 2,170 Hz/Px; ip angle = 45°; acquisition time = 3 min 30 s). Ultra-short TE MRS data were acquired before, during, and after the motor sequence learning task using the STimulated Echo Acquisition Mode (STEAM) sequence (TR/TE = 5,000/5.68 ms; mixing time = 40 ms; vector size = 2,048; bandwidth = 4,000 Hz/Px; average = 64) with VAriable Power RF pulses with Optimized Relaxation delays (VAPOR) water suppression 25,69 . The STEAM sequence was combined with outer volume suppression to improve localization performance. A 4average water reference signal was acquired for eddy current correction and absolute quanti cation of the metabolites. Before data acquisition, all rst-and second-order shim terms were automatically adjusted with the fast automatic shim technique using echo-planar signal readout for mapping along with projections (FASTMAP) 70,71 . In addition, B1 transmit eld strength for localization pulses and VAPOR water suppression was adjusted for individual participants.

MRS data analysis
Raw MRS data were post-processed using MATLAB R2018a (The MathWorks, Inc., MA, USA; RRID: SCR_001622). Motion-corrupted data were removed to improve the spectral quality. To quantify the proportion of gray matter (GM), white matter (WM), and cerebrospinal uid (CSF) fractions in the volume of interest, segmentation in SPM was applied to the T1w MPRAGE images. Eddy current correction and frequency correction were performed using a water reference scan, and the zero and rst-order phases of the array coil were aligned using the cross-correlation method of MRspa (RRID: SCR_017292).
Subsequently, LCModel version 6.3-1N 75,76 (Stephen Provencher, Inc., ON, Canada; RRID: SCR_014455) analysis was used to quantify the concentration of neurochemicals within the chemical shift range of 0.5 to 4.1 ppm 76 . Other parameters in the LCModel were as reported previously 77 . The concentrations of GABA and Glu were normalized to that of total creatine (tCr). The change in glutamate to GABA ratio (Glu/GABA) after the motor sequence learning task was calculated using the following equation: where Glu/GABApre and Glu/GABApost indicate the Glu/GABA ratio at pre-task and post-task, respectively. The distribution of GABA and Glu concentrations was visualized using the RainCloudPlots Python-script 78 (https://github.com/RainCloudPlots/RainCloudPlots). A repeated-measures ANOVA was carried out in SPSS, with the concentrations of GABA and Glu at different time points (pre-task, duringtask, and post-task) as a factor. The Cramer-Rao lower bounds (CRLB), water linewidth at FWHM, and signal-to-noise ratio (SNR) were used for the quality control of spectra 76 . The CRLB and SNR were calculated using LCModel, and water linewidth was obtained by tting to the additional water spectrum using MATLAB. Data were excluded when CRLB > 15 % (n = 1), linewidth > 19 Hz (n = 1), or SNR < 30. A repeated-measures ANOVA was performed on the CRLB, water linewidth, and SNR time points (pre-task, during-task, and post-task) with a within-subjects factor using SPSS. fMRI preprocessing All fMRI data were processed using the functional pipeline (fMRIVolume) of the minimal HCP preprocessing pipeline 72 . This pipeline included the following steps: motion correction, gradient magnetic eld nonlinearity distortion correction, eld map-based distortion correction (Topup) 68 , nonlinear registration into 3T MNI structure data, and grand-mean intensity normalization. Finally, volume-based smoothing with a 5-mm full width at half maximum (FWHM) Gaussian kernel was applied.
Task fMRI data analysis Task fMRI data analysis was performed using Statistical Parametric Mapping (SPM12; RRID: SCR_007037) in MATLAB R2018a. A general linear model (GLM) was tted to the fMRI data for each participant 79,80 . The fMRI time series for preparation phases 2 s before task execution and execution phases were modeled with boxcar functions convolved with the canonical hemodynamic response function. Each block consisted of six execution-related and preparation-related regressors. The design orthogonality between the execution and preparation phases was −0.0137 ± 0.054 for block 1, −0.0141 ± 0.054 for block 2, −0.0137 ± 0.054 for block 3, and −0.0139 ± 0.054 for block 4 (mean ± SD). Temporal high-pass ltering with a cutoff frequency of 1/128 Hz was applied. Using a rst-order autoregressive model, the serial autocorrelation was estimated from the pooled active voxels with the restricted maximum likelihood procedure and subsequently used to whiten the data 81 . Several nuisance covariates, including six head motion parameters and CSF time-series, were incorporated into the model. The parameter estimates for each execution-related and preparation-related regressors were assessed using constant and prede ned linear contrasts. Increasing contrast vectors were de ned numerically as an increment of one per block, maintaining the mean equal to zero.
For group-level analysis of task fMRI data, one-sample t-tests of participants' contrast images were performed 82 . The resulting set of voxel values for each contrast constituted the SPM{t}. We calculated the T-score of linear increment in preparation-related activity in right M1 in non-speci c learning. The statistical threshold was set at p < 0.05, FWE-corrected at the voxel-level 83 , unless otherwise speci ed.

Anatomical labeling and visualization
MRIcron (RRID: SCR_008264) was used to display fMRI activation maps on a standard brain image. The Automated Anatomical Labeling atlas was used for anatomical labeling 84 .

Resting-state fMRI data analysis
Resting-state functional connectivity analysis was conducted using the CONN toolbox version 17 in SPM12 85 (RRID: SCR_009550). An anatomical component-based noise correction method (aCompCor) 86 was applied to remove the ve components of signals from WM, CSF, and residual head motion-related signals through linear regression. A temporal bandpass ltering of 0.008-0.090 Hz was applied.
Seed-to-voxel correlation analysis was performed at the individual level. We selected the preparationrelated increased voxels in M1 (MNI: x = 36, y = −25, z = 51), determined in the second-level analysis of task fMRI (FWE voxel-level corrected p < 0.05), as a seed region of interest (ROI) (Fig. 6A). An individual seed-based functional connectivity map was obtained by computing Pearson's correlation coe cients between the time-series from the M1 seed ROI and the time-series of all other voxels across the whole brain. Fisher's r-to-z transformation was used to convert the correlation coe cients into z-scores. M1seeded functional connectivity changes were integrated using the following equation using AFNI version 18.1.32. (RRID: SCR_005927): where Connectivity pre and Connectivity post are the pre-task and post-task functional connectivity values, respectively.
We calculated the changes in functional connectivity within ROIs of the SMN and FPN de ned from CONN's ICA analyses of the HCP dataset of 497 individuals. The SMN includes the supplementary motor cortex and bilateral sensorimotor cortex, whereas the FPN consists of the bilateral LPFC and PPC. The correlations between Glu/GABA changes within M1 and M1 seed-based functional connectivity changes were analyzed using linear regression analysis.

Declarations Data availability
The datasets generated in this study are available from the corresponding author on reasonable request.