Brain-wide dynamic coactivation states code for hand movements in the resting state

Significance

The brain at rest replays activity patterns observed during task performance, but the function of this replay remains unclear. Here, we test whether resting-state patterns encode information states by comparing activity during ecological and naïve hand movements with resting-state activity. During movement, the primary motor cortex dynamically coactivated with sensory and association regions. Strikingly, similar patterns were also present during rest, with greater task-rest similarity for ecological movements. These results reveal that movement representations are distributed across the cortex, even at rest. This widespread coding may support prediction or long-term memory processes.

Abstract

Resting brain activity, in the absence of explicit tasks, appears as distributed spatiotemporal patterns that reflect structural connectivity and correlate with behavioral traits. However, its role in shaping behavior remains unclear. Recent evidence shows that resting-state spatial patterns not only align with task-evoked topographies but also encode distinct visual (e.g., lines, contours, faces, places) and motor (e.g., hand postures) features, suggesting mechanisms for long-term storage and predictive coding. While prior research focused on static, time-averaged task activations, we examine whether dynamic, time-varying motor states seen during active hand movements are also present at rest. Three distinct motor activation states, engaging the motor cortex alongside sensory and association areas, were identified. These states appeared both at rest and during task execution but underwent temporal reorganization from rest to task. Thus, resting-state dynamics serve as strong spatiotemporal priors for task-based activation. Critically, resting-state patterns more closely resembled those associated with frequent ecological hand movements than with an unfamiliar movement, indicating a structured repertoire of movement patterns that is replayed at rest and reorganized during action. This suggests that spontaneous neural activity provides priors for future movements and contributes to long-term memory storage, reinforcing the functional interplay between resting and task-driven brain activity.

Functional MRI (fMRI) is a valuable tool for investigating brain activity by tracking changes in blood oxygenation level–dependent (BOLD) signals over time, which serve as proxies for neuronal activity (1, 2). Task-based fMRI is employed to discern patterns of brain activation while individuals engage in specific tasks, by comparing task-evoked signals to those from a baseline control condition. When no task is being performed, such as during the resting state, fMRI signals reveal fluctuations that are temporally synchronized across different regions of the brain, a phenomenon known as functional connectivity (FC) (1, 3). Intriguingly, when FC is averaged over time, the resulting topography of temporally synchronized brain regions forms networks referred to as resting-state networks (RSNs), which resemble the topography of networks activated during task-based fMRI (4), although there is no agreement on the distinctiveness of these networks.

Physiological noise from respiratory or cardiac fluctuations or from movement artifacts can potentially cause problems in interpreting resting-state spontaneous brain signals obscuring components of the signal that are neuronal in origin (5). However, over more than two decades of research, it has become evident that spontaneous brain signals are organized spatially and temporally, representing a fundamental aspect of the brain’s functional architecture (6, 7). The similarity between task activation patterns and RSN topography suggests an important, yet undefined, role in shaping patterns of brain activation.

Early hypotheses suggested that spontaneous activity merely reflected poorly controlled cognitive processes that occur when subjects are not performing any task (rest) (8). However, studies on anesthetized individuals or during sleep have shown these patterns to be independent of ongoing cognition and awareness (9, 10). Under these conditions, spontaneous activity patterns closely approximate the anatomical connectivity (10). In contrast, in the awake state, the alignment between anatomy and function diminishes, and spontaneous activity patterns correlate with cognitive and behavioral variables (11–13), as well as with behavioral deficits across various pathological conditions (stroke, AD, and brain tumors) (14, 15). The connection with behavior, which is measured during tasks, implies that RSNs shape task-related patterns as if they were “priors” (16) or “scaffolds” for task-induced activation patterns (17).

More recently, two hypotheses have been proposed about the function of spontaneous activity. Laumann and Snyder proposed that spontaneous activity is closely linked to offline plasticity. In other words, resting-state signals contribute to the consolidation of “learning” novel task patterns through mechanisms like Hebbian learning and homeostatic plasticity (18). Pezzulo, Zorzi, and Corbetta proposed a “representation” model in which spontaneous activity fluctuations reflect “priors” of generative models for common perceptual, motor, cognitive, and interoceptive states (19). Both hypotheses indicate a role of spontaneous brain activity in coding for behaviorally relevant patterns of sensory, motor, or cognitive events, hence a role in the representation of information (20–23).

In this study, we test whether motor-related information about hand movements is represented dynamically in spontaneous activity brain patterns. We examine two predictions. First, we test whether “dynamic” states coding for hand movements also occur during the resting state. While previous studies primarily focused on the similarity in temporally averaged topography (1), we analyze temporally resolved spatial patterns and dynamics, including the duration of states and the transitions between states. It has been proposed that spontaneous brain activity might be primarily composed of brief bursts of activity that generate cascades of activity in connected regions (neuronal avalanches). These bursts can start in different regions and link different networks at different times underlying functional relationships that conventional static (time-averaged) data analysis does not detect (24). The representation theory (19) predicts that even before executing the motor task, resting-state activity should contain patterns resembling those occurring during movement, as spontaneous activity acts as a prior for movement. The learning theory predicts that resting-state patterns should resemble motor patterns after performing the motor tasks reflecting the influence of learning.

To ensure anatomical specificity, our analysis focuses on the activity of the hand motor region and correlated brain-wide coactivation patterns (CAPs), both during actual hand movements and rest. We employ the CAP analysis (25) to pinpoint brain-wide states that consistently activate or deactivate with the motor cortex. The CAP analysis identifies peaks of activity in one region (the motor cortex) and determines which other regions in the brain also show temporally synchronous peaks. A clustering analysis then identifies the most common patterns of coactivation over the whole brain (CAPs). Computational studies show that focusing on the peaks of activity explains the majority of the variability of spatial patterns obtained with standard temporal correlation FC (26, 27).

Second, we investigate whether the similarity between hand movement and rest CAPs varies depending on the familiarity of the movement. In previous work, we showed that common ecological hand movements “replay” more frequently in resting-state activity than patterns for a novel uncommon movement, both in the motor cortex (20) and in regions of the dorsal attention network (21). Here, we extend this analysis to brain-wide CAPs during movement and rest. Accordingly, CAPs specific to common ecological movements will occur more frequently, as compared to a novel uncommon movement, both before and after task execution. This is because priors in spontaneous activity represent the most frequent stimuli, actions, or cognitive operations of our life repertoire (19). Additionally, the learning theory (18) predicts limited similarity between resting-state and novel nonecological movements before the task but greater similarity after the task.

Results

Healthy observers (n = 13) were scanned with fMRI at rest (three 5-min blocks of visual fixation) before and after the execution of block-design hand movement tasks. The sample size for this study was estimated based on previous studies comparing patterns of spontaneous neural activity to patterns evoked by stimuli (23, 28, 29). Four different hand movements were performed, with each movement repeated for 10 s alternating with 20 to 24 s of visual fixation. Each movement was repeated randomly three times in each run, and five hand movement runs were scanned sequentially with a short break between runs. Three hand movements were selected for their frequency of occurrence in daily life [extend, grip (30), and pinch], while a fourth movement (wrist shake) was novel and not ecological. A CAP analysis identified CAPs with the hand motor region of interest (ROI) (Fig. 1) of the Yeo atlas, which overlapped with the mean motor region of activation [as in (20)] and showed a movement representation difference in (21). The CAP analysis was conducted separately for task blocks and resting-state runs.

Fig. 1.

Flow chart of the data analysis. (A) Preprocessed task and rest fMRI data were used to identify coactivation patterns (CAPs) with a specific motor region (from ref. 21). (B) The BOLD signal time series shown are from one task run of a participant. During this task run, the participant was asked to perform the shake, extend, grip, shake, extend, pinch, pinch, pinch, shake, grip, grip, and extend movements. The time periods during which the movement was being performed were shaded in purple. (C) Then, from the full time series only those frames exhibiting motor signals above a specific threshold were retained (shaded blue area) and fed into a K-means algorithm. (D) A consensus clustering approach was used to identify the optimal number of clusters. (E) Task-rest analysis. The identified CAPs were compared across resting-state and task fMRI conditions to assess the similarity patterns among maps. Voxel-wise spatial differences between CAPs were also assessed. (F) Statistical analysis for task-rest analysis. (G) Representation comparison. The frames identified as CAPs showing during the most activated TRs [shaded green area, this range was selected according to our previous study as it was the most sensitive range for discriminating between the different hand movements and showed a larger BOLD signal change (20)] in a specific block were averaged voxel-wise for each hand movement and each CAP to compute averaged task-evoked CAPs (e.g., CAP1-shake). This procedure yielded vectors for each CAP, one for each hand movement. A similar voxel vector procedure was adopted for resting-state data leading to multiple vectors, one for each TR. Averaged task-evoked CAPs (e.g., CAP1-shake) were correlated with resting-state data, resulting in a distribution of correlation coefficients, which is centered at zero and is spread with long positive and negative tails. The task-evoked-pattern-to-rest correlation time series shown are from one CAP and movement combination (e.g., CAP1-shake). The histogram of correlation coefficients is colored with different shades of blue due to resampling. To compare the strength of the correlation, a cumulative distribution function (CDF) of the squared Pearson’s values was computed and 90th percentile cutoff values were identified for comparison.

Task fMRI CAPs.

In the first analysis, we defined dynamic states during the movement tasks. BOLD signal frames (i.e., volumes) with a t-value above 1 [in line with previous studies (25, 31, 32)] within the motor ROI were selected as the objects for the CAP analysis. For each selected frame, the coincident whole-brain pattern of coactivation was recorded. Around 20% of task frames (Mean ± SD = 19.69 ± 2.43%) were retained and the consensus clustering algorithm run on the coactivation maps for the selected frames settled on K = 3 as the optimal number of CAPs’ clusters (Fig. 2). From CAP1 to CAP3, the number of frames assigned to the CAP decreased. Group-averaged CAPs are reported in Fig. 3. Details for the clusters’ anatomical locations are reported in SI Appendix, Table S1. The first CAP (CAP1) showed motor coactivation within the precentral gyrus, postcentral gyrus, supplementary motor area, paracentral lobule, and regions encompassing the anterior DAN and VAN (e.g., middle frontal gyrus, inferior/superior parietal gyrus, and supramarginal gurus). A motor co-deactivation was reported in visual areas and regions belonging to the DMN (e.g., the posterior part of the cingulate gyrus and precuneus). The second CAP (CAP2) showed coactivation in the motor cortex and visual regions, while co-deactivation occurred in DMN regions (e.g., precuneus, medial prefrontal cortex, and anterior and posterior part of cingulate gyrus). Finally, the third CAP (CAP3) showed coactivation within the motor cortex, DMN (e.g., medial prefrontal cortex, anterior and posterior part of the cingulate gyrus), and visual occipital regions, along with a co-deactivation in posterior regions of DAN (e.g., inferior/superior parietal gyrus), and VAN. CAP1 was highly similar to the contrast map of the motor task (averaging across all movements) vs. the fixation baseline previously reported (21) (r = 0.71, P < 0.001). To avoid confusing the CAP names obtained from resting state and task data, we labeled CAPs from the task condition in alphabetical order (A, B, C). Specifically, task CAP1, CAP2, and CAP3 were labeled CAP A, CAP B, and CAP C, respectively.

Fig. 2.

Identification of CAPs. Top panel: Stable clusters of CAPs were identified with k = 3 for both task (Left panel) and resting state (Right panel) conditions. K-means clustering was performed across 20-folds (red-white-blue colored), with the number of clusters (K) ranging from 3 to 8. For each fold, 80% of the data were randomly selected. Stability values, ranging from 0 to 1, are displayed for each fold and ordered in ascending order. Higher stability values indicate more stable clusters (highlighted with a black rectangle). Bottom panel: In the Top rectangle, each point refers to a single frame assigned to a specific CAP during task fMRI; most of the retained frames are aligned with the task blocks (task execution). In the Bottom rectangle, each colored point refers to a single frame assigned to a specific CAP during the resting state; the retained frames show a random pattern compared to task runs.

Fig. 3.

CAPs analysis. Left panel: Maps of CAPs from task data (Top rows) and resting-state data (Bottom rows). Top Right panel: Spatial similarity between group-averaged CAPs computed during task and rest conditions. Bottom Right panel: Spatial differences between CAPs computed during task and rest conditions. Differences were expressed as paired-sample t test between rest and task CAPs grouped according to spatial similarity (A vs α; B vs β; C vs γ) and plotted at the surface level.

We repeated the CAP analysis for task data with different T threshold values (T = 0.8, 1.2, and 1.4) to show that the results generalized over thresholds. Consensus clustering results showed that K = 3 was the optimal number of CAPs’ clusters for different T thresholds (SI Appendix, Fig. S1). For each T threshold condition, three group-averaged CAPs were computed and compared to the three group-averaged CAPs obtained using the default T threshold (T = 1) to investigate the spatial similarity of CAPs between different T thresholds. A high similarity (r > 0.9) was reported (SI Appendix, Fig. S1). Moreover, we confirmed 3 clusters as the optimal cluster number by comparing CAPs obtained with different K values (K = 2, 4, and 5) to CAPs obtained when K equals 3. Details of the map are reported in SI Appendix, Fig. S2.

Furthermore, we also used the left-hand ROI (SI Appendix, Fig. S3) from (33) as an ROI to repeat the CAP analysis. Spatial similarity between the group-averaged CAPs using the hand ROI and the motor parcel (in the main analysis) was high (r > 0.99, P < 0.001), suggesting robust CAPs.

To show that we can validate these CAPs even in a smaller sample size, we randomly selected 5 participants from the whole dataset and repeated the CAP analysis by using the same motor ROI and the same T threshold (T = 1). The result of consensus clustering settled on K = 3 as the optimal K number and the spatial correlation between CAPs across the whole group and subgroup of participants was high (CAP A r = 0.85, CAP B r = 0.93, and CAP C r = 0.94, respectively with P < 0.001 for all comparisons) (SI Appendix, Fig. S4).

CAPs Selectivity for Different Movements.

Frames selected as CAPs were aligned to the task blocks (Fig. 2) and mostly corresponded to the periods showing the highest BOLD signal change (SI Appendix, Fig. S5A), i.e., range of 10 to 18 s after the beginning of the movement trial) (20). An ANOVA showed significant movement effects for CAP A (task CAP1) [F(3, 36) = 3.28, P = 0.032, η_p² = 0.215] and CAP C (task CAP3) [F(3, 36) = 4.19, P = 0.012, η_p² = 0.259] in terms of the number of frames assigned as a CAP (i.e., occurrence) during each movement. Post hoc analysis showed that frames classified as CAP A were more common during “extend” (Mean ± SD = 9.76 ± 4.41) than “pinch” movements (Mean ± SD = 6.35 ± 3.03) (Bonferroni corrected P = 0.0297). Also, frames classified as CAP C were more common during grip (Mean ± SD = 6.63 ± 2.68) than shake (Mean ± SD = 4.21 ± 2.42) (Bonferroni corrected P = 0.0419) or pinch (Mean ± SD = 4.21 ± 2.58) (Bonferroni corrected P = 0.0424) movement (SI Appendix, Fig. S5B). No significant movement difference occurred for frames classified as CAP B (task CAP2) [F(3, 36) = 1.72, P = 0.18]. In general, this analysis shows that familiar and novel movements yielded similar spatiotemporal task patterns, i.e., a similar occurrence of high-magnitude frames used to compute CAPs. However, in the case of CAP C the grip (familiar) movement yielded more high-magnitude frames than the shake (novel) movement.

Rest fMRI CAPs.

In the second analysis, we examined whether CAPs occur at rest and whether they resemble those found during the motor tasks. About fourteen percent of pretask resting-state frames (Mean ± SD = 14.31 ± 1.78%) were retained and K = 3 was also the optimal K (Fig. 1). Rest-state CAPs are shown in Fig. 3. Details for the clusters’ anatomical locations are reported in SI Appendix, Table S2. Notably, frames selected as CAPs in the resting state were distributed more randomly across time, as compared to task scans in which the CAPs were present only during the movement periods (Fig. 2). This indicates that CAPs appear randomly at rest, but then phase reset when movements start.

Interestingly, there was a moderate spatial correlation between task and pretask resting-state group-averaged CAPs with r ranging from r = 0.65 to r = 0.76 for different CAPs (r > 0.6, P < 0.001, Fig. 3), and it was possible to find a one-to-one correspondence between different task and pretask resting state group-averaged CAPs. Pretask resting-state CAP1 correlated with task CAP2 (i.e., CAP B) (r value = 0.76, P < 0.001); pretask resting-state CAP2 corresponded to task CAP1 (i.e., CAP A) (r value = 0.72, P < 0.001); and pretask resting-state CAP3 corresponded to task CAP 3 (i.e., CAP C) (r value = 0.65, P < 0.001). To show the one-to-one correspondence by mirrored codes but without confusion, resting state CAPs were labeled using the Greek alphabet (α, β, ϒ). Specifically, CAP A from the task condition is matched with CAP α from the rest condition, and so on. Thus, pretask resting-state CAP1, CAP2, and CAP3 were labeled as CAP β, CAP α, and CAP γ, respectively.

A similar analysis was carried out in the posttask resting state where fourteen percent of frames (Mean ± SD = 14.17 ± 1.58%) were retained for CAPs generation. There was a strong spatial correlation between task and posttask resting state group-averaged CAPs with also a one-to-one correspondence (r > 0.7, P < 0.001). Based on the spatial similarity to task CAPs, rest CAPs were also labeled as CAP α, CAP β, and CAP γ, respectively. Finally, CAPs from the pretask resting state and posttask resting state showed a strong spatial correlation (r > 0.8, P < 0.001, Fig. 4).

Fig. 4.

CAP differences between pre- and posttask rest. Left panel: Spatial changes after task performance between resting-state CAPs (red highlights higher values (coactivation) at posttask resting state; blue highlights higher values at pretask resting state). Central panel: Coactivation magnitude difference between pre- and posttask resting state within motor regions from Gordon’s atlas (33). These changes correspond to CAPs in the Left panel. *marks significant difference. Right panel: Spatial similarity matrix between pre- and posttask resting-state averaged CAPs.

Sensitivity analyses were also performed for pretask resting state data as for task data by repeating the CAP analysis with different T threshold values (T = 0.8, 1.2, and 1.4) and different K values (K = 2, 4, and 5). We reported that 3 clusters as the optimal cluster number when T equals 0.8 and 1.4 (SI Appendix, Fig. S1). We also confirmed this optimal K (K = 3) by comparing CAPs obtained with different K values (K = 2, 4, and 5) to CAPs obtained when K equals 3 (SI Appendix, Fig. S6). Thus, the optimal K value of 3 is quite robust, although when T equals 1.2, the optimal K value is 4 while the second optimal K value is 3.

Additionally, an independent set of resting-state scans was obtained from a different cohort of subjects (22, 34), and the CAPs analysis was repeated using the same motor ROI (SI Appendix, Fig. S4). The spatial correlation between CAPs across different groups of subjects was high [CAP α r = 0.72, CAP β r = 0.86, and CAP γ r = 0.76, respectively (P < 0.001)].

In summary, these findings show that during the resting state, even before the execution of the hand movements, brain-wide coactivation states with the motor cortex are very similar to those present during the task. Moreover, these states persist at rest after the execution of the different motor tasks and these resting motor-related patterns were consistent across different groups of subjects. Overall, these findings indicate a surprising similarity of brain-wide patterns of activation at rest and during hand movements. It is therefore critical to understand what differences if any exist both spatially and temporally between these patterns.

CAPs Spatial Changes from Rest to Task.

Spatial differences between pretask resting-state CAPs and task-state CAPs were compared voxel-wise across the whole brain at the group level. For all three CAPs, going from rest to task, we found stronger activation in the left dorsal motor cortex (hand region), contralateral to the performing hand, and bilaterally in the temporoparietal junction (TPJ) based on a connectivity-based parcellation atlas (35). There were also robust deactivations after multiple comparison corrections (Gaussian Random Field (GRF) theory correction: voxel-wise P <0.001, cluster-wise P < 0.025, two-tailed) in the contralateral right motor hand area, and in several regions along the motor strip (Fig. 3). To quantify these effects, we used ROIs based on the atlas from (33) (SI Appendix, Fig. S3). This analysis confirmed significant task-minus-rest activation in the left-hand region (P < 0.001), and significant deactivation in the right-hand region (P < 0.005), bilateral foot (P < 0.001), bilateral mouth (P < 0.05), and bilateral intereffector regions (P < 0.05) (SI Appendix, Fig. S7 and Table S3). Similar results were obtained when using the task CAPs as references to identify CAPs from resting-state data and compare pretask resting-state vs task-state CAPs’ magnitudes in the motor cortex. Details in SI Appendix, Fig. S8 and Table S4.

CAPs Spatial Changes from Pretask to Posttask Rest.

Spatial differences between pretask resting-state CAPs and posttask resting-state CAPs were also compared voxel-wise across the whole brain at the group level. Briefly, after task performance, there was an overall deactivation across the motor cortex as compared to pretask resting-state magnitudes. All ROIs in the motor cortex showed a consistent deactivation (P < 0.05) except for the mouth region in the CAP B vs β, and the intereffector regions in the CAP C vs γ (Fig. 4 and SI Appendix, Table S3).

Similar results were obtained when using the task CAPs as references to identify resting-state CAPs and to compare pre- vs. posttask resting-state CAPs’ magnitudes in the motor cortex (SI Appendix, Fig. S9 and Table S4).

In summary, these spatial analyses of the magnitude of response in the motor cortex clearly show that performing the motor task does not change the resting functional organization of the motor system, but it decreases the magnitude of activation across the motor cortex when comparing pre- to posttask resting activity.

CAPs Temporal Changes from Rest to Task to Rest.

Given the relative stability of rest-task spatial patterns of coactivation, it is important to ask whether the execution of the task reorganizes the duration of brain states and the transition between brain states. Accordingly, we compared three temporal metrics and their modulation in going from rest to movement, and then back to rest. Duration quantifies the duration of CAP persistence once it starts; out-degree represents the probability of transitioning from a CAP to any other CAPs, whereas in-degree represents the probability of entering a CAP from any other CAPs. These metrics were computed considering baseline frames over time. Baseline refers to the frames not selected as CAPs.

To investigate the difference in temporal metrics, we ran two different models. The first ANOVA model investigated the temporal metrics accounting for CAPs and three time points (pretask resting state, task state, and posttask resting state). There was a significant main effect of time in duration [F(2, 48) = 5.32, P = 0.012, η_p² = 0.307], out-degree [F(2, 48) = 33.68, P < 0.001, η_p² = 0.737], and in-degree [F(2, 48) = 33.68, P < 0.001, η_p² = 0.737]. Performing the motor task caused CAPs to last longer (longer duration: Bonferroni-corrected P = 0.0495), being activated more frequently (higher in-degree: Bonferroni-corrected P < 0.001)), but also to shift to other CAPs more frequently, as compared to pretask resting state (Fig. 5 and SI Appendix, Table S5). Similarly, task periods showed a higher in-degree (Bonferroni corrected P < 0.001), as compared to the posttask resting state. Notably, temporal metrics for pre- vs. posttask resting-state showed a lower in-degree in posttask resting-state CAPs (Bonferroni corrected P < 0.001) (SI Appendix, Table S5).

Fig. 5.

Temporal changes from rest to task to rest. The Y axis represents the values of specific temporal metrics (panel A: duration; panel B: out-degree; panel C: in-degree) computed from the CAPs. The X axis represents the three different conditions (pretask resting state, task state, posttask resting state). *marks significant differences (P < 0.05) (black marks code for significant differences in temporal metrics across all the CAPs, while colored marks represent significant differences in temporal metrics for the corresponding colored CAPs between different time conditions.)

There was also a significant interaction of different CAPs by time in the probability of exiting a CAP state vs. any other CAP state [out-degree: F(2, 48) = 7.90, P < 0.001, η_p² = 0.397]. A post hoc analysis showed a higher out-degree for CAP B (β) and CAP A (α) during task states than pre- or posttask rest states (all comparisons Bonferroni corrected P < 0.001).

Performing the motor task decreased the frequency of CAP shifts (lower in- and out-degree) in the posttask rest state as compared to the pretask rest state. This effect was significant in an ANOVA comparing directly the temporal metrics across two time points (pre- and posttask rest) as well as in the above three time point ANOVA (pre-, task, posttask rest) (SI Appendix, Table S5).

When we used task CAPs as references to identify the corresponding CAPs from pre- and posttask resting-state data and performed the same analysis, we validated the above results. Details in SI Appendix, Table S5.

In summary, as compared to spatial CAPs that were very similar going from rest to task, and vice versa, temporal metrics were strongly modulated by performing the motor task. CAPs lasted longer and shifted more frequently. We also found some effects of movement performance on posttask resting temporal metrics.

CAPs for Familiar vs. Novel Movements and Similarity to Resting-State Patterns.

In the last analysis, we test whether the similarity between task-related and resting-state brain-wide CAPs varies depending on the familiarity of the movement. The representation theory posits that, both before and after task execution, brain-wide CAPs specific to familiar ecological movements will resemble more closely resting-state patterns. This is because priors primarily reflect common movement patterns. Additionally, the learning theory predicts limited similarity between resting-state and novel nonecological movements before the task, but greater similarity after the task.

To investigate this question, in addition to using the whole motor hand area to generate CAPs, we focused on subregions of the motor cortex that were specific for either familiar or novel movements. We chose for this analysis grip vs. shake movements that yielded in a previous publication the most difference within the motor hand area (20). Grip or shake-specific ROIs were generated based on the T-score activation map (i.e., contrast map of grip or shake block movements minus fixation baseline) at multiple thresholds (t = 6, 8,10,12) (Fig. 6A). Then, we extracted CAPs brain-wide starting from either the whole hand motor ROI or grip (shake) specific ROIs. Brain-wide CAPs generated from each ROI were further optimized by keeping only those frames that occurred during either grip or shake movements. Hence, we extracted CAPs 1–3 generated only from grip frames, and CAPs 1–3 generated only from shake frames. The corresponding CAPs 1–3 maps for grip and shake, respectively, were averaged within each subject to obtain a brain-wide grip-specific or shake-specific CAPs 1–3. At this point every voxel in the brain contains an average signal magnitude value corresponding to either grip or shake movements, i.e., they are multivoxel pattern vectors associated with a specific movement. Next, we measured frame-by-frame the spatial similarity between task (grip or shake specific CAP 1/2/3) and resting state multivoxel patterns by computing the percentage of frames with a similarity above the 90th percentile of all r-squared values. This threshold (90th percentile) was selected according to previous literature (20–22, 36). These cutoff values were then used for statistical evaluation (Fig. 1).

Fig. 6.

Movement representation comparison between grip and shake. (A) Motor parcel and movement-specific regions of interest (ROIs). The t-map of the contrast for each movement-vs-baseline (i.e., Grip-vs-baseline and Shake-vs-baseline) from ref. 21 was thresholded with a T score from 6 to 12 with steps of 2 and used as ROIs. (B) r² cutoff values for the patterns from grip and shake hand movements considering both pretask and posttask resting state time points. We computed the CDF of r² values across resting frames, i.e., the square of the Pearson correlation values that were computed to assess the similarity of task and rest multivoxel patterns. A cutoff score was defined as the r² value corresponding to the 90% point on the CDF (see Fig. 1 for a CDF plot). Higher cutoff values signify greater task-rest similarity. For CAPs identified by different thresholds, Grip showed a higher value. (C) r² cutoff values for the different hand movements’ patterns computed separately for the pretask and posttask resting state. For CAPs identified by different thresholds and different time points, Grip showed a higher value.

Two different models were tested. The first model (three-way interaction) investigated the similarity pattern accounting for movements (grip and shake), CAPs (CAPs 1–3), and time (pretask resting state and posttask resting state). When using the whole motor parcel as the ROI to identify CAPs, we found a near significant main effect of movement [F(1, 24) = 3.72, P = 0.078] with cutoff values for grip and shake 0.099 and 0.093, respectively. When we used the movement-specific ROI to identify CAPs, a similar difference trend between grip and shake was observed [t-score threshold = 6: F(1, 24) = 3.61, P = 0.082, grip cutoff value = 0.098, shake = 0.092, respectively; t-score threshold = 8: F(1, 24) = 1.02, P = 0.332, grip cutoff value = 0.098, shake = 0.093; t-score threshold = 10: F(1, 24) = 8.26, P = 0.014, η_p² = 0.408, grip cutoff value = 0.10, shake = 0.094; t-score threshold = 12: F(1, 24) = 9.64, P = 0.009, η_p² = 0.446, grip cutoff value = 0.101, shake = 0.092] (Fig. 6B and SI Appendix, Table S6). No interaction effect and main effects of time and CAPs were reported (P > 0.1).

In the second model, we investigated this question for pretask and posttask resting state data, separately. We found that the spatial similarity between pretask rest frames and task-derived CAPs was significantly greater for CAPs generated from the grip condition than from the shake condition [t-score threshold = 10: F(1, 24) = 9.18, P = 0.010, η_p² = 0.433, grip cutoff value = 0.097, shake = 0.090; t-score threshold = 12: F(1, 24) = 9.63, P = 0.009, η_p² = 0.445, grip cutoff value = 0.097, shake = 0.090], but these effects were not significant during the posttask resting state [t-score threshold = 10: F(1, 24) = 3.79, P = 0.075, grip cutoff value = 0.103, shake = 0.097; t-score threshold = 12: F(1, 24) = 4.60, P = 0.053, grip cutoff value = 0.104, shake = 0.095] (Fig. 6C). More details in SI Appendix, Table S6. No interaction effect of movement × CAP and no main effect of CAP were reported (P > 0.1).

These findings show that the brain-wide cortical pattern of activation (CAP) generated by an ecological movement (grip) shows greater spatial similarity to the patterns observed in the pretask resting state than a novel nonecological pattern. Moreover, in the posttask resting frames, the cutoff values for grip and shake were both marginally increased, and the greater spatial similarity of task CAPs to posttask resting frames for the grip than shake condition was not significant in the posttask resting state.

Discussion

The function of spontaneous activity is unknown as is its role in supporting behavior. Here, we examined whether time-varying spontaneous activity brain states code for behaviorally relevant information in the motor domain. We tackled this question by first computing the most common activation states, synchronous with the primary human sensory-motor cortex across the brain (CAPs), during the execution of hand movements. Next, we tested their similarity with patterns similarly computed, but independently, at rest. To test whether spontaneous activity patterns represent priors about behaviorally relevant information or code off-line learned patterns, we compared rest CAPs before and after movement task execution, and for familiar vs. novel movements. The prior model predicts rest-task similarity in activation patterns even before movement execution; and the off-line learning model predicts a significant change of resting patterns before/after task execution, especially for novel movements.

We identified three brain states synchronous with the motor cortex during hand movements. The first CAP—CAP A (α)—corresponds to the mean pattern of activation when subtracting the baseline, and it was the most frequent. Anatomically, it includes precentral (motor), postcentral (sensory), and dorsal-ventral fronto-parietal regions of the dorsal and ventral attention network with deactivation of regions of the default mode network. It therefore includes the classical pattern of activation-deactivation reported in many studies between sensory-motor-attention regions and the default network (37–39). The second CAP—CAP B (β)—corresponds also to activated sensory-motor and visual regions with a corresponding deactivation of default regions. Therefore, CAP A (α) and B (β) correspond to the previously reported gradient of functional organization with sensory-motor-visual cortices on one end, and default regions on the other (40). The third CAP—CAP C (γ)—is the most surprising since it shows joint activation of the sensory-motor cortex with default regions and deactivation of dorsal fronto-parietal regions of the DAN (Fig. 3).

Remarkably, these same brain states were present at rest both before and after task execution. In fact, it was possible to match one-by-one task and rest patterns, even though pre- and posttask rest patterns were more similar to each other than either one with task patterns. The strong correlation between pre- and posttask resting states indicates the stability of these patterns across time.

The similarity of movement-rest CAPs cannot be explained by the presence of ongoing fluctuations both in rest and movement that could cause similar CAPs. In fact, Fig. 2 clearly shows that ongoing “spontaneous” motor-related patterns were present throughout the duration of rest scans while they were aligned with movement blocks during the task. This result indicates that the performance of hand movements induced a phase reset of motor-related states that were randomly present during rest but synchronized to the movement during the task.

Additionally, the presence of movement-related patterns in pretask rest scans cannot be due to practice effects since subjects were essentially naïve to the details of the task except for being shown the correct hand sequences before the experiment. Furthermore, we confirmed the same movement-related patterns at rest in a separate group of subjects that had not performed the task (SI Appendix, Fig. S4). Our interpretation is that the movement patterns were intrinsic and ongoing in the resting state. This interpretation also explains the results in the independent group of subjects.

How does then performing a task modify ongoing brain states? The execution of blocks of repeated hand movements modulated activity in the primary motor cortex where the task-relevant region (contralateral hand) increased its response while task-irrelevant regions (ipsilateral hand region, mouth, foot, and intereffector) bilaterally decreased their response. In addition, there was a clear dynamic change in brain states when going from rest to task with each brain state lasting longer and switching more frequently. These findings suggest that planning and executing hand movements strongly relies on the underlying ongoing brain’s functional network architecture whereby activity is modulated at the effector level, while ongoing brain states are dynamically rearranged to support task performance. These findings complement previous observations on task-rest network similarities that were mainly obtained by averaging activity over time (41–44).

Two results support the view that the spontaneous activity patterns we have demonstrated represent ongoing priors (18, 19). First, we observed motor-related patterns in pretask resting state scans (see above). More importantly, we confirmed at the whole brain level a result that we have already observed in the motor cortex (20) and within the attention networks (21) on the same dataset. Namely, the rest-task similarity of dynamic patterns is higher for a common ecological hand movement (grip) than a novel nonecological movement (shake) (Fig. 6). This result is in line with the idea that ongoing spontaneous brain states represent information states that code for common environmental, body, and cognitive signals and that they represent priors of the brain’s generative model (19).

The similarity between rest and task patterns observed here takes a particular form that we have previously observed in our studies of task condition-specific correspondences between averaged multivoxel patterns from visual and motor tasks and multivoxel patterns from single rest frames (20–22) and that has been also observed in cat and monkey studies (45, 46). Spatial patterns of activity within an area and across areas resembling behaviorally relevant task-evoked or stimulus-evoked patterns occur transiently at rest. They appear as moments in time in which the spatial patterns of activity within an area [e.g., primary motor cortex (20)], within regions of the visual cortex responding best to specific classes of stimuli [e.g., face-preferring regions (22)], within a network (e.g., the DAN or VAN) (21), or over the whole brain (this paper) closely match averaged task-evoked patterns, in the present case, patterns evoked by hand movement. Interestingly, rest-task similarity can be positive or negative, manifesting as a symmetric distribution over rest frames of positive and negative task-rest correlation values, which is centered at zero but is spread significantly wider, i.e., with longer positive and negative tails, for specific task conditions in specific regions [e.g., grip vs. shake in the motor cortex (20) or faces vs place in face-preferring regions (22)]. These results were also observed in the CAPs analysis of the present study. Specifically, we found a symmetric distribution of positive and negative similarities between the task-derived CAPs and single resting frames (Fig. 6) and also found that the spread of the distribution of similarities was condition-specific (grip vs. shake, Fig. 6). Therefore, the current study likely tapped into similar processes as our previous studies but with respect to multiple underlying whole-brain patterns that dynamically varied during task performance, rather than with respect to a single time-averaged pattern in a single region or set of regions. Most importantly, unlike our previous work, the current work also showed the similarity/correspondence of multiple task states with multiple stable and replicable resting states that were independently defined, i.e., the resting-generated CAPs.

We acknowledge that our studies have sampled a small number of task conditions. For example, even though the tested hand movements are very common, it is possible that a higher overall correlation may be obtained when testing other movements. Moreover, we acknowledge that priors for real hand movements may encompass information such as visual or somatosensory information related to object visual processing, and/or visual or somatosensory feedback from the hand, which were not present in our simplified hand movements. Further research is needed to determine whether different types of movements would yield similar interpretations. With this caveat, however, the symmetric positive-negative distributions for rest-task pattern similarity may reflect how information can be coded in the brain. Computational models indicate that both increases and decreases of activity can be relevant for coding information. Some information states may be “explicit” and yield a positive match, while others may represent the same information, but be “latent” or “dormant” and yield a negative match (47). Additionally, positive and negative patterns may be linked by a local excitation–inhibition balance perhaps mediated by horizontal connections within cortical regions. Finally, the limited sample size was balanced by the number of fMRI volumes collected during the experiment, aimed at balancing the sample size and scan time (48).

Additionally, we observed slight changes in the spatial and temporal distribution of resting CAPs before and after the motor task. The similarity between rest and task patterns increased posttask for hand movements, suggesting a potential learning effect. Further studies are required to confirm this finding.

In conclusion, this study reveals that spontaneous brain activity dynamically encodes information states related to hand movements across the whole brain, especially for ecological common movements. It provides two mechanistic insights into how spontaneous activity relates to task-evoked activity. One is that the dynamic organization of activity at rest is a strong spatiotemporal prior for the patterns of activation during task (in this case hand movement), the other is that motor replays are not limited to motor regions but occur widely across the brain. This dynamic replay mechanism may underlie prediction of upcoming movements and long-term memory. Similar replay mechanisms are likely to occur for sensory, cognitive, and memory states representing a putative explanation for the function of spontaneous brain activity.

Materials and Methods

Subjects, Procedures, and Acquisition.

In this study, we used the same dataset as in our previous publications (20, 21). The experimental protocol was approved by the Institutional Review Board (IRB) of Washington University in St. Louis School of Medicine. All methods used in the current study were performed in accordance with the relevant guidelines and regulations of the ethical review board. All participants provided written informed consent before the study and were compensated for their participation. Two participants were excluded because of different scan parameters. Thirteen participants (6 females, all right-handed) were included in this study. During task scans, participants performed a block-design hand-movement task, including four different hand movements: i) grip (i.e., starting from a mid-opening position, closing hand with fingers 2 to 5 close in a grasping movement in opposition to the thumb); ii) extend (i.e., starting from a mid-opening position, opening hand with fingers 2 to 5 extend in one direction while the thumb extends in the opposite direction); iii) pinch (i.e., starting from a mid-opening position, closing hand only with the thumb and index finger); and iv) shake (i.e., starting from a mid-opening position, the wrist flexes and moves back-and-forth in adduction and abduction while keeping the fingers still). Grip and extend correspond to the most common movements (30), while shake is a less ecological movement. Before and after task scans, the same participants were also scanned in three five-minute resting state scans. BOLD time series were acquired using a 3-T Siemens MR scanner (TR = 1 s; voxel size = 3 mm³ isotropic). Structural and functional preprocessing was performed using the FreeSurfer and FS-FAST processing stream (surfer.nmr.mgh.harvard.edu), fully described in refs. 20, 21 (details also in SI Appendix).

fMRI Coactivation Pattern Analysis.

We implemented the CAP analysis using the TbCAPs toolbox (25) based on Matlab version 2019b (https://it.mathworks.com/). A specific parcel in the motor cortex (Fig. 1), obtained from Yeo’s 7-network atlas Schaefer 100 parcel version (49), was selected as a ROI to compute CAPs. This parcel overlapped with the motor activation region reported in ref. 20 and showed a significant movement representation difference in ref. 21. CAP analyses were performed for task data, pretask resting-state data, and posttask resting-state data, separately.

First, we used task fMRI data to compute CAPs. Specifically, all the task fMRI data were z-scored (according to the mean and SD of each frame) and a gray matter mask was applied to retain information only from the cortex. Task fMRI data were then thresholded with a T score (computed from the z-score maps) of the signal within the ROI higher than 1 and frames were also scrubbed if the framewise displacement (FD) value was higher than 0.5 mm. Thus, frames above the threshold were classified as “active” and retained, used for further CAPs analysis. Since there is no consensus about the number of CAPs, K-means clustering was run from K = 3 to K = 8 to find an optimal K through the consensus clustering procedure implemented in the CAPs Toolbox. Second, the retained frames were clustered into the optimal K number of CAPs. Finally, several temporal metrics linked with CAPs were computed: 1) duration; 2) in-degree; and 3) out-degree.

The same analyses were performed for pretask and posttask resting-state data, independently, following the same procedure. For posttask data, the optimal K identified in the prerest task data was applied to compute the CAPs. Finally, the group-averaged CAPs were computed at the vowel-wise level.

To validate the above analysis, we also used CAPs obtained from task data as references to identify the corresponding CAPs from pretask resting-state data and posttask resting-state data.

Spatial Similarity and Difference.

We tested whether group-averaged pretask resting-state CAPs and task-state CAPs showed a similar pattern using Pearson’s correlation, suggestive of a similar connectivity dynamic. Matched CAPs from pretask and posttask resting states were compared at the voxel level through a paired-sample t test. GRF corrected for P < 0.001 at the voxel level, and P < 0.025 at the cluster level, two-tailed.

Movement-Related CAP Comparison Between Pretask Resting-State and Posttask Resting-State.

Further, we tested whether CAPs play a representational role, coding for movement-related information. A multivoxel analysis was performed according to the same procedure applied in our previous publications (20, 21). Grip was selected as a more ecological movement, while Shake was selected as a less ecological (control) movement.

For each participant and each movement, an averaged brain-wide CAPs (built from all the task runs) was computed, separately. Specifically, if a frame during the grip movement trial was assigned as CAP1, this frame was identified as Grip-CAP1. These frames assigned to the specific movement-related CAP were averaged to create a movement-related CAP vector for each participant. In total, 2 × K vectors for each participant (K is the number of CAPs). The length of these vectors represented the number of voxels across the cerebral cortex. A vector of the same length was computed for each resting-state frame. Movement-related CAPs were correlated with each frame of the resting-state data using Pearson’s correlation. To compare the strength of these correlations, a cumulative distribution function (CDF) was computed for r-squared values, and the 90th percentile cutoff value was identified. These cutoff values were finally inserted as the dependent variable in an ANOVA.

This analysis was repeated selecting different movement-preferred ROIs, which are created from the t-map of the contrast for each movement-vs-baseline (i.e., Grip-vs-baseline and Shake-vs-baseline) from (21) with a different range of T score threshold (from t = 6 to t = 12 with a step of 2) (Fig. 6).

Statistical Analysis.

First, we tested whether different movements had different occurrences of CAPs during motor performance. Specifically, for each type of CAP, ANOVA was performed, respectively, with movements as an independent variable, participants as a random variable, and the CAP occurrences during each movement as the dependent variable.

Second, to investigate whether there was temporal reorganization among CAPs, ANOVAs were used for different temporal metrics, where time and CAPs were included as independent variables, participants as a random variable, and each temporal metric as the dependent variable.

Finally, to investigate whether hand movement-related CAPs were represented in spontaneous activity and were changed after the performance of the motor task, two different models were performed. In the first model, ANOVA was performed for different ROIs (canonical motor parcel and activation ROI with different T thresholds) respectively, with CAPs, movements, and time as independent variables, participants as a random variable, and the 90th cutoff values as the dependent variable. In the second model, for different ROIs, we tested the CAP-movement interaction independently for each timepoint (pre- and posttask), where CAPs and movements were included as independent variables, participants as a random variable, and the 90th cutoff values as the dependent variable.

Sensitivity Analyses.

First, different thresholds were assessed to establish the robustness of the results. Specifically, the analyses were repeated considering an incremental T-score cutoff from 0.8 to 1.4 with 0.2 steps for both resting state and task data. Second, we compared the spatial similarity between CAPs obtained with different K values. CAPs obtained with different K values were compared with CAPs obtained from the main analysis that involved an optimal K value based on consensus clustering. Third, we reran the main analysis in a small subcohort (n = 5) of individuals randomly selected from the main cohort to confirm the CAPs patterns. Fourth, to assess the impact of parcels we reran the main analysis using the hand region from (33) as the ROI. Finally, an additional dataset of structural MRI and fMRI data from 32 healthy adult participants (21 female; mean 27.6) (22, 34) was retrospectively included in this study to validate the resting-state CAPs.

Supplementary Material

Appendix 01 (PDF)

Data, Materials, and Software Availability

The dataset analyzed in this study is stored at cnda.wustl.edu. Restrictions apply to the availability of the dataset which was analyzed in the current study under license from the Central Neuroimaging Data Archive (CNDA) Center in Washington University in St. Louis. The dataset is not publicly available; however, it could be made available upon reasonable usage request from the corresponding author along with the permission of the CNDA Center (cnda.wustl.edu). All other data are included in the article and/or SI Appendix.