Motor sequence learning task.

Subjects were scanned while performing a motor sequence learning (MSL) task that was adapted from the sequential finger-tapping paradigm developed by Karni and colleagues [40] and coded in Cogent2000 (http://www.vislab.ucl.ac.uk/cogent.php) using MATLAB (Mathworks Inc., Sherbom, MA). A custom MR-compatible ergonomic response pad comprising four push buttons located in a horizontal row was used. First, the experimenter demonstrated the task using a slide show presentation and explained all instructions prior to participants being placed in the scanner. Once in the scanner, the session included a brief pre-training phase, where subjects were asked to repeatedly and slowly perform a 5-item sequence of finger movements (4-1-3-2-4, where 1 stands for the index finger and 4 for the little finger) using their non-dominant hand until they were able to consecutively reproduce three correct 5-item sequences. Subjects were not scanned during this pre-training, nor was it included in subsequent analyses. The pre-training procedure was simply intended to ensure that subjects had placed their hand correctly on the keypad, understood the instructions, explicitly memorized the sequence of finger movements, and could perform the 5-item sequence correctly.

For the training portion of the MSL task used in subsequent data analyses, participants were then instructed to perform the same explicitly learned sequence by tapping the finger movements as quickly as possible, while making as few errors as possible during blocks of practice. When an error occurred, participants were instructed to continue at the beginning of the 5-item sequence to ensure that practice continued in an uninterrupted fashion within each practice block. The training and retest practice sessions consisted of 14 blocks of practice (indicated by a green cross displayed in the center of the screen). Each block terminated after 60 key presses (equivalent to twelve repetitions of the sequence). Each practice block was separated by a 15 s rest period (indicated by a red cross) during which subjects were asked to keep their fingers immobile. The individual key presses were un-cued (i.e., at a self-initiated pace). Given the explicit and simple nature of the sequence, all subjects performed the sequence with a very high level of accuracy ≥ 83%, corresponding to more than 10/12 correct sequences per block. The timing of all key presses were also recorded and speed was measured by the inter-key-press interval for correct responses only. To be consistent with previous studies [2,16,41,42], gains in performance from training to retest (taken as an indicator of offline memory processing, hence consolidation) were assessed by the difference in the mean speed of the last four blocks in the training session and the mean speed of the first four blocks in the retest session.

Motor performance control task.

The same four-button MR-safe response box was used for the CTRL task. In the MSL task, the 5-item sequence was produced by pressing each button individually, whereas in the CTRL task, all four buttons were repeatedly pressed at the same time, and at the same average pace as the MSL task. This task was designed to have the same motor performance characteristics of the MSL task (e.g., same number of finger flexion movements, same average inter-key press interval, all in the same amount of time), but importantly, without any sequence to learn and for motor movements that were self-initiated (as opposed to externally cued). Similar to the MSL task, the CTRL task was administered in two phases. Prior to being placed in the scanner, the experimenter demonstrated the task using a slide show presentation and verbally gave all instructions. Participants were then placed in the scanner. The session began with a pre-training phase where subjects were not scanned and instructed to press all four keys simultaneously following the rhythm of an auditory tone (presented monotonically at 3 Hz) as long as a green cross was displayed on the screen. This first step of the pre-training phase consisted of three blocks of practice. Each block terminated after 60 simultaneous 4-key presses. Practice blocks were separated by 15-second blocks of rest (indicated by a red cross displayed on the screen). This first pre-training step was intended to entrain subjects to the average speed of performance (~3 Hz) normally observed for the MSL task [43]. The exact same procedure was used in the second step of the pre-training phase, but this time, in the absence of the audio tones. Here, participants were instructed to maintain the same rhythm as performed in the prior three blocks of practice in the first step of the pre-training phase. The speed of performance (in Hz) was visually displayed on-screen in red text during rest blocks in order to provide feedback to the subjects on their self-initiated speed of performance. Unknown to the subjects, however, this step of the pre-training phase was terminated once performance was maintained at 3 Hz (± 0.25 Hz) for three consecutive practice blocks. This pre-training phase ensured that subjects could reliably press all four keys simultaneously at the target rhythm. Similar to the MSL task, this pre-training was not included in subsequent analyses.

For the actual practice sessions with the CTRL task, participants were instructed to follow the same rhythm as practiced during the pre-training phase, and to rest during the presentation of the red cross. Again, each of the 14 practice blocks terminated after 60 simultaneous 4-key presses, and each intervening rest period lasted 15 seconds. Similar to the MSL task, the evening CTRL training session and morning CTRL retest session were identical. To be consistent with the MSL task, performance on the CTRL task was measured as the inter-response interval between consecutive key presses (i.e., simultaneous flexion of all four fingers). The onset of the first of four finger presses was recorded and used in subsequent analyses in the event that the four fingers did not precisely touch their respective keys instantaneously.

Behavioral analyses.

To assess the nature of the initial learning on the two tasks (MSL and CTRL) in the training session, best-fit curve estimation procedure, available in SPSS that uses least-squares regression to test the goodness of fit for linear and non-linear models [2,44] were used separately for each task to illustrate the nature of the performance changes at training and to visualize whether performance during training and retest follow the hypothesized pattern of results. In addition, a repeated-measures block (blocks 1–14) by task (MSL, CTRL) ANOVA was used to directly compare the changes in performance over the course of the training sessions. Finally, paired t-tests were used to test for offline gains in performance across the sleep interval from the end of training (mean of last 4 blocks) to the beginning of retest (mean of first four blocks) for the MSL and CTRL tasks.

Recording parameters.

EEG was recorded using an MR-compatible EEG cap (Braincap MR, Easycap, Herrsching, Germany) with 64 ring-type electrodes and two MR-compatible 32-channel amplifiers (Brainamp MR plus, Brain Products GmbH, Gilching, Germany). EEG caps included 62 scalp electrodes referenced to FCz. Two bipolar ECG recordings were taken from V2-V5 and V3-V6 using an MR-compatible 16-channel bipolar amplifier (Brainamp ExG MR, Brain Products GmbH, Gilching, Germany). Using high-chloride abrasive electrode paste (Abralyt 2000 HiCL; Easycap, Herrsching, Germany), electrode-skin impedance was reduced to < 5 KOhm. To reduce movement-related EEG artifacts, subjects' heads were immobilized in the head-coil by surrounding the subject’s head with foam cushions. EEG was digitized at 5000 samples per second with a 500-nV resolution. Data were analog filtered by a band-limiter low pass filter at 250 Hz and a high pass filter with a 10-sec time constant corresponding to a high pass frequency of 0.0159 Hz. Data were transferred via fiber optic cable to a personal computer where Brain Products Recorder Software, Version 1.x (Brain Products, Gilching, Germany) was synchronized to the scanner clock. EEG was monitored online with Brain Products RecView software using online artifact correction.

Preprocessing.

EEG data were preprocessed by a low-pass filter (60 Hz), down-sampled to 250 samples/sec and re-referenced to averaged mastoids. Scanner artifacts were removed using the “fMRI Artifact rejection and Sleep Scoring Toolbox (FASST)” (ver. 0.302; http://www.montefiore.ulg.ac.be/~phillips/FASST.html) for MATLAB (Mathworks, Natick, Massachusetts, USA) [45], using an adaptive average subtraction method. Ballistocardiographic artifacts were then removed using an algorithm based on a combination of artifact template subtraction and event-related independent component analysis [46] for artifacts time-locked to the R-peak of the QRS complex of the cardiac rhythm.

Spindle detection.

Following gradient artifact and ballistocardiographic artifact correction, EEG recordings were sleep stage scored according to standard criteria [47] to identify periods of NREM sleep, free of any movement artifact, during which the EEG and fMRI data were analyzed. There were no significant differences in sleep architecture or sleep spindle characteristics between the MSL and CTRL night (all p > 0.05 from paired t-tests). Sleep spindles were automatically detected at Fz, Cz and Pz (re-referenced to average mastoids) in order to include spindles from derivations where spindles are typically maximal, as described previously, using a previously published and validated method [2,48,49]. Briefly, similar to the root mean square and other related transformations commonly used to transform the raw EEG signal prior to spindle detection [50,51], this method uses a complex demodulation transformation [52] to extract the power for each data point in the frequency range of interest (e.g., 11–17 Hz). This approach has the advantage of extracting the signal of interest with a very good signal to noise ratio, and of yielding only positive data point values, hence making event detection straightforward. Also, similar to other methods that employ an individualized amplitude threshold [18] calculated from a percentile score of the whole recording (e.g., 95%), the present data from each channel and for each participant were individually normalized using a Z-score transformation derived from a 60-second sliding window to account for changes in spindle-related activity over time [53,54]. Similar to Ray et al [49], events were detected on the transformed signal using a cut-off z-score = 2.33, equivalent to the 99th percentile. We excluded any spindles that were less than 0.50 sec and included only spindles that occurred in artifact-free NREM sleep that were at least 2 seconds apart to minimize overlap in the HRF of individual spindle events.

Recording parameters.

As described in detail previously [2] brain images were acquired using a 3.0T TIM TRIO magnetic resonance imaging system (Siemens, Erlangen, Germany) and a 12-channel head coil. In all subjects, a structural T1-weighted MRI image was acquired using a 3D MPRAGE sequence (TR = 2300 ms, TE = 2.98 ms, TI = 900 ms, FA = 9°, 176 slices, FoV = 256×256 mm², matrix size = 256×256×176, voxel size = 1×1×1 mm³). Multislice T2*-weighted fMRI images were acquired during the practice sessions of the MSL and CTRL tasks with a gradient echo-planar sequence using axial slice orientation (TR = 2160 ms, TE = 30 ms, FA = 90°, 40 transverse slices, 3 mm slice thickness, 10% inter-slice gap, FoV = 220×220 mm², matrix size = 64×64×40, voxel size = 3.44×3.44×3 mm³). Importantly, the sequence parameters were chosen so that the gradient artifact would be time stable, and the lowest harmonic of the gradient artifact (18.52 Hz) would occur outside the spindle band (11–17 Hz). This was achieved by setting the MR scan repetition time to 2160 ms, such that it matched a common multiple of the EEG sample time (0.2 ms), the product of the scanner clock precision (0.1 μs) and the number of slices (40 slices) used. Imaging parameters were the same during post-training sleep where EEG measurements were simultaneously recorded with fMRI acquisitions.

Image preprocessing.

Functional volumes were preprocessed and analyzed using SPM8 (http://www.fil.ion.ucl.ac.uk/spm/software/spm8/; Welcome Department of Imaging Neuroscience, London, UK) implemented in MATLAB (ver. 7.14 R2012a) for OS X (Apple, Inc. Cupertino, CA). As described in detail previously [2], functional scans of each session were realigned using rigid body transformations, iteratively optimized to minimize the residual sum of squares between the first and each subsequent image separately for each session. A mean realigned image was then created from the resulting images. The structural T1-image was coregistered to this mean functional image using a rigid body transformation optimized to maximize the normalized mutual information between the two images. Coregistration parameters were then applied to the realigned BOLD time series. The coregistered structural images were segmented into grey matter, white matter and cerebrospinal fluid. An average subject-based template was created using DARTEL in SPM8. All functional and anatomical images were spatially normalized using the resulting template, which was generated from the structural scans. Finally, spatial smoothing was applied on all functional images (Gaussian kernel, 8 mm full-width at half-maximum [FWHM]).

The analysis of fMRI data, based on a mixed effects model, was conducted in 2 serial steps, accounting respectively for fixed and random effects, described in the following sections.

Practice sessions.

For each subject, changes in brain regional responses were estimated by a model including the responses to the task (practice and rest) in each practice session (training and retest). These regressors consisted of box cars convolved with the canonical hemodynamic response function. Movement parameters were not included in the model as Johnstone and collaborators have demonstrated that it is not recommended in a block design [55]. Slow drifts were removed from the time series using a high pass filter with a cut-off period of 128 s. Serial correlations in the fMRI signal were estimated using an autoregressive (order 1) plus white noise model and a restricted maximum likelihood (ReML) algorithm. Linear contrasts were then performed for the MSL and CTRL sessions, to investigate changes in activation within each session (i.e., last 7 blocks > first 7 blocks) and to compare the difference between tasks (main effect of practice during training [MSL>CTRL]_Training and retest [MSL>CTRL]_Retest sessions). The decision to divide each 14-block practice session into 7 blocks (i.e., in half) was in order to maximize the amount of available data included in each contrast so that there was sufficient statistical power for this comparison. These linear contrasts generated statistical parametric maps [SPM(T)]. The resulting contrast images were then further spatially smoothed (Gaussian kernel 6mm FWHM) and entered in a second-level analysis, corresponding to a random effects model, hence accounting for inter-subject variance.

In the second level analyses, one-sample t-tests were performed to test the linear contrasts described above. These activation maps constituted t-statistics maps [SPM(T)] thresholded at p < 0.001 (uncorrected for multiple comparisons). All statistical inferences were performed at a threshold of p<0.001 (unc.) over the entire volume and effects significant at p<0.05, family wise error corrected over the entire volume were indicated.

Sleep sessions.

For data acquired during the simultaneous EEG-fMRI sleep recordings, within-session series of consecutive fMRI volumes sleep stage scored as NREM sleep according to standard criteria [47] were selected from each complete fMRI time series (e.g., MSL and CTRL post-training sleep sessions). To be included in the fMRI analysis, the EEG had to be visibly movement-free and be a segment of uninterrupted NREM sleep longer in duration than 55 volumes (i.e., ~120 seconds or longer; corresponding to the minimum amount of sleep that was needed to perform the automated spindle detection), resulting in inclusion of 36% of the total recorded data (i.e., 11,466 of 31,852 MRI volumes during non-REM sleep). Each time series corresponding to NREM sleep that met these criteria were entered into the GLM as a separate session so that no gaps existed in the design matrix. For each subject, brain responses were estimated in an event-related design using a fixed-effects general linear model including responses time-locked to spindle events (11–17 Hz detected at Fz, Cz and Pz that did not overlap in time; by taking the onset for the first spindle for overlapping events). Given the limited amount of sleep, it was not possible to further subdivide spindles into slow (e.g., 11–14 Hz at Fz) and fast (e.g., 14–17 Hz at Pz) spindle types without loosing subjects due to missing data or insufficient number of spindle onsets (e.g., 30 events per spindle type, per subject, per recording). Thus, only results from full bandwidth spindles in NREM sleep were reported. Consistent with similar previous studies [36,39,56,57], the vectors including spindle events were convolved with the canonical hemodynamic response function (HRF) as well as with its temporal and dispersion derivatives. Nuisance variables in the model included the movement parameters estimated during realignment (translations in x, y, and z directions and rotations around x, y, and z axes), the squared value of the movement parameter, the first derivative of each movement parameter, and the square of the first derivative of each movement parameter, in addition to the mean white matter intensity and mean cerebral spinal fluid intensity for each subject. Slow wave activity is a defining characteristic of NREM sleep [47]. This activity was accounted for by including spectral power (μV²) in the delta band (0.5–4 Hz) for each TR window (2160 ms) as a variable of no interest, convolved with the hemodynamic response function. Slow drifts were removed from the time series using a high pass filter with a cut-off period of 128 s. Serial correlations in the fMRI signal were estimated using an autoregressive (order 1) plus white noise model and a restricted maximum likelihood (ReML) algorithm. Linear contrasts tested the main effect spindle events for the MSL and CTRL night and between the two post-training nights (MSL vs. CTRL) for the canonical HRF. These analyses generated statistical parametric t maps [(SPM(T)]. The resulting contrasts images were then further smoothed (FWHM 6 mm Gaussian Kernel) and entered in a second-level analysis.

The group-level analysis consisted of one sample t-tests for each contrast of interest (i.e., MSL, CTRL and MSL>CTRL contrasts). The error covariance was not assumed independent between regressors and a correction for non-sphericity was applied. To investigate the relationship between the magnitude of the spindle-dependent activation and overnight changes in performance, offline gains in performance for each subject (difference between the mean of the last 4 blocks of practice and the mean of the first 4 blocks of retest) were entered gains scores as a covariate of interest in the described GLM. These activation maps constituted maps of the t statistic [SPM(t)] testing for the main effect for each contrast of interest. All statistical inferences were performed at a threshold of p<0.001 (unc.) over the entire volume and effects significant at p<0.05, family wise error corrected over the entire volume were indicated.

Overlap between practice and sleep sessions.

To more clearly illustrate the overlap of activations between the training and sleep sessions, the conjunction was taken as the minimum t-statistic using the global null hypothesis (see [58,59] over: 1) a t-map testing for the main effect of condition (MSL) during the training session, and 2) a t-map testing for the main effect of spindle events during the MSL night. These two statistical maps were thresholded so that when factored together, the resulting image was equivalent to a combined threshold of p < 0.001 [59]. The resulting mask image presents the areas that were consistently high and jointly activated in the training and spindle maps. It should be noted that a significant conjunction does not mean activations from both contrasts were individually significant (i.e., a conjunction of significance). Note that the minimum t-values do not have the usual Student’s t-distribution and small minimum t-values can be highly significant.

In order to investigate the degree of similarity between the training and spindle maps, we computed the spatial correlation between the two unthresholded activations maps with which the conjunction analysis was performed. The significance of the correlation was tested using a permutation-based approach from 5000 spatial permutations. This approach allows for an obtained value to be tested against a distribution of correlations generated from the permutation procedure. In addition, in order to confirm that any overlap observed in the conjunction analysis was not due to random overlap between the two maps, we generated 30 spatial permutations of randomized spindle maps. We then correlated these maps with the training map to ensure that there was no relationship between the training session and potentially random spindle-related activations (mean: r = -0.00007, p = 0.54, SD: r = 0.001, p = 0.31).

Finally, to confirm that activations time-locked to spindles and correlated with gains in performance were not simply an epiphenomena of NREM sleep, we generated the same number of random events as spindles in each segment of NREM sleep for all subjects during both the MSL night. These onsets took place between 5 and 15 seconds either before or after each originally detected spindle onset, with the condition that the random onsets did not overlap with any other spindle event. We re-ran the exact same GLM as conducted for spindle onsets, with the only difference being that the randomly generated onsets were included, as opposed to the spindle onsets.

Training session.

As expected, MSL performance in the training session improved rapidly across blocks of practice (Fig 2A) compared to the motor control (CTRL) task (Block [blocks 1–14] x Task [MSL, CTRL] ANOVA interaction effect: F(13,156) = 4.94, p < 0.001), and followed a prototypical learning curve (quadratic inverse function: R² = 0.95, F(1,12) = 238.45, p < 0.001). By contrast, despite a gradual linear increase in speed, performance on the CTRL task did not follow the same typical learning curve (linear, straight-line function: R² = 0.96, F(1,12) = 311.96, p < 0.001) during the initial training session, nor during the retest session the following day.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Behavioral data.

A: Mean (+/- SEM) performance speed (inter-key press interval) during training and retest sessions for the explicit motor sequence learning (MSL) and control (CTRL) tasks. B: Gains in performance (reflecting consolidation) was measured using the mean (+/- SEM) inter-key-press interval in the last 4 blocks of the training vs. the first 4 blocks of retest. As expected, subjects showed significant gains in performance in the MSL task only (* indicates p < 0.05).

https://doi.org/10.1371/journal.pone.0174755.g002

Offline gains.

Accumulating evidence has shown that offline gains in explicit motor sequence performance occur in the absence of additional practice, mostly after a period of sleep [8,60,61]. Likewise, in the present study, a session (M last 4 blocks training, M first 4 blocks retest) by task (MSL, CTRL) within subjects ANOVA revealed that performance improved from training to retest in the MSL but not the CTRL task (F(1,12) = 22.65, p <0.0001). Post hoc t-tests showed that offline gains in performance on the MSL task (Fig 2B) were observed from the end of training (mean of last four blocks) to the beginning of retest (mean of first four blocks; t(12) = 3.00, p = 0.011), whereas no significant gain in performance was observed for the CTRL task (t(12) = -1.31, p = 0.214). Of note, the effect for the MSL task is significant whether as many as 7 blocks of practice are included in the means (t(12) = 3.38, p = 0.005), or as few as 3 blocks (t(12) = 2.58, p = 0.024), but not for the CTRL task. Together, these results suggest that sequence memory, but not performance on the CTRL task, was consolidated overnight.

Training session.

Similar to previous studies [13,60], a distinct MSL-related pattern of brain activity was observed during the training session, as shown by a significant increase of activity in regions that was greater in the MSL compared to the CTRL condition (i.e., the MSL>CTRL contrast, Fig 3A, see 2.1 in Table 2). These included the premotor cortex (BA 6), parietal cortex (BA 7), somatosensory cortex (BA 1–3), and cerebellum (CB; lobule VI).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Cerebral activation during training, sleep spindles and retest.

A: MSL-specific (MSL>CTRL) brain activation during the training session. (inset): Increased activation bilaterally in the striatum from the beginning of training (first 7 blocks) as compared to the end of the training session (last 7 blocks) on the MSL task. B: Activations time-locked to sleep spindles following MSL. (inset): Conjunction of within training session-related and spindle-related activations. C: MSL-specific (MSL > CTRL) brain activation during the retest session. Results displayed at p<0.001, uncorrected.

https://doi.org/10.1371/journal.pone.0174755.g003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Cerebral activations during practice sessions.

Statistically significant functional imaging results from one-sample t-tests performed on training and retest sessions (see Figs 3, 4A & 5).

https://doi.org/10.1371/journal.pone.0174755.t002

Within training session.

To follow-up the Increased activation of the striatum, in particular, evolved over the course of the training session (i.e., MSL>CTRL contrast in the last 7 blocks training > first 7 blocks training; Figs 3A (inset), 4A and 5, see 2.2 in Table 2, see Materials and methods for details), suggesting that the striatum was activated more for the MSL condition in the last half of the training session, when behavioral performance became asymptotic (no significant effect of practice block at the end of training session: F(1,12) = 1.26, p = 0.29), as compared to the initial fast-learning phase, at the beginning of training period. This analysis also revealed other regions involved in MSL [21] were activated during the training session, including the parietal cortex (BA 7), cerebellum (VI) and the brainstem (e.g., pons).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Cerebral activations during practice.

Change in striatal activation from the training to the retest session in the MSL vs. CTRL condition (A), was associated with overnight gains in performance (B). Results displayed at p<0.001, uncorrected. Signal intensity expressed in arbitrary units (a.u.) obtained from raw beta weights.

https://doi.org/10.1371/journal.pone.0174755.g004

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Cerebral activations during training and retest.

Striatal activity within the training session (yellow) and the retest session (red). Following sleep, striatal activity was reorganized in different sub regions, as opposed to a strengthening of activity in the same regions. Results displayed at p<0.001, uncorrected.

https://doi.org/10.1371/journal.pone.0174755.g005

To follow-up this analysis, we conducted a region of interest (ROI) based analysis to test the specificity of the MSL-related within session changes in the striatum. We extracted the beta values from the left and right striatum and conducted a 2 (MSL, CTRL) x 2 (first 7 blocks, last 7 blocks) repeated-measures ANOVA comparing activations in the MSL vs. CTRL condition across sessions. We observed a significant interaction in the left (F(1,12) = 5.54, p = 0.036) but not the right striatum which was greater for the MSL vs. CTRL at the end of training compared to the beginning of training. Pairwise follow-up within-session Bonferroni corrected comparisons revealed that the within-session effect was significant in the MSL but not the CTRL condition for the left (p<0.001 vs. p = 0.198, respectively) and right (p<0.001 vs. p = 0.422, respectively) striatum. Thus, suggesting that striatal activation was increased in the MSL but not the CTRL condition during the training session.

Retest session.

After a night of sleep, during the retest session, significant motor sequence-related (MSL>CTRL contrast) activity was again observed in the same regions as the training session (e.g., BA 6, BA 7 and the cerebellum), except that sequence-specific activation (i.e., the MSL>CTRL retest contrast) of the striatum was prominently enhanced at retest during practice the following morning (see Fig 3C, see 2.3 in Table 2). Additional activations at retest that were not significant during the training session were also observed that support the slow phase [62] of motor learning (e.g., primary motor cortex, anterior cingulate and additional recruitment of parietal regions).

Within retest session.

There was no change in striatal activity within the retest session itself with additional practice from the start (MSL>CTRL first 7 blocks) as compared to the end (MSL>CTRL last 7 blocks) of the retest session (see 2.4 in Table 2), suggesting that after sleep, the striatal memory trace was stable (i.e., consolidated).

Training vs. retest session.

After a night of sleep, activity in the region of the striatum that was activated during the training session was significantly increased from the training to the retest session in the MSL vs. CTRL conditions (Fig 4A, see 2.5 in Table 2), amongst other structures such as the hippocampus, supplementary motor area and somatosensory cortex. Importantly, changes in striatal activity from training to retest were correlated with overnight gains in performance (Fig 4B; r(11) = 0.56, p = 0.047).

In addition, there was a reorganization of cerebral activity, particularly in the striatum, from rostrodorsal (associative) to more caudoventral (sensorimotor) regions that was remarkably similar to a previous report [21]. This reorganization took place from the training (MSL>CTRL) to the retest (MSL>CTRL) session (Fig 5; showing the overlay of both sessions). Taken together, these results suggest that sub-regions of the striatum that show increased activation during the MSL training session (as compared to the CTRL condition) are not simply increased further after a period of sleep as one might predict. Rather, there is within-striatal reorganization of activity, and the extent of changes in striatal activity is associated with MSL.