Enhanced diversity on connector hubs following sleep deprivation: Evidence from diffusion and functional magnetic resonance imaging

Sleep deprivation has been demonstrated to exert widespread and intricate impacts on the brain network. The human brain network is a modular network composed of interconnected nodes. This network consists of pro-vincial hubs and connector hubs, with provincial hubs having diverse connectivities within their own modules, while connector hubs distribute their connectivities across different modules. The latter is crucial for integrating information from various modules and ensuring the normal functioning of the modular brain. However, there has been a lack of systematic investigation into the impact of sleep deprivation on brain connector hubs. In this study, we utilized functional connectivity from resting-state functional magnetic resonance imaging, as well as structural connectivity from diffusion-weighted imaging, to systematically explore the variation of connector hub properties in the cerebral cortex after one night of sleep deprivation. The normalized participation coefficients (PCnorm) were utilized to identify connector hubs. In both the functional and structural networks, connector hubs exhibited a significant increase in average PCnorm, indicating the diversity enhancement of the connector hub following sleep deprivation. This enhancement is associated with increased network cost, reduced modularity, and decreased small-worldness, but enhanced global efficiency. This may potentially signify a compensatory mechanism within the brain following sleep deprivation. The significantly affected connector hubs were primarily observed in both the Control Network and Salience Network. We believe that the observed results reflect the increasing demand on the brain to invest more effort at preventing performance deterioration after sleep loss, in exchange for increased communication efficiency, especially involving systems responsible for neural resource allocation and cognitive control. These results have been replicated in an independent dataset. In conclusion, this study has enhanced our understanding of the compensatory mechanism in the brain response to sleep deprivation. This compensation is characterized by an enhancement in the connector hubs responsible for inter-modular communication, especially those related to neural resource and cognitive control. As a result, this compensation comes with a higher network cost but leads to an improvement in global communication efficiency, akin to a more random-like network manner.


Introduction
Sleep is a fundamental requirement for maintaining the normal functioning of human physiology and psychology.It also plays a crucial role in promoting cognitive performance such as learning and memory (Maquet, 2001;Paller et al., 2021).In recent years, the prevalence of insufficient sleep has escalated, emerging as a global concern (Zhao et al., 2021).This escalating trend poses substantial health challenges and carries significant economic ramifications that demand serious attention (Hafner et al., 2017;Hillman and Lack, 2013).
Sleep deprivation has widespread but complex effects on cognitive functions (Krause et al., 2017).For basic cognitive processes such as vigilance and attention, sleep deprivation often leads to slower reaction times and increased variability (Garcia et al., 2021;Hudson et al., 2020).However, there is a lack of converging evidence regarding the impact of sleep deprivation on many higher-order cognitive functions, such as abstract reasoning and task switching, which seem to be less vulnerable to insufficient sleep (Khazaie et al., 2010;Killgore, 2010;Pesoli et al., 2022).While consensus remains elusive, the effects of sleep deprivation on cognitive functions appear to resist simple summation as "global impairment."To gain a more profound comprehension of this intricate phenomenon, further examination of the neural system responses to deprivation is essential.
For the brain's neural system supporting cognitive activities, hubs play a pivotal role in facilitating communication and neural integration (Palva, 2018;van den Heuvel and Sporns, 2013).Hubs, in the brain, are characterized by their extensive connectivity (Oldham and Fornito, 2019).Over the past decade, the primary type of hub (Marcel and Martijn, 2013), known as the connector hub (see Fig. 1 for an illustration of the toy concept of a connector hub), has garnered attention from researchers due to its essential role in modular biological networks and cognition (Bertolero et al., 2018;Gordon et al., 2018).Connector hubs, distinguished by their considerable connectivity diversity, are integral to a wide spectrum of cognitive functions and serve as essential role in cross-module communication in tasks involving multiple cognitive components (Bertolero et al., 2015).The perturbation of hub nodes as a consequence of sleep deprivation has been observed (Pan et al., 2023).However, there is a lack of literature on the impact of sleep deprivation on connector hubs.Sleep deprivation's disruption on modular networks has been a subject of prior investigation (Ben Simon et al., 2017;Wang et al., 2015).Further investigation into how connector hubs, which serve as the backbone and modulators of modular networks (Bertolero et al., 2018), are influenced in this context is highly warranted.Investigating this topic will serve as a bridge to comprehend the impact of sleep deprivation on brain neural activity and cognitive processes.
The interaction and communication between neural units are increasingly recognized as the foundation of brain physiology and cognition (Bernstein-Eliav and Tavor, 2022).This recognition has encouraged researchers to explore cognitive neuroscience more extensively from the perspective of brain connectivity rather than focusing solely on the activity of individual brain regions.Functional connectivity (FC), derived from the coupling of blood oxygenation level-dependent (BOLD) signals, and structural connectivity (SC) obtained through white matter fiber tracking are commonly employed tools in neuroimaging research.A prevalent approach in FC research is the resting-state paradigm, which is utilized to explore the brain's spontaneous thoughts and has been demonstrated to have a close association with cognitive task processing (Cole et al., 2016).While SC serves as the physical foundation for FC (Straathof et al., 2019), the two still manifest distinct patterns associated with cognition (Zimmermann et al., 2018).To achieve a comprehensive understanding of the effects of sleep deprivation on brain inter-regional interactions, a fusion of both modalities would be beneficial.
Therefore, this study comprehensively examined the impact of sleep deprivation on connector hubs through two modalities: resting-state FC and SC.We anticipated that FC would exhibit significant alternations in the properties of connector hubs, while SC, representing the structural foundation influenced by sleep deprivation, would also demonstrate alternations in connector hub properties in a consistent direction.The strength of FC is typically intertwined with SC strength but also exhibits variations.Regions characterized by dense SC tend to manifest dense FC as well, yet FC are also observed between regions with sparse or absent SC, possibly due to indirect driving by structural connectivity (Damoiseaux and Greicius, 2009).Therefore, we expected that both FC and SC would capture alterations in connector hub properties in a consistent direction, with the effects of FC potentially being more sensitive or capable of detecting widespread hubs affected by sleep deprivation.Additionally, we systematically explored the associations Y. Tian et al. NeuroImage 299 (2024) 120837 between the changes in connector hubs and other global network properties to discern its physiological significance, including network cost, communication efficiency, network segregation, and small-worldness.Lastly, from the perspective of large-scale brain networks, we investigated the major functional networks that are primarily affected by sleep deprivation in relation to the identified connector hubs.This study would provide a deeper understanding of the alterations in brain neural activity and network properties under conditions of sleep loss.

Participants
The main dataset (referred to as the discovery dataset in the manuscript) was derived from our sleep deprivation project, as previously reported by Tian et al. (2022).The experiment received approval from the Ethics Committee of Faculty of Psychology, Southwest University.All participants gave written informed consent.A total of 33 participants were recruited for the study, with ages ranging from 18 to 30.Among them, 19 participants were women.
The data utilized for replicating the study's findings (referred to as the replication dataset in the manuscript) was obtained from the Stockholm Sleepy Brain Project: Effects of Sleep Deprivation on Cognitive and Emotional Processing in Young and Old.This dataset is accessible on OpenNeuro (https://openneuro.org/datasets/ds000201).The project was preregistered with the identifier (NCT number: NCT02000076) on clinicaltrials.govand received approval from the Regional Ethics Review Board of Stockholm.The project provided openly accessible data from 86 participants, with ages ranging from 20 to 75, including 44 females.All participants provided written informed consent.

Protocol
Discovery dataset: The participants were randomly assigned to undergo both well-rested and sleep deprivation sessions in a counterbalanced order, with an interval of approximately one week between each session (refer to Fig. 2).During the well-rested session, participants slept in their own homes and then visited the laboratory upon awakening to undergo magnetic resonance imaging (MRI) scans.In contrast, during the sleep deprivation session, participants were required to remain awake under the supervision of experiment assistants in a laboratory with controlled lighting conditions (~100 lux).Participants were permitted to partake in relaxation activities such as listening to music, reading, and watching movies.However, they were explicitly instructed to avoid engaging in strenuous physical exercises or playing video games.Scans were conducted following a full night of sleep deprivation, about 8:00 to 10:00 a.m.For this study, we utilized data from the well-rested session and the session following 24 h of total sleep deprivation, including resting-state functional MRI (rs-fMRI), structural imaging, and diffusion-weighted imaging (DWI) data.Multiple task sessions were also collected during the MRI scans.However, due to exceeding the scope of this study, detailed description of this aspect will not be provided here.Additional details can be found in the previous study (Tian et al., 2022).The overall analysis flowchart can be referred to as Fig. 3.
Replication Dataset: Participants were randomized and subjected to well-rested and sleep deprivation sessions in a counter-balanced order with an interval of approximately 1 month.The participants slept in their own homes during both sessions.Under the sleep deprivation condition, the participants were instructed to go to bed 3 h before the time they would usually get up and then get up at their normal time (i.e., partial sleep deprivation).MRI imaging was performed in the evening following sleep deprivation or normal sleep.For this study, we utilized data from the well-rested session and the session following partial sleep deprivation, including resting-state functional MRI (rs-fMRI) and structural imaging.However, the diffusion-weighted imaging (DWI) data, as it was collected only for one session, was not used for replication purposes.More detailed descriptions of this project can be found in previous studies (Nilsonne et al., 2017;Tamm et al., 2017) and the open science framework (https://osf.io/zuf7t).

Acquisition of MRI data
Discovery datasets: The MRI images were acquired using a General Electric (GE) DISCOVERY MR750 3.0 Tesla scanner equipped with an 8channel head coil.Participants' heads were immobilized with a sponge to minimize movement artifacts.Noise-reducing earplugs and headphones were used to reduce scanner noise interference.Resting-state images were acquired using Echo Planar Imaging (EPI) with the following parameters: TR = 2000 ms, TE = 30 ms, flip angle = 90 • , acquisition matrix = 64 × 64, in-plane resolution = 3.0 × 3.0 mm 2 , FOV = 240 × 240 mm 2 , 33 slices with a thickness of 3.5 mm and 0.7 mm gap, resulting in a scan time of 410 s.A total of 205 images were acquired.During the resting-state scan, participants were instructed to keep their eyes open and focus on a white cross presented at the center of a black screen, while avoiding engaging in deliberate systematic thinking.

Fig. 2.
The protocol for the discovery dataset.To mitigate potential sequence effects, the order of the two visits to the laboratory was counter-balanced.Questionnaires included demographic information, the Amsterdam Resting-State Questionnaire et al.PVT stands for the Psychomotor Vigilance Test.The statistical details of these information were previously reported in our prior work (Tian et al., 2022).Data were collected across three sessions during the experiment.
Subsequently, high-resolution T1-weighted images were acquired using a gradient-echo (GR) sequence.The scanning parameters were as follows: TR/TE = 6.6/2.9 ms, FOV = 240 × 240 mm 2 , flip angle = 12 • , acquisition matrix = 256 × 256 mm 2 , slice thickness/gap = 1/0 mm, 192 slices, and a scanning time of 5 min.For DWI, a spin-echo EPI sequence with a high angular resolution diffusion imaging (HARDI) protocol was employed.The imaging protocol included both b = 0 s/ mm 2 and b = 1000s/mm 2 images, with the latter acquired along 64 gradient directions.The DWI acquisition included slices with a thickness/gap of 2/0 mm.The TE was 81.4 ms, flip angle was 90 • , TR was 8.724 s, and the acquisition matrix was 112 × 112.The total scanning time for the DWI sequence was approximately 10 min and 46 s.A total of 75 slices were acquired.
Replication dataset: The data was collected using a GE Discovery MR750 3T scanner.Resting-state EPI images were acquired with the following parameters: a flip angle of 75 • , TE of 30 ms, TR of 2.5 s, FOV of 288 × 288 mm 2 , slice thickness of 3 mm with no gap, and a total of 49 slices.The scanning duration for the echo-planar images was 8 min and 45 s.T1-weighted anatomical scans were acquired using a sagittal BRAVO sequence with a FOV of 240 × 240 mm 2 , a flip angle of 11 • , a slice thickness of 1 mm with no gap, and a total of 180 slices.The scanning time for the T1-weighted scans was approximately 3 min and 59 s.

Image preprocessing
First, we conducted quality screening of functional images using MRIQC (Esteban et al., 2017).Resting-state images from both the Discovery dataset and the Replication dataset were excluded based on the same motion criteria: average Frame Displacement (FD) > 0.2 mm or more than 30 % of time points with FD > 0.2 mm (Power et al., 2012).According to the quality report generated by MRIQC v21.0.0rc2, three participants were excluded from the Discovery dataset, resulting in a final sample size of 30 participants.For the Replication dataset, we utilized the pre-generated quality report based on MRIQC v0.9.6 available on the website (https://openneuro.org/datasets/ds000201/versiFig. 3. Analysis flowchart of the study.The DWI data of the replication dataset was collected for only one session and, therefore, was not utilized for replication purposes.The metrics in this figure were computed individually for each participant.Rs-fMRI: resting state functional magnetic resonance imaging; DWI: diffusion weighted imaging; FC: functional connectivity; SC:structural connectivity.ons/1.0.3).The report indicated the exclusion of 23 participants, leaving a total of 57 participants for further analysis.Additionally, in the analysis of the DWI data from the Discovery dataset of 33 participants, one participant was excluded due to the presence of image artifacts, resulting in a sample size of 32 participants available for the analysis of structural connectivity.
The resting-state data from the discovery dataset and replication dataset underwent preprocessing using the standardized pipeline of fMRIPrep v23.0.0 (Esteban et al., 2019), which is based on Nipype v1.8.5 (Gorgolewski et al., 2011).The preprocessing pipeline mainly included: generating a reference volume and its skull-stripped version, estimating head motion parameters, performing slice-time correction, and further resampling the images back to their original, native space.Subsequently, the preprocessed functional images were resampled onto the fsaverage surface space using information obtained from the preprocessing and cortical reconstruction of the anatomical images.For a comprehensive description of the resting-state image preprocessing methodology, please refer to the supplementary material.
The DWI data underwent preprocessing using the standardized pipeline of micapipe v0.1.5(Cruces et al., 2022).The preprocessing of DWI data involved several steps, including the following: alignment of all DWI scans through rigid-body registration and concatenation; denoising using MP-PCA and Gibbs ring correction; correction of eddy current-induced distortions and motion; non-uniformity bias field correction; linear registration of the b0 image to the structural image (nativepro); estimation of response functions for cerebrospinal fluid (CSF), gray matter, and white matter using spherical deconvolution; estimation of fiber orientation distributions using spherical deconvolution; computation of a non-linear transformation from the DWI structural image aligned to the b0 scan; and application of an inverse non-linear transformation to the 5-tissue-type images for anatomically-constrained tractography.For a comprehensive description of the DWI preprocessing, please refer to the work by Cruces et al. (2022) and the online document available at https://micapipe.readth edocs.io/.

Connectivity matrix construction
We employed the Schaefer 200-surface parcellation to construct cortical FC and SC matrices.This parcellation scheme offers a finegrained coverage of the cortical surface, providing detailed information about spatial network organization and demonstrating excellent compatibility with various neuroimaging modalities (Schaefer et al., 2018).
The preprocessed resting-state images were denoised and used to construct connectivity matrices using built-in house code (https://gith ub.com/WolkeTian/postproc_fmriprep).
The denoising process employed the following methods: regression of unsteady states at the beginning of the scan; ICA-aroma to specifically designed to eliminate motion artifacts (Pruim et al., 2015); aCompCor50, a variant of the aCompCor method that leverages principal components derived from white matter and CSF signals capable of explaining the top 50 % of variances to mitigate physiological noise (Muschelli et al., 2014); and high-pass filtering using a discrete cosine filter with a 128 s cut-off.Global signal regression (GSR) was not implemented due to its potential to introduce spurious negative connectivities (Murphy et al., 2009;Saad et al., 2012) and exacerbate the distance-dependence of correlations between motion and functional connectivity (Parkes et al., 2018).The connectivity matrices were calculated by computing the Pearson correlation between the average denoised BOLD signals of each cortical region, followed by Fisher's z transformation.
The structural connectivity matrix was computed using the standardized pipeline of micapipe v0.1.5(Cruces et al., 2022).Tractography was performed using the second-order integration over fiber orientation distributions (iFOD2) algorithm, incorporating 10 million streamlines, and employing 3-tissue anatomically constrained tractography (Smith et al., 2012;Tournier et al., 2010).Subsequently, spherical deconvolution informed filtering of tractograms (SIFT2) was applied to reconstruct whole brain streamlines weighted by cross-sectional multipliers (Smith et al., 2015b).The reconstructed cross-section weighted streamlines were then mapped to the Schaefer 200 cortical parcellation (Smith et al., 2015a) and warped to the native space of the DWI.The connection weights between nodes were defined as the weighted streamline count.

Graph theory metrics
Proportional thresholding of the connectivity matrix is a common step to achieve equal sparsity for subsequent graph analysis.However, it has been pointed out that the differences in overall FC among samples during the application of proportional thresholding may artificially induce distortions in network organization (van den Heuvel et al., 2017).Following the recommendations of van den Heuvel et al. ( 2017), we incorporated the fully weighted matrix for subsequent graph analysis.Considering the ambiguity of the physiological significance of negative values in the FC matrix, as well as the fact that many graph algorithms do not explicitly handle negative values, we followed the common practice in the field (Bertolero et al., 2018) and excluded negative values from the FC matrix for each participant.As SC naturally inherently does not contain negative values, no adjustment was needed.
Additionally, considering the popularity of using proportional thresholding, we also employed the common thresholding standard (cost 0.15) to binarize each full connectivity matrix (Redcay et al., 2013) for validation purposes.
The graph-theoretical computations in this study were primarily performed using the Brain Connectivity Toolbox, abbreviated as BCT (Rubinov and Sporns, 2010).
Community detection and modularity were calculated using the 'community_louvain' function from the BCT, which is a highly efficient and accurate multi-iterative extension of the Louvain algorithm designed to maximize the modularity.
The normalized participation coefficient (PCnorm), as proposed by Pedersen et al. (2020), was utilized to identify connector hubs.Unlike the conventional participation coefficient calculation, PCnorm reduces the influence of intramodular connectivity and is not affected by the size or connectedness of modules (Pedersen et al., 2020).A higher PCnorm value indicates that the node exhibits a greater connectivity diversity.The formula for this method can be referenced in the supplementary materials or Pedersen et al. (2020).
The calculation of cost was based on the 'density_und' function from the BCT, which computes the fraction of present connectivities to possible connectivities.Weight information was disregarded.
The calculation of global efficiency was performed using the 'effi-ciency_wei' function from the BCT.It computes the average of the inverse shortest path length.
The small-worldness quantifies the balance between local clustering and shortest path length within the graph.It was calculated using three functions from the BCT: 'randmio_und' to generate random graphs while preserving degree distribution through 100 rewiring iterations, 'dis-tance_wei' to compute the average shortest path length of the random and real networks, and 'clustering_coef_wu' to compute the average clustering coefficient of the random and real networks.Small-worldness is calculated by dividing the ratio of the average clustering coefficient of the real network to that of the random network by the ratio of the average shortest path length of the real network to that of the random network.

Statistical analysis
Before contrasting the attributes of connector hubs, we first assessed the potential of PCnorm as a stable and reliable physiological metric.We examined the consistency of PCnorm distributions in functional modalities across different datasets.A Pearson correlation analysis was conducted on the group average PCnorm values of 200 cortical parcellations, derived from the FC matrices of the well-rested session in both the discovery and replication datasets.Additionally, we explored the degree of functional-structural coupling of PCnorm distributions.For the discovery dataset, we calculated the Pearson correlation between the group average PCnorm values of 200 parcellations obtained from both FC and SC matrices.
PCnorm was utilized as a metric to measure the inter-modular connectivity diversity of nodes and to identify connector hubs.For each dataset and modality, we identified connector hubs as the nodes with the highest connectivity diversity at the group level, specifically, the top 5 % (i.e., the top 10) of nodes based on PCnorm calculated during the wellrested session.Then, we conducted a paired t-test to contrast the average PCnorm of connector hubs between the sleep-deprived session and the well-rested session, investigating the overall effect of sleep deprivation on connector hubs.The paired t-test was also conducted on connector hubs designated at 10 % and 15 % thresholds.Given the prevalent use of proportional thresholding to define graphs, we employed the proportional threshold method to define the binary matrix for calculating PCnorm, with a proportional threshold set at a cost of 0.15.The same threshold was used to identify connector hubs (the top 5 % nodes of PCnorm) and paired t-test was performed on the average PCnorm of connector hubs to examine whether different graph definition methods yielded consistent results.
Subsequently, we investigated the alternations in global network properties associated with alternations in connector hub properties.We firstly employed paired t-tests to examine changes in global network metrics before and after sleep deprivation, including network efficiency, cost, modularity, and small-worldness.Then, we used Pearson correlation analysis to explore the relationship between these changes and the changes in the average PCnorm of connector hubs.Correlation analysis was also used to investigate whether the baseline average PCnorm of connector hubs was associated with changes during sleep deprivation.False discovery rate (FDR) correction was used for multiple comparison.
After examining the overall effects of connector hubs, we shifted our focus to specific connector hubs affected by sleep deprivation.We conducted paired t-tests to compare the PCnorm of each connector hub before and after sleep deprivation.Correction for multiple comparisons was performed using FDR.
To illustrate the statistical results more clearly, we used q-values to represent FDR-corrected p-values in the subsequent results section.Meanwhile, p will represent raw p-values.
To further validate our findings and assess their dependency on the brain template employed, we conducted additional analyses using the Destrieux atlas (Destrieux et al., 2010).The algorithms and analytical procedures employed remained consistent throughout the process.The top 5 % (i.e., the top 10) of nodes based on PCnorm calculated during the well-rested session were identified as connector hubs.

Psychomotor vigilance test
The psychomotor vigilance test (PVT), a highly sensitive experimental paradigm for assessing the effects of sleep deprivation (Dinges et al., 1997), was employed in this study.The PVT used in this study is a 10-minute standard version implemented using E-prime 2.0, with an inter-stimulus interval ranging from 2 to 10 s.We utilized two widely recognized metrics from this paradigm, namely reaction time (RT) and lapses, to investigate the behavioral associations with connector hub properties.RT was computed as the median reaction time from valid trials falling between 100 ms and 500 ms.lapses were determined by counting the number of trials ranging from 500 ms to 3000 ms (trials exceeding 3000 ms were considered invalid).
The effects of impaired vigilance following sleep deprivation (significantly increased reaction time and elapses) have been reported in our previous work (Tian et al., 2022).In this study, we conducted exploratory analysis to examine whether there was a significant association between the diversity enhancement on connector hubs and the decline in vigilance.Specifically, Pearson correlation analyses were performed to investigate the relationship between the alterations in average PCnorm of connector hubs and the alterations in reaction time and lapses measured by the PVT paradigm.

Reliability analysis of PCnorm
To assess the potential of PCnorm as a stable and reliable functional metric in the manuscript, we conducted a Pearson correlation analysis on the group average PCnorm values of 200 cortical parcellations, derived from the FC matrices of the well-rested session in both the discovery and replication datasets.The results showed a strong positive correlation (r = 0.730, p < 1 × 10 -33 , df = 198), indicating excellent consistency in the distribution of PCnorm across the datasets.For the discovery dataset, we calculated the Pearson correlation between the group average PCnorm values of 200 parcellations obtained from both FC and SC, yielding a correlation coefficient of r = 0.141, p = 0.034, df = 198 (Fig. 4).
Furthermore, to ensure that the results of the SC contrasts were not influenced by the unreliability of the SC matrix estimation methods, we examined the correlation between the SC matrices of each participant in the discovery dataset across their two sessions.The results showed a mean correlation coefficient of r = 0.965 ± 0.018 (std), indicating a high level of estimation stability in our SC estimates.

The spatial distribution of connector hubs
Connector hubs consistently identified in the functional modalities of both the discovery and replication datasets include the insula and the medial parietal of the Salience Network (SN), as well as the temporal pole of the limbic system (Table 1).However, the parietal occipital cortex and the postcentral cortex of the Dorsal Attention Network (DAN), the lateral prefrontal cortex (lPFC) of the Control Network (CN), as well as the precentral cortex and frontal operculum of the SN were exclusively identified in the FC analysis of the discovery dataset.In the FC analysis of the replication dataset, the findings that diverge from those of the discovery dataset are the precuneus of CN, temporal parietal cortex, as well as the lPFC and medial posterior prefrontal cortex of SN.Interestingly, except for the lPFC, all these regions correspond to the connector hubs identified in the SC analysis of the discovery dataset.The connector hubs identified by SC exhibit disparities with FC, particularly regarding the presence of somatosensory/motor areas, as well as the medial frontal of SN, ventral lateral prefrontal cortex of CN, and ventral prefrontal cortex of Default Mode Network (DMN).Refer to Table 1 for details.

Contrast of connector hub
For the FC of the discovery dataset, the connectivity diversity (average PCnorm) of connector hubs showed a significant enhancement after deprivation (t = 7.416, p < 1 × 10 -7 , CI = [0.039,0.069], df = 29).Similarly, the SC of the discovery dataset showed a significant increase in the average PCnorm of the connector hubs after deprivation (t = 2.763, p = 0.010, CI = [0.002,0.011], df = 31).The FC of the replication dataset also reproduced the similar results (t = 8.406, p < 1 × 10 -10 , CI = [0.027,0.044], df = 56).All tests were two-tailed, and the 95 % confidence interval (CI) is provided.Please refer to Fig. 5 for a graphical representation.For further verification, given the prevalent use of proportional thresholding to define graphs in this field, we employed a population proportional threshold (cost 0.15) to define the binary matrix and obtained consistent results (see Table S1).
Additionally, given that classifying connector hubs relies on this arbitrary threshold (5 %), we also examined the results when the threshold was set to the top 10 % and top 15 % of nodes.Consistently, the results replicated a significant increase in the average PCnorm of the connector hubs (see Table S2).

Association with other graph metrics
In examining the associations between the diversity enhancement of connector hubs and alterations in global network properties during deprivation, we initially conducted Paired-T tests on these network metrics, encompassing global efficiency, network cost, modularity, and small-worldness.Following FDR correction, significant alterations were observed in the discovery dataset's FC metrics following deprivation.These changes include an increase in global efficiency (t = 2.865, q = 0.023, CI = [0.004,0.024], df = 29), a rise in network cost (t = 4.213, q = 0.001, CI = [0.026,0.076]), a decrease in modularity (t = − 4.406, q = 0.001, CI = [− 0.054, − 0.020]), and a reduction in small-worldness (t = − 3.946, q = 0.002, CI = [− 0.080 − 0.0254]).However, for the FC in the replication dataset and the SC in the discovery dataset, no significant results were observed after correction.The results are presented in Figure S1.
Subsequently, we examined the associations between the diversity enhancement of connector hubs and alterations in these global network metrics during deprivation.We observed a significant correlation between the enhanced diversity of connector hubs during sleep deprivation and the increased global efficiency (r = 0.454, q = 0.015), elevated network cost (r = 0.620, q < 0.001), reduced modularity (r = 0.718, q < 1 × 10 -4 ), and reduced small-worldness (r = 0.489, q = 0.008).The qvalues represents the FDR-corrected p-values.These findings were consistent in both the FC of the discovery dataset and the replication dataset (Fig. 6).In contrast, the SC analysis did not show an enhancement in connector hub diversity correlated with increased global efficiency (r = − 0.074, q = 0.739), network cost (r = − 0.142, q = 0.471), and reduced small-worldness (r = 0.019, q = 0.918) after sleep deprivation.However, we did find a significant correlation with decreased modularity (r = 0.694, q < 0.001).Additionally, we investigated the baseline diversity of connector hubs during the well-rested session and observed that individuals with lower baseline diversity exhibited a more pronounced enhancement after sleep deprivation.This correlation was supported by the SC analysis of the discovery dataset (r = − 0.493, q = 0.006), the FC of discovery dataset (r = − 0.743, q < 1 × 10 -4 ) and the FC of the replication dataset (r = − 0.527, q < 1 × 10 -4 ).Please refer to Fig. 6 for a visual representation of these findings.

The network in which the enhanced connector hubs reside
Subsequently, we conducted a contrast analysis of connectivity diversity in each connector hub (see Table 1) and examined the networks to which these significant nodes belong.The network assignment in this section is referenced from Schaefer et al. (2018).The results showed a widespread enhancement of connector hubs in the SN in the FC analysis across the dataset, but no such observation was made in the SC analysis.On the other hand, consistent enhancement of connector hubs in the CN was observed in both the SC of discovery dataset and FC analyses of both datasets.All results were subjected to FDR correction.Please refer to Fig. 7.
Then we verified the correlation between the enhancement of connector hubs and the alternations of global network metrics detected in FC.The results were consistent with those found using the Schaefer template.Specifically, the enhancement of connector hubs showed significant positive correlations with the increase in global efficiency (r = 0.644, q < 0.001), the increase in cost (r = 0.693, q < 1 × 10 -4 ), the reduction in small-worldness (r = 0.576, q < 0.001), the reduction in modularity (r = 0.709, q < 1 × 10 -4 ), and a significant negative correlation with the baseline average PCnorm intensity (r = − 0.723, q < 1 × 10 -4 ).The q-values represented the FDR corrected p-values (Figure S3).

The exploratory analysis of behavioral associations
In previous research, the effects of impaired vigilance following sleep deprivation (significantly increased RT and lapses) have been reported (Tian et al., 2022).In this study, we examined whether there was a statistically significant association between the enhancement of connectivity diversity on connector hubs and the decline in vigilance.Our findings did not uncover any statistically significant associations, whether considering RT or lapses, after applying multiple comparison corrections.

Discussion
Our study unveiled a potential compensatory mechanism in the brain in response to sleep deprivation.This compensation is characterized by an enhancement in the connector hubs responsible for inter-modular communication, accompanied by an improvement in global efficiency, despite coming at the cost of increased network wiring.Detailed, our study determined the enhancement in the diversity of connector hubs after sleep deprivation, observed in both functional and structural modalities.Further investigation demonstrated a significant correlation between the increased network efficiency and cost, as well as the reduced small-worldness and modularity of the brain networks after sleep deprivation, with the enhanced diversity of connector hubs.Lastly, from the perspective of large-scale brain networks, we investigated which networks were primarily involved in the enhancement of diversity.We found widespread and consistent enhancement of connector hubs in the SN in the functional modality, while both functional and structural modalities exhibited significant impact on connector hubs within the CN.

Function-structure coupling of PCnorm
The correlation between the PCnorm distributions obtained from FC and SC suggests the presence of possible but weak anatomical constraints (r = 0.141, p = 0.034, df = 198).It has been observed that functional/structural coupling is weaker in association cortices, while single-modal network functional/coupling relationships, such as those in visual cortices, tend to be stronger (Preti and Van De Ville, 2019).Considering that PCnorm primarily measures cross-modal communication and that nodes with higher PCnorm are mainly located in association cortices, this outcome may be predominantly driven by the weaker functional/structural coupling in association cortices.

The spatial distribution of connector hubs
In the functional modalities, both the discovery and replication datasets supported the involvement of the insula, the medial parietal, as well as the temporal pole of the limbic system.These brain regions are closely associated with emotion and affective processing (Gasquoine, 2014;Heinzel et al., 2005;Olson et al., 2007).The consistent findings in these regions may suggest that, despite personalized configurations of functional networks and connector hubs (Pines et al., 2022;Wang et al., 2018), there is a greater commonality in the role of central brain network hubs in emotion and affect processing (Jastorff et al., 2015).As mentioned above, while brain networks and connector hubs exhibit personalized configurations across individuals-this may be a major reason for the inconsistency in the specific locations of connector hubs found in the FC of the discovery and replication datasets-these nodes were commonly found within the association cortices, including Salience Network (SN), Control Network (CN), Dorsal Attention Network (DAN) and the temporal parietal.The locations primarily distributed in the association cortices, which are mainly involved in integrating sensory information or executing various functions, also align with the role of connector hubs in facilitating cross-modal communication.On the other hand, inconsistencies in experimental manipulations, scanning equipment, and experimental locations across datasets may also contribute to the observed discrepancies in the distribution of connector hubs.In addition to the consistent findings of extensive involvement of the association cortices with FC, the inconsistencies observed in SC primarily involve the participation of the Sensorimotor Network (SMN).In early life, FC exhibits stronger coupling with SC, and this coupling tends to weaken with development, particularly with the emergence of rich functional connectivities within the association cortices.These connectivities are not strictly dependent on SC foundations (Paus, 2007;Preti and Van De Ville, 2019;Supekar et al., 2010).This may explain the observation in our study where connector hubs of the SMN were identified in SC but not in FC.

Enhanced diversity comes with increased cost
The increase in cross-network interactions has been robustly observed in sleep deprivation studies (Chee and Zhou, 2019;Krause et al., 2017), including enhanced coupling between the SN and the CN, as well as between the SN and the Default Mode Network (DMN) (Lei et al., 2015).This study further supports this finding from the perspective of connector hubs, which are key regions facilitating information exchange between different functional modules in the brain.Our study also uncovered the effects of sleep deprivation on network cost and global efficiency, and replicated the disruption of modularity and small-worldness due to sleep deprivation (Ben Simon et al., 2017;Qi et al., 2021).While no significant findings were observed in the SC and replication dataset, this could be attributed to the experimental design involving partial sleep deprivation in the replication dataset, potentially allowing participants opportunities for sleep that could aid in the   restoration of brain network properties.In contrast, SC may not be as sensitive to detecting changes in global network attributes, possibly due to its relatively stable structural properties.Nevertheless, the significant association between the enhancement of connector hub diversity and the magnitude of changes in these global network properties remained consistently observed across both functional modalities in the two datasets.This association was partially replicated in the structural modality as well.
The enhancement of connector hub diversity suggests that functional networks in the brain tend to make more diverse and abundant connectivities to communicate with each other.In contrast, for cognitive functions, the brain requires continuous information exchange and interactions between multiple individual networks (Schurz et al., 2020;Smith, 2016).The enhancement of connector hub diversity during sleep deprivation, along with the corresponding increase in global efficiency and cost, may indicate a compensatory mechanism within the brain, wherein additional effort is invested to mitigate performance deterioration during extended wakefulness.This could involve allocating more wiring resources for communication between networks, thereby improving communication efficiency and reducing the shortest path length.Evidence from transcriptomic studies also supports the upregulation of genes associated with energy metabolism and synaptic strengthening during short-term sleep deprivation (Cirelli, 2006).This also implies a decrease in network segregation and achieving efficient communication in a manner closer to random networks: shorter path length, higher cost, and lower modularity.This is consistent with the findings of Ben Simon and his colleagues, who observed the loss of functional segregation and a trend towards a more random-like network within the brain under conditions of sleep deprivation (Ben Simon et al., 2017).
Additionally, the lower diversity of connector hubs at well-rested was significantly associated with a greater degree of diversity enhancement after sleep deprivation (see Fig. 6).This suggests that individuals with lower diversity on connector hubs, i.e., those with lower wiring costs, exhibit greater changes in response to sleep deprivation.In contrast, individuals with higher diversity on connector hubs at wellrested show less impact from sleep deprivation.This could be attributed to their inherently higher wiring costs, which may not require significant reconfiguration and compensation after prolonged wakefulness.

Discoveries from structural connectivity
The consistent finding of enhanced diversity on connector hubs in both modalities indicates that the increased cross-network communication after sleep deprivation is not only evident in the abundant indirect connectivities present in FC but also reflected in the direct connectivities supported by fiber bundles.While white matter fiber bundles are considered stable structural features with longer timescales of change, previous research has supported alterations in structural characteristics, such as gray matter morphometry and white matter microstructure, following sleep deprivation (Dai et al., 2018;Long et al., 2020;Voldsbekk et al., 2021).Our study further supports the existence of daily variance in white matter fiber connectivities.On the other hand, despite the similarities with the findings in FC, the findings on SC are still less sensitive.This may be attributed to the relatively limited changes in the properties of direct connectivities represented by SC and the slower time scale at which these changes occur compared to the functional coupling reflected by the bold signals.
The enhancement of overall diversity of identified connector hubs in structural connectivity appears to be primarily driven by the precuneus (refer to Table 1).Serving as the structural core of the brain network (Hagmann et al., 2008), the precuneus exhibits extensive structural connectivities with both cortical and subcortical regions (Cavanna and Trimble, 2006).Known for its extensive heterogeneity in both structure and function, the precuneus comprises multiple subregions and is thought to perform high-order cognitive functions by integrating multimodal information (Tanglay et al., 2022;Yamaguchi and Jitsuishi, 2023).The enhanced diversity of cortical connectivities in the structural core following sleep deprivation drove the primary findings in structural connectivity and likely facilitated the augmentation of functional connector hubs.This is considering the structural connectivities of the precuneus with multiple resting-state networks, such as the DMN, DAN, and SN (Yamaguchi and Jitsuishi, 2023).The compensatory functional activity of the brain following sleep deprivation may be achieved through the structural plasticity of the brain, such as increased cortical connectivities within the structural core.The effect of total sleep deprivation on the precuneus was supported in a previous study investigating the cortical thickness (Elvsåshagen et al., 2017).

Diversity enhancement of connector hub at control network and salience network
The SN is closely associated with the detection and integration of stimuli and the allocation of neural resources (Uddin, 2015), such as emotional regulation, multisensory perception, the switch between the DMN and the CN (Goulden et al., 2014;Pinto et al., 2023).Previous studies have observed an enhancement in the coupling between the SN and the CN, as well as between the SN and the DMN (Lei et al., 2015) along with increased connectivity to the reward system under conditions of sleep loss (Fang et al., 2015).In conjunction with our findings, these observations suggest that during sleep deprivation, the brain would allocate more resources to facilitate communication between the SN and other networks, and becomes more actively involved in the allocation of neural resources to cope with the high homeostatic pressure that may lead to dysfunction under conditions of sleep deprivation.This also supports the role of the SN as a functional network for perceiving and responding to homeostatic demands (Seeley, 2019).
The CN or frontoparietal network plays a crucial role in executive function and cognitive task control, such as attention, reasoning, and complex problem-solving (Menon, 2011;Power and Petersen, 2013).The execution of cognitive tasks requires the collaborative involvement of the CN and other unimodal networks, such as sensory processing regions (Wu et al., 2020).Therefore, the increased cross-network communication of the CN after deprivation may indicate the brain's effort to ensure normal cognitive functioning.In our study, a novel finding is that SC also supports the changes in connector hub properties within the CN.This finding supports the plasticity of brain white matter fibers, enabling the brain networks to self-adjust in response to increasing homeostatic stress.The notion that such plasticity can occur at the scale of a day is supported by the evidence from the effects of short-term mental training on white-matter tracts (Tang et al., 2010).A study conducted by Voldsbekk and her colleagues also supports the impact of sleep deprivation on white matter microstructure (Voldsbekk et al., 2021).

The behavioral significance of enhanced diversity on connector hubs
In this study, we did not detect a significant association between vigilance and the diversity of connector hubs.This lack of association may, in part, be attributed to the possibility that enhanced diversity in connector hubs could reflect a compensatory mechanism within the brain during sleep deprivation.Furthermore, connector hubs play a pivotal role in facilitating cross-module communication, especially in tasks that necessitate the engagement of multiple cognitive components (Bertolero et al., 2015).On the other hand, this may be due to variations in vulnerability to sleep deprivation across different individuals and cognitive variables (Tkachenko and Dinges, 2018).While the direction of impact on both variables appears to be correlated, the magnitude of their respective influences may not necessarily exhibit statistical correlation.Similar phenomena have been observed in the effects of sleep deprivation on sleepiness and vigilance.Despite both being impaired by sleep deprivation, the severity of impairment lacks statistical correlation (Chandler et al., 2013;Tian et al., 2022).It is conceivable that a more sensitive task and a dedicated experimental design may be required to explore the behavioral significance of connector hubs more comprehensively.

Limitations
The study has several limitations.Firstly, the identification of connector hubs was based on the well-rested connectivity matrix, without considering the possible reconfiguration of connector hubs following sleep deprivation.Secondly, given the limited sample size, validation in larger independent datasets is needed to further corroborate the findings related to SC. Lastly, there is an inconsistency in the experimental design between the replication dataset and the discovery dataset.While both datasets involve one night of sleep deprivation, the replication dataset includes a 3 h opportunity for sleep.

Conclusion
In conclusion, our study makes three main contributions.Firstly, we provide evidence from structural and functional connectivity supporting the increased intermodular communication of connector hubs during sleep deprivation.Secondly, we observe that this enhancement may reflect a compensatory mechanism within the brain, as it is associated with increased network costs, reduced functional segregation, and improved global efficiency.Lastly, the affected connector hubs primarily involve the SN and the CN, suggesting a heightened demand for neural resource allocation and cognitive control after sleep deprivation.

Fig. 4 .
Fig. 4. Cross-modal/cross-dataset similarity of PCnorm distribution.A. Correlation analysis on the group average PCnorm values derived from the FC matrices of the well-rested session in both the discovery and replication datasets.B. Correlation analysis on the group average PCnorm values from the FC matrices and the group average PCnorm values from the SC matrices, both sourced from the well-rested session of discovery datasets.The degree of freedom is 198, given the 200 cortical parcellations.PCnorm: normalized participation coefficient.

Fig. 5 .
Fig. 5. Contrast of connectivity diversity on connector hubs.We utilized the MATLAB ksdensity function to perform a probability density estimate for visualization purposes.The horizontal axis represents the subject-wise average PCnorm of connector hubs.FC: functional connectivity; SC: structural connectivity.***: p < 0.001; **: p < 0.01.

Fig. 6 .
Fig. 6.The associations between connector hub diversity (average PCnorm) and various global network metrics.These metrics encompassed global efficiency, network cost, modularity, and small-worldness.The horizontal axis represents the enhancement of average PCnorm in connector hubs (sleep-deprived minus wellrested).The rising cost and rising global efficiency indicate the direction of "sleep-deprived minus well-rested," while the reduced small-worldness and reduced modularity indicate the direction of "well-rested minus sleep-deprived."FC: functional connectivity; SC:structural connectivity; ns: not siginificant.All q-values represented the FDR corrected p-values.

Fig. 7 .
Fig. 7.The network in which the significantly enhanced connector hubs reside.The polar plot represents the quantity of significantly affected connector hubs within the network after sleep deprivation.The network assignment is referenced from Schaefer et al. (2018).FC: functional connectivity; SC: structural connectivity; DorsAttn: dorsal attention network; SomMot: somatosensory/motor network; TempPar: temporoparietal network.