Microstructural dynamics of motor learning and sleep-dependent consolidation: A diffusion imaging study

Summary Memory consolidation can benefit from post-learning sleep, eventually leading to long-term microstructural brain modifications to accommodate new memory representations. Non-invasive diffusion-weighted magnetic resonance imaging (DWI) allows the observation of (micro)structural brain remodeling after time-limited motor learning. Here, we combine conventional diffusion tensor imaging (DTI) and neurite orientation dispersion and density imaging (NODDI) that allows modeling dendritic and axonal complexity in gray matter to investigate with improved specificity the microstructural brain mechanisms underlying time- and sleep-dependent motor memory consolidation dynamics. Sixty-one young healthy adults underwent four DWI sessions, two sequential motor trainings, and a night of total sleep deprivation or regular sleep distributed over five days. We observed rapid-motor-learning-related remodeling in occipitoparietal, temporal, and motor-related subcortical regions, reflecting temporary dynamics in learning-related neuronal brain plasticity processes. Sleep-related consolidation seems not to exert a detectable impact on diffusion parameters, at least on the timescale of a few days.


INTRODUCTION
Confronted with novel environmental stimulations, the brain progressively adapts both its function and structure to more efficiently meet external demands. 1,2In the minutes to hours following exposure to new learning material, synaptic plasticity takes place, involving early cellular determinants of synaptic strength and persistence triggered within individual neurons (e.g., changes in dendritic length, spine density, or synapse formation). 3Additionally, modifications in glial activity can also be observed in response to learning. 46][7] For instance, rapid 8,9 and delayed [9][10][11][12] changes within functional networks in response to motor skill learning have been well documented, highlighting a progressive reorganization from the beginning of the learning episode to delayed retest.As compared with the available data for functional reorganization, evidence regarding structural remodeling is less documented.Indeed, imaging studies probing experience-dependent gray (GM) or white matter (WM) structural changes (for a review, see e.g., 13 ) have been limited by the relatively low sensitivity of structural magnetic resonance imaging (MRI) to detect subtle changes in the underlying brain anatomy.Therefore, only substantial changes over long and sustained training periods could be observed.For instance, the first longitudinal human study reported transient bilateral expansion in GM in the mid-temporal area (hMT/V5) and in the left posterior intraparietal sulcus after months of juggling practice. 14GM changes in the mid-temporal area (hMT/V5) were robustly replicated in older adults using the same protocol. 15Later longitudinal human studies demonstrated learning-induced plasticity changes in GM after a few weeks 16 or days 17 of motor skill practice.Changes in WM integrity were also shown to parallel the increased GM density in the intraparietal sulcus after repeated juggling practice, 17 and correlations were found between WM integrity and GM volume in brain regions functionally engaged in motor sequence performance. 18It is actually known that momentary adaptations in functional connectivity alter structural connections, which in turn affect functional connectivity. 19Lastly, WM 20 and GM 20,21 structural plasticity changes correlate with behavioral measures of improvement, suggesting that both GM and WM are capable of relatively rapid remodeling when acquiring novel information.
Besides a mere effect of time, post-learning sleep mechanisms contribute to memory consolidation 22 by promoting long-lasting changes in neural networks.Notwithstanding robust findings showing functional changes (using e.g., fMRI or EEG) in brain responses and behavior ll OPEN ACCESS after offline periods including sleep in humans (for reviews, see e.g., 22,23 ), concrete evidence for sleep-dependent (micro)structural alterations in learning-related areas remains scarce.According to the synaptic homeostasis hypothesis, 24 locally increased synaptic strength after learning is downscaled by slow oscillations during sleep to a baseline level, which is both energetically sustainable and beneficial for the consolidation of novel memory traces.In line with this theory, a diffusion tensor imaging (DTI) study highlighted decreased diffusivity in cortical GM following extended training, which reverted after recovery sleep. 25At variance, however, animal data showed post-learning sleep deprivation prevented the formation of branch-specific spines, 26 suggesting that sleep rather promotes synapse formation and strengthening.Hence, there is no definitive evidence about the micro-structural modifications that dynamically take place in the human brain during early, then delayed post-training periods, and their modulation by post-learning sleep.
Fortunately, the last decade witnessed the speedy development of advanced MRI methods to combine brain macro-and micro-structure. 27This enabled a dramatically shortened timescale at which one can detect the structural remodeling that accompanies functional neuroplasticity and learning.Significant changes were thus observed over increasingly shorter learning episodes (e.g., 2 h 20,21 up to 1 h 28 ), using standard DTI measures, i.e., mean diffusivity (MD) and fractional anisotropy (FA) that relate to tissue density and fiber organization/directionality, respectively.These rapid-learning-related microstructural changes were initially mostly identified in the limbic system, 20,21,29 but recent studies showed that such changes can also be detected in motor-learning-related regions.For instance, MD reduction was observed after only 45 min of motor practice in the left premotor cortex, the superior part of the cerebellum, and the left middle temporal gyrus. 30Subsequently, decreased MD was identified in the hippocampus and precuneus after only 15 min of motor sequence learning. 31Although these results demonstrate the sensitivity of DTI to detect rapid and subtle changes in the underlying brain structure, there is an inherent limitation in standard DTI measures, such as MD and FA markers, that provide only aggregate information about the underlying cellular processes. 32For example, a reduction in FA may be caused by decreased fiber density, decreased fiber coherence, or increased free water contamination.In this respect, new biophysical models have been proposed to improve specificity by disentangling the contributions from the different compartments (hindered extra-axonal and restricted intra-axonal) of the tissue and the specific geometry of each compartment, 27 e.g., CHARMED (composite hindered and restricted model of diffusion) 33 or NODDI (neurite orientation dispersion and density imaging). 32NODDI is one of the most widely used biophysical models, as it provides a better fit to the diffusion imaging data. 34Indeed, it can model the dispersion/fanning of axonal fibers or dendrites and disentangle the microstructural effects underlying standard DTI metrics. 35,36More specifically, NODDI enables the voxel-wise estimation of three metrics: the neurite density index (NDI), which gauges the packing of neurites including both axons and dendrites, the orientation dispersion index (ODI), which quantifies the angular variability of the neurites, and the free-water fraction (FWF), which assesses the proportion of water molecules moving freely, not hindered in their course, along no preferential direction.
As stated earlier, currently available data indicate that microstructural brain changes underlying motor sequence learning can be evidenced already in the short term, up to tens of minutes. 30,31There is also robust evidence that post-training sleep contributes to the consolidation of motor memories. 23However, concrete evidence for sleep-dependent and offline structural modifications in learning-related networks in human is still lacking.Finally, there is a need to highlight the cellular processes involved in microstructural brain changes using more specific approaches than standard DTI.Therefore, we aimed at exploring using both DTI and multi-compartment diffusion imaging analysis with NODDI (1) the immediate microstructural changes following initial motor learning, (2) the delayed changes influenced by sleep availability on the first post-learning night, and (3) the microstructural modifications during a relearning episode on Day 5, also influenced by the presence or absence of post-learning sleep (Figure 1; for task-related details see Figure 2).

Demographic data
Welch ANOVAs performed separately on age, laterality, sleep quality, and chronotype (see Table S1) with between-subject factor Sleep (SD versus RS) did not reveal any significant differences between the SD and RS groups (all ps > 0.388).

Behavioral data
For motor learning at Day 1, the ANOVA computed on mean reaction time (RT) for sequential blocks performed with within-subject factor Block (1:25, 27:30) and between-subject factor Sleep (RS versus SD) disclosed a main Block effect (F For delayed motor memory consolidation at Day 5, the ANOVA compared mean performance on the 2 last blocks of the learning session (LS D1; blocks 29:30) and the 2 first blocks of the relearning session (RL D5; blocks 33:34) with within-subject factor Day (End LS D1 versus Begin RL D5) and between-subject factor Sleep (RS versus SD) to assess delayed offline gains in performance.The analysis revealed a significant decrease in mean RT between D1 and D5 (F 1, 59 = 13.968,p < 0.001, h p 2 = 0.191).However, both groups exhibited a similar decrease over time with no significant Sleep (F 1, 59 = 0.125, p = 0.725, h p 2 = 0.002) or Day*Sleep interaction (F 1, 59 = 0.317, p = 0.575, h p 2 = 0.005) effect (Figure 3).Also, no significant effect was found concerning changes in accuracy (all ps > 0.343).
Looking at motor sequence relearning at Day  Also, we computed a complementary mixed ANOVA comparing sequential-specific amelioration at the end of learning and at the beginning of relearning (respectively computed as the difference in RT between the mean of the 2 sequential blocks surrounding random block 26 or 35 minus the random block, then divided by the random block and multiplied by 100) with between-subject factor Sleep (RS versus SD).The analysis revealed a significant difference (F 1, 59 = 54.097,p < 0.001, h p 2 = 0.478) with a significantly larger difference between random and sequential bloc at the beginning of RL (À46,653%) compared with the end of LS (À37.219%),confirming the previous findings showing effective delayed memory consolidation.However, no main Sleep (F 1, 59 = 0.237, p = 0.628, h p 2 = 0.004) or Sleep*Day interaction (F 1, 59 = 3.488, p = 0.067, h p 2 = 0.056) effects were found.Finally, all participants reported noticing the sequential nature of the task by the end of the procedure.

Diffusion-weighted imaging data
Learning-related short-term structural changes (Day 1; DWI1 versus DWI2) Cortical ribbon.Looking at DTI parameters, the surface-based statistical analysis disclosed extended clusters in the whole sample following learning with decreased MD bilaterally in the inferior parietal and paracentral gyri, the precuneus, the insula, the precentral, lingual, superior parietal, lateral occipital, superior frontal, postcentral, supramarginal, middle-, superior-, and inferior temporal gyri, the rostral anterior cingulate and fusiform gyri, the cuneus, and the transverse temporal gyrus.Smaller clusters were found in the banks of the superior temporal sulcus, the rostral middle frontal, isthmus cingulate, and parahippocampal gyri as well as in the left caudal anterior cingulate and caudal middle frontal gyri, posterior cingulate and lateral orbitofrontal gyri, and right pars opercularis, pericalcarine gyrus, and pars orbitalis (Figure 4; detailed results and cluster size in Table S2).
The surface-based analysis conducted on NODDI parameters evidenced increased NDI after learning in regions almost perfectly overlapping with regions exhibiting MD changes described earlier.Bilaterally, NDI increased in the lateral occipital, posterior cingulate, caudate middle frontal, inferior parietal, and lingual gyri, the precuneus, the supramarginal, superior frontal, precentral, rostral middle frontal, superior temporal, insular, postcentral, middle temporal, fusiform, and rostral anterior cingulate gyri, the pars opercularis, the lateral orbitofrontal gyrus, the pars orbitalis, the caudal anterior cingulate, medial orbitofrontal, and superior parietal gyri, the pars triangularis, and the transverse temporal gyrus.NDI clusters were also significant in the left banks of the superior temporal sulcus and cuneus and the right parahippocampal, inferior temporal, paracentral, and pericalcarine gyri.FWF decrease was also evidenced in most areas exhibiting MD decrease, such as the bilateral lateral occipital, lingual, precentral, superior parietal and frontal gyri, the cuneus, postcentral gyrus, the insula, fusiform and superior temporal gyrus, the precuneus, the caudal anterior cingulate and paracentral gyri, the left inferior parietal, pericalcarine, supramarginal, isthmus cingulate and rostral middle frontal gyri, and the right caudal middle frontal, transverse temporal, medial orbitofrontal, middle temporal, and posterior cingulate gyri.Finally, we also found ODI changes (both increase and decrease depending on the region), with small clusters in the bilateral postcentral and caudal middle frontal gyri, the precuneus, and the superior temporal and frontal gyri, left posterior cingulate, precentral, rostral middle frontal and lateral occipital gyri clusters, right lateral orbitofrontal, inferior parietal, fusiform, and supramarginal and middle temporal gyri (see Table S2).
As expected, no between-group differences were evidenced at this stage of the procedure for any of the 4 tested metrics.Finally, we computed correlations between microstructural changes observed over the learning session (computed as the differences between cortical maps at t2 and t1 at each vertex for each individual DWI metric) and motor performance improvement during learning (computed as the difference in RT between the mean of the 2 last blocks (29-30) minus the mean of the 2 first blocks (1-2) divided by the mean of the 2 first blocks (1-2) multiplied by 100), as well as correlations with sequential-related amelioration (computed as the difference in RT between the mean of the 2 sequential blocks surrounding random block 26 [thus, 25 and 27] minus random block 26 divided by random block 26 multiplied by 100).No significant correlation emerged from these analyses.
Subcortical ROIs.MANOVAs with within-subject factor Learning (Pre versus Post) were conducted on the DTI parameters for each ROI (see Figure 5).Significant changes were found in both sides of the cerebellar cortex (p corr < 0.001) and the hippocampus (p corr < 0.001), the right thalamus (p corr = 0.002) and caudate (p corr < 0.001), and the left putamen (p corr = 0.001).Post hoc analyses revealed a strong MD decrease in all those regions (all ps < 0.001), suggesting increased tissue density after learning (as compared to baseline).Also, there was increased FA in the left (p = 0.034) and right cerebellar cortex (p = 0.001), the right caudate (p = 0.005), and the right hippocampus (p = 0.035), reflecting increased directionality of diffusion inside of these regions.Additional analyses were performed to verify the absence of a Sleep (RS versus SD) or interaction effect.As expected, no difference existed between groups prior the experimental night (main Sleep effects, all p corr s > 0.016; Learning*Sleep interaction, all p corr s > 0.021).Lastly, correlations between DTI metrics changes observed over the learning session (computed for each individual metric as the ROI value at t2 minus t1) and motor performance improvement during learning (see details in section 2.3.1.1.)were computed in ROIs exhibiting significant changes over the learning session.No correlation was significant for MD (all ps > 0.078) or FA (all ps > 0.302) parameters.Likewise, correlations targeting sequential-related amelioration did not disclose significant effects for MD (all ps > 0.226) or FA (all ps > 0.285).
Similar MANOVAs conducted on NODDI parameters highlighted changes in similar ROIs as with DTI, i.e., both sides of the cerebellar cortex (p corr < 0.001), left (p corr < 0.001) and right hippocampus (p corr = 0.003), right thalamus (p corr = 0.004) and caudate (p corr < 0.001), and left putamen (p corr = 0.001).Post-hoc analyses disclosed decreased FWF in both sides of the cerebellar cortex (p < 0.001), left (p = 0.021) and right (p = 0.006) hippocampus, and right thalamus (p < 0.001) and caudate (p < 0.001), confirming the MD changes observed using DTI and suggesting increased tissue percentage.Increased NDI in the left putamen (p < 0.001) and left hippocampus (p = 0.038) suggests increased number of neurites, whereas increased ODI in the left putamen (p = 0.048) suggests neurite reorganization and remodeling.Here again, no effect of sleep was found (main Sleep effects, all p corr s > 0.030; Learning*-Sleep interaction, all p corr s > 0.034) when including Sleep in the ANOVA (Figure 5).Correlations between changes in NODDI metrics over the learning session (computed for each individual metric as the ROI value at t2 minus t1) and behavioral parameters in ROIs exhibiting significant changes over learning were non-significant, either with motor improvement (all ps > 0.254) or with sequential amelioration (all ps > 0.157) measures.Delayed structural changes and sleep-related effects in pre-learning versus pre-relearning (DWI1 versus DWI3 x RS versus SD) Cortical ribbon.For DTI parameters, session-related main effects were observed in the cortical ribbon with decreased MD in the left lingual and superior temporal gyri, as well as an increased MD in the right pars triangularis, the superior parietal gyrus, and the banks of the superior temporal sulcus (see Figure 4; detailed results and cluster size in Table S3).
Looking at NODDI parameters, we found session-related NDI changes (mostly increases; see Table S3 for detailed directionality per region) in bilateral superior temporal and lateral orbitofrontal gyri; supramarginal, rostral middle frontal, and postcentral clusters; and right superior parietal and lingual gyri, pars orbitalis, fusiform and paracentral gyri, precuneus, and pars opercularis.For FWF changes, clusters mostly showing decreases were found in the bilateral fusiform gyrus; the left lingual, superior temporal, and caudal middle frontal gyri; and the right superior frontal, lateral occipital superior parietal, precentral, and supramarginal gyri.Lastly, ODI changes were found in clusters bilaterally in the superior frontal, postcentral, and inferior parietal gyri; the left lateral occipital, lateral orbitofrontal, fusiform, precentral, and supramarginal gyri; and the right rostral middle frontal and parahippocampal gyri, precuneus, and inferior temporal gyrus.
Sleep-related effects were non-significant, as well as correlations between behavioral performance improvement (computed as the difference in RTs between the mean of the 2 first blocks [33-34] of the Day 5 relearning session minus the mean of the 2 last blocks [29-30] of the Day 1 learning session divided by the mean of the 2 last blocks [29-30] multiplied by 100) and structural changes (computed as the differences between cortical maps at t3 and t1 at each vertex for each individual DWI metric).
Subcortical ROIs.The MANOVAs computed on DTI parameters with within-subject factor Day (D1 versus D5) and between-subject factor Sleep (RS versus SD) did not reveal any significant main effect of Day (all p corr s > 0.061), Sleep (all p corr s > 0.014), or Day*Sleep interaction (all p corr s > 0.058) for any of our ROIs.
Likewise, the MANOVAs computed on NODDI parameters did not disclose main effects of Day (all p corr s > 0.015) or Sleep (all p corr s > 0.034), nor a Day*Sleep interaction effect (all p corr s > 0.099) in all ROIs.It suggests that learning-related changes observed at Day 1 in subcortical ROIs were not maintained at the beginning of Day 5, nor modulated by the presence of the post-learning sleep episode.Consequently, brainbehavior correlations were not computed.
Post-relearning structural changes and sleep-related effects (Day 5; DWI3 versus DWI4 X RS versus SD) Cortical ribbon.The surface analysis comparing the post-(DWI 4 ) versus pre-(DWI 3 ) relearning scans did not highlight sleep-related effects.
For DTI parameters, there was a main relearning effect with MD decreases in bilateral precuneus, banks of the superior temporal sulcus, lateral occipital gyrus, insula, superior and inferior parietal gyri, cuneus, posterior cingulate, post central, superior frontal, supramarginal, lingual, caudal middle frontal, lateral orbitofrontal, precentral, superior temporal, fusiform, rostral middle frontal, and isthmus cingulate gyri.Decreased MD was also observed in the left pericalcarine and transverse temporal gyri and in the right medial orbitofrontal, middle temporal, paracentral, and parahippocampal gyri (see Figure 4; detailed results and cluster size in Table S4).
For NODDI parameters, increased NDI was found bilaterally in precuneus, lateral occipital, caudal middle frontal, rostral middle frontal, superior frontal and parietal, insular, superior temporal, lateral orbitofrontal, and posterior cingulate gyri, the banks of the superior temporal sulcus, the precentral, inferior parietal, and middle temporal gyri, the pars opercularis and paracentral gyrus.NDI also increased in the left parahippocampal, postcentral, supra marginal, entorhinal, and inferior temporal gyri and in the right medial orbitofrontal and fusiform gyri, the cuneus, and isthmus cingulate gyrus.Conversely, FWF decreased bilaterally in the cuneus, lateral occipital gyrus, precuneus, postcentral, superior parietal, precentral, superior frontal, and insular gyri; in the left pericalcarine, lingual, inferior parietal rostral middle frontal, medial-and lateral orbitofrontal, and posterior cingulate gyri; and in the right middle temporal paracentral, supramarginal, and superior temporal gyri; and the banks of the superior temporal sulcus and caudal anterior cingulate gyrus.Finally, ODI changes were evidenced in the bilateral precuneus; postcentral, lateral occipital, and superior frontal gyri; the left entorhinal gyrus and the right precentral, inferior, and superior parietal gyri; cuneus, insula, paracentral, rostral middle frontal, and middle temporal gyri.
Correlations between changes in diffusion parameters (computed as the differences between cortical maps at t4 and t3 at each vertex for each individual DWI metric) and performance changes over the relearning episode (computed as the difference in RT between the mean of the 2 last blocks [51-52] minus the mean of the 2 first blocks [33-34] divided by the mean of the 2 first blocks [33-34] multiplied by 100) and sequential-related improvement (computed as the difference in RT between the mean of the 2 sequential blocks surrounding random block 35 [thus, 34 and 36] minus random block 35 divided by random block 35 multiplied by 100) were non-significant, neither using DTI nor using NODDI parameters.
Subcortical ROIs.The MANOVA computed on DTI parameters with within-subject factor ReLearning (Pre versus Post) and between-subject factor Sleep (SD versus RS) highlighted a significant main ReLearning effect bilaterally in the cerebellar cortex (p corr < 0.001), the left (p corr = 0.002) and right putamen (p corr < 0.001), and the left thalamus (p corr < 0.001).Post-hoc analyses showed that these changes were driven by decreased MD (all ps < 0.001), suggesting tissue densification.No changes in FA were observed (all ps > 0.052).There was no main Sleep (all p corr s > 0.022) or ReLearning*Sleep interaction (all p corr s > 0.020) effect, suggesting that post-learning sleep did not modulate the relearning of previously studied material.Correlations between the MD difference between t4 and t3 and relearning-related motor changes were nonsignificant (all ps > 0.560), as well as with sequence learning parameters (all ps > 0.268).
The MANOVA computed on NODDI parameters disclosed a main ReLearning effect with changes in the left thalamus (p corr < 0.001), right putamen (p corr = 0.003), and right hippocampus (p corr = 0.007).Post-hoc tests revealed decreased FWF in the left thalamus only (p = 0.018), increased NDI in the right putamen (p < 0.001) and hippocampus (p = 0.018), and increased ODI in the left thalamus (p < 0.001) and right hippocampus (p = 0.020).Again, there were no main Sleep (all p corr s > 0.040) or ReLearning*Sleep interaction (all p corr s > 0.063) effects.Structural changes observed over the relearning session (t4 versus t3) in the regions mentioned earlier did not correlate with motor changes (all ps > 0.564) or sequence learning indices (all ps > 0.141).

Task-fMRI data
A two-sample t test conducted at the random effect (RFX) level did not disclose between-group (SD versus RS) difference in BOLD responses to sequential versus random blocks (all p FWEcorr > 0.626).
A one sample t test comparing BOLD responses to sequential versus random blocks (over both groups) disclosed higher activity in the right (p FWEcorr = 0.001) and left precuneus (p FWEcorr = 0.001), the left hippocampus (p FWEcorr = 0.017), and the left (p FWEcorr = 0.041) and right (p FWEcorr = 0.050) caudate for sequential than random blocks.Conversely, higher activity for random than sequential blocks was found in the right superior (p FWEcorr = 0.001) and middle temporal gyrus (p FWEcorr = 0.001), the right (p FWEcorr < 0.011) and left middle frontal gyrus (p FWEcorr = 0.005), and the left post central gyrus (p FWEcorr = 0.023).

Additional (control) analyses Alertness before task performance
An ANOVA computed on reciprocal reaction time (RRT) in the PVT5 (1/RT) 37

Sleep quality and duration during the protocol
The ANOVA performed on sleep quality with within-subject factor Night

Sleep quality and duration on the experimental RS night
Separate ANOVAs were carried out in the RS group only to verify if the quantity/quality of the experimental night was equivalent to the previous and next night.The ANOVA on sleep quality with within-subject factor Night (3:5) revealed a main Night (F 2, 58 = 11.700,p < 0.001, h p 2 = 0.287) effect with a lower quality on the experimental night compared with the previous (p corr = 0.005) and next (p corr < 0.001) ones, which can be explained by sleeping with the hd-EEG setting on that experimental night.The ANOVA on sleep duration disclosed no significant difference between the 3 nights (F 2, 58 = 2.813, p = 0.068, h p 2 = 0.088).

Vigilance and sleepiness during experimental SD night
The ANOVA performed on sleepiness KSS scores with within-subject factor Time (hourly score between 22 h and 8 h) evidenced a Time effect (F 3.686, 106.883 = 24.080,p < 0.001, h p 2 = 0.454) with increasing sleepiness all over the night.Similarly, the ANOVA computed on vigilance RRT (PVT 10 ) scores with within-subject factor Time (bihourly score between 22 h and 6 h) evidenced a Time effect (F 1.585, 45.977 = 16.668,p < 0.001, h p 2 = 0.365) with decreasing alertness over the night.

DISCUSSION
In the present study, we aimed, using diffusion-weighted imaging (DWI), at exploring the dynamic microstructural reorganization happening in the short-term following learning and its long-term modulation by the presence or absence of sleep during the first post-training night.Also, we wanted to investigate the microstructural changes within minutes of practice on a previously experienced material.Behaviorally, motor performance improved during learning and reached the same level for both groups at the end of Day 1.A similar pattern was observed during relearning at Day 5 both for the sleep-deprived (SD) and regular sleep (RS) group.At morning retest on Day 2, RTs were slower in the SD than the RS group, which was predictable after a full night spent awake.Besides mere motor learning, RT significantly increased on random as compared with sequential blocks, indicating that both groups learned the repeated sequence adequately and similarly, this knowledge persisted 3 days later, and that delayed practice at relearning on previously acquired material led to further performance improvement gains accumulating when re-practicing the task.No post-training sleep effect was detected.The evolution of performance with time and practice is consistent with previous findings. 38,391][42] Especially when focusing on the SRTT literature, sleep-related effects were not consistently observed.For example, differentiated neural responses were detected using fMRI between post-learning sleep and sleep deprivation conditions 3 days after the experimental night but in the absence of observable behavioral differences. 43][46] At the structural level, we found rapid-learning-related modifications in brain's microstructure after 1 h of motor sequence training (versus baseline), in line with previous reports showing changes even after more restricted learning times. 30,31In the cortical ribbon, changes were observed in a large part of the occipitoparietal and temporal area, respectively, known for subtending visuomotor and somatosensory transformation and integration processes. 47,48Decreased MD was found in extended clusters in these areas, potentially reflecting increased tissue density.Also, decreased FWF in similar but more restricted territories suggests increased tissue proportion.Various cellular processes activated during learning may be responsible for increased tissue proportion and density.For instance, animal studies found gliogenesis as one of the key mechanisms activated during post-learning neocortical remodeling. 49,502][53] To a lesser extent, neurogenesis might also be involved although neocortical neurogenesis remains disputed nowadays. 49,54,55However, increased NDI in regions exhibiting MD and FWF changes indicates that at least part of this reorganized/created cortical tissues may be neurites.Finally, ODI changes in a few clusters, again in analogous regions, also suggest tissue reorganization in the brain tissue microstructure, even if at a smaller scale.Interestingly, although MD decreases have been repeatedly reported in the neocortex following learning, 28,30,56 a previous study using NODDI to investigate spatial learningrelated structural plasticity observed FWF decreases but no NDI increase contrary to our results. 56Besides task differences, this might be due to the fact that we used the revised version 57,58 instead of the original NODDI model 32 used in this preceding experiment, probably giving us the opportunity to gain in preciseness when it comes to disentangling the different components influencing MD alterations.Then, in subcortical regions of interest, decreased MD was found in the left putamen, right thalamus and caudate, bilateral hippocampus, and cerebellar cortex, together with increased FA in the right caudate and cerebellar cortex.Additionally, analyses on NODDI parameters highlighted reduced FWF in bilateral thalamus, hippocampus, and cerebellar cortex, similarly to MD changes, enhanced NDI in left putamen, and hippocampus and increased ODI in left hippocampus.MD reductions in the hippocampus and cerebellum after motor practice confirm previously reported motor-learning-related microstructural changes developing in the short term. 30,31Also, sprouting of new mossy fiber terminals was observed following learning in mice hippocampus. 59Increased NDI in the hippocampus and other ROIs is in line with these findings, whereas modifications in FA suggest increased directionality and changes in ODI rapid-motor-learning-related reorganization.Altogether, modifications in structural parameters suggest a rapid-motor-learning-related remodeling in most of our ROIs and a set of neocortical regions encompassing a large part of the occipitoparietal and temporal cortices, reflecting learning-related neuronal brain plasticity.
Then, we hypothesized that delayed post-learning changes would be modulated by the presence/absence of post-learning sleep and that this would be reflected in the first DWI acquisition on Day 5 compared with baseline.Small but significant clusters persisted in the cortical ribbon, but no learning-related changes were found in ROIs in the long term.Also, sleep did not modulate these changes: both groups exhibited the same pattern at Day 5 (DWI 3 ) and Day 1 (DWI 1 ).As discussed earlier, it is possible that consolidation for a sequential motor SRTT (or at least this one) relies more on time than sleep, both at the behavioral and structural level.Alternatively, it cannot be excluded that microstructural modifications take more time than a few days to be fully consolidated and observable or that SD participants benefitted from the recovery nights after the RS/SD experimental night to catch up with the sleep group before the delayed scan, as such effects were reported at the behavioral level. 60Also, the morning retest on Day 2 might have helped the SD group to catch up by reactivating the memory trace; however, this has been found to be without consequences in a similar protocol, at least at the behavioral level. 39We also obtained diffusion MRI scans at the outset of the RS/SD night experimental night in the morning of Day 2 (not reported here), but these were not suitable for investigating this issue due to contamination by the lack of sleep [61][62][63] and circadian confounds that have been found to influence structural measures taken 12 h apart. 64,65astly, we hypothesized that microstructural parameters would change again when re-exposed to the initial motor task and that these changes could be modulated by the sleep opportunity on the post-learning night, as shown in a prior study investigating topographical learning. 56In the cortical ribbon, the pattern was similar to the one observed during the first learning episode but less extended, and this was not affected by the sleep manipulation.At the functional level, cortical involvement was shown to increase with sequential proficiency, whereas hippocampus and dorsomedial striatum progressively disengage. 12In the case of the relearning session at Day 5, the structural network was already partially shaped during the first learning episode at Day 1 and thus only needed to be refined and fine-tuned during re-exposure, which likely explains the smaller amplitude of the modifications.Finally, correlations between DWI metrics and behavioral measures were all non-significant, at variance with previous reports disclosing correlations between behavior and WM 20 or GM 20,21 modifications.However, correlations in those past studies were found exclusively in the hippocampus for GM 20,21 and in the fornix for WM 20 after 2 h of spatial training.Both areas are heavily solicited and recruited when performing spatial tasks.In our case, a more widespread network seemed to be recruited to answer the motor task demands, maybe explaining the fact that no strong correlation emerged between structural changes in specific regions and the evolution of motor performance.Although the recruitment of such a broad territory in response to motor learning and consolidation might be surprising at first, the relevant literature has highlighted the great diversity of brain structures involved in skill learning 66 whose involvement and interplay evolve with time and practice. 67Indeed, results from integrative and interdisciplinary approaches suggest that motor skills require the dynamic interaction of different neural networks.These networks are organized in loops usually described as formed by the frontoparietal cortices, the basal ganglia (BG), and the cerebellum (CB) 66,68,69 but their involvement varies over time.For instance, at the cortical level, the acquisition of a motor sequence during a visuomotor task seems to involve frontal regions-dorsolateral prefrontal cortex (DLPFC) and presupplementary motor area (preSMA)-67 but with a shift toward parietal areas-intraparietal sulcus (IPS) and precuneus-during later stages such as the delayed retrieval of a visuomotor sequences. 67,70t the end of the protocol, participants practiced random and sequential blocks in the fMRI setting to partially disentangle the brain areas subtending motor and sequential components of performance in the SRTT.Functional MRI results evidenced higher sequence-related activation in the bilateral precuneus, caudate, and left hippocampus and a stronger involvement of the visuo-motor and executive areas during random practice.Interestingly, the precuneus also exhibited structural modifications in MD, NDI, FWF, and ODI following the learning and relearning session as well as persisting NDI and ODI changes at Day 5 in the right precuneus only.The right caudate showed MD decrease and FA increase after learning, and the hippocampus showed bilateral MD and FWF decrease, right FA increase, and left NDI increase following learning and right NDI increase following relearning.These results confirm not only the repeatedly observed overlap between structurally and functionally engaged regions, 18,19,28 suggesting that momentary adaptations in functional connectivity alters structural connections, which in turn affect functional connectivity, 18 but they also corroborate the fact that our results are sequence-specific at least in the precuneus, caudate, and hippocampus, rather than attributable to motor practice only.
To sum up, we observed important and rapid tissue remodeling in response to sequential motor learning and relearning in occipitoparietal and temporal regions, as well as in ROIs involved in motor processing.However, we found limited persistence of those changes 3 days after initial learning.The use of the revised NODDI model in combination with conventional DTI brought us one step closer to image the cellular mechanisms and more specifically, the axonal and dendritic remodeling present in humans in response to learning independently from the glial adaptation that happens simultaneously to answer the task demands.However, no sleep effect was detected 3 days later, suggesting that either sleep-related cellular changes are too subtle to be identified at that macroscopic level using non-invasive DWI measures, or at that timescale, or that consolidation in the SRTT might not be sleep-dependent in these particular conditions.These issues should be investigated in further studies.Nonetheless, the use of DTI combined with NODDI or other complex biophysical models opens the way to reunite the cellular processes underlying learning-and sleep-related remodeling observed only in animals until now, with non-invasive brain imaging techniques applicable to humans.

General procedure
To prevent for a hormonal bias on motor performance, sleep and consolidation, women were tested during their luteal phase. 74,75All participants were explicitly asked to refrain from drinking caffeine or other stimulating drinks on testing days, and to maintain a regular sleep-wake schedule for the entire duration of the experiment.The regularity of the sleep-wake schedule was controlled for the entire procedure using self-reported daily sleep logs for sleep quality and duration (St.Mary's Hospital sleep questionnaire 76 ) and visual inspection of actimetric recordings (ActigraphTM wGT3X-BT, Pensacola, FL, USA). Figure 1 illustrates the experimental design.Three days before the first testing day (Day 1), participants came to the lab for a habituation night sleeping with a 256-channels high-density EEG (hd-EEG).During the Day 1 session held around 16:30, a first baseline diffusion weighted MRI (DWI1) was acquired.Immediately after, and outside of the scanner, participants were trained for 30 blocks (approximate duration 1h) on a motor serial reaction time task (SRTT) adapted from 77,78 (see section 8.3.2. for more details).Thirty minutes after the end of the SRTT session, post-training diffusion weighted MRI (DWI2) was acquired again.Participants were then informed about their assignment to one of the two possible conditions for the post-learning night, i.e., Regular Sleep (RS) or Sleep Deprivation (SD).Around 21:30, participants from the RS group were equipped with the 256-channels hd-EEG, and then slept from their regular bedtime in the sleep laboratory for the whole night (approx.8-9 hours).SD participants spent the night awake in the laboratory for a period of 10h (maximum two participants at a time), during which they were allowed quiet activities (e.g., playing board games, read, watching non-arousing movies) under the experimenter's supervision.Keyboard typing activities were forbidden to prevent motor interferences.Isocaloric food portions and water ad libitum were available all night.SD participants filled in hourly the Karolinska Sleepiness Scale (KSS 79 ) to document the evolution of their sleepiness, 80 and performed every 2 hours on the 10-min version of the Psychomotor Vigilance Task (PVT 81 ) to track vigilance modifications over the SD night.On Day 2 around 09:30, all SD and RS participants performed a short SRTT-retest (2 sequential blocks; approx. 2 min) and were sent back home for the 3 following nights with the instruction to keep a regular sleep-wake schedule and avoid daytime naps.In the afternoon of Day 5 (same time as Day 1 to control for circadian effect on diffusion images 64,65 and cognitive performance 82 ), diffusion weighted MRI (DWI3) was acquired first.Then, participants were trained again for 20 blocks on the previously learned SRTT sequence (approximate duration 40 min), followed 30 min after the end of practice by diffusion weighted MRI (DWI4), and a final task-related fMRI acquisition during which they alternated motor practice on sequential and random SRTT blocks (see section 8.3.2.).To control for between-groups or -sessions changes in behavioral alertness, participants performed the 5-min version of the PVT 83 before each SRTT session.

Motor learning
We used a 6-choice version of the SRTT (Figure 2; adapted from 77,78 ) coupled with auditory tones, running on PsychoPy3 v2020.2.10 (Nottingham, UK).Auditory coupling was established for the sake of another experiment.Participants were seated in front of a computer screen and asked to place their 6 middle fingers (index, middle and ring fingers of each hand) on 6 response keys matching the 6 squares horizontally arranged on the screen.They were given the instruction to press as fast and as accurately as possible the corresponding key every time a visual cue appeared at one of the 6 positions on the screen.Every time a key was pressed, a sound (beep tone) coupled with the key/position was emitted, and the next trial was presented after 500 msec.Ninety-six trials were presented during one block, and all blocks were separated by a short resting period (self-defined duration).The 96 trials in each block could either repeat a 12-element sequence (5-3-1-6-2-4-1-5-2-3-6-4) or be a pseudo-random succession of cues (the only restriction being that the same key is never pressed twice in a row).The learning session comprised 30 blocks (all sequential except block 26 random), the morning retest 2 sequential blocks and the relearning session 20 blocks (all sequential except blocks 3 and 16 random -also numbered as blocks 35 and 48).Participants were not notified in advance about the sequential nature of the task.Random blocks were inserted to enable discrimination of respectively the sequential and motor contribution to performance improvement.Indeed, reaction time is expected to increase when there is no predictability in the succession of the stimuli (random block), indirectly demonstrated the learning of the regularity in sequential blocks.
For the task-based fMRI session at the end of the experiment, aimed at evidencing motor learning-related networks, participants performed in the scanner the same SRTT task as described above for 30 blocks (20 sequential; 10 random) presented in a semi-random order with no more than 3 times the same block type in a row.To increase variability in the fMRI block design, half of the blocks counted 24 keypresses, the other half 36.Each block was separated from the next by a randomly determined rest duration ranging 5 to 15 s.

Behavioral data analyses
At the behavioral level, SRTT performance was assessed for each block computing the mean reaction time (RT) and accuracy (percentage of correct triplets throughout the 96 keypresses, as humans show a natural tendency to divide behavioral sequences in chunks 84,85 up to 3 elements 86 ).Frequentist statistics were computed using JASP version 0.15 (JASP Team (2021)).Welch t-tests and Welch ANOVAs were always preferred to Student t-tests and classical One-way ANOVAs considering their increased power in case of heterogeneity of variance, that Levene's test for equality of variances often fails to detect. 87,88When normality was violated, Mann-Whitney U-tests were performed.Degrees of freedom were corrected with Greenhouse-Geisser sphericity correction in case Mauchly's sphericity test indicated violated assumption.Bonferroni correction for multiple comparison was applied when post-hoc tests were conducted.All tests are based on a two-sided significance level set at p < 0.05.

MRI data processing Anatomical processing
Raw T1-weighted images were corrected for bias field signal using BiasFieldCorrection_sqrtT1wXT2w script from https://github.com/Washington-University/HCPpipelines/tree/master, as described in the minimal preprocessing pipelines for the Human Connectome Project. 89Segmentation was performed on the T1-weighted images using FastSurferCNN, 90 an advanced deep learning model trained to replicate Freesurfer DKT's segmentation.It segments the brain into 95 cortical and subcortical regions following the Desikan-Killiany-Tourville protocol. 91,92Cortical surface reconstruction was performed on T1-weighted images using FastSurfer 90 which is an extensively validated pipeline to efficiently mimic Freesurfer recon [93][94][95] by leveraging FastSurferCNN output.interest were tested computing linear contrasts (i.e., difference between sequential and random blocks), generating statistical parametric maps.The cut-off period for high-pass filtering was set at 256 s as successive blocks were interleaved with 5-15 s breaks.Individual summary statistic images were then spatially smoothed (6 mm FWHM Gaussian kernel).Next, individual statistics images were introduced in a secondlevel analysis to evaluate differences in sequential vs. random brain response between the SD and RS groups, corresponding to a random effects (RFX) model.The resulting set of voxel values for each contrast constituted a statistical t-map (SPM(T)).Statistical inferences were obtained after correction for multiple comparisons at the voxel level (Family Wise Error (FWE) correction p < 0.05) in the whole brain.Labels were obtained using the MNI152 atlas.

Figure 1 .
Figure 1.Experimental designThree days before the first testing day, all participants come to the lab for a first habituation night under hd-EEG.On Day 1, participants undergo a first diffusionweighted imaging (DWI1) session before being trained for 1 h on a motor sequence learning task (SRTT) in which each response key is associated with a specific auditory tone.Thirty minutes later, a second DWI (DWI2) session follows.During the subsequent night, subjects are either fully sleep deprived (1; SD) or get a regular night of sleep (2; RS) under hd-EEG.On the next morning, they all have a short behavioral retest.After 3 nights of regular sleep at home, participants come back for a 3 rd DWI session (DWI3).Next, they are trained again on the SRTT for 40 min followed by a last DWI session (DWI4) and then task-related fMRI during which they were asked to perform the same SRTT task in the scanner for 20 min with a pseudo-random alternation between sequential and random blocks.

Figure 2 .
Figure 2. Serial reaction time task (SRTT)Volunteers are seated in front of the computer screen with their 6 fingers (no pinkies and thumbs) placed on the 6 keys matching the 6 positions on the screen.Every time one of the 6 squares lights up, the participant has to press as fast and accurately as possible the corresponding key.Each time a key is pressed, an auditory tone matching this position/key is played before the apparition of the next cue (ISI = 500 msec).One block is composed of 96 trials (either sequential or pseudo-random).

Figure 3 .
Figure 3. Performance evolution (speed) in the SRTT over the entire protocol Mean RT (msec) G standard deviation plotted for all blocks executed over the 3 different testing days.On day 1, participants performed 30 blocks (D1; block 26 pseudo-random).After the experimental night taking place between Day 1 and Day 2, performance was assessed in the morning for two blocks (D2).After three recovery nights at home, volunteers performed the task again for 20 blocks (D5; block 35 and 48 pseudo-random).***p < 0.001.

Figure 5 .
Figure 5. Changes in DTI and NODDI parameters in ROIs Boxplots representing (top row) the difference between post-and pre-learning (learning effect; DWI 2 -DWI 1 ), plotted for all regions in which MANOVA effects were significant, (middle row) lack of statistical differences or sleep-related interaction between baseline scan and the first scan at Day 5 (DWI 3 -DWI 1 ), and (bottom row) the difference between post-and pre-relearning for all regions in which MANOVA effects were significant (DWI 4 -DWI 3 ).*p < 0.05, **p < 0.01, ***p < 0.001 for univariate post-hoc tests performed on separate metrics.

( 1 : 3 , 5 : 7 ;
night 4 being the experimental night) and between-subject factor Sleep (SD versus RS) disclosed a main effect of Night (F 5, 290 = 34.055,p < 0.001, h p 2 = 0.370) and Night*Sleep interaction (F 5, 290 = 3.314, p = 0.006, h p 2 = 0.054) with a significantly lower sleep quality and this even more for the RS group, on the first night corresponding to our habituation night at the lab, but a significantly higher sleep quality on the 5 th night, thus the night following the experimental night spend at the lab.However, no main effect of Sleep (F 1, 58 = 2.916, p = 0.093, h p 2 = 0.048) was highlighted.A similar ANOVA conducted on sleep duration with within-subject factor Night (1:3, 5:7) and between-subject factor Sleep (SD versus RS) revealed neither main Sleep effect (F 1, 58 = 0.332, p = 0.567, h p 2 = 0.006) nor Night*Sleep interaction (F 3.986, 231.196 = 1.176, p = 0.322, h p 2 = 0.020).There was a main Night effect (F 3.986, 231.196 = 8.832, p < 0.001, h p 2 = 0.132) with reduced sleep duration on the 5 th night (i.e., the night after the experimental RS/SD night, p corr s < 0.001).
with within-subject factor PVT Session (D1, D2, D5) and betweensubject factor Sleep (SD versus RS) found no main Sleep effect (F 1, 59 = 2.492, p = 0.120, h p 2 = 0.041) but significant main PVT Session expected, post hoc tests disclosed similar alertness in both groups before the learning session at Day 1 (p corr = 1.000) or the relearning session at Day 5 (p corr = 1.000) and a strong difference (p corr < 0.001) at morning retest at Day 2 at the outset of the experimental night, with significantly lower alertness in the SD than in the RS group.
72tailed methods are provided in the online version of this paper and include the following: Sixty-one young, healthy participants (31 females) aged 18-29 years (mean age GSD = 21.31G2.26)providedawritteninformedconsent to participate in this study approved by the Lie `ge University Hospital Ethics Committee (approval #2020/138).They were all free of any neurological/psychiatric history, had a body mass index <28, were EEG & MRI compatible, exhibited moderate to neutral chronotype (mean score GSD = 54.07G7.84,min=32,max= 73, Morningness-Eveningness Questionnaire71), and had good sleep quality (mean score GSD = 3.33 G 1.24, min = 0, max = 6, Pittsburgh Sleep Quality Index72).Musicians and computer scientists who might exhibit high-level hand motor dexterity, smokers and individuals exposed to jetlag within the past 3 months were excluded.Right-and left-handers (mean score GSD = 4.98 G 6.39, min = À10, max = 10, Edinburgh Inventory 73 ) were both included considering the bimanual character of our motor task (see below).After being pseudo-randomly assigned to one of the 2 groups to maintain gender balance, the sleep deprivation (SD) group counted 31 participants while the regular sleep (RS) group counted 30 participants (see TableS1for details).