Abstract
The connectome of the brain has a great impact on the function of the brain as the structure of the connectome affects the speed and efficiency of information transfer. As a highly energy-consuming organ, an efficient network structure is essential. A previous study has shown consistent overall brain connectivity across a large variety of species. This connectivity conservation was explained by a balance between interhemispheric and intrahemispheric connections; that is, spices with highly connected hemispheres appear to have weaker interhemisphere connections. This study examines this connectivity trade-off in the human brain using diffusion-based tractography and network analysis in the Human Connectome Project (970 subjects, 527 female). We explore the biological origins of this phenomenon, heritability, and the effect on cognitive measures.
The proportion of commissural fibers in the brain had a negative correlation to hemispheric efficiency, pointing to a trade-off between inner hemispheric and interhemispheric connectivity. Network hubs including anterior and middle cingulate cortex, superior frontal areas, medial occipital areas, the parahippocampal gyrus, post- and precentral gyri, and the precuneus had the strongest contribution to this phenomenon. Other results show a high heritability as well as a strong connection to crystalized intelligence. This work presents cohort-based network analysis research, spanning a large variety of samples and exploring the overall architecture of the human connectome. Our results show a connectivity conservation phenomenon at the base of the overall brain network architecture. This network structure may explain much of the functional, behavioral, and cognitive variability among different brains.
SIGNIFICANCE STATEMENT The network structure of the brain is at the basis of every brain function as it dictates the characteristics of information transfer. Understanding the patterns and mechanisms that guide the connectome structure is crucial to understanding the brain itself. Here we unravel the mechanism at the base of the connectivity conservation phenomenon by exploring the interaction between hemispheric and commissural connectivity in a large-scale cohort-based connectivity study. We describe the trade-off between the two components and examine the origins of the trade-off and observe the effect on cognitive abilities and behavior.
Introduction
The white matter of the brain can, quite intuitively, be modeled as edges in a network as it constitutes the connections among different nodes in the brain (Bullmore and Bassett, 2011). This network, the connectome, is assumed to have a great impact on brain functionality as it dictates the paths of information transfer in the brain (Bullmore and Sporns, 2012; Mišić et al., 2014).
An earlier study showed that the brain network connectivity is balanced and conserved across the brains of the mammalian class (Assaf et al., 2020). That study, exploring a large variety of mammalian brains, revealed that the overall connectivity, as described by the characteristic path length, is consistent across species, despite the large variance in brain volume and network structure. This connectivity consistency was manifested by a balance between interhemispheric and intrahemispheric connections. That is, the brain networks of species with fewer commissural connections seemed to compensate with shorter paths between node pairs, pointing to a more efficient network. This trade-off was also observed in human subjects with agenesis of the corpus callosum, presenting a strong hemispheric network, thus suggesting a mechanism that compensates for the lack of commissural connectivity (Owen et al., 2013a,b). This inherent balance among the networks may explain the connectivity conservation phenomenon across species.
In this study, we examined the connectivity conservation phenomenon in a large database of human subjects. Although we must take into account the interspecies variability that is significantly larger than interhuman variability, functional studies show that different subject connectomes have different and individually unique connectivity patterns (Finn et al., 2015). Thus different human subjects are still expected to show a variety of network structures while conserving the overall network connectivity, independently of brain size or the number of tracts.
Unlike with animals, analyzing this phenomenon in the well-mapped and long-studied human brain allows us to compare network properties at the node level rather than globally. More specifically, by using the same coordinate system for all human subjects, we can pinpoint the anatomic and functional origins of the connectivity conservation, an exploration that could not be done across species.
The brain network has a small-world structure, as was demonstrated by functional (Achard et al., 2006; Achard and Bullmore, 2007; van den Heuvel et al., 2008), structural (He et al., 2007), and diffusion tensor imaging (Sporns et al., 2004). This structure has been linked to many aspects of brain function as it could be affected by disease (Stam et al., 2007; Bassett and Bullmore, 2009; Stam, 2014), aging (Zuo et al., 2017) and pharmacological interventions (Achard and Bullmore, 2007). As network efficiency is highly dependent on this small-world structure (Latora and Marchiori, 2001), it may be expected to have a key role in many brain functions. Brain network efficiency, both structural and functional, was shown to affect cognition and other brain functions as loss of network efficiency in aging or disease was associated with a decline in cognitive abilities (Vlooswijk et al., 2011; Wen et al., 2011; Wiseman et al., 2018), whereas local and global efficiency both predict intelligence (Fischer et al., 2014; Pineda-Pardo et al., 2016). As such, it can be assumed that a trade-off between hemispheric efficiency and commissural connectivity affects cognition. The analysis of connectivity conservation across humans could be used to reveal cognitive implications of this structural phenomenon, leading to a greater understanding of brain organization and the effect on brain function.
In this work, we explore the connectivity conservation phenomenon in a large cohort of human subjects. We use the Human Connectome Project (HCP) young adult dataset (Van Essen et al., 2012) to observe interhemispherc and intrahemispheric connectivity across subjects and examine the possible structural origins for this balance. We used network analysis measures to locate brain regions that contribute to the balance as the more central, rich-club nodes. We also point to a possible hereditary source for this phenomenon, as well as the effect on cognitive abilities.
Materials and Methods
Data
We used diffusion and T1 data for 970 healthy adults (age 22–37 years, mean 28.8, 527 females) from the HCP 1200 young adult release (Van Essen et al., 2012). This includes high-quality T1w structural images as well as multishell diffusion scans. Full protocol details are available in the HCP reference manual (https://www.humanconnectome.org/storage/app/media/documentation/s1200/HCP_S1200_Release_Reference_Manual.pdf).
Personal information for each subject consisted of data from the National Institutes of Health (NIH) Toolbox questionnaires (NIH Toolbox (https://www.healthmeasures.net/explore-measurement-systems/nih-toolbox); Gershon et al., 2010) measuring cognitive abilities. In addition to that, family information was also collected, and genetic samples were taken for some participants. Of those, 105 pairs of subjects were genetically confirmed as monozygotic (MZ) twins, and 65 as dizygotic (DZ) twins, 95 were nontwin siblings and 270 were not related or third siblings of twin pairs. For heritability analysis, we only used data from twin pairs with genetic confirmation.
Experimental design
Preprocessing
The HCP provides minimally preprocessed images (Glasser et al., 2013). This preprocessing pipeline includes intensity normalization across runs, topup and eddy corrections, gradient nonlinear correction, and registration of diffusion image to MPRAGE using the Functional MRI of the Brain (FMRIB) linear image registration tool boundary-based registration (BRB)+bbregister. The full pipeline is available online (https://github.com/Washington-University/Pipelines).
To reduce the potential effect of brain volume, we resampled all images to a similar number of voxels per brain, that is, the resized brain mask for each subject had the same number of voxels as the smallest brain mask in the dataset. Then five tissue types (CSF, cortical gray matter, subcortical gray matter, white matter, and nonbrain tissue) were segmented on the structural image using the FMRIB Software Library (FSL) algorithm on MRtrix3 (Zhang et al., 2001; Jenkinson et al., 2002; Smith et al., 2004, 2012; Patenaude et al., 2011; Tournier et al., 2019).
Cortical parcellation
For anatomic parcellation, we used a modified automated anatomical labeling (AAL) atlas (Tzourio-Mazoyer et al., 2002). Subcortical and cerebellum areas were removed from the original atlas. As connectivity measures such as node degree are statistically biased toward larger areas, we divided the remaining brain regions into 150 similarly sized areas in each hemisphere (Cammoun et al., 2012) using a k means clustering algorithm. These areas were divided so that each new label consists of voxels from a single original AAL label, but the larger an original label was, the more areas it was divided into. This new (AAL300) atlas was registered to each subject's native space using the FMRIB FSL nonlinear registration tool FNIRT.
Diffusion-based tractography
Multishell multitissue constrained spherical deconvolution (CSD) (Tournier et al., 2004, 2019; Jeurissen et al., 2014; Dhollander et al., 2016), followed by anatomically constrained deterministic tractography, was performed using MRtrix3. To eliminate biases because of the large intersubject variability in brain volumes, we normalized the tractography parameters by each subject's brain volume.
Two tractograms were computed. In one, we filtered tracts to include only those that both start and end in or near voxels included in the cortical atlas with a search radius of two voxels to account for registration errors (Yeh et al., 2019). The remaining streamlines were then further filtered using the anatomically constrained spherical-deconvolution informed filtering of tractograms (SIFT) algorithm (Smith et al., 2013) to a set number of streamlines for each subject. On the second tractogram (full brain tracts), we used tracts from the entire brain mask and then filtered them using the SIFT algorithm. Here we removed the corticospinal tracts using a manually drawn “not” mask. This mask was created in Montreal Neurological Institute (MNI) space and registered to each subject's native space using FNIRT.
Network matrices
We computed connectivity matrices for individual brain networks. For each pair of brain regions in the modified AAL300 atlas, a connection was added if any fibers connecting the two areas were found. Streamlines with end points outside the atlas mask were discarded if no atlas area was found within a two voxel radius. We set the distance between each pair of nodes as the reciprocal of normalized number of fibers found between the two nodes, under the assumption that this weighting represents the probability of information transfer between those nodes (Sotiropoulos and Zalesky, 2019). This was done for a random subset of 15,000 tracts of the atlas-based tractogram, over 100 permutations. Our final connectivity matrix consisted of the mean weight of each edge over all permutations.
To compute a group-averaged connectivity matrix, we averaged the connection strength for each pair of nodes across subjects. Connections that appeared in 75% of matrices or less were discarded.
For normalizations and statistical analysis, we generated 1000 random matrices. The common model for random brain networks is the Maslov–Sneppen algorithm (Maslov and Sneppen, 2002) as it preserves the degree distribution in the network. However, a small-world network such as the brain is defined by a mean path length that is close to random characteristics (Milgram, 1967). As efficiency, which is a main measure used in this article, is computed based on the path length, the efficiency of a graph created by the Maslov–Sneppen algorithm will be very similar to that of the original graph. Thus, we created the random networks that preserves the group-average matrix edge weight distribution and not the degree distribution.
Statistical analysis and network analysis measures
Rich club and s-coreness
The rich-club coefficient is a global metric describing the extent to which nodes with high degrees are also connected between them. S-coreness is a weighted version of the coreness metric. For this metric, the maximal subgraph in which all nodes have a strength of k or higher is computed, and the coreness for each node is the highest k for which it can be found in the subgraph.
Both the rich-club coefficient
As random graphs also display some rich-club properties, we estimated
Mean shortest path and efficiency
To estimate the hemispheric connectivity, we calculated the network efficiency as described by Latora and Marchiori (2001). As the efficiency measure is normalized by the efficiency of an ideal graph, we first normalized each subject's connectivity matrix between 0 and 1 so that the ideal graph efficiency would be 1. The shortest path length (l) between each pair of nodes was computed using the Dijkstra (1959) algorithm, with the distance between each pair of connected nodes set as
Commissural ratio
We estimated the commissural connectivity in two different ways. The percentage of commissural connections was calculated as the number of streamlines passing through the subject's corpus callosum. We created a 2D midsagittal corpus callosum mask for each subject. The masks were created using automatic region-of-interest selection based on a fiber orientation map and then manually checked and corrected them based on the T1 and fiber orientation distribution (FOD) maps where necessary. We then filtered the full brain tracts to only include tracts passing through the corpus callosum mask. To reduce the chance of including false-positive tracts, we also discarded tracts that recrossed the midsagittal plane outside the corpus callosum, as well as tracts that crossed between nonhomologous areas as most tracts in the corpus callosum are relatively symmetrical (Hofer and Frahm, 2006). The proportion of commissural connections was computed as the percent of streamlines that pass through this mask divided by the total number of streamlines for this subject.
As this proportion might be biased by tracking parameters, we used an additional estimate of the commissural ratio. This measure is the quotient of the square root of the midsagittal area of the corpus callosum (based on the corpus callosum mask) divided by the cubic root of the brain volume (based on the brain mask). This measure, although more commonly used (Manger et al., 2010), is affected by image resizing and also represents the corpus callosum area, without taking into account the fiber density in this area.
Hemispheric efficiency and commissural ratio correlation
Node and edge removal
To locate brain regions with a high effect on connectivity conservation, we removed each node (i) from the network and recalculated the hemispheric efficiency without that node (
As strong connections are expected to have a greater effect on connectivity conservation, we also calculated the network efficiency value for networks with an increasing pruning threshold. That is, edges were removed by connectivity strength, with the weaker connections removed first. The hemispheric efficiency/commissural ratio correlation was computed after each pruning.
Heritability
Heritability measures quantify the extent to which people's genes account for differences in their traits. It is a statistical concept describing the level of variation in a specific trait that can be attributed to genetic variation. In this work, we estimated heritability using two measures. First, Falconer's formula (Falconer, 1960) defines heritability (
Canonical correlation analysis modeling for cognitive data and network measures
To examine the possible connections to cognitive abilities, we performed canonical correlation analysis (CCA). For this analysis, subjects with missing data in any of the measures used were discarded. As monozygotic twins were found to have a very similar network configuration, one twin of each monozygotic pair was also removed from the analysis.
An 819 × 5 (subjects × graph measures) matrix
A second matrix was created to represent the cognitive measures for each subject. Both corpus callosum size and functional and structural global network efficiency have been previously reported as associated with general intelligence (Luders et al., 2007; Kim et al., 2016; Pineda-Pardo et al., 2016). To examine this connection, we used cognitive measures associated with intelligence from the NIH Toolbox Cognition composite scores (Heaton et al., 2014). Each feature was standardized (z-score) across subjects.
To assess whether areas that have a strong effect on the connectivity conservation phenomenon also contribute to cognitive abilities, two similar, 819 × 17 (subjects × graph measures) matrices (
All network, node, and cognitive measures used to create the matrices for the CCA
Canonical correlation was cross-validated by removing a random 25% of the subjects as test data. We extracted the canonical weights using the remaining 75% and then computed the canonical correlations for the test dataset, using these canonical weights. This process was repeated over 1000 permutations, choosing a random set of subjects for the test dataset each time. The r values reported here are the averaged out-of-sample correlations over all permutations.
Significance for each correlation was assessed by randomizing the subject's order in
To establish the effect of the conservation phenomenon on the canonical correlation, we examined which graph measures contributed most to the correlation. For that, we computed the canonical loadings of each measure. The canonical loadings are the correlation between the measure itself and the canonical components so that the closer they are to 1, the more effective the measure is on the overall correlation.
Results
Brain network characteristics
The network of the brain is expected to present a rich-club organization (van den Heuvel and Sporns, 2011; Grayson et al., 2014). As such, it should have a core of nodes with high degrees that are highly interconnected and constitute the main connection between the lower degree nodes. This brain organization has been examined and described previously, along with a characterization of specific nodes that form the rich-club network hubs (van den Heuvel and Sporns, 2011; Ball et al., 2014; Mišić et al., 2014).
Brain volume distribution and impact on connectivity and the rich-club coefficient over different degree thresholds. A, The rich-club coefficient for the HCP dataset subjects (dark blue), randomized networks (light blue), and the normalized value (red). The k values with p < 0.05 are marked in a gray background on the graph as the k values for the rich-club nodes. B, Brain volume seemed to have a small effect on overall connectivity in the brain, with longer characteristic path lengths for larger brains, although the difference in connectivity was minor. C, A histogram of the variations in brain size. The variability in brain volumes was large, with the largest brain nearly 1.8 times larger than the smallest one. Note that we use weighted networks (weighted by 1/# fibers) and thus the actual efficiency values are on an arbitrary scale.
The connectomes extracted for our subjects presented a rich-club organization. The rich-club coefficient
Our recent study, focusing on the mammalian connectome, showed that although white matter volume appeared to be a function of brain volume, overall connectivity did not vary much across species (Assaf et al., 2020). This appeared to be consistent despite the large variance in brain volume and network structure, where the overall brain connectivity across mammals decreased by only 40% from the largest brain to the smallest brain, although brain volume decreased by more than four orders of magnitude.
Here, we examined the overall connectivity in the human brain and the dependency on brain volume. Overall brain network efficiency was estimated as the mean of the reciprocals of the distances (measured by the connection strength) between all node pairs. Pearson's correlation between whole-brain efficiency and brain volume for all the HCP subjects showed only a weak correlation (
Hemispheric efficiency correlation to the commissural ratio. Correlation between the average of the efficiencies computed for each hemisphere and the commissural ratio was significant with
A correlation between hemispheric efficiency and commissural ratio
To examine the connection between commissural connection strength and hemispheric connectivity, we estimated Pearson's correlation between the commissural ratio and hemispheric efficiency. A strong and significant correlation was found between the two measures (
The parcellation used in this analysis is an atlas-based anatomic parcellation. To negate possible biases of the specific parcellation (de Reus and van den Heuvel, 2013), we also created an anatomy-free parcellation using k-means clustering of all tract endpoints in each hemisphere for each subject separately. This created three unique parcellations for each subject, with 150, 300, or 500 areas per hemisphere that are not based on known anatomic or functional priors. Connectivity matrices were created based on these parcellations, and efficiency was computed for each hemisphere with similar results.
To ensure the correlation is not a result of our protocol, we extracted new connectomes for each subject using different numbers of tracts and received similar results. Normalizing the connection strength for each edge between 0 and 1 and setting the distance between nodes as
As connection strength was set by the number of streamlines found connecting every two areas, weaker connections, that is, connections consisting of fewer streamlines, have a higher probability to be false positives. To negate possible biases from false tracts, we tested the effect of pruning by testing various pruning thresholds and found that connectivity conservation is robust across most thresholds.
Node-based analysis
Different areas have a different effect on overall brain connectivity and on the balance between interhemispheric and intrahemispheric connections. To examine whether specific regions in the brain contribute more to the overall trade-off, we reran the analysis on a leave-one-out basis, removing each node from the network and recalculating the hemispheric efficiency to observe the effect of removal of that node on the trade-off between efficiency and the commissural ratio.
Removing the hippocampus, the superior motor areas, and the ACC from the network resulted in a lower correlation, suggesting that they are strong contributors to the effect. When removing other nodes such as occipital and temporal areas around the insula, we observed a higher negative correlation, suggesting they are nodes with connections that reduce the effect. These regions consist of a large part of the rich club and is mostly characterized by a high centrality in the network measured by the s-coreness (van den Heuvel and Sporns, 2011), which is a measure of the centrality of the node in the network. A significant correlation (
Contribution of nodes to the overall trade-off. Nodes with the strongest effect on the correlation between hemispheric efficiency and commissural ratio as found in the leave-one-out analysis. A, Nodes in each hemisphere colored by the
Heritability
Some elements in the connectome structure have a genetic source, whereas others are affected by the environment during brain development. As the connectivity conservation phenomenon seems to be a fundamental principle of the network architecture, and is inherent across species, we expected this trade-off to have an apparent heritable source. Participants were divided into monozygotic and dizygotic twins, pairs of siblings, and third siblings of twin pairs. Participants in the last group were randomly divided into unrelated pairs. We used Tukey's HSD to examine the differences in network efficiency and in the commissural ratio between these related pairs. The difference in the commissural ratio in monozygotic twins was significantly lower than that of all other pairs (MZ/DZ, p = 0.004; MZ/not twin, p = 0.04; MZ/not siblings, p = 0.001). However, the difference in efficiency between monozygotic twin pairs was only significantly lower than that of the nontwin and unrelated pairs (p = 0.04 and p = 0.001, respectively; Fig. 4).
The difference in hemispheric efficiency and the commissural ratio between sibling pairs. A, B, The difference in (A) hemispheric efficiency and (B) commissural ratio between sibling pairs. The x-axis signifies the mean absolute difference between siblings across groups. Corpus callosum size and diffusion measures are hypothesized to be affected by heritability factors.
Both measures were highly heritable, with Falconer's
Although this might point to a stronger environmental effect on the hemispheric efficiency than the commissural ratio, the connection between the two measures, that is, the basis of the connectivity conservation phenomenon presented in this work, seems to be genetically influenced. The cross-twin-cross-trait correlation between the commissural ratio and efficiency was significant for monozygotic twins (
Canonical correlation analysis for cognitive abilities
To examine the cognitive relevance of the connectivity conservation phenomenon, we ran a CCA analysis. This method examines the cross-covariance between two sets matrices (structured subjects × variables, with different variables in each matrix) by finding pairs of linear combinations for each set of variables in each matrix that have the largest correlation between pairs and with the minimum correlation between linear combinations of the same set. Here we examined the correlation between a cognitive measures matrix (
The first canonical pair for
Canonical correlation between connectivity measures and crystallized intelligence. Color and bar heights represent the r value of the correlation between the cognitive measure and the canonical variate. A, Canonical loadings for each cognitive measure used for the initial CCA. Measures associated with crystalized cognition are marked in red. B, Correlation between the canonical variates for the overall measures and the cognitive ones. C, Canonical loadings for the nodes with the strongest contribution to the connectivity conservation phenomenon. Nodes are presented by order of contribution and colored by the
We further examined the canonical correlation between connectivity measures (
When examining the canonical loadings for
Last, we examined the canonical correlation for
Discussion
In this study, we explored the existence and impact and of a connectivity conservation phenomenon in a large cohort of humans as seen before across mammals. We further study the mechanisms at the base of the connectivity conservation across brains.
Network analysis is an emerging theme in neuroimaging in the past few years (Bassett and Sporns, 2017). Recent works focus on specialized network structures or specific populations. As such, many works have explored modular structures of brain networks, local subnetworks, and differences in community structure (Valencia et al., 2009; Meunier et al., 2010; Sporns and Betzel, 2016; Puxeddu et al., 2020), while others have examined the dynamics of the network (Bassett et al., 2011; Beaty et al., 2016; Sporns, 2018). We present a cohort-based network analysis research, spanning a large variety of samples and exploring the overall architecture of the human connectome. This large-scale overview provides a new outlook on the network structure of the brain.
We show a trade-off between the association and commissural fibers, where brains with higher intrahemispheric connectivity have fewer connections between hemispheres, explaining the connectivity consistency seen across subjects and species. Our results pinpoint the biological origins of this phenomenon, with central rich-club nodes at the basis. Furthermore, we explore possible hereditary sources of the phenomenon and the effect on cognitive abilities. Our findings may shed light on the factors that guide brain network architectures and the effect of these factors.
Previous works show that white matter volume is dependent on brain size (Zhang and Sejnowski, 2000). However, our results show that brain efficiency is independent of brain volume and thus on white matter volume. Furthermore, our results show a correlation between the efficiency of the hemispheric network and the proportion of commissural fibers in the brain, indicating a conserved balance between these two factors. These findings strengthen the hypothesis that brain networks are organized in a way that preserves overall consistent connectivity.
Examining this phenomenon within species, and especially within the human brain, allows us to more easily examine the contribution of the network structure to connectivity conservation. As the variability in the human brain structure is smaller than it is among different species, this comparison supports the results because it shows that the vast differences in brain structure and network architecture across species are not the underlying cause of this trade-off. The human brain is also well researched and well mapped, allowing us to look into the effect of different structures on this phenomenon on a comparable parcellation.
When examining the brain regions that may have a stronger contribution to the connectivity conservation, our results show a strong effect of the hippocampus, the superior motor and occipital areas, and the areas surrounding the insula. Most of these areas are considered as network hubs or rich-club nodes (van den Heuvel and Sporns, 2011, 2013), that is, highly interconnected areas that form the connections between other nodes. As such, these areas are expected to be central in connectivity mechanisms.
We also explored the heritability of commissural and hemispheric connections using twin and sibling data. Earlier results observed a high heritability of the corpus callosum (Johnson et al., 1994; Woldehawariat et al., 2014). Many studies have also examined the heritability of diffusion measures and streamline count in the brain (Thompson et al., 2013; Budisavljevic et al., 2016; Yeh et al., 2016) as well as the functional connectome (Miranda-Dominguez et al., 2018). However, the heritability of the structural connectome and the overall connectivity measures does not seem to be widely investigated.
In the developing brain, individual differences in both functional and structural connectivity are reflected in the brain function. These differences may be inherent in the network structure, pointing to a genetic source, or change over time, and reflect plasticity due to each individual's specific experiences. Here, although the factors leading a specific brain to be on each end of the connectivity balance scale are unknown, a basis for the organizing mechanism for the whole network is seen. By showing heritability in both commissural ratio and hemispheric efficiency, as well as a strong cross-twin cross-trait correlation, our work points to a hereditary source for connectivity conservation. This, along with findings across different mammalian species, shows that this phenomenon is not environment or even species specific.
The brain network structure, consisting of a network of hubs strongly connected together, creates a modular structure that is at the base of network efficiency (van den Heuvel and Sporns, 2011). Functional studies have shown that performance in crystalized cognitive tasks is associated with the modular state of the functional network (Bertolero et al., 2015; Shine et al., 2016). These works suggest that the modular architecture of the network, and thus the efficiency, has an effect on the individual's cognitive abilities (Barbey, 2018).
The connectome architecture displays individually unique characteristics across subjects (Finn et al., 2015). These features have been seen to predict cognitive functions (Cai et al., 2021), and some were hypothesized to be affected by individual experiences (Marques et al., 2015). Our results also highlight a connection between the connectivity architecture and intersubject variability in cognitive abilities. Our analysis points to a connection between the components of the connectivity conservation phenomenon and crystallized intelligence. We also observe a weaker connection between the residuals and crystallized intelligence measures, suggesting that incompliance with the interhemispheric and intrahemispheric balance might affect cognitive abilities. It is difficult to infer the direction of connection between specific variables in the CCA. However, correlation direction suggests that lower hemispheric efficiency is associated with a higher crystallized cognition score. Corpus callosum size and diffusion measures have been previously found to be associated with intelligence and other cognitive functions (Hines et al., 1992; Clarke and Zaidel, 1994; Luders et al., 2007). Furthermore, brain efficiency was associated with intelligence (Kim et al., 2016; Ma et al., 2017), and deficits in higher cognitive abilities have been associated with deficiencies in brain network structure (Brown and Paul, 2000; Rudie et al., 2012) and efficiency (Du et al., 2019). These findings may underline our results, as the cognitive variance between subjects with different values of hemispheric connectivity is apparent. Our findings may shed light on the mechanism behind the intersubject cognitive variability.
The connectome extraction process has many parameters to be considered, and the resulting connectome is somewhat sensitive to those parameters. To avoid bias based on brain volume or resolution, all parameters were normalized to the volume of each brain. The number of nodes and the variability of these nodes in size may also create a bias in the connectivity measures (Cammoun et al., 2012) as well as using an anatomic parcellation. To avoid a bias toward larger areas, similar size areas were used. For the anatomic parcellation bias, the same analysis was run with a non-prior-based atlas, with different numbers of areas. All parcellations led to similar results, rejecting the possibility of the parcellation affecting the outcome. Another limitation of using tractography to map the corpus callosum tracts is the relatively large number of false positives in the tract-creating process. Manual validation and filtering of the corpus callosum tracts can easily locate many of those false streamlines but is more susceptible to variance because of subjective decisions. Here, as a compromise between the need for manual validation and conformity among samples, we mapped the common false tracts manually and created a set of objective rules to guide automatic filtering.
Our findings suggest that a trade-off between hemispheric connectivity and interhemispheric connections is the basis of the consistent connectivity of the brain network over different architectures and brain volumes. These results provide a unique opportunity to examine the network characteristics underlying this phenomenon and map the biological origin. Furthermore, our results examine the cognitive implications of this phenomenon.
As a connectivity phenomenon that spans across subjects, connectivity conservation should be examined in many scopes. First, as brain connectivity changes drastically over these years, it is important to examine changes in overall brain connectivity over time and the predictability of the adults brain network structure from that of a young age. Next, as functional networks are highly correlated to structural networks, exploring the phenomenon and the implication in the functional networks is essential, especially as the dynamics of the functional network and the ability of the network to change from a more connected to a more modular state (Barbey, 2018) may in part explain the intersubject differences in cognitive abilities. Furthermore, although this study has examined a large cohort of subjects, the difference in connectivity conservation in smaller populations, such as different age groups or subjects with pathologies, may show a different effect in the network structure. Last, our findings regarding cognitive measures connected to the phenomenon could lead to the exploration of this connectivity conservation mechanism in specialized networks or the changes in the connection structure induced by learning.
Although further studies should continue to explore this phenomenon and look into the cognitive implications, this study provides a deeper insight into the basis of the network structure of the brain and characterizes the architecture of the hemispheric network in the brain.
Footnotes
This work was supported by Israel Science Foundation Grant 1303/20 (to Y.A.), National Science Foundation–Israel Binational Science Foundation Collaborative Research in Computational Neuroscience Grant 2018711 (to Y.A. and Y.Y.), and British Council Israel Research and Academic Exchange Grant 43BX (to Y.A.). Data were provided by the Human Connectome Project, Washington University and University of Minnesota Consortium (1U54MH091657), funded by the 16 National Institutes of Health that support the Blueprint for Neuroscience Research, and by the McDonnell Center for Systems Neuroscience at Washington University. We thank Prof. Olaf Sporns from Indiana University for discussions.
The authors declare no competing financial interests.
- Correspondence should be addressed to Ronnie Krupnik at ronniek{at}mail.tau.ac.il