Skip to main content
eLife logoLink to eLife
. 2020 May 13;9:e55241. doi: 10.7554/eLife.55241

A critical re-evaluation of fMRI signatures of motor sequence learning

Eva Berlot 1, Nicola J Popp 1, Jörn Diedrichsen 1,2,3,
Editors: Marius V Peelen4, Floris P de Lange5
PMCID: PMC7266617  PMID: 32401193

Abstract

Despite numerous studies, there is little agreement about what brain changes accompany motor sequence learning, partly because of a general publication bias that favors novel results. We therefore decided to systematically reinvestigate proposed functional magnetic resonance imaging correlates of motor learning in a preregistered longitudinal study with four scanning sessions over 5 weeks of training. Activation decreased more for trained than untrained sequences in premotor and parietal areas, without any evidence of learning-related activation increases. Premotor and parietal regions also exhibited changes in the fine-grained, sequence-specific activation patterns early in learning, which stabilized later. No changes were observed in the primary motor cortex (M1). Overall, our study provides evidence that human motor sequence learning occurs outside of M1. Furthermore, it shows that we cannot expect to find activity increases as an indicator for learning, making subtle changes in activity patterns across weeks the most promising fMRI correlate of training-induced plasticity.

Research organism: Human

eLife digest

It has famously been claimed that it takes 10,000 hours to become an expert at something. But while most of us will never become concert pianists, we can all learn new motor skills and improve existing ones – from touch-typing to tennis – by practicing. What happens in the brain to produce these improvements in performance?

Researchers have tried to answer this question by scanning the brains of people as they practice motor skills, but the results have proved inconsistent. Some studies find that specific brain areas become more active as people practice. This could indicate that these areas are ‘storing’ new skills. But others report that brain activity decreases with practice. This might indicate that practice instead makes certain brain areas work more efficiently. It is also unclear where in the brain these learning-related changes occur. Some studies suggest that most occur in the primary motor cortex, or M1 – the area that sends commands to muscles. Others suggest that most changes take place outside of M1, in areas that plan movements.

Berlot et al. set out to resolve these inconsistencies by scanning the brains of healthy volunteers as they learned to play six 9-digit sequences on a keyboard. Each volunteer completed about 4,000 training trials over 5 weeks, and had their brain scanned four times. As the weeks passed, the volunteers became faster and more accurate at playing the sequences. However, the activity of their primary motor cortex did not change. By contrast, the activity of areas involved in planning movements decreased throughout training. The patterns of activity for each individual sequence reorganized throughout learning in the areas outside of the M1. This happened most quickly during the early stages of training when the volunteers showed the fastest improvements in performance.

Overall, these findings suggest that when we learn a new skill, activity in the brain areas supporting that skill decrease as the brain becomes more efficient. Increases in brain activity are thus unlikely to reflect the acquired skill. Instead, more subtle changes, in which the brain uses more specific patterns of activity to encode the skill, may underlie improved performance. This may also be true for other types of learning, such as acquiring a new language.

Introduction

Humans have the remarkable ability to learn complex sequences of movements. While behavioural improvements in sequence learning tasks are easily observable, the underlying neural processes remain elusive. Understanding the neural underpinnings of motor sequence learning could provide clues about more general mechanisms of plasticity in the brain. This motivation has led numerous functional magnetic resonance imaging (fMRI) studies to investigate the brain changes related to motor sequence learning. However, there is little agreement about how and where in the brain learning-related changes are observable. Previous studies include reports of signal increases across various brain regions (Floyer-Lea and Matthews, 2005; Grafton et al., 1995; Hazeltine et al., 1997; Karni et al., 1995; Lehéricy et al., 2005; Penhune and Doyon, 2002), as well as signal decreases (Jenkins et al., 1994; Peters et al., 2017; Toni et al., 1998; Ungerleider et al., 2002; Wiestler and Diedrichsen, 2013), nonlinear changes in activation (Ma et al., 2010; Xiong et al., 2009), spatial shifts in activity (Lehéricy et al., 2006; Steele and Penhune, 2010), changes in multivariate patterns (Wiestler and Diedrichsen, 2013; Wymbs and Grafton, 2015), and changes in inter-regional functional connectivity (Bassett et al., 2015; Bassett et al., 2011; Doyon et al., 2002; Mattar et al., 2016). Additionally, some experiments have matched the speed of performance (Karni et al., 1995; Penhune and Doyon, 2002; Steele and Penhune, 2010; Lehéricy et al., 2005; Seidler et al., 2002; Seidler et al., 2005), while others have not (Bassett et al., 2015; Lutz et al., 2004; Wiestler and Diedrichsen, 2013; Wymbs and Grafton, 2015). Given that fMRI analysis has many degrees of freedom, these inconsistencies may not be too surprising. However, the implicit pressure in the publication system to report findings may also have contributed to a lack of coherency. To address this issue, we designed a comprehensive longitudinal study of motor sequence learning that allowed us to systematically reinvestigate previous findings. In order to increase transparency, we pre-registered the design, as well as all tested hypotheses on the Open Science Framework (Berlot et al., 2017; https://osf.io/etnqc), and make the full dataset available to the research community (Berlot, 2020; OpenNeuro accession number ds002776).

The main aim of our study was to systematically evaluate different ideas of how learning-related changes are reflected in the fMRI signal. In the context of motor sequence learning, the most commonly examined brain region is the primary motor cortex (M1). Previous reports of increased M1 activation after long-term learning have been interpreted as additional recruitment of neuronal resources for trained behavior, taken to suggest the skill is represented in M1 (Floyer-Lea and Matthews, 2005; Karni et al., 1995; Karni et al., 1998; Lehéricy et al., 2005; Penhune and Doyon, 2002; for a review see Dayan and Cohen, 2011; Figure 1a). Since then, several pieces of evidence have suggested that sequence-specific memory may not reside in M1 (Beukema et al., 2019; Wiestler and Diedrichsen, 2013; Yokoi and Diedrichsen, 2019). However, some of these reports studied skill acquisition over a course of a few days, while human skill typically evolves over weeks (and months) of practice. Therefore, including several weeks of practice might be more suitable to test whether, and at what time point, M1 develops skill-specific representations.

Figure 1. Potential fMRI signatures of learning in a specific brain area.

Figure 1.

Each panel shows hypothetical activation for the six trained sequences (red) and the six untrained sequences (blue) in the space of two hypothetical voxels. (a) Activation could increase during learning across voxels, indicating additional recruitment of resources involved in skilled behavior. (b) Activation could decrease across voxels, implying that the region performs its function more efficiently. (c) Some voxels (x-axis) could increase activation with training, while others (y-axis) could decrease. This would lead to a shift of the overall activity pattern in the region without an overall net change in activation. (d) Activation patterns specific to each trained sequence could undergo more change than untrained sequences, reflective of plastic reorganization of the sequence representation. Arrow length in the figure indicates the amount of reorganization. (e) One specific form of such reorganization would be increasing dissimilarities (pattern separation) between activity patterns for individual trained sequences.

Outside of M1, learning-related activation changes have been reported in premotor and parietal areas (Grafton et al., 2002; Hardwick et al., 2013; Honda et al., 1998; Penhune and Doyon, 2002; Tamás Kincses et al., 2008; Vahdat et al., 2015), with activation increases commonly interpreted as increased involvement of these areas in the skilled behavior. Yet, recent studies have mostly found that, as the motor skill develops, activation in these areas predominantly decreases (Penhune and Steele, 2012; Wiestler and Diedrichsen, 2013; Wu et al., 2004). Such reductions are harder to interpret as they could reflect a reduced areal involvement in skilled performance or, alternatively, more energy efficient implementation of the same function (Figure 1b; Picard et al., 2013; Poldrack et al., 2005). To complicate things further, regional activity increases and decreases could occur simultaneously in the same area (Figure 1c; Steele and Penhune, 2010). In such a scenario, the net activation in the region would not change, yet, the trained sequences would engage slightly different subpopulations of the region than untrained sequences.

A variant of this idea is that each specific sequence becomes associated with dedicated neuronal subpopulation (and hence fMRI activity pattern). Such a representation would form the neural correlate of sequence-specific learning – the part of the skill that does not generalize to novel, untrained motor sequences (Karni et al., 1995). Sequence-specific activation patterns should change early in learning (Figure 1d), when behavior improves most rapidly, and stabilize later, once the skill has consolidated and an optimal pattern is established (Peters et al., 2017). One possible way in which sequence-specific patterns could reorganize is by becoming more distinct from one another (Figure 1e; Wiestler and Diedrichsen, 2013). Having a distinctive code for each sequence might be of particular importance to the system in a trained state, allowing it to produce different dynamical sequences, while avoiding confusion or ‘tangling’ of the different neural trajectories (Russo et al., 2018).

To systematically examine the cortical changes associated with motor sequence learning, we carried out a longitudinal study over 5 weeks of training with 4 sessions of high-field (7 Tesla) fMRI scans. Behavioural performance in the first three scanning sessions was imposed to the same speed of performance. This allowed us to inspect whether examined fMRI metrics reflect brain reorganization, independent of behavioral change. However, controlling for speed incurs the danger of not tapping into neural resources that are necessary for skilled performance (Orban et al., 2010; Poldrack, 2000). We therefore compared the fMRI session with paced performance at the end of behavioural training with one acquired with full speed performance (Figure 2). This manipulation allowed us to systematically assess the role of speed on the fMRI metrics of learning, thereby addressing an important methodological problem faced by virtually every study on motor learning.

Figure 2. Experimental design and paradigm.

(a) Apparatus and task. Participants were trained to perform six 9-item sequences on a keyboard device. For each finger press, the corresponding digit on the screen turned green (correct) or red (incorrect). During fMRI scans 1–3, an expanding pink line under the numbers indicated the pace at which participants had to press the keys. See Figure 2—figure supplement 1 for trial structure during scanning sessions. (b) Training protocol lasted for 5 weeks, and included four behavioral test sessions (yellow underlay) and four scans (grey underlay). Scans 1–3 were performed at a paced speed, while scan 4 performance was full speed (fs). (c) Average group performance executing trained sequences across the training sessions, measured in seconds. The average movement time (MT) decreased with learning. Shaded area denotes between-subject standard error. (d) Performance during scanning sessions and behavioral tests, measured in seconds. Performance of trained sequences improved across all subsequent behavioral test sessions. Performance improved also for untrained sequences from week 2 onwards, suggesting some transfer in learning, but performance was still faster for trained sequences, indicating sequence-specific learning. Error bars indicate between-subject standard error. Stars denote significance levels lower than p<0.001.

Figure 2.

Figure 2—figure supplement 1. Experimental trial structure during scanning sessions.

Figure 2—figure supplement 1.

Each trial consisted of a preparation period, execution period and inter-trial-interval (ITI), during which the feedback was presented on correctness of the trial. Each sequence was presented twice in a row. Periods of rest were added in-between the trials.

Results

Speed of sequence execution increases with learning

We trained 26 participants to perform six 9-digit sequences with their right hand on a keyboard device (Figure 2a). During training, they received visual feedback (green for correct and red for incorrect presses) and were rewarded for both accuracy and speed (see Materials and methods). Over the course of 5 weeks, participants practiced ~4000 trials (Figure 2b). This led to substantial performance improvement, with the average movement time (MT) to complete a sequence decreasing from an initial 3.2 s to 1.2 s at the end of the training (Figure 2c). The training regime was complemented with behavioral assessments on four occasions designed to specifically assess participants’ performance on trained sequences relative to untrained sequences (Figure 2d, yellow underlay). Prior to training (test day 1), the speed of sequence execution did not differ between trained and untrained sequences. For all subsequent sessions, MTs were significantly faster for trained than untrained sequences (p<0.001), implying sequence-specific learning. Additionally, performance of trained sequences improved between all subsequent sessions, even after week 3 (week 3–5: t(25)=5.49, p=1.1e−5). Thus, participants’ performance of trained sequences improved across the five weeks.

To assess fMRI changes with learning, participants underwent four fMRI scans (1st scan: before the main training; 2nd scan: week 2; 3rd and 4th scan: week 5), performing both trained and untrained sequences (Figure 2d – grey underlay). The first session served as a baseline, where all of the presented sequences (trained and untrained) were novel to participants. Both trained and untrained sequences were always cued by presenting the corresponding digits on the screen (Figure 2a). During the first three sessions, participants were paced with a metronome so that all sequences, trained and untrained, were performed at the same speed as in the first scan. Performance in the fourth session was at maximum speed, resulting in significantly lower MTs for trained compared to untrained sequences (Figure 2d). To assess different neural signatures of observed behavioral learning, we first examined how the overall evoked activation changed over weeks of training for the same speed of movement.

Overall activation does not change in M1

First, we re-investigated the classical finding that activity, measured as the percent BOLD signal change relative to rest, increased in M1 for matched performance after long-term training (Karni et al., 1995; Figure 1a). Our task elicited activation in a range of cortical areas (Figure 3a for session 1 – that is, prior to learning). A region of interest (ROI) analysis of the hand area of M1, contralateral to the performing hand, however, showed no significant change across weeks (F(2,50)=0.44, p=0.85). There was a significant main effect of sequence type (F(1,25)=6.32, p=0.019), but none of the post-hoc t-tests revealed a significant difference. Additionally, the interaction between the two factors was also not significant (F(2,50)=0.17, p=0.84).

Figure 3. Overall activation and changes with learning in defined regions of interest.

(a) Average activation during production of any sequence in scanning session 1 (prior to learning) in the hemisphere contralateral to the performing hand. Activation was contrasted against resting baseline. On the right, activation map is presented on a flattened surface, corresponding to surface maps in other figures. (b) Changes in activation across predefined areas – primary motor cortex (M1), primary somatosensory cortex (S1), premotor dorsal area (PMd) and superior parietal lobule – anterior (SPLa). No significant changes in activation were observed in M1 or S1 across weeks or between trained and untrained sequences (* indicates p<0.01). Error bars indicate between-subject standard error. See Figure 3—figure supplement 1a for results with error trials excluded. (c) The cosine angle dissimilarity between average trained and untrained sequence across scanning weeks. The cosine angle increased significantly across weeks in PMd, SPLa and S1, but not M1 (* indicates p<0.05). Error bars indicate between-subject standard error. See Figure 3—figure supplement 1b for results with error trials excluded.

Figure 3.

Figure 3—figure supplement 1. Changes in activation and mean pattern dissimilarity across predefined areas, estimated only on correct trials.

Figure 3—figure supplement 1.

(a) Changes in activation across predefined areas – primary motor cortex (M1), primary somatosensory cortex (S1), premotor dorsal area (PMd) and superior parietal lobule – anterior (SPLa). As in the analysis with all trials included, no significant changes in activation were observed in M1 or S1 across weeks or between trained and untrained sequences (* indicates p<0.01). Error bars indicate between-subject standard error. (b) The cosine angle dissimilarity between average trained and untrained sequence across scanning weeks. The cosine angle increased significantly across weeks in PMd, and SPLa, but not M1. In S1, average dissimilarity estimated on correct trials only increased significantly across weeks 2–5, but not 1–2 (* indicates p<0.05). Error bars indicate between-subject standard error.

The absence of overall activity changes, however, should not be taken as evidence for an absence of plasticity in the region. It is possible that some subregions of M1 increased in activation for learned sequences, while other decreased, as suggested by Steele and Penhune, 2010. Such mixed changes would result in a shift of the overall pattern, which would lead to an increase in the angle between the mean activity pattern for trained and untrained sequences (Figure 1c). Because we calculated the angle between activity patterns for each participant separately, this criterion does not assume that the observed shift is spatially consistent across individuals – any idiosyncratic shift could be detected. Therefore it serves as a sensitive statistical criterion to detect shifts in spatial location of activation, which were previously reported only descriptively (Steele and Penhune, 2010).

However, in M1, the averaged cosine angle (Figure 3c) remained unchanged across the weeks (F(2,50)=1.71, p=0.19), indicating that the average activity pattern remained comparable across trained and untrained sequences. In sum, we found no evidence for activation increases (Karni et al., 1995), decreases, or relative shifts in activation patterns (Steele and Penhune, 2010) in M1.

Learning-related activation changes in premotor and parietal areas

To investigate activation changes in areas outside of M1, we calculated changes in activity between the weeks in a map-wise approach (Figure 4a). Over the three measurement time points, we found no reliable activation increases in any cortical area that was activated by the task in week 1. Instead, we observed widespread learning-related reductions in activity in premotor and parietal areas (Figure 4a), in line with our pre-registered prediction. These activation reductions were observed across both subsequent sessions (i.e. weeks 1–2, weeks 2–5) for trained and untrained sequences, with bigger reductions for trained sequences. In weeks 2 and 5, trained sequences elicited overall lower activity than untrained sequences (Figure 4b; see Figure 4—figure supplement 1 for statistical maps). These learning-related reductions in activity were also statistically significant in our predefined ROIs in premotor (dorsal premotor cortex – PMd) and parietal cortices (anterior superior parietal lobule – SPLa) (Figure 3b): In a 3 (week) x 2 (sequence type) ANOVA on observed activation both main effects and interaction were highly significant in PMd (week: F(2,50)=17.47, p=1.77e−6; sequence type: F(1,25)=11.86, p=2.03e−3; interaction: F(2,50)=13.22, p=2.46e−5) as well as in SPLa (week: F(2,50)=19.14, p=6.73e-7; sequence type: F(1,25)=19.36, p=1.77e−4; interaction: F(2,50)=21.59, p=1.74e−7). In contrast, no main effect of week was observed in S1 (F(2,50)=1.82, p=0.17). Neither was there a significant effect of sequence type (F(1,25)=0.19, p=0.66), or interaction between the two factors (F(2,50)=2.01, p=0.14). This pattern of results on changes in overall activation remained unchanged after excluding error trials from the analyses (see Figure 3—figure supplement 1a). Thus, we observed widespread activation decreases with learning across secondary and association cortical areas.

Figure 4. Changes in average activation across the cortical surface.

(a) Average change in activation across subsequent sessions. Activation was measured as difference in percent signal change relative to the resting baseline. Activation decreased (blue shades) in motor-related regions across sessions during sequence execution. (b) Contrast of activation for trained vs. untrained sequences per scanning session. In weeks 2 and 5, trained sequences elicited lower activation in motor-related regions than untrained sequences (blue shades; see Figure 4—figure supplement 1 for t-maps and statistical quantification of activation clusters). Areas with observed increases in activation for trained sequences (red shades) lie in the default mode network that showed on average lower activity during task than rest.

Figure 4.

Figure 4—figure supplement 1. Statistical maps for the trained vs. untrained contrasts on elicited activation in each session.

Figure 4—figure supplement 1.

Trained >untrained is shown in red, untrained >trained in blue. Maps were thresholded at a t25 = ± 2.5, p<0.01 uncorrected for a two-sided t-test. Tables show peak t-value and size (in cm2) for each super-threshold cluster (indicated by numbers) for maps of week 2 and 5. pt_unc is the uncorrected p-value for the peak of each cluster. Family-wise error corrected p-values were determined using permutation testing for the peak t-value (pt_corr) and cluster size (pcluster).

In a few smaller areas, activation increased with learning (red patches in Figure 4a–b). This was observed uniformly in areas with activity at or below baseline – thus these changes reflect decreased suppression of activity rather than increases. It is likely that these activity increases are not task relevant, but instead reflect the increasing automaticity and lower need for central attentional resources with learning (see Discussion).

We also examined whether there were, in addition to the overall activity decreases, shifts in the average activity patterns in the predefined regions of interest (Figure 1c). As for M1, we calculated the cosine angle dissimilarity (see Materials and methods) between the average activity patterns for trained and untrained sequences, separately for each scanning session. Figure 5a shows cosine angle dissimilarities between trained and untrained sequences in PMd, displayed using multidimensional scaling (MDS). Patterns for trained sequences moved away from the starting point over weeks, and became more different from untrained patterns. Both in parietal and premotor areas there was clear evidence for a shift – cosine angular dissimilarity between the average trained and untrained sequence activation increased significantly across weeks (PMd: F(2,50)=23.63, p=5.98e−8; SPLa: F(2,50)=23.19, p=7.49e−8) (Figure 3c). S1 also showed a significant increase in cosine dissimilarity between trained and untrained patterns with learning (F(2,50)=8.68, p=5.79e−4). These changes, however, were much less pronounced than those observed in premotor and parietal areas. This observed increase in dissimilarity between average trained and untrained pattern in PMd and SPLa, and to a lesser extent in S1, was also observed when analyzing only trials with correct performance (see Figure 3—figure supplement 1b).

Figure 5. Relative change in evoked activation.

Figure 5.

(a) Multidimensional scaling plot of cosine angle dissimilarities for trained and untrained sequences in premotor dorsal area (PMd) across weeks 1–5. Each dot represents a single sequence, and dots are connected for each session and sequence type separately. Trained sequences on average become more distant from untrained sequences with learning. Untrained sequences on average also progress across weeks, but less than trained sequences. (b) Normalized activation plots for trained sequences in week 1 (blue) and 5 (red). The arrows and brackets indicate the direction and range of activation cross-sections presented in c). Areas: dorsal premotor cortex (PMd), primary motor cortex (M1), primary somatosensory cortex (S1), superior parietal lobule (SPL). (c) Cross-section of elicited activation for trained sequences in defined areas, in weeks 1 (blue) and 5 (red).

To investigate whether the observed changes in the overall activity patterns in premotor and parietal areas were spatially consistent across individuals, we normalized (z-scored) activation maps in each region and assessed the relative contribution of subregions to overall activation in weeks 1 and 5 (Figure 5b). Comparing the pattern of activation revealed that before training (week 1, blue) sequences elicit relatively more activation in rostral parts of the premotor and supplementary motor areas, and that activity was more caudal after training (week 5, red; Figure 5c displays the cross-section of relative activation changes). Some differences were also observed in the posterior parietal cortex, with activation shifting from more posterior to anterior subregions after learning (Figure 5c). Altogether, these results show that with learning, the execution of sequences relies on slightly different subareas within premotor and parietal regions.

Sequence-specific activity patterns reorganize early in learning

Our analyses so far have been concerned with changes in the overall pattern of trained vs. untrained sequences, and showed widespread reductions in activation and some more subtle changes in relative location. The sequence-specific performance advantage, however, indicates that the brain must represent specific sequences – i.e., there should be activity patterns that are unique to each individual sequence. Sequence-specific learning should then be reflected in changes of these sequence-specific activity patterns with learning (Figure 1d). Consistent with previous results (Wiestler and Diedrichsen, 2013; Yokoi and Diedrichsen, 2019), we detected sequence-specific activity patterns, i.e. activity patterns that differentiate between the tested motor sequences, in various cortical regions, even in session 1 (Figure 6a). This allowed us to assess their reorganization across sessions.

Figure 6. Sequence-specific activity patterns reorganize across sessions.

(a) Cortical surface map of crossnobis dissimilarities between activity patterns for different sequences in session 1. These regions encode which sequence is executed by the participant. (b) Evidence of models of correlation values between r = 0 and r = 1 for sequence-specific patterns across weeks 1–2 (solid) and 2–5 (dashed), separately for trained (red) and untrained (blue) sequences. Evidence was assessed with a type-II log-likelihood, relative to the average log-likelihood across models. Shaded areas indicate standard error across participants. Difference between log-likelihoods can be interpreted as a log-Bayes factor, with a difference of 1 indicating positive evidence. Vertical lines indicate the winning correlation model for trained (red) and untrained (blue) patterns across weeks 1–2. Black dots are projections of the two winning models onto the correlation function of trained sequences across weeks 1–2. The horizontal lines from the two black dots indicate the likelihood of the trained data under the two models, which was tested in a crossvalidated t-test. See Figure 6—figure supplement 1a for the same analysis with error trials excluded. (c) Map displaying the correlation of the winning model for trained and untrained sequences across weeks 1–2 and 2–5. The correlation of the winning correlation model is shown in all tessels where the difference between evidence for winning model vs. worst-fitting model exceeds log-Bayes factor of 1 (averaged across participants). See Figure 6—figure supplement 1b for the difference in best model correlation between trained and untrained sequences, and an indication of tessels where the difference is significant, as based on the crossvalidated t-test. (d) Crossnobis dissimilarities between trained and untrained sequence pairs across weeks. No significant effect of week, sequence type or their interaction was observed in any of the regions. Error bars indicate standard error across participants.

Figure 6.

Figure 6—figure supplement 1. Changes in pattern correlation across weeks.

Figure 6—figure supplement 1.

(a) Evidence for models of correlation values between r = 0 and r = 1 for sequence-specific patterns across weeks 1–2 (solid) and 2–5 (dashed), estimated only on trials with correct performance. Correlations of trained patterns are in red, untrained in blue. Evidence was assessed with a type-II log-likelihood, relative to the average log-likelihood across models. Shaded areas indicate standard error across participants. Vertical lines indicate the winning correlation model for trained (red) and untrained (blue) patterns across weeks 1–2. Black dots mark the log-likelihood of the trained sequence across weeks 1–2 under the winning models. Horizontal lines from the two black dots indicate the difference in likelihood of the trained data under the two models, tested in a crossvalidated t-test (* indicates p<0.05 for one-sided t-statistics). (b) Difference between correlation of the winner models for trained and untrained sequences, as presented in Figure 6c. Blue indicates a lower correlation across weeks for trained than untrained patterns of activity. The correlation difference values are plotted in tessels where the difference in model evidence was significant, as based on the cross-validated t-test (for two-sided p<0.05).

Our pre-registered hypothesis (https://osf.io/etnqc) was that earlier in learning sequence-specific activity patterns would change more for trained than untrained sequences, and would stabilize later in learning. In contrast to the other ideas tested in this paper, this was a novel hypothesis and not based on previous reports. Specifically, we predicted that the correlation of each sequence-specific pattern between weeks 1 and 2 should be lower for trained as compared to untrained sequences. The problem with performing a simple correlation analysis on the patterns, however, is that the estimated correlation will be biased by noise – that is, more within-session variability for one set of sequences will result in a lower correlation (Diedrichsen et al., 2018). To address this problem, we used the pattern component modelling (PCM) framework which explicitly models and estimates the signal and noise for each session explicitly. Using this approach, we estimated the likelihood of participants’ data under a series of models, each assuming a true correlation in the range between 0 (uncorrelated patterns) and 1 (perfect positive correlation; see Materials and methods for details). Figure 6b shows the log-likelihood for each specific correlation model relative to the mean across all models. In SPLa, the most likely correlation of the activity patterns for the trained sequences between weeks 1 and 2 was r = 0.37. For week 2–5, the likelihood peaked at r = 0.6. In contrast, the likelihood functions for untrained sequences indicated that the most likely model was between r = 0.6–0.7 for both week 1–2 and 2–5. The advantage of this analysis is that we can be sure that the observed low correlation in week 1–2 for trained sequence was not due to increased noise. In fact, if the noise in one or both sessions was too high, then the model would be unable to distinguish between any of the correlation models – i.e., the likelihood curve would be a flat line.

To statistically assess the difference in correlations across trained and untrained sequences, we compared the likelihood of the data of trained sequences between two models: the best-fitting model for the trained sequences (r = 0.37 in SPLa) and the correlation model best fitting the data of untrained sequences (r = 0.6) (black dots and projections onto y-axis in Figure 6b). To avoid double-dipping, the ‘best-fitting’ model was chosen on 25 participants (n-1) and the likelihood assessed on the left-out subject (see Materials and methods). The difference in model evidence was significant for correlation between weeks 1–2 in SPLa (t(25)=2.88, p=4.0e−3). In contrast, no difference in correlation was observed later in learning, between weeks 2 and 5 (t(25)=1.21, p=0.24). A similar pattern of results was observed in PMd, with correlation of trained sequences significantly lower than that of untrained sequences between weeks 1 and 2 (t(25)=2.93, p=3.6e−3), but not between weeks 2 and 5 (t(25)=0.88, p=0.39). No such change in correlation across weeks 1–2 was observed in M1 (t(25)=0.43, p=0.34). In S1, the effect was just significant (t(25)=1.72, p=0.049). To ensure that the observed lower correlation for trained patterns was not due to larger difference in error rate between weeks 1 and 2 for trained than for untrained sequences, we repeated the analysis excluding error trials. The pattern of results remained unchanged in PMd and SPLa (see Figure 6—figure supplement 1a), with lower correlation for trained than untrained patterns across weeks 1–2. In S1, after accounting for error trials, the correlation across weeks 1–2 did no longer differ between trained and untrained patterns. Overall, we found significant evidence that sequence-specific trained patterns in SPLa and PMd reorganize more in weeks 1–2 as compared to the untrained sequences, and stabilize later on with learning, in line with our new pre-registered prediction.

To determine more generally where in the neocortex sequence-specific plasticity could be detected, we fit PCM correlation models to regularly tessellated regions spanning the cortical surface. Figure 6c displays the correlation with the highest evidence for activity patterns across weeks 1–2 and 2–5; separately for trained and untrained sequences. In general, the highest correlations were found in core sensorimotor areas. Across weeks 1–2 for trained sequences, correlations were significantly lower in a number of dorsal premotor, inferior frontal, and parietal regions (Figure 6c). Across the cortex, correlation for trained patterns increased for weeks 2–5, resulting in similar values which did not differ significantly between trained and untrained sequences for most tessels (see Figure 6—figure supplement 1b). Together, these results confirmed that sequence-specific activation patterns in secondary association areas show less stability early in learning, but stabilize later on.

Can we obtain further insight into how the sequence-specific patterns change in these areas? One specific preregistered prediction was that there would be an increase in distinctiveness (dissimilarity) between fMRI patterns underlying each trained sequence (Wiestler and Diedrichsen, 2013; Figure 1e). To test this hypothesis, we calculated crossnobis dissimilarities (Walther et al., 2016) between sequence-specific activations, separately for trained and untrained sequences. In contrast to our prediction, no significant change in dissimilarity across weeks was observed in any of the predefined regions (Figure 6d). This suggests that the reorganization observed for trained sequences early in learning did not increase the average distinctiveness of the sequence-specific patterns.

Trained sequences elicit distinct patterns during full speed performance

In the last part of the experiment, we asked whether some of the negative findings (e.g. no changes in M1, no increase in dissimilarities for trained sequences) might have been due to the fact that participants were paced at a relatively slow speed. Matching the speed across sessions allows for the comparisons of changes in neural activity for exactly the same behavioral output (Karni et al., 1995; Lehéricy et al., 2005). However, it could be that controlling for speed impairs our ability to study brain representations of motor skill; simply because after learning, the system is not challenged enough to activate the neuronal representations supporting skilled performance. Consequently, several studies have not (Bassett et al., 2011; Wymbs and Grafton, 2015), or not strictly (Wiestler and Diedrichsen, 2013), matched performance across sessions or levels of training. To examine the effect of performance speed, we added a fourth scanning session (full speed - fs), just a day after from the third session in week 5, in which participants were instructed to perform the sequence as fast as possible.

Performance during the 4th scan was 1010 ms faster than in the first session (t(25)=15.7, p=1.82e−14) and also 338 ms (t(25)=9.92, p=4.58e−10) faster for trained than for untrained sequences. Averaged over trained and untrained sequences, we found that the faster performance in this session led to an increase in activity across premotor and parietal areas (Figure 7a,b). Although trained sequences were executed faster than untrained sequences, activation was still lower for trained compared to untrained sequences, similar to what we observed for paced performance (Figure 7c; see Figure 7—figure supplement 1a for statistical maps). In M1 and S1, we found no difference in activation between trained and untrained sequences (Figure 7a; M1: t(25)=1.78, p=0.09; S1: t(25)=1.69, p=0.10). Overall, the pattern of results for evoked activation did not change qualitatively when participants performed at full speed.

Figure 7. Speed-related changes in activation and dissimilarities.

(a) Overall activation in week 5 in paced and full speed sessions for trained (red) and untrained (blue) sequences. Activation was measured as percent signal change over resting baseline (* indicates p<0.05). Error bars indicate standard error across participants. (b) Increase in activation for full speed compared to paced speed in percent signal change, averaged across trained and untrained sequences. Red colors indicate an increase in activity during full speed performance compared to paced performance. Blue colors indicate higher activation during paced compared to full speed performance. (c) Difference in activation elicited for trained relative to untrained sequences, during the paced and full speed sessions (see Figure 7—figure supplement 1a for statistical maps). Trained > untrained is shown in red, untrained > trained in blue. (d) Average crossnobis dissimilarity between sequence-specific patterns in paced and full speed sessions for trained (red) and untrained (blue) sequences. Dissimilarities are significantly larger for trained, as compared to untrained patterns in PMd during the full-speed session (* indicates p<0.05). Error bars indicate standard error across participants. (e) Difference between crossnobis dissimilarities across full speed and paced sessions, averaged across trained and untrained sequences. Higher dissimilarities for full speed than paced session are shown in red, whereas blue/green hues indicate higher dissimilarities during paced than full speed session. (f) Difference in dissimilarities for trained relative to untrained sequences, during the paced and full speed sessions. Trained > untrained is shown in red, untrained > trained in blue/green. Trained sequences elicited higher dissimilarities than untrained in full speed, but not paced session (see Figure 7—figure supplement 1b for statistical t-maps).

Figure 7.

Figure 7—figure supplement 1. Statistical maps for trained vs. untrained contrasts in week 5 (paced) and 5* (full speed) sessions.

Figure 7—figure supplement 1.

Trained > untrained is shown in red, untrained > trained in blue. (a) Statistical contrast for average activation. Maps were thresholded at a t25 = ± 2.5, p<0.01 uncorrected for a two-sided t-test. Tables show peak t-value and size (in cm2) for each super-threshold cluster. pt_unc is the uncorrected p-value for the peak of each cluster. Family-wise error corrected p-values were determined using permutation testing for the peak t-value (pt_corr) and cluster size (pcluster). (b) Statistical contrast for average dissimilarity of sequence-specific activity pattern. Map was thresholded at t25 = ± 1.7, p<0.05, uncorrected. Statistical quantification using permutation tests is in the table below each map.

Next, we examined whether the brain representations of individual sequences are similarly engaged at slow and fast speeds. The correlation between sequence-specific patterns was relatively high (r = 0.62) across our regions of interest. We found no differences between the different regions (F(3,75)=1.47, p=0.23), or sequence types (trained vs. untrained: F(1,25)=0.25, p=0.62). Thus, the sequence-specific representations activated during performance at high skill level (full speed) are at least partly activated even when performance slowed down.

Having established that the mean activation results are replicated across paced and full-speed performance, and that similar sequence-specific representations are activated in both cases, we tested whether activation patterns for different trained sequences are more distinct during full speed performance, as reported in Wiestler and Diedrichsen, 2013. Overall, crossnobis dissimilarities increased at full speed for trained sequences in PMd and SPLa (Figure 7e). No such changes were found in M1 or S1. Moreover, trained sequences showed larger dissimilarities than untrained at full-speed performance across premotor and parietal cortices (Figure 7f), which was not the case for the last paced session. In our predefined ROIs, this difference was significant for PMd (Figure 7d), but also parietal areas showed significantly higher dissimilarities between trained sequences at full speed (Figure 7—figure supplement 1b). This suggests that while activity patterns at full speed are correlated to those during paced performance, they are more distinguishable for trained sequences.

Could this effect be driven by behavioral performance, with trained sequences performed more differently at full speed (i.e. different speeds across trained sequences), while untrained sequences were performed at a more equal speed? To test for this, we calculated crossnobis dissimilarities between movement times associated with different trained and untrained sequences. The dissimilarities based on speed of performance did not differ significantly across trained and untrained sequences (t(25)=0.57, p=0.57). Therefore, increased dissimilarity of trained compared to untrained patterns in premotor and parietal areas could not be explained by a difference in execution speed. Instead, this effect likely reflects changes in activity patterns underlying full speed skilled performance.

Striatal activity patterns for trained sequences manifest at full speed performance

We observed learning-related changes in cortical association areas, but not in the primary motor cortex. Of course, learning could also be driven by neuronal changes in subcortical brain regions (Ashby et al., 2010; Graybiel, 2000; Graybiel and Grafton, 2015; Hikosaka et al., 1999; Yin et al., 2009). The striatum in particular has been proposed as a structure where motor skills are stored (Kawai et al., 2015; Lehéricy et al., 2006). Inspecting changes in overall activity across sessions, we observed no difference in activity between trained and untrained sequences in either putamen or caudate nucleus (Figure 8a).

Figure 8. Striatal changes in activation and dissimilarities with learning.

Figure 8.

(a) Overall activation (percent signal change over resting baseline), for trained (red) and untrained (blue) sequences. Activation did not differ across sessions, or sequence types in the striatum. Error bars indicate the standard error across participants. (b) Activation during performance of trained sequences in the striatum across weeks 1, 5 (paced speed) and 5 (full speed – fs), averaged across sequences and participants. (c) Crossnobis dissimilarities between activation patterns of sequence pairs, calculated separately for trained and untrained patterns. Dissimilarities were not significantly different for trained or untrained sequences during paced performance. At full speed, sequence-specific activity patterns amongst trained sequences differed significantly in both caudate nucleus and pallidum (* indicates p<0.05). Error bars indicate the standard error across participants.

Previous experiments have reported that with learning, activation moves from more ‘cognitive’ areas of the striatum (i.e. caudate nucleus) to more ‘motor’ areas (i.e. putamen) (Coynel et al., 2010; Lehéricy et al., 2005; Reithler et al., 2010). Our data fail to replicate this result: Both the visual inspection (Figure 8b), and statistical quantification of the mean pattern difference for trained and untrained sequences across sessions revealed no such learning-specific shift of mean striatal activation pattern with learning.

Lastly, we examined if the striatum represents individual sequences. During the paced sessions, activity patterns for different sequences were not distinguishable in either caudate nucleus or putamen (Figure 8c). However, during full speed performance trained sequences elicited distinct activity patterns in both regions (i.e. crossnobis dissimilarity >0: caudate nucleus: t(25)=2.27, p=0.032; putamen: t(25)=2.44, p=0.022; Figure 8c). This effect was specific to the trained sequences, with untrained sequences still exhibiting undistinguishable patterns of activity at full speed. Thus, we found some evidence that trained motor sequences are represented in the form of distinct activity patterns in the striatum during full speed skilled performance.

To examine whether the speed purely pulls the signal out of the noise better, or qualitatively changes the representation, we, similarly to the analyses in the cortical regions, performed the PCM correlation model across the paced and full speed sessions in week 5. The correlation between sequence-specific patterns in both regions was higher than 0 (putamen: t(25)=9.56, p=8.0e−10; caudate: t(25)=6.37, p=1.1e−6), but lower than 1 (putamen: t(25)=8.85, p=3.6e−9; caudate: t(25)=5.86, p=4.1e−6). Similarly as for the cortical regions, we found no differences between the caudate nucleus and putamen (F(3,75)=0.19, p=0.66), or sequence types (trained vs. untrained: F(1,25)=0.05, p=0.83). Thus, the sequence-specific representations activated during performance at high skill level (full speed) are at least partly activated even when performance slowed down. This suggests that moving faster engages similar representations as moving slower, but helps to increase the signal-to-noise ratio.

Discussion

Here we present a large longitudinal motor sequence learning study that allowed us to systematically investigate several previously proposed fMRI signatures of motor learning, including one new hypothesis concerning the change in multivariate activity patterns with learning. The existing literature, with its diversity of experimental protocols and analysis approaches, does currently not provide a consistent picture of learning-related changes. This inconsistency is exacerbated by the fact that most papers prioritize making new claims over re-examining previously established findings. Consequently, it is very hard to assess the replicability of most past findings. We address this issue here by a) producing a well-powered, longitudinal data set that tackles some of the methodological inconsistencies (i.e. speed matching), b) pre-registering both design and hypotheses, and c) making data and analysis pipelines openly available, such that other hypotheses and analyses techniques can be freely tested.

Our findings reveal that parietal and premotor areas show widespread decreases in overall activation, as well as reorganization of sequence-specific patterns early in learning. Additionally, we observed that sequence specific patterns in these areas (as well as the striatum) were more distinct during full speed performance. In contrast to this set of results, none of these learning-specific metrics were detected in M1, even after 5 weeks of training.

On the one hand, our lack of any observable change in M1 activation contradicts some prior results, where increased activation in M1 was observed for matched performance after learning (Karni et al., 1995; Matsuzaka et al., 2007; Penhune and Doyon, 2002; Steele and Penhune, 2010; Vahdat et al., 2015), and does not align with reports of M1 stimulations influencing consolidation or storage of motor skills (in motor sequence tasks: Kang and Paik, 2011; Nitsche et al., 2003; Reis et al., 2009; Waters-Metenier et al., 2014; in other motor tasks: Classen et al., 1998; Galea et al., 2011; Hadipour-Niktarash et al., 2007). We also found no support for a combination of increases and decreases of activation with training, which would lead to an overall change of the mean activity pattern (Steele and Penhune, 2010).

Instead, our results suggest that the pattern of neural activity in M1 does not change as participants become more skilled at producing motor sequences. This is consistent with a recent line of evidence demonstrating that M1 does not change activation with learning (Huang et al., 2013), and primarily encodes single movement elements, rather than sequences (Yokoi et al., 2018; Russo et al., 2019). Somewhat more surprisingly, we also observed no difference in overall M1 activation during full speed performance, when performance was considerably faster for trained sequences. This suggests that the activity increases related to faster movement speeds are compensated for by the shorter duration spent on the task.

Primary somatosensory cortex in many ways paralleled the results observed in M1. We observed no overall activation change, or change in the sequence-specific pattern correlation across sessions. The only exception was the observed shift in the mean activation pattern across sessions. One possible explanation is that feedback-related sensory activity in S1 undergoes some plastic changes with learning. This is consistent with a recent study demonstrating that S1, but not M1, is involved during consolidation of motor skills (Kumar et al., 2019; for a review on somatosensory plasticity in motor learning see Ostry and Gribble, 2016).

In contrast to the limited evidence of learning-related changes in primary somatosensory and primary motor areas, higher order association areas (e.g. parietal and premotor cortices) displayed an array of learning-related changes. First, activation decreased in areas involved in sequence execution, with larger decreases for trained as compared to untrained sequences. This result contrasts with other previous studies reporting increases in activation in premotor areas with learning (Grafton et al., 2002; Honda et al., 1998; Penhune and Doyon, 2002; Vahdat et al., 2015). Partially responsible for these inconsistencies may be a publication bias, favoring reports of signal increases over signal decreases with learning. For example, a recent metanalysis reanalyzed evidence for signal increases in the main text, while moving the (matched) evidence for signal decreases into the supplementary materials (Hardwick et al., 2013). Our data corroborate a number of recent studies reporting reduced activation in task-evoked premotor and parietal areas (Steele and Penhune, 2010; Wiestler and Diedrichsen, 2013; Wu et al., 2004).

The only activation increases for trained relative to untrained sequences were observed in areas that were suppressed below baseline during sequence execution. This has also been previously reported in a motor sequence learning study (Tamás Kincses et al., 2008), where deactivation was larger during performance of trained than random sequences. These areas include the precuneous, temporal parietal junction and the cingulate, regions commonly assigned to the default mode network (Raichle et al., 2001; Shulman et al., 1997). This group of regions is more activated during rest than during task performance, and has been associated with functions such as episodic memory retrieval and attention to internal states (Andrews-Hanna et al., 2010; Gusnard et al., 2001). Our observation of decreased inhibition of the default mode network likely reflects central attentional resources being freed up, allowing participants to engage in other mental processes (e.g., daydreaming) while performing the task. Thus, this release from initial deactivation is possibly task-irrelevant, reflecting increased automaticity with learning (Shamloo and Helie, 2016).

Overall, changes in average activation are relatively hard to interpret, as they could reflect a combination of numerous factors. As a more direct fMRI metric of plasticity, we suggest to inspect changes in the sequence-specific activity patterns, since these constitute a more likely fMRI correlate of the sequence-specific performance advantage observed after training. In this project, this provided us with two key insights of how activation patterns reorganize in association areas with learning. First, activity patterns associated with each of individual trained sequences, changed to a greater extent earlier in learning, and stabilized later. This finding resonates with several animal studies suggesting that the emergence of skilled behavior is associated with early plasticity and later stabilization of neuronal activity patterns (Makino et al., 2017; Peters et al., 2017). Here we report a similar effect in humans, and advance these findings by demonstrating that this reorganization occurs at the level of sequence-specific patterns. In past studies using rodent models, sequence-specific patterns could not be dissociated from the overall activity pattern, as the animals were only trained on production of a single sequence. Additionally, by pacing participants’ speed, we were able to cleanly dissociate changes in the organization of activity patterns from changes in the behavioral performance or variability. Second, activation patterns became more distinct for trained sequences at full speed. This indicates that the engagement of specific neuronal subpopulations for different sequences is particularly important when pushing the limit of performance.

While our study focused on the role of cortical areas in motor sequence learning, we also examined activation in the striatum, which has been suggested to play a critical role in skilled performance (Graybiel and Grafton, 2015; Kawai et al., 2015; Otchy et al., 2015). In contrast to previous fMRI studies (Coynel et al., 2010; Lehéricy et al., 2006; Reithler et al., 2010), we did not find clear evidence for differences in overall activity, or shifts of the overall activity pattern with learning. Nonetheless, we observed distinguishable striatal activation patterns for different trained sequences at full speed, in line with a recent report showing distinguishable striatal patterns for performance of consolidated motor sequences (Pinsard et al., 2018). While by itself the finding of differential sequence-specific activity patterns is not evidence for a causal role of the striatum in the production of skilled behaviors, it is a necessary condition for such a functional role. Therefore, our results here are in line with the proposed involvement of the striatum in motor sequence learning. Additionally, our results suggest that full speed performance might be particularly important for further studies of striatal multivariate activation to counteract the generally lower signal-to-noise ratio in this region.

An important feature of our design was that we collected imaging data in the trained state, both when performance was clamped to the initial speed, and when participants performed as fast as possible. Previous studies have usually included only one of these two options, making direct comparisons difficult (see Lutz et al., 2004 for an examination of various execution speeds on BOLD activity and Orban et al., 2010 in a motor learning context). Our results provide two important insights: first, in terms of the overall fMRI activation, the pattern of results remained the same for paced vs. full speed performance. This indicates that, in this specific case, the increased motor demands and the decreased time on task averaged out. In general, however, these two factors may not balance perfectly – therefore paced performance may be a better choice when comparing overall activation across sessions. Second, even though slow and paced performance in the trained state activated sequence-specific activation patterns, these were much stronger when performing at maximal speeds. Thus, for questions regarding the fine-grained patterns, it might be more suitable to challenge the system fully.

Of course, our list of inspected fMRI metrics of learning was not exhaustive. For instance, we did not investigate whether various fMRI correlates of learning predict behavioral outcomes, or how functional connectivity and network metrics change with learning, partly because of the absence of specific predictions. Pre-registration of hypothesis are especially important for these analyses, since the search space of possible tests becomes exponentially larger (e.g. correlating all possible brain metrics with all possible behavioral metrics; or using various metrics to assess inter-regional relationships). However, we hope that our dataset can serve as a resource for other researchers to (re-)test novel predictions about learning related changes.

Conclusion

The search for neural substrates of learning is a daunting task: the acquisition of longitudinal data sets is work intensive, and the large dimensionality of possible brain metrics makes the search difficult (Poldrack, 2000). Historically, the question was simplified by studying activation increases in single areas as proxies for motor ‘engram’ localization (Lashley, 1950Berlot et al., 2018). Here we found no evidence for such activation increases; instead we observed widespread and distributed decreases in activation across cortical areas. In contrast, subtler changes in the distributed patterns of fMRI activity have the potential to provide more direct metrics of plasticity. Increased pattern reorganization (across weeks), and larger pattern separation for trained sequences was found across prefrontal, parietal, and striatal regions. These metrics may be useful as general fMRI correlates of neural reorganization beyond the domain of motor learning.

Materials and methods

Participants

Twenty-seven volunteers participated in the experiment. One of them was excluded because field map acquisition was distorted in one of the four scans. The average age of the remaining 26 participants was 22.2 years (SD = 3.3 years), and the sample included 17 women and 9 men. All participants were right-handed and had no prior history of psychiatric or neurological disorders. They provided written informed consent to all procedures and data usage before the study started. The experimental procedures were approved by the Ethics Committee at Western University.

Apparatus

Participants performed finger sequences with their right hand on an MRI-compatible keyboard (Figure 2a), with keys numbered 1–5 for thumb-little finger. The keys had a groove for each fingertip and were not depressible. The force of isometric finger presses was measured by the force transducers (FSG-15N1A, Sensing and Control, Honeywell; dynamic range 0–25 N) mounted underneath each key with an update rate of 2 ms. A key press was recognized when the sensor force exceeded 1 N. The measured signal was amplified and sampled at 200 Hz.

Learning paradigm

Participants were trained to execute six 9-digit finger sequences over a period of five weeks (Figure 2a). They were randomly split into two groups, with trained sequences of one group constituting the untrained sequences for the other group and vice versa. Finger sequences of both groups were matched as closely as possible in terms of the starting finger, number of finger repetitions in a sequence and first-order finger transitions. This counterbalancing between the groups ensured that any of the observed results were not specific to a set of chosen trained sequences.

In the pre-training session prior to the first scan (Figure 2b), participants were acquainted with the apparatus and task performed during scanning. Sequences executed during this pre-training session were not encountered later on in the experiment.

During the training sessions, participants were trained to perform the six sequences as fast as possible. They received visual feedback for the correctness of their presses with digits turning green for a correct finger press and red for an incorrect one. After each trial, participants received points based on the accuracy and their movement time (MT – time from the first press until the last finger release in the sequence; Figure 2c). Trials executed correctly and faster than participant’s median MT from the previous blocks were rewarded with 1 point. If participants performed correctly and 20% faster than the median MT from previous blocks, they received 3 points. If they made a mistake or performed below their median MT, they received 0 points. Participants performed each sequence twice in a row: digits were written on the screen for the first execution, but removed for the second execution so that participants had to perform the finger sequence from memory. Training sessions were broken into several blocks, each consisting of 24 trials (4 trials per trained sequence), with time between blocks to rest. At the end of each block, participants received feedback on their error rate, median MT and points obtained during the block. If participants performed with an error of <15% and faster than the previous median MT, the MT threshold was updated. This design feature was chosen to maintain participants’ motivation to execute the sequences as fast as possible, within the allowed error range.

During the behavioral test sessions (Figure 2d), participants executed both the trained sequences they were trained on, as well as matched untrained sequences, with all sequences randomly interspersed. As in training, each sequence was performed twice in a row – however, the 9-digit sequence numbers were presented on the screen present on both executions. Therefore, the requirement to remember the sequences from the first to second execution, which was present in training sessions, was omitted for test sessions. For this reason, performance in training and test sessions (Figure 2c–d) cannot be directly compared.

As an additional feature of the four behavioral test sessions, we examined participants’ performance with their left hand. Specifically, we tested them on execution of intrinsically-matched trained sequences (i.e. producing the same finger combinations), extrinsically-matched trained sequences (i.e. producing the same external consequences using mirrored fingers) and random sequences. This was added to probe to what extent learning generalized to the other effector in intrinsic or extrinsic coordinate frames, at different stages of learning. Per session, participants performed 4 repetitions of each trained sequence in intrinsically-matched space and 4 repetitions in extrinsically-matched space.

Experimental design during scanning

Participants underwent four scanning sessions (Figure 2d) – with the first one before learning regime started, the second after a week and two more scans after completion of the 5 training weeks. Each scanning session consisted of eight functional runs. We employed an event-related design, randomly intermixing execution of trained and untrained sequences. Each sequence was repeated twice in a row with digits present on the screen during both executions. Thus, there was no need to memorize either trained or untrained sequences from first to second execution in the scanner. Each sequence was repeated for a total of six times in every run. Each trial started with 1 s preparation time, during which the sequence was presented on the screen. After that time, a ‘go’ signal was displayed as short pink line underneath the sequence numbers. In scanning sessions 1–3, this line started expanding below the written numbers, indicating the speed at which participants were required to press along. In scanning session 4, only a short line was presented in front and underneath the sequences. When the line disappeared, this signaled a ‘go’ cue for participants to execute the presented sequence as fast as possible. The execution phase including the feedback on overall performance lasted for 3.5 s, and the inter-trial interval was 0.5 s (see Figure 2—figure supplement 1). Each trial lasted for 5 s. Participants always received 3 points upon correct execution of the sequence, and 0 points otherwise. Five periods of rest, each 10 s long, were added randomly between trials in each run to provide a better estimate of baseline activation. Participants performed the task inside the scanner for approximately 75 min. After each scanning session, they filled out a recall and recognition questionnaires on trained and untrained sequences performed inside the scanner.

Image acquisition

Data was acquired on a 7-Tesla Siemens Magnetom scanner with a 32-receive channel head coil (8-channel parallel transmit). Anatomical T1-weighted scan was acquired at the beginning of the first scanning session, using a magnetization-prepared rapid gradient echo sequence (MPRAGE) with voxel size of 0.75 × 0.75×0.75 mm isotropic (field of view = 208×157 x 110 mm [A-P; R-L; F-H], encoding direction coronal). Functional data were acquired using a sequence (GRAPPA 3, multi-band acceleration factor 2, repetition time [TR]=1.0 s, echo time [TE]=20 ms, flip angle [FA]=30 deg). We acquired 44 slices with isotropic voxel size of 2 × 2 × 2 mm. For estimating magnetic field inhomogeneities, we additionally acquired a gradient echo field map. Acquisition was in the transversal orientation with field of view 210 × 210 × 160 mm and 64 slices with 2.5 mm thickness (TR = 475 ms, TE = 4.08 ms, FA = 35 deg). To monitor the use of 7T for human research, participants filled out a questionnaire rating their levels of dizziness, wellbeing, alertness and focus after each imaging session.

First-level analysis

Functional data were analyzed using SPM12 and custom written MATLAB code. Functional runs were corrected for geometric distortions using fieldmap data (Hutton et al., 2002), and head movements during the scan (3 translations: x, y, z; 3 rotations: pitch, roll, yaw), and aligned across sessions to the first run of the first session. The functional data were then co-registered to the anatomical scan. No smoothing or normalization to an atlas template was performed.

Preprocessed data were analyzed using a general linear model (GLM; Friston et al., 1994). Each of the performed sequences was defined as a separate regressor per imaging run, resulting in 12 regressors per run (6 trained, 6 untrained sequences), together with intercept for each of the functional runs. All instances of sequence execution were included into estimating regressors, regardless of whether the execution was correct or erroneous (see section Treatment of error trials below). The regressor was a boxcar function starting at the beginning of the trial and lasting for trial duration. The boxcar function was convolved with a hemodynamic response function, with a time to peak of 5.5 s, and time to undershoot of 12.5 s. We adjusted the hrf onset individually per participant. For that, we defined a combined region of interest between PMd and M1, and averaged the response across all voxels in the combined ROI for all performed sequences (i.e. trained and untrained sequences together) in session 1. We then performed a grid-search with delay values of 0, 0.5 and 1 s, and chose the one that maximally fit the evoked response for each subject. The same delay was used across all sessions. Ultimately, this analysis resulted in one activation estimate (beta image) for each of the 12 conditions per run, in each scanning session.

Surface reconstruction and regions of interest

We reconstructed individual subjects’ cortical surfaces using FreeSurfer (Dale et al., 1999). All individual surfaces were aligned to the FreeSurfer’s Left-Right symmetric template (Workbench, 164 k nodes) via spherical registration. To detect sequence representation across the cortical surface, we used a surface-based searchlight approach (Oosterhof et al., 2011), where for each node we selected a circular region of 120 voxels in the grey matter. The resulting analyses (dissimilarities between sequence-specific activity patterns, see below) was assigned to the center node. As a slightly coarser alternative to searchlights, we performed regular tessellation of cortical surface into 162 tessels per hemisphere. This allowed us to fit correlation models (see below) across the cortical surface, while not being as computationally intensive as searchlight analyses.

We defined four regions of interest to cover primary somato-motor regions as well as secondary associative regions. M1 was defined using probabilistic cytoarchitectonic map (Fischl et al., 2008) by including nodes with the highest probability of belonging to Brodmann area (BA) 4, while excluding nodes more than 2.5 cm from the hand knob (Yousry et al., 1997). Similarly, S1 was defined as nodes related to hand representation in BA 1, 2 and 3. Additionally, we included dorsal premotor cortex (PMd) as the lateral part of the middle frontal gyrus. The anterior part of the superior parietal lobule (SPLa) was defined to include anterior, medial and ventral intraparietal sulcus. We also defined caudate nucleus and putamen as striatal regions of interest. The definition was carried out in each subject using FSL’s (Jenkinson et al., 2012) subcortical segmentation.

Changes in overall activation

We calculated the average percent signal change for trained and untrained sequences (averaged across the 6 trained and 6 untrained sequences) relative to the baseline for each voxel. The resulting volume map was projected to the surface for each subject, and a group statistical t-map was generated across subjects. Statistical maps were thresholded at p<0.01, uncorrected, and the family-wise error corrected p-value for the size of the peak activation and activation cluster size was determined using a permutation test. Specifically, we ran 1000 simulations where we randomly flipped the sign of the contrast for subjects (chosen at random out of 226 possible permutations). The rationale behind this is that under the null hypothesis, there should be no difference between the two conditions, and the sign of each contrast should be interchangeable. As for the data, we thresholded the statistical map from each permutation, and recorded the peak t-value (across the map) and the size of the largest cluster. The real data was then compared against this distribution to assess the probability of the observed t-value and cluster-size under the null hypothesis.

Additionally, we assessed changes in percent signal in predefined regions of interest (M1, S1, PMd, SPLa). This was performed in the native volume space of each subject. To do so, we averaged the percent signal change of voxels belonged to a defined region per subject and quantified activation changes across subjects using ANOVAs and t-tests across subjects.

Besides overall activation, we also examined relative changes in elicited activation for trained sequences across sessions. This was done by normalizing (z-scoring) the percent signal change surface maps across voxels, separately for each subject. Normalization was applied both map-wise (for Figure 5b), as well as for each of the pre-defined ROIs separately (for cross-sections in Figure 5c). Statistical assessment of the difference between relative evoked activation pattern for trained vs. untrained sequence was carried out by calculating cosine angle dissimilarities between the mean evoked patterns. Cosine angle dissimilarity was chosen because it is not sensitive to overall magnitude in activation, and therefore assesses the difference in the relative activation distribution.

Dissimilarities between sequence-specific activity patterns

To evaluate which cortical areas display sequence-specific encoding, we performed a searchlight analysis calculating the dissimilarities between evoked beta patterns of individual sequences. Beta patterns were first multivariately prewhitened (standardized by voxels’ residuals and weighted by the voxel covariance matrix), which has been found to increase the reliability of dissimilarity estimates (Walther et al., 2016). We then calculated the cross-validated squared Mahalanobis dissimilarities (i.e. crossnobis dissimilarities) between evoked sequence patterns (66 dissimilarity pairs for 6 trained and 6 untrained sequences). These dissimilarities were then averaged overall, as well as separately for pairs within trained sequences, and within untrained sequences. This metric was used both for searchlight analysis and calculation of metric within predefined regions (cortical and striatal). The cortex surface maps contrasting dissimilarities between trained and untrained sequences were corrected for multiple comparisons using permutations, as described above for percent signal change surface maps.

Pattern component analyses: modelling sequence-specific correlation across sessions

Correspondence of sequence-specific patterns across sessions was quantified using pattern component modelling (PCM; Diedrichsen et al., 2018). This framework is superior at estimating correlations than simply performing Pearson’s correlation on raw activity patterns, or even in a crossvalidated fashion. The main problem with estimating correlations on data is that activation patterns are biased by noise, which varies across scanning sessions, and would therefore underestimate the true correlation. PCM separately models the noise and signal component, and can in this way combat the issue more than by simply performing crossvalidation. We designed 30 correlation models with correlations between 0 and 1 in equal step sizes and assessed the group likelihood of the observed data under each model.

Subsequent group inferences were performed using crossvalidated approach on assessing individual log-Bayes factors (model evidence). A crossvalidated approach was used to ensure that our choice of ‘best-fitting models’ and the evidence associated was independent and did not involve double-dipping. Specifically, we used n-1 subjects to determine the best-fitting models for trained and untrained patterns and recorded the log-Bayes factors for those two correlation models on the left-out subject. This was repeated across all subjects and a t-test was performed on the recorded log-Bayes factors (i.e. out-of-sample model evidences). The same evaluation was performed for pre-defined regions of interest (Figure 6b), as well as a regular tessellation across the cortical surface (Figure 6c).

Treatment of error trials

As in behavioral sessions, participants were instructed to keep their error rate below 15% also inside the scanner. This was on average achieved, with the following error rate for trained vs. untrained sequences across the 4 scanning sessions: Week 1: 0.14 ± 0.02 vs. 0.15 ± 0.02, week 2: 0.08 ± 0.01 vs. 0.14 ± 0.02, week 5: 0.06 ± 0.01 vs. 0.09 ± 0.01, speeded scan week 5: 0.14 ± 0.01 vs. 0.13 ± 0.01. The number of errors varied significantly across sessions, and between sequence types. A session x sequence type ANOVA was significant for week (F(3,75)=9.19, p=2.97e−5), sequence type (F(1,25)=11.16, p=2.63e−3), as well as for their interaction (F(3,75)=8.39, p=7.00e−5). Post-hoc t-tests revealed that the error rate differed between trained and untrained sequences in week 2 (t(25)=4.20, p=2.95e−4) and 5 (t(25)=4.81, p=6.1e−5), but not in week 1 and for the speeded session of week 5. To control for difference in error rate, we performed an additional first-level analysis with error trials excluded to ensure that our results were not due to inclusion of errors. Indeed, our results did not differ qualitatively when excluding errors, therefore we here report only the analyses with all trials included.

Code availability

The code is available at https://github.com/eberlot/motor_sequence_learning.git (Berlot, 2020; copy archived at https://github.com/elifesciences-publications/motor_sequence_learning/).

Acknowledgements

The work was supported by an Ontario Trillium Scholarship to EB, an NSERC Discovery Grant (RGPIN-2016–04890) to JD, and the Canada First Research Excellence Fund (BrainsCAN). We thank Giacomo Ariani, Tamar Makin, Andrew Pruszynski, and Atsushi Yokoi for helpful comments on an earlier version of the manuscript.

Funding Statement

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Contributor Information

Jörn Diedrichsen, Email: jdiedric@uwo.ca.

Marius V Peelen, Radboud University, Netherlands.

Floris P de Lange, Radboud University, Netherlands.

Funding Information

This paper was supported by the following grants:

  • Natural Sciences and Engineering Research Council of Canada NSERC Discovery Grant (RGPIN-2016-04890) to Jörn Diedrichsen.

  • Canada First Research Excellence Fund BrainsCAN to Jörn Diedrichsen.

  • Ontario Trillium Foundation Graduate Student Scholarship to Eva Berlot.

Additional information

Competing interests

Reviewing editor, eLife.

No competing interests declared.

Author contributions

Conceptualization, Formal analysis, Validation, Investigation, Visualization, Methodology, Writing - original draft, Writing - review and editing.

Conceptualization, Investigation, Methodology, Writing - review and editing.

Conceptualization, Supervision, Funding acquisition, Validation, Investigation, Methodology, Project administration, Writing - review and editing.

Ethics

Human subjects: Informed consent and data usage agreement was obtained from participants prior to the onset of the study. It was emphasized that participants could withdraw from the study at any timepoint. The experimental procedures were approved by the Ethics Committee at Western University (HSREB File Number: 107061).

Additional files

Transparent reporting form

Data availability

fMRI data and analysis pipelines have been deposited to OpenNeuro, under the accession number ds002776. Analysis code is available on GitHub at https://github.com/eberlot/motor_sequence_learning.git (copy archived at https://github.com/elifesciences-publications/motor_sequence_learning).

The following dataset was generated:

Berlot E, Nicola JP, Diedrichsen J. 2020. Motor sequence learning. OpenNeuro. ds002776

References

  1. Andrews-Hanna JR, Reidler JS, Sepulcre J, Poulin R, Buckner RL. Functional-anatomic fractionation of the brain's default network. Neuron. 2010;65:550–562. doi: 10.1016/j.neuron.2010.02.005. [DOI] [PMC free article] [PubMed] [Google Scholar]
  2. Ashby FG, Turner BO, Horvitz JC. Cortical and basal ganglia contributions to habit learning and automaticity. Trends in Cognitive Sciences. 2010;14:208–215. doi: 10.1016/j.tics.2010.02.001. [DOI] [PMC free article] [PubMed] [Google Scholar]
  3. Bassett DS, Wymbs NF, Porter MA, Mucha PJ, Carlson JM, Grafton ST. Dynamic reconfiguration of human brain networks during learning. PNAS. 2011;108:7641–7646. doi: 10.1073/pnas.1018985108. [DOI] [PMC free article] [PubMed] [Google Scholar]
  4. Bassett DS, Yang M, Wymbs NF, Grafton ST. Learning-induced autonomy of sensorimotor systems. Nature Neuroscience. 2015;18:744–751. doi: 10.1038/nn.3993. [DOI] [PMC free article] [PubMed] [Google Scholar]
  5. Berlot E, Popp N, Diedrichsen J. Neural Correlates of Long-Term Motor Sequence Learning. Open Science Framework; 2017. https://osf.io/ueswg/ [Google Scholar]
  6. Berlot E, Popp NJ, Diedrichsen J. In search of the engram, 2017. Current Opinion in Behavioral Sciences. 2018;20:56–60. doi: 10.1016/j.cobeha.2017.11.003. [DOI] [Google Scholar]
  7. Berlot E. motor_sequence_learning. 00b2e9aGithub. 2020 https://github.com/eberlot/motor_sequence_learning
  8. Beukema P, Diedrichsen J, Verstynen TD. Binding during sequence learning does not alter cortical representations of individual actions. The Journal of Neuroscience. 2019;39:6968–6977. doi: 10.1523/JNEUROSCI.2669-18.2019. [DOI] [PMC free article] [PubMed] [Google Scholar]
  9. Classen J, Liepert J, Wise SP, Hallett M, Cohen LG. Rapid plasticity of human cortical movement representation induced by practice. Journal of Neurophysiology. 1998;79:1117–1123. doi: 10.1152/jn.1998.79.2.1117. [DOI] [PubMed] [Google Scholar]
  10. Coynel D, Marrelec G, Perlbarg V, Pélégrini-Issac M, Van de Moortele PF, Ugurbil K, Doyon J, Benali H, Lehéricy S. Dynamics of motor-related functional integration during motor sequence learning. NeuroImage. 2010;49:759–766. doi: 10.1016/j.neuroimage.2009.08.048. [DOI] [PMC free article] [PubMed] [Google Scholar]
  11. Dale AM, Fischl B, Sereno MI. Cortical Surface-Based analysis. NeuroImage. 1999;9:179–194. doi: 10.1006/nimg.1998.0395. [DOI] [PubMed] [Google Scholar]
  12. Dayan E, Cohen LG. Neuroplasticity subserving motor skill learning. Neuron. 2011;72:443–454. doi: 10.1016/j.neuron.2011.10.008. [DOI] [PMC free article] [PubMed] [Google Scholar]
  13. Diedrichsen J, Yokoi A, Arbuckle SA. Pattern component modeling: a flexible approach for understanding the representational structure of brain activity patterns. NeuroImage. 2018;180:119–133. doi: 10.1016/j.neuroimage.2017.08.051. [DOI] [PubMed] [Google Scholar]
  14. Doyon J, Song AW, Karni A, Lalonde F, Adams MM, Ungerleider LG. Experience-dependent changes in cerebellar contributions to motor sequence learning. PNAS. 2002;99:1017–1022. doi: 10.1073/pnas.022615199. [DOI] [PMC free article] [PubMed] [Google Scholar]
  15. Fischl B, Rajendran N, Busa E, Augustinack J, Hinds O, Yeo BT, Mohlberg H, Amunts K, Zilles K. Cortical folding patterns and predicting cytoarchitecture. Cerebral Cortex. 2008;18:1973–1980. doi: 10.1093/cercor/bhm225. [DOI] [PMC free article] [PubMed] [Google Scholar]
  16. Floyer-Lea A, Matthews PM. Distinguishable brain activation networks for short- and long-term motor skill learning. Journal of Neurophysiology. 2005;94:512–518. doi: 10.1152/jn.00717.2004. [DOI] [PubMed] [Google Scholar]
  17. Friston KJ, Holmes AP, Worsley KJ, Poline J-P, Frith CD, Frackowiak RSJ. Statistical parametric maps in functional imaging: a general linear approach. Human Brain Mapping. 1994;2:189–210. doi: 10.1002/hbm.460020402. [DOI] [Google Scholar]
  18. Galea JM, Vazquez A, Pasricha N, de Xivry JJ, Celnik P. Dissociating the roles of the cerebellum and motor cortex during adaptive learning: the motor cortex retains what the cerebellum learns. Cerebral Cortex. 2011;21:1761–1770. doi: 10.1093/cercor/bhq246. [DOI] [PMC free article] [PubMed] [Google Scholar]
  19. Grafton ST, Hazeltine E, Ivry R. Functional mapping of sequence learning in normal humans. Journal of Cognitive Neuroscience. 1995;7:497–510. doi: 10.1162/jocn.1995.7.4.497. [DOI] [PubMed] [Google Scholar]
  20. Grafton ST, Hazeltine E, Ivry RB. Motor sequence learning with the nondominant left hand a PET functional imaging study. Experimental Brain Research. 2002;146:369–378. doi: 10.1007/s00221-002-1181-y. [DOI] [PubMed] [Google Scholar]
  21. Graybiel AM. The basal ganglia. Current Biology. 2000;10:R509–R511. doi: 10.1016/S0960-9822(00)00593-5. [DOI] [PubMed] [Google Scholar]
  22. Graybiel AM, Grafton ST. The striatum: where skills and habits meet. Cold Spring Harbor Perspectives in Biology. 2015;7:a021691. doi: 10.1101/cshperspect.a021691. [DOI] [PMC free article] [PubMed] [Google Scholar]
  23. Gusnard DA, Akbudak E, Shulman GL, Raichle ME. Medial prefrontal cortex and self-referential mental activity: relation to a default mode of brain function. PNAS. 2001;98:4259–4264. doi: 10.1073/pnas.071043098. [DOI] [PMC free article] [PubMed] [Google Scholar]
  24. Hadipour-Niktarash A, Lee CK, Desmond JE, Shadmehr R. Impairment of retention but not acquisition of a visuomotor skill through time-dependent disruption of primary motor cortex. Journal of Neuroscience. 2007;27:13413–13419. doi: 10.1523/JNEUROSCI.2570-07.2007. [DOI] [PMC free article] [PubMed] [Google Scholar]
  25. Hardwick RM, Rottschy C, Miall RC, Eickhoff SB. A quantitative meta-analysis and review of motor learning in the human brain. NeuroImage. 2013;67:283–297. doi: 10.1016/j.neuroimage.2012.11.020. [DOI] [PMC free article] [PubMed] [Google Scholar]
  26. Hazeltine E, Grafton ST, Ivry R. Attention and stimulus characteristics determine the locus of motor-sequence encoding a PET study. Brain. 1997;120:123–140. doi: 10.1093/brain/120.1.123. [DOI] [PubMed] [Google Scholar]
  27. Hikosaka O, Nakahara H, Rand MK, Sakai K, Lu X, Nakamura K, Miyachi S, Doya K. Parallel neural networks for learning sequential procedures. Trends in Neurosciences. 1999;22:464–471. doi: 10.1016/S0166-2236(99)01439-3. [DOI] [PubMed] [Google Scholar]
  28. Honda M, Deiber MP, Ibáñez V, Pascual-Leone A, Zhuang P, Hallett M. Dynamic cortical involvement in implicit and explicit motor sequence learning a PET study. Brain. 1998;121:2159–2173. doi: 10.1093/brain/121.11.2159. [DOI] [PubMed] [Google Scholar]
  29. Huang Y, Zhen Z, Song Y, Zhu Q, Wang S, Liu J. Motor training increases the stability of activation patterns in the primary motor cortex. PLOS ONE. 2013;8:e53555. doi: 10.1371/journal.pone.0053555. [DOI] [PMC free article] [PubMed] [Google Scholar]
  30. Hutton C, Bork A, Josephs O, Deichmann R, Ashburner J, Turner R. Image distortion correction in fMRI: a quantitative evaluation. NeuroImage. 2002;16:217–240. doi: 10.1006/nimg.2001.1054. [DOI] [PubMed] [Google Scholar]
  31. Jenkins IH, Brooks DJ, Nixon PD, Frackowiak RS, Passingham RE. Motor sequence learning: a study with positron emission tomography. The Journal of Neuroscience. 1994;14:3775–3790. doi: 10.1523/JNEUROSCI.14-06-03775.1994. [DOI] [PMC free article] [PubMed] [Google Scholar]
  32. Jenkinson M, Beckmann CF, Behrens TE, Woolrich MW, Smith SM. FSL. NeuroImage. 2012;62:782–790. doi: 10.1016/j.neuroimage.2011.09.015. [DOI] [PubMed] [Google Scholar]
  33. Kang EK, Paik NJ. Effect of a tDCS electrode montage on implicit motor sequence learning in healthy subjects. Experimental & Translational Stroke Medicine. 2011;3:2–7. doi: 10.1186/2040-7378-3-4. [DOI] [PMC free article] [PubMed] [Google Scholar]
  34. Karni A, Meyer G, Jezzard P, Adams MM, Turner R, Ungerleider LG. Functional MRI evidence for adult motor cortex plasticity during motor skill learning. Nature. 1995;377:155–158. doi: 10.1038/377155a0. [DOI] [PubMed] [Google Scholar]
  35. Karni A, Meyer G, Rey-Hipolito C, Jezzard P, Adams MM, Turner R, Ungerleider LG. The acquisition of skilled motor performance: fast and slow experience-driven changes in primary motor cortex. PNAS. 1998;95:861–868. doi: 10.1073/pnas.95.3.861. [DOI] [PMC free article] [PubMed] [Google Scholar]
  36. Kawai R, Markman T, Poddar R, Ko R, Fantana AL, Dhawale AK, Kampff AR, Ölveczky BP. Motor cortex is required for learning but not for executing a motor skill. Neuron. 2015;86:800–812. doi: 10.1016/j.neuron.2015.03.024. [DOI] [PMC free article] [PubMed] [Google Scholar]
  37. Kumar N, Manning TF, Ostry DJ. Somatosensory cortex participates in the consolidation of human motor memory. PLOS Biology. 2019;17:e3000469. doi: 10.1371/journal.pbio.3000469. [DOI] [PMC free article] [PubMed] [Google Scholar]
  38. Lashley K. Society for Experimental Biology, Physiological Mechanisms in Animal Behavior. Academic Press; 1950. Insearch of the engram; pp. 454–482. [Google Scholar]
  39. Lehéricy S, Benali H, Van de Moortele PF, Pélégrini-Issac M, Waechter T, Ugurbil K, Doyon J. Distinct basal ganglia territories are engaged in early and advanced motor sequence learning. PNAS. 2005;102:12566–12571. doi: 10.1073/pnas.0502762102. [DOI] [PMC free article] [PubMed] [Google Scholar]
  40. Lehéricy S, Bardinet E, Tremblay L, Van de Moortele PF, Pochon JB, Dormont D, Kim DS, Yelnik J, Ugurbil K. Motor control in basal ganglia circuits using fMRI and brain atlas approaches. Cerebral Cortex. 2006;16:149–161. doi: 10.1093/cercor/bhi089. [DOI] [PubMed] [Google Scholar]
  41. Lutz K, Koeneke S, Wüstenberg T, Jäncke L. Asymmetry of cortical activation during maximum and convenient tapping speed. Neuroscience Letters. 2004;373:61–66. doi: 10.1016/j.neulet.2004.09.058. [DOI] [PubMed] [Google Scholar]
  42. Ma L, Wang B, Narayana S, Hazeltine E, Chen X, Robin DA, Fox PT, Xiong J. Changes in regional activity are accompanied with changes in inter-regional connectivity during 4 weeks motor learning. Brain Research. 2010;1318:64–76. doi: 10.1016/j.brainres.2009.12.073. [DOI] [PMC free article] [PubMed] [Google Scholar]
  43. Makino H, Ren C, Liu H, Kim AN, Kondapaneni N, Liu X, Kuzum D, Komiyama T. Transformation of Cortex-wide emergent properties during motor learning. Neuron. 2017;94:880–890. doi: 10.1016/j.neuron.2017.04.015. [DOI] [PMC free article] [PubMed] [Google Scholar]
  44. Matsuzaka Y, Picard N, Strick PL. Skill Representation in the Primary Motor Cortex After Long-Term Practice. Journal of Neurophysiology. 2007;97:1819–1832. doi: 10.1152/jn.00784.2006. [DOI] [PubMed] [Google Scholar]
  45. Mattar M, Wymbs NF, Bock AS, Aguirre GK, Grafton ST, Bassett DS. Predicting future learning from baseline network architecture. bioRxiv. 2016 doi: 10.1101/056861. [DOI] [PMC free article] [PubMed]
  46. Nitsche MA, Schauenburg A, Lang N, Liebetanz D, Exner C, Paulus W, Tergau F. Facilitation of implicit motor learning by weak transcranial direct current stimulation of the primary motor cortex in the human. Journal of Cognitive Neuroscience. 2003;15:619–626. doi: 10.1162/089892903321662994. [DOI] [PubMed] [Google Scholar]
  47. Oosterhof NN, Wiestler T, Downing PE, Diedrichsen J. A comparison of volume-based and surface-based multi-voxel pattern analysis. NeuroImage. 2011;56:593–600. doi: 10.1016/j.neuroimage.2010.04.270. [DOI] [PubMed] [Google Scholar]
  48. Orban P, Peigneux P, Lungu O, Albouy G, Breton E, Laberenne F, Benali H, Maquet P, Doyon J. The multifaceted nature of the relationship between performance and brain activity in motor sequence learning. NeuroImage. 2010;49:694–702. doi: 10.1016/j.neuroimage.2009.08.055. [DOI] [PubMed] [Google Scholar]
  49. Ostry DJ, Gribble PL. Sensory plasticity in human motor learning. Trends in Neurosciences. 2016;39:114–123. doi: 10.1016/j.tins.2015.12.006. [DOI] [PMC free article] [PubMed] [Google Scholar]
  50. Otchy TM, Wolff SBE, Rhee JY, Pehlevan C, Kawai R, Kempf A, Gobes SMH, Ölveczky BP. Acute off-target effects of neural circuit manipulations. Nature. 2015;528:358–363. doi: 10.1038/nature16442. [DOI] [PubMed] [Google Scholar]
  51. Penhune VB, Doyon J. Dynamic cortical and subcortical networks in learning and delayed recall of timed motor sequences. The Journal of Neuroscience. 2002;22:1397–1406. doi: 10.1523/JNEUROSCI.22-04-01397.2002. [DOI] [PMC free article] [PubMed] [Google Scholar]
  52. Penhune VB, Steele CJ. Parallel contributions of cerebellar, striatal and M1 mechanisms to motor sequence learning. Behavioural Brain Research. 2012;226:579–591. doi: 10.1016/j.bbr.2011.09.044. [DOI] [PubMed] [Google Scholar]
  53. Peters AJ, Lee J, Hedrick NG, O'Neil K, Komiyama T. Reorganization of corticospinal output during motor learning. Nature Neuroscience. 2017;20:1133–1141. doi: 10.1038/nn.4596. [DOI] [PMC free article] [PubMed] [Google Scholar]
  54. Picard N, Matsuzaka Y, Strick PL. Extended practice of a motor skill is associated with reduced metabolic activity in M1. Nature Neuroscience. 2013;16:1340–1347. doi: 10.1038/nn.3477. [DOI] [PMC free article] [PubMed] [Google Scholar]
  55. Pinsard B, Boutin A, Gabitov E, Lungu O, Benali H, Doyon J. Consolidation alters motor sequence-specific distributed representations. bioRxiv. 2018 doi: 10.7554/eLife.39324. [DOI] [PMC free article] [PubMed]
  56. Poldrack RA. Imaging brain plasticity: conceptual and methodological issues--a theoretical review. NeuroImage. 2000;12:1–13. doi: 10.1006/nimg.2000.0596. [DOI] [PubMed] [Google Scholar]
  57. Poldrack RA, Sabb FW, Foerde K, Tom SM, Asarnow RF, Bookheimer SY, Knowlton BJ. The neural correlates of motor skill automaticity. Journal of Neuroscience. 2005;25:5356–5364. doi: 10.1523/JNEUROSCI.3880-04.2005. [DOI] [PMC free article] [PubMed] [Google Scholar]
  58. Raichle ME, MacLeod AM, Snyder AZ, Powers WJ, Gusnard DA, Shulman GL. A default mode of brain function. PNAS. 2001;98:676–682. doi: 10.1073/pnas.98.2.676. [DOI] [PMC free article] [PubMed] [Google Scholar]
  59. Reis J, Schambra HM, Cohen LG, Buch ER, Fritsch B, Zarahn E, Celnik PA, Krakauer JW. Noninvasive cortical stimulation enhances motor skill acquisition over multiple days through an effect on consolidation. PNAS. 2009;106:1590–1595. doi: 10.1073/pnas.0805413106. [DOI] [PMC free article] [PubMed] [Google Scholar]
  60. Reithler J, van Mier HI, Goebel R. Continuous motor sequence learning: cortical efficiency gains accompanied by striatal functional reorganization. NeuroImage. 2010;52:263–276. doi: 10.1016/j.neuroimage.2010.03.073. [DOI] [PubMed] [Google Scholar]
  61. Russo AA, Bittner SR, Perkins SM, Seely JS, London BM, Lara AH, Miri A, Marshall NJ, Kohn A, Jessell TM, Abbott LF, Cunningham JP, Churchland MM. Motor cortex embeds Muscle-like commands in an untangled population response. Neuron. 2018;97:953–966. doi: 10.1016/j.neuron.2018.01.004. [DOI] [PMC free article] [PubMed] [Google Scholar]
  62. Russo AA, Khajeh R, Bittner SR, Perkins SM, Cunningham JP, Abbott LF, Churchland MM. Neural trajectories in the supplementary motor area and primary motor cortex exhibit distinct geometries compatible with different classes of computation. bioRxiv. 2019 doi: 10.1101/650002. [DOI] [PMC free article] [PubMed]
  63. Seidler RD, Purushotham A, Kim SG, Uğurbil K, Willingham D, Ashe J. Cerebellum activation associated with performance change but not motor learning. Science. 2002;296:2043–2046. doi: 10.1126/science.1068524. [DOI] [PubMed] [Google Scholar]
  64. Seidler RD, Purushotham A, Kim SG, Ugurbil K, Willingham D, Ashe J. Neural correlates of encoding and expression in implicit sequence learning. Experimental Brain Research. 2005;165:114–124. doi: 10.1007/s00221-005-2284-z. [DOI] [PubMed] [Google Scholar]
  65. Shamloo F, Helie S. Changes in default mode network as automaticity develops in a categorization task. Behavioural Brain Research. 2016;313:324–333. doi: 10.1016/j.bbr.2016.07.029. [DOI] [PubMed] [Google Scholar]
  66. Shulman GL, Fiez JA, Corbetta M, Buckner RL, Miezin FM, Raichle ME, Petersen SE. Common blood flow changes across visual tasks: ii. decreases in cerebral cortex. Journal of Cognitive Neuroscience. 1997;9:648–663. doi: 10.1162/jocn.1997.9.5.648. [DOI] [PubMed] [Google Scholar]
  67. Steele CJ, Penhune VB. Specific increases within global decreases: a functional magnetic resonance imaging investigation of five days of motor sequence learning. Journal of Neuroscience. 2010;30:8332–8341. doi: 10.1523/JNEUROSCI.5569-09.2010. [DOI] [PMC free article] [PubMed] [Google Scholar]
  68. Tamás Kincses Z, Johansen-Berg H, Tomassini V, Bosnell R, Matthews PM, Beckmann CF. Model-free characterization of brain functional networks for motor sequence learning using fMRI. NeuroImage. 2008;39:1950–1958. doi: 10.1016/j.neuroimage.2007.09.070. [DOI] [PubMed] [Google Scholar]
  69. Toni I, Krams M, Turner R, Passingham RE. The time course of changes during motor sequence learning: a whole-brain fMRI study. NeuroImage. 1998;8:50–61. doi: 10.1006/nimg.1998.0349. [DOI] [PubMed] [Google Scholar]
  70. Ungerleider LG, Doyon J, Karni A. Imaging brain plasticity during motor skill learning. Neurobiology of Learning and Memory. 2002;78:553–564. doi: 10.1006/nlme.2002.4091. [DOI] [PubMed] [Google Scholar]
  71. Vahdat S, Lungu O, Cohen-Adad J, Marchand-Pauvert V, Benali H, Doyon J. Simultaneous Brain-Cervical cord fMRI reveals intrinsic spinal cord plasticity during motor sequence learning. PLOS Biology. 2015;13:e1002186. doi: 10.1371/journal.pbio.1002186. [DOI] [PMC free article] [PubMed] [Google Scholar]
  72. Walther A, Nili H, Ejaz N, Alink A, Kriegeskorte N, Diedrichsen J. Reliability of dissimilarity measures for multi-voxel pattern analysis. NeuroImage. 2016;137:188–200. doi: 10.1016/j.neuroimage.2015.12.012. [DOI] [PubMed] [Google Scholar]
  73. Waters-Metenier S, Husain M, Wiestler T, Diedrichsen J. Bihemispheric transcranial direct current stimulation enhances effector-independent representations of motor synergy and sequence learning. The Journal of Neuroscience. 2014;34:1037–1050. doi: 10.1523/JNEUROSCI.2282-13.2014. [DOI] [PMC free article] [PubMed] [Google Scholar]
  74. Wiestler T, Diedrichsen J. Skill learning strengthens cortical representations of motor sequences. eLife. 2013;2:e00801. doi: 10.7554/eLife.00801. [DOI] [PMC free article] [PubMed] [Google Scholar]
  75. Wu T, Kansaku K, Hallett M. How self-initiated memorized movements become automatic: a functional MRI study. Journal of Neurophysiology. 2004;91:1690–1698. doi: 10.1152/jn.01052.2003. [DOI] [PubMed] [Google Scholar]
  76. Wymbs NF, Grafton ST. The human motor system supports Sequence-Specific representations over multiple Training-Dependent timescales. Cerebral Cortex. 2015;25:4213–4225. doi: 10.1093/cercor/bhu144. [DOI] [PMC free article] [PubMed] [Google Scholar]
  77. Xiong J, Ma L, Wang B, Narayana S, Duff EP, Egan GF, Fox PT. Long-term motor training induced changes in regional cerebral blood flow in both task and resting states. NeuroImage. 2009;45:75–82. doi: 10.1016/j.neuroimage.2008.11.016. [DOI] [PMC free article] [PubMed] [Google Scholar]
  78. Yin HH, Mulcare SP, Hilário MR, Clouse E, Holloway T, Davis MI, Hansson AC, Lovinger DM, Costa RM. Dynamic reorganization of striatal circuits during the acquisition and consolidation of a skill. Nature Neuroscience. 2009;12:333–341. doi: 10.1038/nn.2261. [DOI] [PMC free article] [PubMed] [Google Scholar]
  79. Yokoi A, Arbuckle SA, Diedrichsen J. The role of human primary motor cortex in the production of skilled finger sequences. The Journal of Neuroscience. 2018;38:1430–1442. doi: 10.1523/JNEUROSCI.2798-17.2017. [DOI] [PMC free article] [PubMed] [Google Scholar]
  80. Yokoi A, Diedrichsen J. Neural organization of hierarchical motor sequence representations in the human neocortex. Neuron. 2019;103:1178–1190. doi: 10.1016/j.neuron.2019.06.017. [DOI] [PubMed] [Google Scholar]
  81. Yousry TA, Schmid UD, Alkadhi H, Schmidt D, Peraud A, Buettner A, Winkler P. Localization of the motor hand area to a knob on the Precentral Gyrus a new landmark. Brain. 1997;120:141–157. doi: 10.1093/brain/120.1.141. [DOI] [PubMed] [Google Scholar]

Decision letter

Editor: Marius V Peelen1
Reviewed by: Aaron L Wong2

In the interests of transparency, eLife publishes the most substantive revision requests and the accompanying author responses.

Acceptance summary:

This longitudinal fMRI study used a large sample, preregistered methods, and advanced analyses to test previous, sometimes conflicting, findings in the domain of motor sequence learning. The study provides convincing evidence for motor sequence learning outside of primary motor cortex. The work is a tour-de-force involving four scanning sessions and thousands of behavioural practice trials.

Decision letter after peer review:

Thank you for submitting your article "A critical re-evaluation of fMRI signatures of motor sequence learning" for consideration by eLife. Your article has been reviewed by three peer reviewers, one of whom is a member of our Board of Reviewing Editors, and the evaluation has been overseen by Floris de Lange as the Senior Editor. The following individual involved in review of your submission has agreed to reveal their identity: Aaron L Wong (Reviewer #2).

The reviewers have discussed the reviews with one another and the Reviewing Editor has drafted this decision to help you prepare a revised submission.

Summary:

This paper reports a longitudinal fMRI study investigating motor sequence learning. The study aimed to verify previous (sometimes conflicting) findings using a relatively large sample and preregistered methods. The study also tested new hypotheses and directly compared speeded and paced performance of trained and untrained sequences.

The authors find that, consistent with their previous work, they see no evidence of sequence-specific changes in M1; in association cortices, they observe average activity decreases associated with sequence learning, with sequence-specific activity patterns changing primarily early in learning.

All reviewers were impressed with the quality of the design and analysis of the study, and agreed that it will provide an important contribution to the literature. The work is a tour-de-force involving four scanning sessions and thousands of behavioral practice trials. It is great, therefore, that this valuable dataset will be made publicly available, such that other researchers can test their own hypotheses as well.

Essential revisions:

1) While the authors have carefully controlled for movement speed, it isn't clear whether they also took into consideration two other related factors that likely affect performance and neural activity: the error rate, and the amount of reward (i.e., points) earned. Given that MT in the first test session is typically faster than during initial training, individuals seem likely to make a fair number of errors in Week 1 that would largely decrease with training (since MT in week 2 is now faster in training than during test, it seems likely that the largest change in error/reward rate occurs between Weeks 1-2). It is important to know what performance accuracy was like across weeks, if incorrect trials are excluded from the fMRI analysis, and whether error/reward rate might have contributed to changes in neural activity (e.g., in the striatum).

2) On a related note, it is strange that individuals were performing nearly 1 second faster during Test 1 compared to the start of training (i.e., Figure 2 comparing panel C and D) despite a lack of pacing in both cases. Was this slowdown systematic across individuals, and what would have caused it? Likewise, it is curious that MTs in the initial test session is so much faster than the initial training session (but in following weeks MT is on par or slower in the testing session compared to training). This seems to call into question the accuracy of the baseline measurements taken in Week 1 prior to training. It would be helpful for the authors to comment on this.

3) Finally, it is curious that some patterns of activity only become evident at full speed (e.g., in the striatum). Does this mean that, akin to a model where neural activity scales with movement speed, moving faster simply pulls the same signal out of the noise better (which seems to be true in cortex)? Or does this imply that moving at peak speed actually changes the patterns of neural activity in some way? The authors address this question for cortex (subsection “Trained sequences elicit distinct patterns during full speed performance”, third paragraph), but should perform a similar analysis for the striatum to know how similar sequence representation patterns are throughout the brain. This would also speak to the authors' discussion of the role of the striatum in motor sequence learning.

eLife. 2020 May 13;9:e55241. doi: 10.7554/eLife.55241.sa2

Author response


Essential revisions:

1) While the authors have carefully controlled for movement speed, it isn't clear whether they also took into consideration two other related factors that likely affect performance and neural activity: the error rate, and the amount of reward (i.e., points) earned. Given that MT in the first test session is typically faster than during initial training, individuals seem likely to make a fair number of errors in Week 1 that would largely decrease with training (since MT in week 2 is now faster in training than during test, it seems likely that the largest change in error/reward rate occurs between Weeks 1-2). It is important to know what performance accuracy was like across weeks, if incorrect trials are excluded from the fMRI analysis, and whether error/reward rate might have contributed to changes in neural activity (e.g., in the striatum).

Thank you for raising this point. For the fMRI analysis presented in the original submission, we did not exclude errors trials, which has been clarified in text now. This is an approach we’ve commonly taken in our previous papers (e.g. Yokoi and Diedrichsen, 2019; Yokoi et al., 2018; Ejaz et al., 2015, etc.), as the dominant form of error in this paradigm are small execution errors (i.e., finger failed to reach force threshold), such that most of the normal planning- and execution related activity is likely preserved. Exclusion of errors introduces an unequal trial number across conditions and runs, which leads to decreased power and can also introduce biases into the multivariate analysis. However, it is a serious concern and we therefore have conducted a range of additional analyses to confirm that errors did not affect our cross-session correlation analyses.

First off, the reward rate inside the scanner was purely dependent on correctness of execution – i.e. if inside the scanner participants performed the sequence correctly, they received 3 points, and 0 otherwise. This differs from the approach we took for reward system in training sessions, where the additional criterion was the speed of performance, to keep participants motivated throughout learning. Since scans 1-3 were performed at a paced speed, we omitted this speed-dependence on reward in all scanning sessions. This also prevented a difference in reward / points given between paced and full speed sessions.

During behavioral training, participants were instructed to go as fast as possible, while keeping the error rate below 15%. Errors were more frequent in the first two sessions (which was mainly due to the second trial where they had to perform sequences from memory – see response to the 2nd point), but overall hovered around 15%, as shown in Author response image 1A. As in behavioral sessions, participants were told to keep their error rate below 15% also in scanning sessions. However, the number of errors executed varied significantly across sessions, and between sequence types (see Author response image 1B). A session x sequence type ANOVA was significant for both session (F(3,75)=9.21, p=2.93e-5) and sequence type (F(1,25)=11.15, p=2.64e-3). There was also a significant interaction (F(3,75)=8.42, p=6.75e-5). Post-hoc t-tests revealed that the error rate differed between trained and untrained sequences in week 2 (t(25)=4.22, p=2.84e-4) and week 5 (t(25)=4.81, p=6.1e-5), but not in week 1, or in the speeded session of week 5 (5* in Author response image 1B). Comparing error rate across sessions, week 1 had overall more errors than week 2 and 5 (t(25)>=2.94, p<=0.007). Similarly, errors were more frequent in speeded session than the paced sessions (week 2 and 5) (t(25)>=2.40, p<=0.024). The speeded session and week 1, however, did not differ significantly in terms of number of errors.

Author response image 1. Error rate in behavioral and scanning sessions.

Author response image 1.

a) Percentage of erroneous trials for trained sequences throughout behavioral training days. b) Percentage of errors for trained (red) and untrained (blue) sequences in the scanning sessions across weeks 1, 2 and 5 (1, 2, 5 at paced speed, 5* at full speed).

Since errors indeed differed between the paced session in week 1 and 2, and were different for trained than untrained sequences, the concern you are bringing is valid – it could be that differences in pattern correlation across sessions are due to differences in the number of errors across the two sessions. To addressed this concern, we completely re-did the fMRI analysis, starting with a first level general linear model that estimated the activity patterns for each sequence and run using only the correct trials. Error trials were modelled by a separate regressor that was not used in subsequent analyses. We then re-did the PCM correlation analysis on these newly estimated betas. The general pattern of observed correlation was very similar to the previous analysis (Author response image 2A – with error trials, Author response image 2B – without error trials).

Author response image 2. Relative log-evidence (log marginal likelihood) for different correlation models across sessions (week 1-2 in solid lines, week 2-5 in dotted lines) for trained (red) and untrained (blue) activity patterns.

Author response image 2.

Regions: PMd (dorsal premotor cortex), SPLa (superior parietal lobule – anterior part), M1 (primary motor cortex), S1 (primary somatosensory cortex). Difference between log-likelihoods can be interpreted as log-Bayes factor, with a difference of 1 indicating positive evidence. Vertical lines indicate the winning correlation model for trained (red) and untrained (blue) patterns across weeks 1-2. Black dots are projections of the two winning models onto the correlation function of trained sequences (weeks 1-2). The horizontal lines from the two black dots indicate the likelihood of the trained data under the two models, which was tested in a cross validated t-test. a) Evidence based on preprocessing, reported in manuscript (here just adjusted in y-axis. b) Evidence for correlations after removing error trials from a general linear model estimation.

To statistically quantify the results, we performed a crossvalidated t-test, as in the original submission. We used a one-sided threshold for significance since our hypothesis was directional, predicting that trained patterns would display a lower correlation across weeks 1-2 than untrained ones. As with the original analysis, we observed that trained patterns displayed significantly lower correlation across weeks 1 and 2 than untrained patterns in PMd (t(25)=1.72, p=.048) and SPLa (t(25)=1.71, p=0.049). This difference between trained and untrained patterns across weeks 1-2 was not significant for M1 (t(25)=0.87, p=0.40), or S1 (t(25)=0.53, p=0.15). Additionally, as reported previously, this difference in correlations between trained and untrained patterns was no longer present across weeks 2 and 5 (PMd: t(25)=0.08, p=0.46; SPLa: t(25)=0.96, p=.17). Although the significance of the effects is not as strong as with the errors included, the mean estimates of the most likely correlation did not change systematically when excluding errors. Therefore, we believe that this reduction in significance is caused mainly by the increased variability of the estimates caused be the inclusion of fewer trials into the analysis.

We then systematically inspected the other reported results, to determine if they could have been affected by difference in error rate between trained and untrained sequences in the paced sessions in weeks 2 and 5. This observed pattern of errors could have contributed to differences in overall evoked activity in premotor and parietal areas. Namely, we observed reduced activation for trained sequences in weeks 2 and 5 (Author response image 3A), where also the number of errors was lower for trained sequences. Even after removing error trials, activation was still significantly lower in PMd and SPLa in weeks 2 and 5 for trained than untrained sequences (Author response image 3A – with error trials, Author response image 3B – without error trials), suggesting that the observed reduced activation was not due to lower error rates for trained sequences.

Author response image 3. Elicited activation, measured as percent signal change over resting baseline across weeks for trained (red) and untrained (blue) sequences.

Author response image 3.

Regions: M1 (primary motor cortex), S1 (primary somatosensory cortex), PMd (dorsal premotor cortex), SPLa (superior parietal lobule – anterior part). a) Estimated elicited activation based on preprocessing, reported in manuscript. b) Estimated elicited activation after removing error trials from a general linear model estimation.

Similarly, the results concerning the cosine angle dissimilarity between average trained and untrained sequences were largely unaffected by difference in error rate in weeks 2 and 5 (Author response image 4) – for both PMd and SPLa the cosine angle increased significantly across weeks. The only difference was that previously significant increase in average cosine angle from week 1 to 2 in S1 was no longer significant (t(25)=1.97, p=0.060). Already in the original analysis the increase across weeks was smaller in S1 than in PMd and SPLa.

Author response image 4. Dissimilarity between average trained and untrained sequence, measured with cosine angle.

Author response image 4.

Regions: M1 (primary motor cortex), S1 (primary somatosensory cortex), PMd (dorsal premotor cortex), SPLa (superior parietal lobule – anterior part). a) Estimated elicited activation based on preprocessing, reported in manuscript. b) Estimated elicited activation after removing error trials from a general linear model estimation.

Finally, since the error rate did not differ significantly between trained and untrained sequences the speeded session of week 5, our results of higher dissimilarities for trained than untrained sequences at full speed cannot be due to difference in error or reward. Similarly, in the striatum, we observed significant dissimilarities only for trained sequences in session 4, which thus cannot be driven by error/reward rate.

We have now clarified treatment of errors in our analysis pipeline in the manuscript.

Changes made:

“All instances of sequence execution were included into estimating regressors, regardless of whether the execution was correct or erroneous (see section Treatment of error trials below).”

“Treatment of error trials

As in behavioral sessions, participants were instructed to keep their error rate below 15% also inside the scanner. […] Indeed, our results did not differ qualitatively when excluding errors, therefore we here report only the analyses with all trials included.”

We have also added an explicit description of reward rate for experimental design during scanning.

Changes made:

“Participants always received 3 points upon correct execution of the sequence, and 0 points otherwise.”

Additionally, we have commented on the results of analyses without error trials in the main Results section, and referred the reader to the newly added Figure 3—figure supplement 1, and Figure 6—figure supplement 1A.

Changes made:

“This pattern of results on changes in overall activation remained unchanged after excluding error trials from the analyses (see Figure 3—figure supplement 1A).”

“This observed increase in dissimilarity between average trained and untrained pattern in PMd and SPLa, and to a lesser extent in S1, was also observed when analyzing only trials with correct performance (see Figure 3—figure supplement 1B).”

“To ensure that the observed lower correlation for trained patterns was not due to larger difference in error rate between weeks 1 and 2 for trained than for untrained sequences, we repeated the analysis excluding error trials. […] In S1, after accounting for error trials the correlation across weeks 1-2 did no longer differ between trained and untrained patterns.”

2) On a related note, it is strange that individuals were performing nearly 1 second faster during Test 1 compared to the start of training (i.e., Figure 2 comparing panel C and D) despite a lack of pacing in both cases. Was this slowdown systematic across individuals, and what would have caused it? Likewise, it is curious that MTs in the initial test session is so much faster than the initial training session (but in following weeks MT is on par or slower in the testing session compared to training). This seems to call into question the accuracy of the baseline measurements taken in Week 1 prior to training. It would be helpful for the authors to comment on this.

The main reason for the slowdown was the fact that in training, participants performed the sequences from memory. Specifically, each sequence presented twice: during the first execution, the full 9 digits were presented, whereas the second execution needed to be done from memory. We adopted this approach to encourage participants to form an explicit memory of the sequences, as we have done in several previous studies (Yokoi and Diedrichsen, 2019; Yokoi et al., 2018; Popp et al., 2019). This means that, especially early in learning, participants had to simultaneously memorize a 9-digit sequence while performing it for the first time, and then execute it from memory the second time around. In contrast, during testing and scanning sessions, where both trained and untrained sequences had to be performed, the full 9-digit sequences were always presented, thereby removing memory requirements. Thus, any comparison between training (Figure 2C) and testing (Figure 2D) needs to be made with caution.

The extra load to memorize the sequence slowed participants down substantially on the first day. Participants were significantly slower on the second, memorized, execution (Author response image 5A; t(25)=3.17, p=.004), and made more errors on the second execution (Author response image 5B). However, it is very likely that participants also slowed down the first execution to be able to better encode the sequence. Therefore, test session 1 better indicates the speed participants were able to perform at motorically. Already on training day 2, there was no difference in speed for first and second execution.

Author response image 5. Changes in movement time and error in behavioral training sessions, split between executions 1 and 2.

Author response image 5.

a) Movement time to execute trained sequences across training days, split between execution 1 – digits presented on the screen, and execution 2 – from memory.

b) Error rate, split between executions 1 and 2 for training days.

While our behavioral training sessions were designed to motivate participants to learn specific trained sequences, behavioral test sessions were in place to compare performance of trained to untrained sequences. Those two types of sessions therefore required different design choices. Importantly, none of our results or conclusions rely on the behavioral performance during training. We have now clarified some of this confusion in the main text.

“During the behavioral test sessions (Figure 2D), participants executed both the trained sequences, as well as matched untrained sequences, with all sequences randomly interspersed. […] Therefore, the requirement to remember the sequences from the first to second execution, which was present in training sessions, was omitted for test sessions. For this reason, performance in training and test sessions (Figure 2C-D) cannot be directly compared.”

3) Finally, it is curious that some patterns of activity only become evident at full speed (e.g., in the striatum). Does this mean that, akin to a model where neural activity scales with movement speed, moving faster simply pulls the same signal out of the noise better (which seems to be true in cortex)? Or does this imply that moving at peak speed actually changes the patterns of neural activity in some way? The authors address this question for cortex (subsection “Trained sequences elicit distinct patterns during full speed performance”, third paragraph), but should perform a similar analysis for the striatum to know how similar sequence representation patterns are throughout the brain. This would also speak to the authors' discussion of the role of the striatum in motor sequence learning.

This is a very interesting point. It could be that while in the cortical regions the patterns during full speed performance just become clearer, yet still qualitatively similar to those in the paced performance, the striatum shows a qualitative shift in responses from paced to full speed performance. As you are mentioning, our analysis of dissimilarities alludes to this possibility since sequence-specific striatal patterns were only significant for trained sequences during full speed execution. A correlation of sequence-specific patterns across the paced and full speed sessions in week 5 would allow us to address this question more directly. The results of the analysis are presented in Author response image 6. Similar to the regions in the cortex, there was some correlation of activity patterns across paced and full speed sessions – this correlation was clearly higher than 0 (putamen: t(25)=9.56, p=8.0e-10; caudate: t(25)=6.37, p=1.1e-6), indicating some consistency across different speeds. On the other hand the correlation was, even after accounting for measurement noise, lower than 1 (putamen: t(25)=8.85, p=3.6e-9; caudate: t(25)=5.86, p=4.1e-6), suggesting a modest change across speeds. Thus, although the correspondence wasn’t perfect, the results nevertheless suggests that, similarly to cortical regions, performance at full speed makes fine-grained changes in activity more pronounced, i.e. pulls the signal out of the noise.

Author response image 6. Evidence for different correlation models, across paced and full speed sessions in week 5, for trained (red) and untrained (blue) patterns in putamen and caudate.

Author response image 6.

Model evidence was measured as a log-Bayes factor relative to the average log-Bayes across all models (horizontal line at 0). The red and blue lines are drawn for correlation model with highest evidence for trained and untrained patterns, respectively. Vertical lines indicate winning model for trained and untrained sequences. Horizontal lines indicate the difference in log-likelihood of the trained data under these two model, which was evaluated in a cross validated fashion.

Also similar to what we’ve observed in cortical regions, there was no significant evidence that the correlation differed between trained and untrained sequences. Specifically, the evidence of the trained patterns under the winning model for trained and untrained sequences (assessed in cross validated fashion) was not significant different – this was the case for both the putamen (t(25)=0.58, p=.57) and caudate (t(25)=1.43, p=.17). While based on the figure, there seems to be an interaction in the observed correlation shapes for trained and untrained sequences across the two regions, this was also not significant, as determined by a region x sequence type ANOVA (region: F(1,25)=0.19, p=.66; sequence type: F(1,25)=0.054, p=.83, interaction: F(1,25)=0.92, p=.35).

We have now added this analysis to our Results section on the striatal changes with learning.

Changes made in the Results section:

“To examine whether the speed purely pulls the signal out of the noise better, or qualitatively changes the representation, we, similarly to the analyses in the cortical regions, performed the PCM correlation model across the paced and full speed sessions in week 5. […] This suggests that moving faster engages similar representations as moving slower, but helps to increase the signal-to-noise ratio.”

Changes made in the Discussion:

“Additionally, our results suggest that full speed performance might be particularly important for further studies of striatal multivariate activation to counteract the generally lower signal-to-noise ratio in this region.”

Associated Data

    This section collects any data citations, data availability statements, or supplementary materials included in this article.

    Data Citations

    1. Berlot E, Nicola JP, Diedrichsen J. 2020. Motor sequence learning. OpenNeuro. ds002776

    Supplementary Materials

    Transparent reporting form

    Data Availability Statement

    fMRI data and analysis pipelines have been deposited to OpenNeuro, under the accession number ds002776. Analysis code is available on GitHub at https://github.com/eberlot/motor_sequence_learning.git (copy archived at https://github.com/elifesciences-publications/motor_sequence_learning).

    The following dataset was generated:

    Berlot E, Nicola JP, Diedrichsen J. 2020. Motor sequence learning. OpenNeuro. ds002776


    Articles from eLife are provided here courtesy of eLife Sciences Publications, Ltd

    RESOURCES