https://orcid.org/0000-0002-4299-1434
Abstract
In the last decade, the exclusive role of the hippocampus in human declarative learning has been challenged. Recently, we have shown that gains in performance observed in motor sequence learning (MSL) during the quiet rest periods interleaved with practice are associated with increased hippocampal activity, suggesting a role of this structure in motor memory reactivation. Yet, skill also develops offline as memory stabilizes after training and overnight. To examine whether the hippocampus contributes to motor sequence memory consolidation, here we used a network neuroscience strategy to track its functional connectivity offline 30 min and 24 h post learning using resting-state functional magnetic resonance imaging. Using a graph-analytical approach we found that MSL transiently increased network modularity, reflected in an increment in local information processing at 30 min that returned to baseline at 24 h. Within the same time window, MSL decreased the connectivity of a hippocampal-sensorimotor network, and increased the connectivity of a striatal-premotor network in an antagonistic manner. Finally, a supervised classification identified a low-dimensional pattern of hippocampal connectivity that discriminated between control and MSL data with high accuracy. The fact that changes in hippocampal connectivity were detected shortly after training supports a relevant role of the hippocampus in early stages of motor memory consolidation.
Introduction
Ever since the discovery of memory systems, the study of the neural mechanisms supporting the formation of declarative—e.g. episodic- and nondeclarative—e.g. motor—memories has progressed somewhat in parallel, with the former viewed as strongly dependent on the hippocampus and the latter on the cerebellum and the basal ganglia (Squire and Zola 1996; Poldrack and Packard 2003; Doyon and Benali 2005; Doyon et al. 2009, 2018; Albouy et al. 2013a). This distinction established a deeply rooted dichotomy in the field of neuroscience, with motor learning viewed as implicit and automatic, and episodic learning as explicit and heavily dependent on consciousness and attention.
In the last few years, however, this view has started to change. Studies in humans have identified the contribution of the hippocampus in the encoding and consolidation of newly learned predefined sequences of motor actions (motor sequence learning [MSL]) (Grafton et al. 2002; Albouy et al. 2013b; Albouy et al. 2015; Boutin et al. 2018; Buch et al. 2021). Recently, using functional magnetic resonance imaging (fMRI), we found evidence linking the hippocampus to memory encoding during MSL (Jacobacci et al. 2020). Improvements in performance during learning in this task take place during the quiet rest periods interleaved with practice, rather than during execution (Bönstrup et al. 2019). Using fMRI, we found that whereas cortico-cerebellar and cortico-striatal systems were most active during task execution, the hippocampus was most active during the rest periods interspersed with practice, a phenomenon reminiscent of the behavioral benefits produced by task replay in rodents (Pfeiffer and Foster 2013). Critically, activity in the left anterior hippocampus peaked during initial learning and predicted improvements in performance, pointing to a role of this structure in memory reactivation.
Yet, motor skill can also develop offline as memory stabilizes (Robertson et al. 2004; Walker and Stickgold 2004). To determine whether the hippocampus participates in the consolidation of motor sequence memories, here we examined the changes in functional connectivity that took place offline through a 24 h period. We used resting-state fMRI (rs-fMRI), as it allows identifying spontaneous oscillations in brain activity in the absence of external stimulation or movement (Beckmann et al. 2005), thereby avoiding potential confounds that may arise as training proceeds (Della-Maggiore et al. 2015). To this aim, we analyzed the rs-fMRI data acquired as part of our previous study (Jacobacci et al. 2020) before, 30 min and 24 h post learning the motor sequence (MSL) or an active control (CTL) task. This longitudinal approach allowed us to evaluate whether learning-induced changes in hippocampal connectivity are transient or persistent, thus reflecting its involvement in early or late phases of motor memory consolidation.
To examine the impact of motor learning on functional connectivity and topology, we performed a network analysis using three different approaches that allowed us to track changes in the functional connectome following data-driven (first two approaches) and hypothesis-driven (last approach) strategies. We first explored global network changes by means of two graph-theoretical measures, namely global efficiency and modularity, each of which characterizes, respectively, specific aspects of information integration and segregation processes across the entire functional network. We examined their relevance to motor skill learning and consolidation by using a longitudinal design that allowed us to quantify changes in these metrics relative to a pre learning condition, and throughout a 24-h window. Next, to identify the relevant connections driving learning-related changes in global topology, we centered the analysis at the edge-level, considering each edge between pairs of nodes of the whole brain. Finally, to corroborate further the relevance of specific hippocampal connections to sequence learning, we performed a multivariate analysis using machine learning techniques to identify patterns or feature combinations of hippocampal edges that serve to discriminate between pre and post learning hippocampal-network states.
Materials and methods
Several aspects of the subjects, study design, and data preparation steps used in this functional network study were already described in a recent multimodal MRI study (Jacobacci et al. 2020), but are repeated here for completeness.
Participants
In all, 21 subjects between 19 and 31 years of age (11 female; ages: mean ± SD = 24.4 ± 2.7) participated in this study. All participants were healthy volunteers with no self-reported history of psychiatric, neurological, or cognitive impairment, nor any history of sleep disturbances. None of them had previous experience playing a musical instrument. Subjects were asked to abstain from alcohol and caffeine the day before and during the experiment to avoid sleep disturbances; caffeine may alter the temporal dynamics of the BOLD response (Liu et al. 2004). They were also instructed to maintain their regular sleep habits throughout the duration of the experiment. Also, none of the subjects worked night shifts or were engaged in transmeridian trips before or during the study. All volunteers were right-handed as assessed by the Edinburgh handedness inventory (Oldfield 1971). Participants provided written consent and were paid for their participation. The experimental procedure was approved by the local Ethics Committee (Hospital de Clínicas, University of Buenos Aires), and performed according to the Declaration of Helsinki.
Experimental design and procedure
The study followed a longitudinal design. The same participants trained on three different sensorimotor tasks on three separate sessions that were scheduled one week apart: an MSL task, a visuomotor adaptation task, and an active control (CTL) task involving no learning. The order of practice on these tasks was counterbalanced. Only the data corresponding to the MSL and CTL tasks are reported here, however. Diffusion-weighted (DWI), T1-weighted (T1w), and blood-oxygen-level-dependent (BOLD) fMRI images were acquired during the study. DWIs and resting-state BOLD images (rs-fMRI) were obtained before practice (at baseline), 30 min and 24 h post training. BOLD images were also obtained during MSL training. Note that, in this study, we only report the results of analyses using rs-fMRI data (Fig. 1A); the DWI and task-related fMRI data have been published in a previous study from our laboratory (Jacobacci et al. 2020).
Experimental approach. (A) Experimental procedure. Subjects underwent a rs-fMRI baseline session before performing the MSL and CTL tasks, 30 min and 24 h after training. Each rs-fMRI session lasted 10 min. (B) Brain parcellation and connectivity analysis. Shown are the cortical and subcortical parcellations that were used to define the set of regions of interest (left panel), the construction of the whole-brain functional connectivity matrices (center panel), and the three connectivity analyses performed using these matrices (right panel). As specified in Material and methods, global and edges analyses were carried out on the whole brain connectivity, whereas the main supervised classification was performed on the hippocampal connectivity profiles only.
Experimental paradigms
The MSL task required subjects to press a series of four keys using the fingers of the nondominant (left) hand following a sequence of movements of five elements (4-1-3-2-4, with 4 being the index finger and 1, the pinky finger, see Fig. 1A) (Debas et al. 2010; Albouy et al. 2015). Subjects were instructed to execute the sequence as quickly as possible without making errors. In the motor practice session, which lasted around 15 min, subjects executed the motor sequence in a self-paced manner in 15 blocks of 60 key presses separated by rest periods of 25 s each. A Test session, which comprised another eight blocks of practice of the same task, was carried out 24 h later to assess overnight offline gains in performance associated with sleep consolidation; the Test took place after the MRI acquisition. During practice of the MSL task, behavioral performance was quantified using the intertap interval, i.e. the time elapsed between successive key presses from correctly executed sequences. Specifically, we computed the difference (delta) in the mean intertap interval between key presses of correct sequences. The sum of changes in behavioral performance across blocks of the learning session (day 1) was computed based on the sum of deltas for all blocks (total learning). To compute overnight offline gains, we calculated the percent difference between the average intertap interval of the first three blocks of the test session (assessed 24 h post learning) relative to the last three blocks of the motor practice/learning session.
An active CTL task was also administered to subjects to account for the sensorimotor aspects of the MSL task. Subjects were instructed to press a button in response to a green target appearing on a screen (P = 0.9), and to withhold from responding whenever the target was red (P = 0.1). The duration of the CTL task lasted approximately the same as the MSL task. Both experimental paradigms were programmed using MATLAB’s Psychophysics Toolbox, Version 3 (Brainard, 1997).
MRI data acquisition, preprocessing, and brain parcellation
Magnetic resonance images (MRI) were acquired with a 3T Siemens Tim TRIO scanner using a 12-channel head RF receive coil (Instituto Angel Roffo, University of Buenos Aires, Argentina). Structural T1w sagital images were acquired using the MP2RAGE WIP sequence (Marques et al. 2010), with the following parameters: repetition time (TR) = 5000 ms; echo time (TE) = 2.89 ms; flip angle (FA) 1 = 4°; FA2 = 5°; inversion time (TI) 1 = 700 ms; TI2 = 2500 ms; bandwidth (BW) = 240 Hz/Pixel; field-of-view (FOV) = 256 × 256 mm2; acquisition matrix = 256 × 256; voxel size = 1 × 1 × 1 mm3; slices = 176; parallel acquisition = GRAPPA mode, acceleration factor 2. T1w were used for registration purposes. As illustrated in Fig. 1(A), rs-fMRIs were acquired before, 30 min and 24 h after performing the MSL and CTL tasks using the multiband-accelerated sequence implemented by the Center for Magnetic Resonance Research, University of Minnesota (Uğurbil et al. 2013; Xu et al. 2013). The following protocol was used for acquisition: voxel size = 3 × 3 × 3 mm3, FOV = 192 × 192 mm2, 42 slices aligned with the AC-PC line, 10% gap, posterior-anterior (P-A) phase encoding direction, TR = 1433 ms, TE = 30 ms, multiband acceleration factor = 2, no PAT, BW = 1502 Hz/Pixel, echo spacing = 0.75 ms, EPI factor = 64, FA = 69°. A gradient echo field map was also acquired to correct for field inhomogeneities following these acquisition parameters: 42 slices, 20% gap, voxel size = 3 × 3 × 3 mm3, FOV = 192 × 192 mm2, P-A phase-encoding direction, TR = 444 ms, TE1 = 4.92 ms, TE2 = 7.38 ms, FA = 44°, BW = 260 Hz/pixel.
The DICOM images were converted into an NifTI format using the dcm2niix software (Li et al. 2016), and the first five volumes of each BOLD acquisition were removed to allow for T1 stabilization. The preprocessing was performed using mainly the Statistical Parametric Mapping software SPM12 (Wellcome Department of Cognitive Neurology, London, UK) (https://www.fil.ion.ucl.ac.uk/spm/). Correction of magnetic field inhomogeneities using phase and magnitude data from a gradient-echo field map was performed along with realignment to the middle volume of each functional time series. This process was also applied to the Single-Band Reference image, a single-volume, high-contrast EPI image acquired in the same space as the fMRI data (Andersson et al. 2003). This was followed by slice-timing correction and bandpass filtering with cutoffs of 0.009 and 0.08 Hz (Biswal et al. 1995).
To reduce physiological noise, tissue masks were used to obtain the time course of the average signal from the white matter and cerebrospinal fluid, and regressed from the BOLD data using the fslmeants and fsl_glm commands from FSL.
The whole brain cortical surface of each subject was parcellated into 100 cortical parcels (Fig. 1B, left panel) using the Schaefer’s seven networks parcellation (Schaefer et al. 2017). This cortical parcellation assigns each region to one of seven resting-state networks according to the Yeo-Krienen atlas (Thomas Yeo et al. 2011): frontoparietal control (FPCN), default mode (DMN), dorsal attention (DAN), salience ventral attention (SVAN), limbic (LIMB), sensorimotor (SMN), and visual (VIS) networks. Additionally, we parcellated subcortical structures including the hippocampus (anterior and posterior), the thalamus, the amygdala, and the basal ganglia (putamen, globus pallidus, caudate, accumbens) into a set of 32 regions according to the Melbourne atlas (Tian et al. 2020) (Fig. 1B, left panel). Schaefer’s parcellation was selected because it is based on functional connectivity data and provides parcels that overlap with well-known resting-state networks (Thomas Yeo et al. 2011) that are consistently associated with different task domains. On the other hand, anatomical atlases perform worse than functional atlases or atlases obtained by means of clustering or linear decomposition methods in classification approaches in which resting-state fMRI data are used as input features (Dadi et al. 2019).
Lastly, FSL’s fslmaths and fslmeants were used to extract the time course of the signal from each parcel.
Data analysis and statistical approach
After normalizing each brain region’s signal using z-scores, we computed the Pearson correlation coefficient across the time series of all possible pairs of regions i and j (Aij) to construct a set of 132 × 132 static, nondirected, weighted, and signed adjacency matrices per session (Fig. 1B, center panel). For the global analysis with graph-theoretical metrics, we then removed the negative weights (i.e. anti-correlations) from the adjacency matrices to preserve only the edges associated with cooperation and integration between different brain regions, and finally thresholded these matrices using a density criterion (Fornito et al. 2016). In contrast, for the edge-level analysis and supervised classification, we kept the fully connected and weighted graphs.
Global-level analysis. In order to explore network-wide changes induced by MSL, we computed two global metrics, namely, global efficiency and modularity. Global efficiency is a metric that measures how efficiently information is exchanged across the network; the latter can be understood as the amount of potential information transfer that can be performed in parallel (Latora and Marchiori 2001). Formally, the global efficiency is the average inverse shortest path length in the network. On the other hand, the concept of modularity refers to a topological property that accounts for the presence of modules or communities of nodes that are densely connected between each other and that are sparsely connected to nodes of other modules. This metric can thus serve to uncover functional building blocks or subnetworks (Sporns and Betzel 2016). The computation of network modularity is part of a data clustering problem in which the network is partitioned into a number of K modules or communities (for a review of the different algorithms available for community detection, see Fortunato 2010 and Fortunato and Hric 2016). Here, we used an agglomerative clustering method, namely the Louvain algorithm, which maximizes the quality function Q and measures the quality of partitions (Blondel et al. 2008). A value of Q near 0 indicates that the network is poorly structured in terms of different communities, whereas a value of Q near 1 indicates a strong community structure. In this context, communities are thus defined as groups of nodes whose observed connection density is maximally greater than what would be expected by chance.
To maximize the separation between putatively real (signal) and spurious (noise) functional links, before computing the global efficiency and modularity, we first thresholded the connectivity matrices using a density-based (or wiring-cost) strategy (Fornito et al. 2016). To this aim, we computed the minimum spanning tree (Kruskal 1956; Alexander-Bloch et al. 2010) of each subject’s 132 × 132 adjacency matrix to ensure that the graphs remain connected, and we then iteratively added the strongest edges until the desired density was reached. The importance of using a density or sparsity-based threshold is twofold. On one hand, it ensures that the number of edges (or network global degree) of each subject’s graph was the same, hence avoiding some statistical biases and potential confounds that can emerge when comparing networks with different number of connections as those derived when using a weight-based thresholding strategy (van Wijk et al. 2010; Ginestet et al. 2011; Ginestet et al. 2014). On the other hand, it provides more stable measures across different density values compared to the use of a weight criterion (Garrison et al. 2015; van den Heuvel et al. 2017). To obtain robust results, instead of choosing a single and arbitrary density value, we computed the global metrics for a set of densities ranging from the 10 to 20% of the strongest positive connections with a step value of a 0.5% density increase at each iteration (21 densities were explored in total). In this way, for each subject and metric of interest, we obtained a curve that characterized the measure’s evolution as a function of network connection density.
Global efficiency and modularity values were expressed/normalized relative to those extracted from random networks that preserved the same number of nodes, edges, and node degrees of the empirical networks (Maslov and Sneppen 2002; Rubinov and Sporns 2011; Váša and Mišić, 2022). For each subject (n = 21), each task (n = 2), and condition (n = 3), we constructed 100 randomly rewired networks (12,600 random networks in total). We then computed the global metrics of interest for each of these random networks and built a null distribution. Finally, the global efficiency and modularity metrics corresponding to each individual were divided by the mean of the corresponding null distribution (Váša and Mišić, 2022).
We performed a statistical comparison of an integrated or summary value for each global measure—global efficiency and modularity—across densities (i.e. across the graph curve) (Ginestet et al. 2014) that we defined as the area under the curve (AUC) of each subject’s respective curves (Bassett et al. 2012; Fornito et al. 2013; Fornito et al. 2016). By computing this integrated value, we thus obtained a threshold-independent summary measure. Statistical analyses were then performed on these AUC’s across groups (MSL and CTL) and time points (baseline, 30 min, 24 h) to investigate longitudinal differences in global efficiency and modularity. Given the longitudinal nature of the study design, we employed a multilevel modeling approach by means of linear mixed-effects models (Snijders and Bosker 2011). To take into account the repeated measures across sessions, random intercepts and random slopes were estimated for each subject. Models’ significance was estimated by means of F-tests. Here we report the resulting value of F-tests and the corresponding P-values. Either network modularity or global efficiency were defined as the dependent variable, and factors task (two levels: MSL, CTL) and session (three levels: baseline, 30 min, 24 h) as independent variables. In addition to the main effects of task and session, we included the task × session interaction term. Finally, we performed post hoc, contrast analyses to identify specific differences between pairs of sessions. The P-values were Bonferroni-corrected for multiple comparisons.
Edge-level analysis. Although we were particularly interested in the hippocampal connectivity, we nevertheless adopted a data-driven approach to examine all the edges of the functional connectome. To identify relevant connections -or edges- driving whole-brain learning-related changes across the three time points (i.e. sessions) and tasks (MSL and CTL), we employed a simple mass-univariate hypothesis testing procedure consisting of conducting paired samples t-tests for each edge taken from the upper triangle of the 132 × 132 adjacency matrices (n = 8646), and correcting the P-values for multiple-comparisons using the false discovery rate (FDR; Benjamini and Hochberg 1995). Furthermore, a more liberal statistical analysis was conducted as a second approach for the CTL group on the edges that were identified as modulated by MSL to provide a more robust corroboration of our findings and determine the specificity of the learning-related changes on individual connections.
Supervised classification based on hippocampal features. A drawback of mass-univariate hypothesis testing procedures is that they only allow to make inferences at the level of individual elements taken in isolation. In addition, statistical significance does not always coincide with predictive relevance (Bzdok et al. 2020). In contrast, machine learning multivariate classification algorithms can be used to recognize and learn complex patterns from the data that contain relevant information to discriminate classes, groups, or conditions of interest and make predictions on held out or new/unseen data.
Here we applied a progressive feature elimination (PFE) approach and a supervised classification with a Random Forest classifier (RFC) (Breiman 2001) to identify hippocampal connectivity patterns that are modified by learning in the early phase of memory consolidation (see Supplementary Fig. S1 for a graphical representation of the employed pipeline). We implemented this framework in a main binary classification with balanced classes to discriminate pre and 30 min post learning hippocampal-network states using the connectivity profiles of the four hippocampal nodes of the network (left and right, anterior and posterior) as features. Although we were interested in the edges between the sensorimotor network from the Schaefer parcellation and hippocampal regions, we nevertheless included the connections with the rest of cortical and subcortical parcels to explore the possible presence of additional relevant information. This a priori selection of edges left us with a total of 524 features (4 hippocampal nodes × 131 nodes).
The PFE (Donnelly-Kehoe et al. 2018; Princich et al. 2021) was used to optimize the accuracy of the classifier and extract the group of N features whose joint contribution captures the between-session changes in hippocampal connectivity. The selection of RFC was based on three premises: (i) we were interested in considering both linear and nonlinear relationships between features; (ii) given that the number of samples was low, the application of regularization techniques should be an optional step; (iii) the interpretability of the relevant features used for prediction should be clear (Palczewska et al. 2013). RFCs are powerful ensemble models that generally introduce randomization in the model fitting process in two different ways: each tree in the forest is fit using a bootstrap sample of the training data, and each node uses a random sampling of features to discriminate the classes of interest. Here we used RFCs with a fixed number of trees (2000) and a random sampling of d features from the full set of m features in each split of a tree, where d = √m. The maximum depth of the trees was not predefined, so nodes were expanded until all leaves were pure (i.e. until all leaves contained less than two samples).
To avoid using information of the same subject both in the training and validation set, potentially biasing prediction accuracy, we decided to employ a leave-one-subject-out cross-validation scheme. At each iteration, the hippocampal connectivity belonging to the baseline and 30 min sessions of a single subject were left-out for prediction, thus using the connectivity matrices of the remaining subjects to train the classifier. The PFE was implemented as follows. On each iteration of the cross-validation, we first performed a single-fit model-based feature selection using the train data and all the features to obtain an “importance ranking” based on the Gini impurity index. This index was the impurity criterion we used to measure the quality of a split in a tree and compute each feature’s relative importance (Breiman 2001). Once we obtained the ranking, we iteratively fitted new models starting with a single feature (that with the highest importance score), and then progressively adding the next features in the ranking to predict the validation dataset. Following this procedure, we thus fitted as many models as features. In this way, we were able to analyze how the classifier’s accuracy evolved according to the different number of features that were used to train the models. Finally, we kept the first N features in the ranking, where N is the optimal number of features such that using more than N fails to improve the classifier’s performance. We used this fixed number of features to compute the accuracy (fraction of correct predictions), precision (the ability of the classifier not to label as positive a sample that is negative), recall (the ability of the classifier to find all the positive samples), the confusion matrix, the ROC curve, and to obtain each subject’s predicted probability of being in each class (pre learning and 30 min post learning). The confusion matrix C is constructed such that Ci,j is equal to the proportion of observations known to be in group i and predicted to be in group j.
To evaluate the performance of the models trained in the cross-validation, we created a null model by shuffling the target labels before running the cross-validation process. The AUC of this null model thus provided an estimate of what we could obtain at a chance level.
The analyses were performed using the Python Scikit-learn package (Pedregosa et al. 2011; Abraham et al. 2014).
Results
Behavior
The learning curve corresponding to the time course of the intertap interval as a function of practice on MSL is depicted in the original publication (Jacobacci et al. 2020). Here, we show the amount of total learning (sum of intertap intervals across the 15 practice blocks) and the overnight offline gains (Fig. 2). On average, subjects reduced the intertap interval duration as reflected in the sum of deltas depicted in Fig. 2(A) (P < 0.0001, one-sample t-test against zero), indicating that they successfully learned the motor sequence. Yet, overall, they didn’t exhibit significant overnight offline gains (Fig. 2B, P > 0.05, one-sample t-test against zero).
Behavioral performance. (A) Amount of total learning (sum of intertap intervals across the 15 blocks of practice). On average, subjects learned the motor sequence, reflected by the increase in the sum of deltas (P < 0.0001, one-sample t-test against zero). (B) Overnight offline gains. Subjects didn’t exhibit overnight offline gains (P > 0.05, one-sample t-test against zero).
Global-level analysis
To find out whether a unique session of MSL can rapidly drive changes in global information transmission across the whole-brain functional connectome, and whether these changes are sustained 24 h after training, we computed two graph-theoretical global metrics: network modularity and network global efficiency, which are indicative of processes of information segregation and integration, respectively.
We first examined potential changes in modularity using a linear-mixed effects multilevel modeling approach (Snijders and Bosker 2011) on the AUC obtained for both tasks and the three sessions (Fig. 3). We found that MSL increased the level of network modularity 30 min post training compared to the CTL group, but returned to baseline levels 24 h post training (Fig. 3B; task-by-session interaction, F(2, 40) = 3.86, P = 0.029, followed by pairwise, paired samples t-tests with Bonferroni correction on the estimated marginal means of each session: baseline vs 30 min session P = 0.025; 30 min vs 24 h session, P = 0.004; baseline vs 24 h session, P = 1). No difference between sessions was observed for the CTL task (Fig. 3B; pairwise paired samples tests across all sessions, P > 0.05).
Global analysis. (A) Group mean modularity values for each task and resting-state session across the explored range of network densities. Shaded areas depict 95% confidence intervals (CI). (B) Modularity expressed as the AUCs +/− SE of the mean, computed based on the data of (A) for each task and resting-state session. Modularity was significantly increased by MSL at 30 min, but decreased to baseline levels at 24 h. (C) Shown are the Pearson correlations between the baseline modularity level for each subject (left panel) and the modularity difference (baseline and 30 min; right panel) with the amount of total MSL learning. Those subjects with lower baseline levels of modularity were better learners than those with relatively higher starting level of modularity (r = −0.5, P = 0.014). Moreover, fast learners showed a stronger increment in brain modularity 30 min post learning (r = 0.48, Pcorrected = 0.044).
Following the examination of changes in modularity, we evaluated whether MSL influenced the global efficiency of the network. We found that the temporal pattern of change in global efficiency was similar for both MSL and CTL, with no significant task x session interaction (F(2,40) = 0.52, P = 0.593), which indicates that the average path length across all pairs of nodes and thus the efficiency of information transfer across the entire system did not change after MSL.
Finally, we examined whether changes in modularity were associated with the amount of learning. To do so, we explored the presence of a linear association between the amount of learning achieved during the practice session and the baseline level of modularity before the MSL practice session, and also between the amount of learning and the change in modularity observed 30 min post learning. We found that network modularity at baseline related negatively to the amount of learning (r = −0.5, P = 0.014), suggesting that a more integrated community-structure may represent an optimal network state that favors the ability to learn a novel motor sequence (Fig. 3C, left panel). This finding can be linked with previous work that showed that visual-motor connectivity at baseline is related with the learning rate (Mattar et al. 2018). On the other hand, we found that the difference in modularity levels between the pre learning and the 30 min post learning session correlated positively with the amount of total learning during the initial session of training (r = 0.48, P = 0.044, Bonferroni-corrected for two comparisons), indicating that better learning was associated with a more segregated community structure (Fig. 3C, right panel). No significant association was found between the level of offline gains observed overnight and the difference in modularity (r = −0.39, P = 0.16, CI [−0.071,0.06]), further strengthening a transient effect of MSL on the global analysis.
Altogether, our results suggest that MSL induced an increment in brain modularity leading to a more segregated functional network in terms of information transmission and processing, that related to the amount of learning. This configuration was transient in that it returned to baseline 24 h post learning. The fact that the baseline level of brain modularity correlated with the ability to learn a novel motor sequence points to this metric as a valuable predictive tool to be considered in clinical—rehabilitation—settings.
Edges analysis
Once we identified how MSL affected global information transmission in the whole-brain network, we explored learning-related changes at the edge level to identify the individual functional interactions between pairs of regions whose weights were modified by learning. For practicality, we split the analysis to identify changes in connectivity relative to the baseline that took place 30 min (baseline vs 30 min) and 24 h post training (baseline vs 24 h, and 30 min vs 24 h).
The mass-univariate analysis of the functional connectome revealed that a set of five edges between the left hemisphere hippocampus and bilateral sensorimotor areas (from now on referred to as the hippocampal-sensorimotor network, Fig. 4, left panel) and a set of seven edges between the bilateral striatum and premotor areas (referred to as striatal-premotor network, Fig. 4, right panel) presented a significant change in their weights 30 min post MSL training (baseline vs 30 min, P < 0.05 after FDR correction for all edges), but not at 24 h (baseline vs 24 h, P > 0.05 for all edges after FDR correction). The same tests conducted for all edges on the CTL task yielded no significant differences (baseline vs 30 min, P > 0.56 for all edges after FDR correction; baseline vs 24 h, P > 0.9 for all edges after FDR correction). Furthermore, a more liberal statistical analysis conducted for the CTL group on the 12 edges modulated by MSL (Fig. 4, left panel) yielded no significant differences (P > 0.24, FDR corrected for 12 edges), hence suggesting that the connectivity changes identified in these two networks were specific to learning.
Edges analysis. MSL-related changes in connectivity detected 30 min post learning in a hippocampal-sensorimotor network (left panel) and in a striatal-premotor network (right panel). Shown are the edges showing a significant difference between the baseline and 30 min conditions. As depicted in the distribution profiles, the functional connectivity of the hippocampal-sensorimotor network decreased whereas that of the striatal-premotor network increased 30 min post learning (baseline vs 30 min, P < 0.05 after FDR correction for all edges; baseline vs 24 h, P > 0.05 after FDR correction for all edges). The cortical parcels are labeled and color coded according to the Schaefer network parcellation.
Figure 4 illustrates the hippocampal and striatal networks following the color coding and functional labels of Schaefer’s parcellation (Schaefer et al. 2017). The hippocampal-sensorimotor network consisted of functional connections between the anterior portion of the left hippocampus, the node with the highest degree (d = 5), and the bilateral primary motor and somatosensory cortex (LH SMN 5 and RH SMN 5 of Fig. 4), bilateral supplementary motor area (LH SMN 6 and RH SMN 8), and the right superior parietal lobule (RH SMN 7). On the other hand, the striatal-premotor network consisted of functional connections between the left pre-premotor area (LH FEF 1), the anterior and posterior putamen and the bilateral thalamus, and between the ventral premotor cortex (LH DAN PrCv1), the left thalamus and the caudate. Critically, MSL modulated these two networks in an antagonistic manner, decreasing the connectivity of the hippocampal network (pointing to a breakdown or disruption in the cooperation and information integration between the hippocampus and SMN regions), and increasing the connectivity of the striatal network (indicating a process of increased information integration and cooperation). Importantly, no significant differences in edges’ weights were observed between the baseline and the 24 h sessions (P > 0.05 after FDR correction for all edges) for any of the tasks, suggesting that changes in connectivity induced by MSL were short lasting.
Taken together, the edge analysis identified transient learning-related changes in functional connectivity 30 min post training in two networks that varied in an antagonistic manner: an anterior, left hippocampal network that decreased its connectivity with sensorimotor regions (Fig. 4, left panel), and a premotor-striatal network that increased its connectivity (Fig. 4, right panel). The fact that changes in connectivity were not present in the CTL task supports the specificity of our findings.
Supervised classification based on hippocampal features
To get further insight into the potential role of the hippocampus in MSL, we performed a predictive analysis using a supervised learning approach. Given that global and edge analyses pointed to the hippocampus as a relevant structure in early stages of memory consolidation, here we probe the hippocampal connectivity with the rest of the brain to see if we could decode a pre learning from a 30 min post learning state. The main objective of this analysis was thus to track and identify hippocampal connectivity patterns that may serve to discriminate between these states and quantify the relevance of each individual feature for this classification.
We performed a model-based feature selection procedure known as PFE (Donnelly-Kehoe et al. 2018; Princich et al. 2021) on the connectivity profiles of the bilateral anterior and posterior regions of the hippocampus (i.e. 4 hippocampal ROIs x 131 nodes corresponding to the rest of the brain network = 524 edges in total) to identify the optimal number of features that distinguished between pre and post learning states (Fig. 5A). This optimal number of features was used to predict held-out data using an RFC, following a leave-one-subject-out cross-validation scheme (see Materials and methods).
Supervised classification based on hippocampal features. (A) Progressive feature elimination (PFE) results showing the classifier accuracy as a function of the number of features used to train the model. The vertical straight line indicates the optimal number of features (n = 4). (B) Area under the curve of the true model (green) and of a null model constructed with shuffled target labels (blue). (C) Confusion matrix computed after the leave-one-subject-out cross-validation performed with the optimal number of features. The Random Forest correctly classified 86% of the baseline data, and 76% of the 30 min post learning data. (D) Predicted probabilities distributions of belonging to the 30 min session for the pre and post learning conditions. (E) Features importance during the progressive feature elimination process. Shown are the seven most important features, ordered by median Gini impurity. The four most important features are colored in green, and involve connections between the anterior hippocampus and the sensorimotor network (SMN). (F) Anatomical illustration of the four most important features/edges identified by the PFE that discriminate the pre from the post learning session.
The PFE process revealed that the highest accuracy in discriminating between pre and post learning sessions based on the functional connectivity of the hippocampus was achieved by four features (Fig. 5A); this represents a significant reduction of the dimensionality of the original feature space from a 524 to a four-dimensional space. Using this optimal number of features we obtained a mean classification accuracy of 80.95% predicting the test set across the leave-one-subject-out cross-validation; the AUC was 0.83, a value that greatly exceeds the 0.47 achieved by a null model trained with shuffled session labels (Fig. 5B). The percentage of correct and incorrect predictions for the baseline and 30 min sessions are depicted in the confusion matrix (Fig. 5C). As observed in the figure, the Random Forest correctly classified 86% of the baseline data, and 76% of the 30 min post learning data. In addition, the pooled predictions across the cross-validation achieved a Recall or Sensitivity of 0.76 and a Precision or Specificity of 0.84, supporting the robustness of the classifier.
The features importance across the PFE showed that the most stable and informative features were connections between the anterior portion of both left and right hippocampus and three bilateral sensorimotor regions (LH SMN 5 and SMN 6, RH SMN 7) (Fig. 5E and F). Interestingly, three of these four relevant features exhibited a significant decrease in functional connectivity 30 min post training in the edges analysis (see Fig. 4, left panel). The most important feature, defined as the one exhibiting the highest median Gini impurity score, in fact involved the left anterior hippocampus, which showed the highest degree (d = 3).
To evaluate the robustness of our results, we performed an additional classification using the connectivity profiles of the SMN nodes (n = 14) as features (14 × 131 features). We found that the most stable and relevant features were the same as the ones identified using the edges involving hippocampal nodes (see Supplementary Fig. S2). In other terms, even when edges connecting sensorimotor regions with the rest of the brain were available to the classifier, both classifications identified hippocampal-sensorimotor edges as the most informative features for discriminating the pre and 30 min post learning sessions.
Furthermore, to determine whether the information contained in this low-dimensional pattern of hippocampal-sensorimotor connections was specific to sequence learning, we ran a cross-validation using the four features identified by the PFE for MSL (Fig. 5D and F) on the CTL data. We found that the classification accuracy on the cross-validation process dropped to a near chance level performance of 52%, thus indicating that the classifier could not distinguish between the baseline and 30 min sessions (see Supplementary Fig. S3).
In sum, the results of our network-based predictive analysis via feature selection and supervised learning show that a single session of MSL rapidly drives functional changes in a low-dimensional pattern of hippocampus–sensorimotor connections that suffices to discriminate with high accuracy a pre learning hippocampal-network state from a post learning hippocampal-network state on held-out data. This low-dimensional pattern of connectivity was specific to MSL and consisted of four features, three of which were identified by the mass-univariate analysis of edges as showing a decrease in connectivity 30 min post learning, thereby strengthening the involvement of the hippocampus in the early consolidation of motor memories.
Discussion
Learning predefined sequences of actions is of everyday importance for humans. Recent studies in patients have reported impairments in encoding and consolidation of MSL when hippocampal integrity is compromised (Döhring et al. 2017; Long et al. 2018; Schapiro et al. 2019). In line with this neuropsychological evidence, we have shown that gains in performance occurring during the initial stages of MSL are linked to increased hippocampal activity reminiscent of memory reactivation (Jacobacci et al. 2020). Using a network neuroscience strategy, here we aimed to determine if the hippocampus was involved in MSL consolidation by tracking its functional connectivity offline early during the consolidation window (30 min post training) and after memory stabilized (24 h post training). By combining edge analysis with a network-based prediction analysis and machine learning techniques, our findings show that MSL drives transient changes in resting-state networks topology and connectivity early during memory consolidation at different levels.
At a global level, using graph-theoretical metrics we found that MSL induces changes in the brain meso-scale community structure. Specifically, we identified an increase in network modularity 30 min after learning a novel motor sequence that returned to baseline levels 24 h later. Our results advance previous work from Sami and Miall (2013) and Basset et al. (2011) on the impact of motor learning on brain modularity. Sami and Miall (2013) explored resting-state changes in this metric 5 min after learning on the serial reaction time task (SRTT) but found no significant changes in the community structure post training. The difference between their results and ours may be explained by the simplicity of the SRTT task and/or the fact that it was performed with the dominant hand; moreover, the post learning scan took place only 5 min post training. On the other hand, Bassett et al. (2011) measured brain modularity during training using our same experimental paradigm, but found no modulation by practice. However, in their study, network modularity was quantified during training (and not offline) and a control—non learning—condition was not included. The fact that the increase in modularity we observed for MSL correlated with the amount of learning, but not with overnight offline gains, provides strong evidence linking network segregation to early memory consolidation of the motor sequence.
Moreover, we found that the baseline level of modularity was a predictor of MSL. This result is in line with the literature pointing to this metric as a potential biomarker of intervention success and cognitive plasticity in executive functions (Gallen and D’Esposito 2019). For example, it has been shown that brain modularity at baseline predicts improvements in attention and executive function after cognitive training in TBI patients (Arnemann et al. 2015), and improvements in gist reasoning in older adults (Gallen et al. 2016). Likewise, a study with young adults who received cognitive training with video games engaging working memory and reasoning processes showed that baseline modularity was associated with gains in performance on untrained tasks (Baniqued et al. 2019). Our results extend the available literature by showing the relevance of the modularity metric as a valuable tool in the motor domain.
To understand further how MSL drives changes in specific functional connections between regions, we conducted both an edge-level analysis in which we analyzed the whole-brain functional connectivity and a multivariate analysis with machine learning techniques in which we focused on the hippocampal connectivity profile. Given that, as shown in our previous work, hippocampal activity increases during MSL and is associated with gains in performance (Jacobacci et al. 2020), we expected its connectivity profile to be substantially modified shortly after learning. In line with the global analysis, we found that MSL drove transient –short-term- changes in functional connectivity 30 min but not 24 h post learning. Interestingly, the edge-level analysis revealed that these changes were characterized by a decrease in the functional connectivity of an anterior left hippocampal-sensorimotor network, and an increase in the connectivity of a striatal-premotor network. Whereas the former may reflect a segregation or breakdown in the cooperation and information transmission induced by sequence learning, the latter may be indicative of integration or cooperation between these regions.
In humans, hippocampal activity increases early during MSL (Fletcher et al. 2004; Albouy et al. 2008; Gheysen et al. 2010; Jacobacci et al. 2020) and decreases as practice progresses (Grafton et al. 1995; Doyon et al. 2002; Schendan et al. 2003; Fletcher et al. 2004; Jacobacci et al. 2020). In contrast, the sensorimotor portion of the striatum (i.e. putamen) is more active during late phases of training, once subjects reach asymptotic performance (Doyon et al. 2003; Wymbs Nicholas et al. 2012). This temporal pattern may reflect the transition from a goal- to a muscle-based encoding frame of reference that takes place with practice and during early consolidation (Cohen et al. 2005; Albouy et al. 2015; Buch et al. 2021). The alternation we observed in resting-state connectivity between cortico-hippocampal and cortico-striatal networks shortly after training is in line with the sequential recruitment of these structures during motor learning. It is important to emphasize that whereas the anatomy of the striatal-premotor network is supported by abundant literature (e.g., Lanciego et al. 2012; Milardi et al. 2019), there is no evidence of direct pathways connecting the hippocampus with sensorimotor regions (Maller et al. 2019). Thus, it is most likely that the hippocampus is associated with the motor cortex indirectly via precuneus-prefrontal pathways (Margulies et al. 2009) connecting the limbic system with higher-order associative motor areas (Picard and Strick 2001).
Finally, via a feature selection and optimization strategy followed by a supervised classification, we demonstrated that a low-dimensional pattern composed by four sensorimotor-hippocampal connections can serve to discriminate with high accuracy a pre learning from a post learning network state, and thus predict if a subject has already performed the MSL task. In agreement with the univariate analysis of edges, the left anterior hippocampus was the most salient region of the connectivity pattern that best classified pre and post learning states (i.e. the region with the highest degree). Critically, we have previously shown that this region is both linked to gains in performance during the rest periods of MSL and undergoes changes in mean diffusivity compatible with structural plasticity 30 min post training (Jacobacci et al. 2020), thereby strengthening the role of this hippocampal region in early stages of procedural memory consolidation.
Collectively, our findings show that learning a novel sequence of finger movements transiently forces the brain topology toward a regime of more local information processing 30 min post training, as reflected by the increment in the level of network modularity. This increase was associated with improvements in performance observed during training, hence pointing toward a learning-specific phenomenon. The reconfiguration of the global topology was accompanied by fast antagonistic connectivity changes in two different neural networks: a sharp drop in the connectivity between the hippocampus and sensorimotor (motor and somatosensory) cortex, which points to a breakdown in the information transmission and integration across this hippocampal-sensorimotor network, and an increase in connectivity of a network comprising premotor and striatal regions. It is important to emphasize that both the transient modulation of the hippocampal-sensorimotor connections identified in the edge analysis and the most relevant feature identified in the supervised classification were centered around the left anterior hippocampus, the same region whose activity increases during the rest periods interspersed with learning a novel sequence of movements (Jacobacci et al. 2020). In conclusion, the fact that changes in hippocampal connectivity were detected shortly after training, but reverted to baseline levels 24 h later, supports the relevant role of the hippocampus in early stages of motor memory consolidation.
Acknowledgements
We thank Agustin Solano and Guillermina Griffa for their insightful comments on the manuscript.
Funding
This work was supported by a collaborative grant from the Quebec Bio-Imaging Network (QBIN, Canada [(RI-05)]), a grant from the Argentinian Ministry of Defense (PIDDEF 2014–2017/#17), grants from the National Agency for the Promotion of Science and Technology (ANPCyT), and the National Agency for the Promotion of Science and Technology (Grant PICT-2015-0488).
Conflict of interest statement: None declared.
Data availability
The dataset and scripts used to process and analyse the data may be accessed upon request.
