Evolution of the Rich Club Properties in Mouse, Macaque, and Human Brain Networks: A Study of Functional Integration, Segregation, and Balance

Abstract

The rich club, as a community of highly interconnected nodes, serves as the topological center of the network. However, the similarities and differences in how the rich club supports functional integration and segregation in the brain across different species remain unknown. In this study, we first detected and validated the rich club in the structural networks of mouse, monkey, and human brains using neuronal tracing or diffusion magnetic resonance imaging data. Further, we assessed the role of rich clubs in functional integration, segregation, and balance using quantitative metrics. Our results indicate that the presence of a rich club facilitates whole-brain functional integration in all three species, with the functional networks of higher species exhibiting greater integration. These findings are expected to help to understand the relationship between brain structure and function from the perspective of brain evolution.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12264-025-01393-5.

Introduction

When the level of connectivity between high-degree nodes in a network is more abundant than that between randomly-chosen node pairs, these nodes and their interconnecting edges form a rich club [1]. The rich club is believed to be one of the key topological properties of complex networks; it is widely present in various fields, including biological systems, social networks, power grids, and transportation systems [1]. Further, a variety of studies have demonstrated that the rich club organization of brain structural networks is universally present across different species [2–4], including Caenorhabditis elegans, Drosophila, mice, cats, monkeys, and humans [5–10].

The brain’s topological network plays a critical role in the organization of its functional processes. Researchers widely recognize two fundamental principles of functional organization: segregation and integration [11]. Functional segregation refers to the differentiation of brain regions in terms of their functions, specific regions being activated to perform particular tasks. Functional integration implies that a specific function may be accomplished through the cooperative effort of several regions or the entire brain. From the perspective of network topology, the modularity of brain regions supports functional segregation, while the connections between different modules via topological hub nodes facilitate functional integration [12–14]. The macroscopic spatial variation of cortical microcircuits can integrate different input information across various time scales. On the other hand, while large-scale hierarchical organization is present in the brains of mice, macaque, and humans, the mouse brain exhibits a less fully differentiated regional heterogeneity [15], which indicates that the organizational patterns of brain functional networks vary across different species. Moreover, as brain volume increases, modularity becomes more pronounced, which is believed to promote a balance between functional integration and segregation, thereby enabling efficient information processing [16, 17]. A rich body of literature has suggested that the balance between functional integration and segregation can predict cognitive abilities well [18, 19]. Specifically, simple tasks are more likely to be associated with functional segregation, whereas higher cognitive tasks, such as learning and memory, require greater integration across the entire brain [20, 21].

Structure is the foundation of brain function. Research has shown that the rich club of brain structural networks plays a key role in brain functions. However, it is still unclear how the rich club supports functional integration and segregation across different species. A study indicates that interspecies similarities are closely associated with the organization of neural circuits [3]. A variety of species’ connectomes exhibit consistent global features, such as densely connected hubs and small-world properties [2]. Under these circumstances, interspecies differences are often reflected in the role of connections in shaping functional aspects [22]. Therefore, it is interesting to analyze the performance of the rich club regions in functional integration and segregation across the brains of mice, macaques, and humans and explore the similarities and differences among the three species. Many previous studies on functional integration and segregation have primarily relied on graph theory methods. For example, metrics such as modularity, local efficiency, system segregation strength, and the number of provincial hubs have been used to represent network segregation [20, 23–25], while global efficiency and the number of connector hubs have been used to represent global integration [20, 23, 24]. However, these metrics cannot effectively model the hierarchical nature of these networks. Here, we introduce the nested-spectral partition (NSP) method based on eigenmodes [26]. Compared to the traditional network analysis metrics, the NSP approach considers the hierarchical nature of the brain, and thus, functional integration, segregation, and balance can be modeled in a multi-level nested manner.

We therefore investigated the similarities and differences in how the rich club supports functional integration and segregation across mice, macaques, and humans. We aimed to investigate three research questions (Fig. 1). Firstly, we used neuronal tracing data for mice and diffusion magnetic resonance imaging data for macaques and humans to validate the presence of a rich club and identify structural hub regions. Next, based on the NSP method, we explored the role of the structural rich club in functional integration, segregation, and balance by using functional MRI (fMRI) data from the three species in the awake resting state. Finally, from the perspectives of both static and dynamic functional network analysis, we compared the similarities and differences in the mechanisms of how the role of the rich club maintains functional integration, segregation, and balance across these three species. This cross-species study is expected to enhance our understanding of how the characteristics of structural brain networks enhance brain function and evolution.

Fig. 1.

Overview of research questions and the main method. This study aims to investigate the three questions (Q1–Q3). The subfigure below outlines the workflow designed to address the questions. First, we detect the rich club organization based on the structural connectivity of mouse, monkey, and human brains. Then, we introduce the sliding time window to improve the method called nested-spectral partition (NSP) based on eigenmodes to explore the role of the rich club in functional integration and segregation. Finally, we compare the similarities and differences in the role of the rich club in brain functional integration, segregation, and balance among mice, monkeys, and humans.

Materials and Methods

Mouse Dataset

The structural connectivity data for the mouse brain was sourced from the Allen Mouse Brain Connectivity Atlas [27]. After injecting recombinant adeno-associated virus labeled with enhanced green fluorescent protein into adult male wild-type C57BL/6J mice, axonal projections from specific brain regions were anterogradely traced in 140 coronal sections, forming a whole-brain connectivity atlas. Our work utilized the connectivity data of 213 brain regions as defined by the directed weighted connectivity matrix from Oh et al. [27]. We analyzed the ipsilateral connections of 213 regions in the right hemisphere. Detailed information on the 213 regions can be found in Supplementary Table S5. Specifically, a connection was defined as normalized connectivity strength:

$$W_{21}^{\mathit{norm}}=\frac{W_{21}^R}{\left|{\mathcal{R}}_1\right|}$$

where $W_{21}^R$ represents the sum of all projection voxels in the target brain region ${\mathcal{R}}_2$ when the source region ${\mathcal{R}}_1$ is used as the anterograde tracer injection site. $\left|{\mathcal{R}}_1\right|$ represents the number of all voxels in the source region ${\mathcal{R}}_1$. The normalized connection strength is obtained by comparing $W_{21}^R$ and $\left|{\mathcal{R}}_1\right|$.

The resting-state fMRI data of awake mice used in this study were from a public dataset [28], which includes data from 255 C57Bl/6J mice. We primarily used data from 15 of these awake mice. They were fixed in the magnetic field coil using a non-magnetic head fixation device compatible with the MRI CryoProbe [29]. The fMRI data were acquired using a Bruker 7T scanner with a gradient-echo echo planar imaging (EPI) sequence, at the resolution of 0.2 × 0.2 × 0.75 mm³ with an image matrix of 64 × 64 × 18, repetition time (TR) 1500 ms, echo time (TE) 15 ms, and including 390 time points.

The resting-state fMRI data of 15 awake mice were preprocessed as follows: (1) Skull removal: Since the original fMRI data included the skull, it was manually stripped off and checked by an anatomical expert. (2) Removal of the first 10 time points to ensure that the magnetic field had reached a stable state, guaranteeing data quality. (3) Slice timing and head motion correction were applied to the remaining time points, and the functional images were registered to the standard Allen Mouse Brain Common Coordinate Framework (CCFv3) [30]. (4) Band-pass filtering at 0.01 Hz–0.08 Hz, followed by spatial smoothing with a Gaussian kernel of 0.3 mm full-width at half maximum. These processes were applied using the SPM12 (Statistical Parametric Mapping version 12, https://www.fil.ion.ucl.ac.uk/spm/) toolbox on the MatLab platform.

To test the reliability of preprocessing results, we selected the anterior cingulate cortex (ACC) as the seed region. In each mouse, the blood oxygen level-dependent (BOLD) signals of all voxels within the ACC were averaged. Pearson correlation coefficients between this average signal and the signals of all other voxels in the brain were calculated to obtain the functional connectivity of the ACC with the whole brain. Fisher’s z-transformation was then applied to achieve normally-distributed Z-maps. For each mouse's Z-map, a one-sample T-test was conducted to identify the co-activation patterns of its default mode network, primarily including the dorsal caudate nucleus, dorsal thalamus, and areas surrounding the hippocampus (Fig. S1A).

Monkey Dataset

Diffusion MRI (dMRI) data for 8 macaques were acquired using a 9.4 T ultra-high-field Bruker scanner in our lab. Detailed acquisition parameters can be found in the Macaque Brainnetome Atlas [31]. After pre-processing steps such as eddy current correction, artifact removal, and denoising, the data were parcellated into 248 cortical structures and 56 subcortical nuclei. Probabilistic fiber tracking was then used to generate individual structural connectivity, with values representing the number of fiber bundles between brain regions, and diagonal elements set to zero. Detailed processing steps can be referred to in the published Macaque Brainnetome Atlas [31]. Finally, the group-level structural connectivity for macaques was obtained by averaging the structural connectivity matrices of the 8 macaques.

The macaque fMRI dataset used in this study was obtained from publicly available data at Newcastle University [32]. The fMRI data for 14 awake macaques were acquired on a Bruker 4.7 T scanner with an EPI sequence, at a resolution of 1.22 × 1.22 × 1.24 mm³ with an image matrix of 88 × 88 × 48, repetition time 2600 ms, echo time 17 ms, and flip angle 90°, including 250 time points. Ten of these macaques underwent two scanning sessions; due to excessive head motion during the first session for some macaques, only the data from the second session of these 10 macaques were used in this study.

Similarly, the fMRI data of awake macaques were preprocessed using Statistical Parametric Mapping (SPM). The first 10 time points were removed, and the remaining 240 underwent slice timing and head motion correction. The resulting functional images were then registered to the Macaque Brainnetome Atlas [31]. After spatial smoothing with a 2-mm full-width at half maximum Gaussian kernel and band-pass filtering at 0.01 Hz–0.08 Hz, the BOLD signals from 304 regions were extracted based on the Macaque Brainnetome Atlas [31].

To validate the effectiveness of the macaque fMRI preprocessing results, the posterior cingulate cortex (PCC) was selected as the seed region. The co-activation pattern of its functional connectivity with the whole brain was examined (Fig. S1B).

Human Dataset

The human brain structural connectivity data were sourced from the Human Connectome Project (HCP) with 100 unrelated subjects. Detailed data description and acquisition parameters can be found in the HCP Reference Manual (https://www.humanconnectome.org/storage/app/media/documentation/s500/HCP_S500_Release_Reference_Manual.pdf). Individual dMRI data were registered to the human Brainnetome atlas [33], and deterministic fiber tracking was applied using DSI Studio (https://dsi-studio.labsolver.org) to obtain structural connectivity matrices representing the number of fiber connections between brain regions. Finally, connections between brain region pairs that had values in >75% of the subjects were retained, and an average was taken to obtain the group-level structural connectivity for the human brain.

The human fMRI dataset used in this study was collected by our research group from 36 normal subjects using GE MR 750 3T multimodal MRI. The resolution of the resting-state fMRI was 3.75 × 3.75 × 4 mm³ with an image matrix of 64 × 64 × 39, including 255 time points. Only the data from the first acquisition time were used for analysis.

Similar to the pre-processing of mouse and monkey fMRI data, the fMRI data of all 36 subjects were preprocessed using SPM software on the MatLab platform. Considering that the scanner may not have reached a stable magnetic field at the beginning of the scan, the first 5 time points were removed during pre-processing. The remaining time points underwent slice timing and head motion correction. The resulting functional images were then registered to the human Brainnetome Atlas [33], with a resampling voxel size of 3 mm³. Subsequently, spatial smoothing was applied using a Gaussian kernel with 3-mm full-width at half maximum. A band-pass filter with a width of 0.01 Hz–0.08 Hz was applied. Finally, the BOLD time series were extracted for 246 brain regions defined by the human Brainnetome Atlas [33].

To test the effectiveness of the fMRI preprocessing results, the PCC was selected as the seed region. In each subject, the functional connectivity between the seed region and the whole brain was calculated, and Fisher’s z-transformation was applied to obtain Z-maps. The one-sample T-test was then conducted to identify the co-activation pattern of the human default mode network (Fig. S1C).

Detection of the Rich Club

The rich club is a topological property of a network [34]. In brain networks, the rich club consists of a subset of highly interconnected nodes. Connections between rich club nodes are typically denser than those between non-rich club nodes, playing a crucial role in global information communication and whole-brain functional integration [8]. The ‘k-core’ method is commonly used to detect rich clubs [1]:

$${\mathrm{\Phi }}^w\left(k\right)=\frac{w_{>k}}{{\sum }_{l=1}^{>k}{w}_l^{\mathit{ranked}}}$$

where k represents the node degree. For weighted networks [35], the edges are first sorted based on their weights to obtain $w^{\mathit{ranked}}$. For each k, a set of nodes with degrees >k is selected, and the number of connections $E_{>k}$ between these nodes is calculated. The total connection weight among these nodes is denoted as $w_{>k}$. The sum of the weights ${\sum }_{l=1}^{>k}{w}_l^{\mathit{ranked}}$ of the top $E_{>k}$ connections in $w^{\mathit{ranked}}$ is calculated. The ratio of these two quantities is the rich club coefficient. For each k, the rich club coefficient is calculated and normalized:

$${\mathrm{\Phi }}_{\mathit{norm}}\left(k\right)=\frac{\mathrm{\Phi }(k)}{{\mathrm{\Phi }}_{\mathit{random}}(k)}$$

where ${\mathrm{\Phi }}_{\mathit{random}}(k)$ represents the rich club coefficient of a random network. If, for certain values of k, ${\mathrm{\Phi }}_{\mathit{norm}}\left(k\right)$ is >1, this indicates the presence of a rich club in the network.

The rich club consists of highly-clustered hub nodes, typically used to evaluate the degree of node clustering. The criteria for distinguishing between hub nodes and non-hub nodes are mainly based on three metrics: node degree, connection strength [36], and betweenness centrality [37]. The most widely used method is to select the top 10%–20% of brain regions with the highest node degree as hub nodes [37–39].

Dynamic Nested Spectral Partition (NSP) Based on Eigenmodes

In fact, like brain structural networks, functional networks are hierarchically organized; however, functional networks are constantly changing. To dynamically investigate the role of the rich club in functional integration and segregation, we modified the NSP method by introducing a sliding time window [18, 26]. This approach allowed us to explore the integration and segregation of functional networks from the perspective of dynamic temporal changes based on feature modal partitioning. The core idea of the NSP method is to apply spectral decomposition to the functional connectivity matrix and to hierarchically partition modules based on eigenvalues and eigenvectors. Specifically, the eigenvalues and eigenvectors of the functional connectivity matrix were calculated and then sorted in descending order by eigenvalues. The eigenvector corresponding to the largest eigenvalue was considered the first-order mode, and so on. Once the modal order is determined, hierarchical module partitioning is applied based on the positive or negative values of the eigenvectors for each mode. In the first-order mode, there is one module that includes all regions of interest (ROIs), with all eigenvector values being either positive or negative. In the second-order mode, there are two modules, with the ROIs having positive eigenvector values forming the first module and those with negative eigenvector values forming the second module. In the third-order mode, there are four modules, where each of the two modules from the second-order mode is further divided into two sub-modules based on the eigenvector values of the regions. This process continues iteratively until each module contains only a single ROI.

By introducing a sliding time-window, we improved the method to dynamic nested-spectral partitioning based on eigenmodes. First, we determined the window width and the sliding step size. Next, for each time window, we calculated the Pearson correlation of the BOLD series to obtain the functional connectivity for that window. Then, we applied the above NSP to each time window separately.

After completing the dynamic module partitioning based on feature modes, we introduced two metrics—$H_{\mathit{in}}$ and $H_{\mathit{se}}$—proposed by Wang et al. to quantify functional integration and functional segregation respectively [18]. Factor $H_i$ that can reflect hierarchical integration and segregation is defined as:

$$H_i=\frac{{\mathrm{\Lambda }}_i^2{M}_i}{N}$$

Where ${\mathrm{\Lambda }}_i$ represents the eigenvalue corresponding to the i-th order mode, and ${\text{M}}_i$ represents the number of modules contained in the i-th order mode. After normalizing by the total number of brain regions N, the evaluation factor for the i-th partition is scaled to the range [0, 1]. Therefore, in lower-order modes, the contribution of ${\mathrm{\Lambda }}_i$ is much greater than the contribution of ${\text{M}}_i$, which reflects more of the whole-brain integration. In contrast, in higher-order modes, the presence of more small modules leads to a greater emphasis on segregation components. However, there may be imbalances in module partitioning at certain modes. For example, in the second-order mode, if the partition results in two modules where the first module contains only one brain region and the second module contains all the remaining regions, such an imbalance can occur. Therefore, a correction factor ${\text{p}}_i$ needs to be introduced to account for the number of brain regions $m_j=N/M_j(j=\text{1,2},3,\cdots ,M_i)$ in each module:

$$p_i=\frac{{\sum }_j\left|m_j,-,\frac{N}{M_i}\right|}{N}$$

At this point, the factor $H_i$ representing hierarchical module functional integration and segregation is given by:

$$H_i=\frac{{\mathrm{\Lambda }}_i^2{M}_i(1-p_i)}{N}$$

In the first-order mode, there is only one module containing the functional connectivity of all brain regions. Therefore, the global integration component $H_{\mathit{in}}$ is defined as follows:

$$H_{\mathit{in}}=\frac{H_1}{N}=\frac{{\mathrm{\Lambda }}_1^2{M}_1(1-p_1)}{N^2}$$

Starting from the second-order mode, each subsequent mode contains multiple modules. Therefore, the global segregation component $H_{\mathit{se}}$ is defined as follows:

$$H_{\mathit{se}}=\sum _{i=2}^N\frac{H_i}{N}=\sum _{i=2}^N\frac{{\mathrm{\Lambda }}_i^2{M}_i(1-p_i)}{N^2}$$

The balance $H_B$ of functional integration and segregation is defined as:

$$H_B=H_{\mathit{in}}-H_{\mathit{se}}$$

When the functional integration component and the functional segregation component are equal, i.e. $H_B=0$, the brain functional network is in a state of balance between integration and segregation. When H_B > 0 the brain is biased towards integration. When H_B < 0, the brain is biased towards segregation.

Statistics

To determine if there is a statistically significant difference between the hub and non-hub regions within the same species regarding functional integration and segregation metrics, we first assessed the normality and homogeneity of variance for both groups in each species. As these assumptions were not met, the Mann-Whitney U test was used to analyze differences in functional integration and segregation metrics. Since this test is based on rank rather than the raw measurements, it is robust to sample size differences between two independent groups [40]. A P-value < 0.05 indicates a significant difference.

To determine whether there are statistically significant differences between different species in terms of functional integration and segregation indicators, considering the small sample sizes for mice, macaques, and humans, the Mann-Whitney U test was also chosen to assess the differences in these indicators between mice-macaques, mice-humans, and humans-macaques. A P-value <0.05 indicates a significant difference between the two groups.

Results

The Rich Club Phenomenon and the Distribution of Hub Nodes

To analyze and compare the role of the rich club in functional integration and segregation across mice, macaques, and humans, it is necessary to first detect and validate the rich club of the structural network. The connectivity mapped through viral neuronal tracers is currently considered the gold standard for structural connectivity analysis. For mice, there is currently detailed projection mapping image data describing the axonal projections labeled by viral tracers. For macaques, although the CoCoMac database integrates several hundred published axonal tracing studies, there are no whole-brain connectivity data from direct tracer injections as in mice [41, 42]. Moreover, neuronal tracing-based whole-brain datasets are not available for humans. Therefore, the structural connectivity of the mouse was based on neuronal tracing data provided by the Allen Institute for Brain Science [27], while the structural connectivity of the macaque and human were derived from diffusion MRI data.

Here, the 'k-core' method was used to validate the rich club phenomenon of brain structural networks, where k represents the number of connections attached to each node [1]. For each k, the tendency of nodes with a degree >k connecting with each other to form a rich club was quantified. A normalized rich club coefficient >1 indicates the presence of a rich club organization within the network. The rich club coefficient and the normalized rich club coefficient of the structural networks in the brains of mice, monkeys, and humans were calculated, and the results are shown in Fig. 2A. For the mouse, the normalized rich club coefficients were ≥1 when $50<k<88$, indicating the presence of the rich club phenomenon in the mouse brain. After validation, we selected the top 15% of all brain regions with the highest node degrees as hub nodes to further explore their distribution in the brains of mice. This is shown in Fig. 2B, where the red spheres represent the hub nodes, and the size of each sphere is proportional to the node degree. In mice, the rich club brain regions are distributed in the neocortex, striatum, globus pallidus, thalamus, and midbrain. Note that, since the anterograde viral tracer was only injected into the right hemisphere, the original data include only the ipsilateral connectivity data from the right hemisphere [27]. In addition, using the probabilistic tractography connectome based on diffusion MRI data [43], we conducted a rich club analysis of the mouse brain, in which the whole-brain connectivity matrix included 296 regions (148 per hemisphere). We found that the rich club organization persisted in the structural network of the mouse brain, and the distribution of hub regions was consistent with the results based on the neuronal tracing data (See Supplementary Tables S1 and S2 for details).

Fig. 2.

The rich club and small-world attributes of the structural networks in mouse, monkey, and human brains. A The rich club coefficient and the normalized rich club coefficient of the structural networks in the brains of mice, monkeys, and humans. In all three, the normalized rich club coefficients are ≥1 when 50 < k < 88, 165 < k < 302, and 30 < k < 189 respectively, indicating the presence of the rich club phenomenon. r-c coeff: rich-club coefficient; nor r-c coeff: normalized rich-club coefficient; r-c level k: rich-club level k. B The distribution of hub nodes in mouse, monkey, and human brains when selecting the top 15% of regions with the highest node degree as rich club regions. The red spheres represent the hub nodes, and the size of the sphere indicates the node degree. The connectivity data for the mouse only includes ipsilateral connections from the right hemisphere. Therefore, the hub regions in the mouse brain are distributed only in the right hemisphere in our results. C The distribution of node degree, clustering coefficient, betweenness centrality, and shortest path length for hub and non-hub brain regions in mice, monkeys, and humans. Due to variations in the modalities of structural connectivity data across species, the magnitudes of different graph theory metrics differ. However, since this experiment focuses on comparing the significant differences between hubs and non-hubs, these variations are unlikely to impact the results.

As noted above, the validation of the rich club organization of the structural networks of the macaque and human brain was derived from diffusion MRI data (details in Methods). We found that in monkeys and humans, the normalized rich club coefficients were ≥1 when $165<k<302$ and $30<k<189$ respectively, indicating the presence of the rich club phenomenon in the brains of the two species. Further, in monkeys, the hub regions were primarily distributed in the prefrontal cortex, paracingulate cortex, anterior cingulate cortex, parietal lobe, temporal cortex, and thalamus. In humans, the hub nodes were mainly located in the lingual gyrus, precuneus, superior frontal gyrus, superior parietal lobule, insula, thalamus, putamen, and hippocampus, and they were symmetrically bilaterally distributed.

Small-World Attributes of Brain Structural Networks

By analyzing the graph theory attributes of hub and non-hub regions in the brains of mice, monkeys, and humans, we found that the clustering coefficients and betweenness centrality of hub nodes were generally higher than those of non-hub nodes, while the shortest path length of hub nodes was significantly shorter than that of non-hub nodes (Fig. 2C). This indicates that the brain structural networks of all three species exhibit small-world attributes [44], further suggesting that the presence of a rich club is important for understanding whole-brain communication and functional integration.

The Role of Rich Club Regions in Whole-Brain Functional Integration, Segregation, and Balance is Similar in Mouse, Monkey, and Human

To study the organization patterns of the functional networks in the brains of mice, macaques, and humans, we applied the NSP method to perform hierarchical modular decomposition of the functional connectivity network based on eigenvalues. The spectral decomposition of the functional connectivity matrix can reveal the primary patterns and characteristics of brain connectivity, enhancing the understanding of brain functional organization [45, 46]. At the first level, all brain regions belong to the same module, representing whole-brain functional integration. In subsequent levels, brain regions are divided into different modules according to their eigenvector values, representing functional segregation.

Within the framework of resting-state fMRI studies, functional connectivity analysis is used to quantify the temporal correlation and dependency between spatially-separated brain regions [47]. For each individual of the three species, we first calculated the static functional connectivity covering all time points of the entire scanning process. To investigate the role of the rich club in functional integration and segregation, we applied the NSP method to static functional connectivity matrices of hub and non-hub regions for each individual, calculating three static metrics: functional integration, functional segregation, and functional balance. However, studies have shown that functional connectivity may dynamically change throughout the entire 12 min–15 min resting-state scanning session, thus a full analysis of the entire scan may not capture the extent of functional activity [48, 49]. The sliding-window technique is the most widely used method for studying dynamic functional connectivity [50, 51]. Considering the effectiveness of functional connectivity obtained during the analysis of dynamic functional integration and segregation, and the fact that the number of effective fMRI time points for monkeys was only 240, we uniformly selected the first 240 time points for all three species to ensure consistency in the analysis. For dynamic analysis, we set the window width to 120 time points with a sliding step size of 1, covering a total of 240 time points. Research has shown that a window size of 120 is reasonable [52].

For mice, a total of 15 subjects were included. Using the NSP method, we calculated the functional integration, segregation, and balance indices for all 213 brain regions in each mouse; this comprised 32 hub regions and 181 non-hub regions (Fig. 3A(a)). For monkeys, a total of 8 subjects were included. We calculated the functional integration, segregation, and balance indices for the entire brain with 304 brain regions, including 46 hub regions and 258 non-hub regions in each monkey (Fig. 3A(b)). The ninth human subject (NC_09) had only 200 time points left after fMRI preprocessing, so this subject was excluded from the analysis. For the other 35 human subjects, we calculated the functional integration, segregation, and balance indices for all 246 brain regions in each subject, consisting of 37 hub regions and 209 non-hub regions (Fig. 3A(c)).

Fig. 3.

The role of rich club in whole-brain functional integration. A From top to bottom, the three subplots represent the individual functional integration, segregation, and balance metrics of whole-brain, hub regions, and non-hub regions for 15 mice, 8 monkeys, and 35 humans, respectively. B The Mann-Whitney U test for the significance of differences in functional integration, segregation, and balance metrics between hub and non-hub regions across the three species.

Because the functional integration and segregation indices of hub and non-hub regions in these three species did not satisfy normality and homogeneity of variance, we used the Mann-Whitney U test to analyze whether there were significant differences in the functional integration and segregation indices between hub and non-hub regions in the three species. The statistical results are shown in Supplementary Table S3 and Fig. 3B. For three species, in the awake resting state, the functional segregation components of rich club brain regions were significantly lower than those of the non-rich club regions, while the functional integration components were higher than those of the non-hub regions (P < 0.05, Mann-Whitney U test). In addition, hub regions exhibited a higher balance of functional integration and segregation compared to the whole brain and non-hub regions, suggesting that the presence of a rich club is beneficial for whole-brain functional integration in different species.

Considering that different thresholds might affect the results when identifying hub regions, we additionally chose the top 10% and 20% of nodes with the highest degree as thresholds and repeated the above analyses (Supplementary Table S3). According to the results, when the threshold was 20%, the p-values of integration indices for monkeys and humans were slightly greater than 0.05. Different thresholds can have some impact on the statistical significance of the differences, but the conclusion remains consistent.

Differences in the Role of Rich Club Regions in Whole-brain Integration, Segregation, and Balance Across the Mouse, Monkey, and Human

After calculating the individual static functional integration, segregation, and balance indices for the whole brain, hub regions, and non-hub regions in 15 awake mice, 8 awake macaques, and 35 humans, we then calculated these three indices based on dynamic functional connectivity and analyzed the differences in static and dynamic whole-brain integration and segregation among mouse, macaque, and human. In this section, we selected the top 15% of nodes with the highest degree to identify hub regions. As described in the earlier subsection on static functional integration and segregation analysis, different thresholds (10%, 15%, 20%) may have some impact on the statistical significance P-value, but do not affect the conclusions. Therefore, in dynamic functional integration and segregation and cross-species analysis, we have chosen 15% as a compromise threshold.

First, we conducted a cross-species analysis of static functional integration and segregation. Specifically, the results were averaged for each individual to obtain group-level calculations. Due to the small number of subjects, we used the Mann-Whitney U test to assess the significance of differences in indices between species: mouse versus monkey, mouse versus human, and monkey versus human. The results are shown in Fig. 4A and Supplementary Table S4. Figure 4A(a) shows the results for all brain regions (213 in mice, 304 in monkeys, and 246 in humans). Figure 4A(b) shows the results for hub regions (32 in mice, 46 in monkeys, and 37 in humans). Figure 4A(c) shows the results for non-hub regions (181 in mice, 258 in monkeys, and 209 in humans).

Fig. 4.

Differences in whole-brain integration and segregation among mice, monkeys, and humans. A From top to bottom, the three subplots represent the statistical significance of the Mann-Whitney U test for static functional integration, segregation, and balance metrics in the three species for the whole brain, hub regions, and non-hub regions, respectively. *P<0.05. B From top to bottom, the three subplots represent the dynamic functional integration, segregation, and balance metrics based on the sliding window technique for the whole brain, hub regions, and non-hub regions in the three species. The solid line represents the average of all individual indices, and the shaded area represents the 95% confidence interval of the mean. The subplot above the line plot displays the statistical differences between different species for each time window. For example, in the first subplot, the green dashed line indicates a significant difference in whole-brain integration between mouse and human in the 70th to 120th-time windows.

Next, we performed a cross-species analysis of dynamic functional integration and segregation. As noted earlier, the time-window width was uniformly set to 120 time-points with a sliding step-size of 1, covering a total of 240 time-points for all three species. For example, within each time-window, functional integration, segregation, and balance values were calculated separately for the whole brain, hub regions, and non-hub regions for each mouse. The results were then averaged across the 15 mice to obtain the group-level indices for that time-window. By sliding the window step-by-step, we obtained the dynamic functional integration and segregation results across the entire fMRI scanning process. In addition, we applied Mann-Whitney U test significance analysis for pairwise comparisons between mouse-monkey, mouse-human, and monkey-human for each window. Figure 4B shows these results, where the solid line represents the average of indices, the shaded area represents the 95% confidence interval of the mean, and statistical significance results are shown above the line plot.

Comparing Fig. 4A and B, the trends in dynamic functional integration and segregation are consistent with those in static functional integration and segregation for the three species. Specifically, the functional segregation of hub regions in mice, monkeys, and humans is significantly lower than that of non-hub regions, indicating that the rich clubs in all three species show a higher balance of functional integration and segregation than the whole brain or non-rich club regions. According to the integration and balance indices shown in Fig. 4A and B, the degree of whole-brain integration increases from mouse to monkey to human. Rich clubs demonstrate a more balanced functional integration and segregation, with higher-level species showing better whole-brain functional integration.

Discussion

In this study, we investigated the similarities and differences in the role of rich club regions in functional integration, segregation, and balance among the mouse, monkey, and human brains. We discovered that the degree of whole-brain integration increases progressively from mouse to monkey to human, with the rich club showing a more balanced role in functional integration and segregation. This finding indicates that the rich club regions play a crucial role in balancing the integration and segregation of the brain's functional networks. Despite the increasing volume and complexity of the brain from mice to monkeys to humans, the rich club's role remains critical in maintaining the functional balance. We suggest that this balance is likely to be an important driver in species evolution, which is expected to help explain the differences in cognitive abilities across species.

The brain is a product of evolution. Our results indicate that the distribution of hub node regions across the whole brain varies between species. Compared to macaques and humans, the mouse neocortex is smaller, non-folded, and less developed [53]. We found that the hub regions in mice are primarily located in subcortical nuclei such as the striatum, as previously reported [5]. The striatum is one of the most active brain regions in mice, with neurons showing high activity during task performance. It plays a crucial role in the life of the mouse and whole-brain communication [54]. Moreover, the rich club regions primarily overlap with association cortices in rodents [55], and the connections between these hub regions are driven by genes associated with energy synthesis and metabolism [5]. The distribution of macaque hub regions we obtained is consistent with the findings from analyses based on the CoCoMac dataset, which is derived from experimental tracing studies [7]. In contrast, the macaque shows a similar rich club distribution in the prefrontal, parietal, and insular cortices, but with less complexity than in humans [8]. In the human brain, these hub regions have significantly expanded functions, participating in complex cognitive tasks [56, 57]. In addition, the rich club in the human brain is bilaterally symmetrical, consistent with previous reports [8]. These differences highlight how the structural organization of rich clubs evolves with cognitive complexity and functional demands across species. Notably, the cingulate cortex, as a rich club region, has been found in the structural networks of mice, macaques, and humans [7, 58, 59]. Although the structural connectivity in mice exhibits stronger lateralization due to the common practice of analyzing data from only one hemisphere, the cingulate cortex still plays a pivotal role in orchestrating social behaviors. [60]. The presence of the cingulate cortex in rich clubs across different species suggests its conserved evolutionary function, as it is thought to participate in high-level cognitive processes, such as decision-making, emotional regulation, and social behavior [61]. This further illustrates that the rich club is a topological feature preserved throughout the evolutionary process.

The rich club is centrally located within the brain's communication pathways due to its topological characteristics in neural networks, making it a potential core of neural integration [8, 62]. Compared to non-rich club regions, rich club regions exhibit higher node degree, connection strength, and betweenness centrality [63]. ‘Neuroeconomics' posits that long-range and short-range connections, as channels for neural communication, are closely related to the trade-off between wiring cost and network efficiency in functional integration and segregation [64]. When short-range connections increase, resulting in lower wiring costs, the brain exhibits more segregated states. Conversely, when long-range connections increase, indicating high-value connections, the brain's integrative capacity is enhanced [63, 65]. The rich club exhibits high connection costs due to its dense connectivity pattern [7, 13]. At the same time, because of its high level of wiring volume and physical connections, it exhibits more direct communication pathways and shorter transmission delays, which are advantageous for global brain integration [65, 66]. Research in the cat brain has found that the rich club can facilitate a dynamic transition from simple modularity to global synchronization, suggesting its role in whole-brain functional integration and higher cognitive functions in mammals [67, 68]. Moreover, disruption of rich club connectivity can lead to cognitive impairments. For instance, reductions in hub connectivity are closely associated with Alzheimer's disease and schizophrenia, further emphasizing its role in maintaining the brain's functional integrity [59, 69, 70]. Based on these mechanistic properties, the role of the rich club in functional integration and segregation does not vary across species in mice, macaques, and humans, consistently supporting whole-brain functional integration.

More importantly, the rich club of mouse, macaque, and human brains shows significant differences in whole-brain functional integration and segregation. The rich club of the mouse brain is primarily distributed in subcortical nuclei like the striatum [5]. Although it can integrate information, its scope is relatively limited, focusing mainly on supporting functional integration directly related to survival [54]. Transcriptional coupling studies have found that the connections within the mouse's rich club are primarily driven by genes associated with the oxidative synthesis and metabolism of ATP, as well as those involved in learning and biological behavior [5, 71]. Rich club regions in the macaque cerebral cortex exhibit a star-like distribution, primarily located at the center of the star, connecting other structural communities [7]. Their key role is to gather information from other segregated network communities and organize it functionally, thus achieving a balance between functional integration and segregation [11, 14]. In addition, the rich club includes many shortcuts that span module boundaries, facilitating long-range signal flow, which underscores its core role in integrative processing [7, 8]. A recent study on the macaque has found that long-range connections are mainly distributed in non-hub regions, which are crucial for appropriate functional segregation and specialization. On the other hand, these long-range connections tightly link nearly all functional modules, contributing to efficient communication and integration [72]. Compared to the mouse, the macaque brain's rich club regions exhibit greater functional stability over time [73]. More specifically, certain neurons within these regions contribute to more complex functional dynamics [74, 75]. One hallmark of human evolutionary success is the complex brain, which can combine past and present information to make predictions about the future and adjust behavior accordingly [76]. The human brain dynamically and selectively reconfigures functional networks in real time based on different tasks to achieve a balance between integration and segregation [20]. Moreover, healthy aging is accompanied by a structural and functional decline in the brain, leading to gradual cognitive deterioration. Specifically, as age increases, dynamic segregation decreases across all functional networks except the cerebellum, while dynamic global integration increases [23, 77]. The human brain's rich club is associated with inter-individual behavioral variations, including cognitive and intellectual performance, which are considered unique to humans [78]. As brain networks evolve from mouse to macaque to human, the level of whole-brain integration progressively increases, and the role of the rich club in functional integration and segregation becomes increasingly important.

Several limitations should be taken into account when interpreting the results of cross-species analyses. The structural connectivity of the mouse brain is based on tracer data, while the structural connectivity of the macaque and human is derived from diffusion MRI data. Although whole-brain connectivity obtained through viral tracers is currently considered the gold standard for connectivity analysis, there are no tracer-based whole-brain datasets available for monkeys and humans due to technical limitations. While previous studies have shown a good correspondence between tracing and diffusion data [79], some inconsistencies may still exist [80]. Therefore, future analyses should aim to use data from the same modality. Moreover, the criteria and threshold selection for identifying rich club regions remain contentious, and statistically-derived thresholds can influence experimental results [36, 37]. In this study, we used degree ranking as the standard, the most commonly applied criterion. To mitigate the impact of a single threshold, we additionally selected two boundary thresholds covered in existing studies and repeated the same experiments [37–39].

In summary, this study validated the cross-species similarities of the rich club in the structural brain networks of mice, macaques, and humans. More importantly, we identified species-specific differences in the role of the rich club in functional integration, segregation, and balance. We suggest that these differences not only reflect the evolutionary complexity of neural networks but also reveal how different species optimize the function of brain structural networks. These results are expected to offer a new perspective for studying brain evolution.

Supplementary Information

Below is the link to the electronic supplementary material.

Conflict of interest

The authors declare that they have no conflict of interest. The HCP project involving human subjects was approved by the Ethics Committee of NIH. The studies involving humans, monkeys, and mice conformed with the Helsinki Declaration of 1975 (as revised in 2008) concerning Human and Animal Rights, and followed our policy concerning Informed Consent as shown on Springer.com. The human dMRI data were provided by the Human Connectome Project, WU-Minn Consortium (Principal Investigators: David Van Essen and Kamil Ugurbil; 1U54MH091657) funded by the 16 NIH Institutes and Centers that support the NIH Blueprint for Neuroscience Research; and by the McDonnell Center for Systems Neuroscience at Washington University.