Brain structural connectivity is a major organizing principle of the nervous system. Estimating interregional neural connectivity, reconstructing geometric structure of fiber pathways, and mapping the network connectivity to corresponding inter-individual variabilities provide fundamental insights in understanding detailed brain topological architecture, function and dysfunction. A large body of research has been devoted to extracting and investigating macro-scale brain networks from diffusion-weighted imaging (DWI) data (Xie et al., 2018; Jiang et al., 2019; van den Heuvel et al., 2019; Bertolero et al., 2019; Elsheikh et al., 2020), and various behavioral, neurological and neuropsychiatric disorders have been linked to the disrupted brain connectivity (Jiang et al., 2019; van den Heuvel et al., 2019). As structural changes of brain connectivity are phenotypically associated with massive complex traits across different categories, the brain-wide connectome has been extensively studied.

It is worth noting that human brain connectome re-configures its network structure dynamically and adaptively in response to genetic, lifestyle, environmental factors (Cohen and D’Esposito, 2016; Cauda et al., 2018), brain development and aging (Sala-Llonch et al., 2015; Alloza et al., 2018; Varangis et al., 2019). However, the underlying neurobiological mechanism from gene to brain connectome, and to cognitive and behavioral outcomes, and whether this mechanism changes over time, remain unclear. To bridge this gap, we perform a genetic study of brain connectome phenotypes on two different age-specific cohorts: one contains healthy young adults (age: 28.7 ± 3.6), and the other contains elderly participants (age: 73.8 ± 7.0). Our goal is to identify genetic factors affecting brain connectivity and examine their consistency and discrepancy between these two age-specific groups.

Emerging advances in multimodal brain imaging, high throughput genotyping and sequencing techniques provide exciting new opportunities to ultimately improve our understanding of brain structure and neural dynamics, their genetic architecture and their influences on cognition and behavior (Shen and Thompson, 2020). Present studies investigating direct associations among human connectomics, genomics and clinical phenotyping are primarily focused on four aspects: 1) estimating genetic heritability of basic connectome measures such as number of fibers, length of fibers and fractional anisotropy (FA) (Jahanshad et al., 2013; Thompson et al., 2013; Elliott et al., 2018); 2) discovering pairwise univariate associations between single nucleotide polymorphisms (SNPs) and imaging phenotypic traits such as above mentioned basic connectome measures at each edge (Jahanshad et al., 2013; Karwowski et al., 2019) and white matter properties at each voxel (Kochunov et al., 2010; Alloza et al., 2018; Guo et al., 2020); 3) discovering pairwise univariate associations between SNPs and clinical phenotypes such as cognitive or behavioral outcomes (Jahanshad et al., 2013; Elsheikh et al., 2020); and 4) discovering pairwise univariate associations between basic connectome measures and clinical phenotypes (Jiang et al., 2019; van den Heuvel et al., 2019).

Among the studies mentioned above, there exist two major limitations. First, these studies were conducted based on basic connectome measures such as number of fibers, length of fibers and FA, but the complex-network attributes were overlooked, which included network segregation, integration, centrality and resilience and important network components such as hubs, communities, and rich clubs (Sporns, 2013). These attributes were extensively adopted to detect network integration and segregation, quantitatively measure the centrality of network regions and pathways, characterize patterns of local anatomical circuitry, and test resilience of networks to insult (Rubinov and Sporns, 2010). Second, these studies performed analyses by examining the association between an independent variable (e.g., SNP) and a dependent variable (e.g., cognitive or behavioral outcome), without taking into consideration the mediator(s) linking these variables (Baron and Kenny, 1986). Mediation analysis can help identify the underlying mechanism of outcome-relevant genetic effects implicitly mediated by neuroimaging phenotypes (e.g., connectome measures). Of note, mediation analysis requires the independent variable to be significantly associated with both the dependent variable and the mediator. This makes applying it in brain neuroimaging studies a challenge due to the modest effect size of an individual genetic variant on both behavioral and imaging phenotypes (Saykin et al., 2015; Cong et al., 2018), as well as limited size of the sample with all diagnostic, imaging and genetic data available.

With the demand of measuring complex-network attributes, a few recent genome-wide association studies (GWAS) (Bertolero et al., 2019; Elsheikh et al., 2020) recognized the first problem mentioned above and adopted quantitative measurement approaches for complex-network attributes, and treated the attributes as neuroimaging traits for the explorations of complex imaging genomic associations. They successfully identified a number of loci susceptible for Alzheimer’s disease (Elsheikh et al., 2020), and demonstrated the associations between loci and segregated network patterns, which may be involved in brain development, evolution, and disease (Bertolero et al., 2019). However, a notable limitation is that these studies only focus on the brain networks of either young or elderly participants, as a result, their study outcomes are lack of validations in multiple data sets. Since there is an age-related discrepancy for genetic effects on human connectome alterations across lifespan (Varangis et al., 2019), it remains an under-explored topic to examine genetic consistency and discrepancy for complex-network attributes among cohorts different in age. Another factor that may cause discrepancy in the network architecture is the hemispheric asymmetry (Jiang et al., 2019), and the hemispheric asymmetry of network organization has been linked to development processes (Zhong et al., 2017) and neuropsychiatric disorders (Sun et al., 2017). It remains a challenge to understand the genetic basis for the network attributes of two hemispheres as they may be distinctively correlated to cognition level, physical and psychological development.

Among a large number of complex-network attributes, it has been well documented in recent literatures (Cohen and D’Esposito, 2016; Xie et al., 2018) that segregation of neural information such as modularity, transitivity, clustering coefficients and local efficiency represent the connectivity of local network communities that are intrinsically densely connected and strongly coupled. A converging evidence (Cohen and D’Esposito, 2016; Karwowski et al., 2019) is shown that local, within-network communication is critical for motor execution, whereas integrative, between-network communication is critical for measuring connectome (Bertolero et al., 2019). Thus, network segregation is thought to be essential for describing and understanding of complex neural connectome systems (Sporns, 2013). In addition, segregation measures are highly reliable and heritable network attributes (Xie et al., 2018), and these measures have been linked to the disruption of neural network connectivity in brain development, evolution, disease (Cohen and D’Esposito, 2016; Mak et al., 2016; Bertolero et al., 2019), and immunodeficiency (Bell et al., 2018). Given the importance of network segregation, in this study, we first focus on quantifying measures of network segregation, analyzing heritability of segregation measures and performing genetic association analyses by treating them as neuroimaging traits. Then, our next priority is to explore the genetic basis for the rest of the complex-network attributes (e.g. integration, centrality and resilience).

To overcome the challenges mentioned above, this study aims to develop and implement computational and statistical strategies for a systematic characterization of structural connectome optimized for imaging genetic studies, and to determine genetic basis of structural connectome. Specifically, the framework is organized and described in Figure 1, and the primary goals are to address the following six critical issues: 1) construction of basic network connectivity with diffusion tractography, 2) systematic extraction of complex-network attributes, 3) heritability analysis of complex-network attributes, 4) genome-wide association studies of quantitative endophenotypes, 5) examination of mediation effect that intermediately bridges genes and outcomes, and 6) identification of outcome-relevant neuroimaging biomarkers. Given the enormously broad scope of brain connectome, our focus is on studying 1) static tractography-based structural connectome and complex-network attributes characterizing segregation, integration, centrality and resilience; 2) genetic consistency and discrepancy for complex-network attributes among cohorts different in age; and 3) mediation effects of network attributes on outcome-relevant genetics.

The major contributions of this study are fivefold:

• New challenges in human connectome: we elucidate the neurobiological pathway from SNPs to brain connectome, and to phenotypic outcomes. By integrating connectomics and genetics, this study provides new genetic mechanism insights into understanding detailed brain topological architecture, and encoding (or mapping) inter-regional connectivity in the genome.

As summarized in Figure 1, prior to GWAS, we first performed heritability analysis for network attributes screening, and only heritable measures of network segregation were treated as iQT for GWAS. Based on experimental outcomes, all of the segregation measures were highly heritable: transitivity and modularity were heritable at network level, clustering coefficient and local efficiency were heritable at all nodes, which suggested segregation measures were suitable for genetic analyses. Then, we performed GWAS of segregation attributes in 275 HCP subjects, and identified several pairwise associations between SNPs and iQTs as listed in Table 2. These GWAS findings were validated in 178 ADNI subjects. As a validation result, we identified several genetic consistency and discrepancy patterns for human connectome in different life stages (as shown in Table 2). As common findings in both HCP young adult and ADNI elderly adult cohorts, rs7937515 was negatively associated with clustering coefficient and local efficiency respectively measured at left middle temporal gyrus. To the best of our knowledge, this was among the first GWAS of human brain high-level network measures across both young and elderly participants. As shown in Figures 3A,C, the minor allele G of rs7937515 was associated with decreased clustering coefficient and local efficiency of the left middle temporal gyrus in both young and elderly participants. As concluded in (Rudie et al., 2013; Keown et al., 2017; Karwowski et al., 2019; Varangis et al., 2019), the weakness of segregated network connectivity (e.g., modularity, clustering coefficient, and local efficiency) was linked to multiple brain disorders such as age-related cognitive declines and autism spectrum disorder. Thus, our GWAS findings for HCP young adults demonstrated that rs7937515 played a risk effect on human network segregation. This neurorisk effect was also confirmed in a targeted genetic association analysis for ADNI elderly participants (as shown in Figures 3B,D), these common discoveries between HCP and ADNI datasets suggested a consistent genetic risk effect across young and old life stages.

This study was further conducted by performing several post-hoc analyses in the following three aspects (shown as bottom sections in Figure 1): 1) examining genetic mechanisms for common outcome measures in the HCP and ADNI studies, and elucidating the mediated effect of iQTs for such outcome-relevant genetic association, 2) discovering outcome-relevant imaging biomarkers, and 3) exploring the genetic mechanisms of other important complex-network attributes.

For the first aspect, our goal was to elucidate the neurobiological pathway from SNPs to brain connectome, and to phenotypic outcome. In addition, we aimed to discover the role of iQTs in the outcome-relevant genetic associations by performing mediation analyses in both HCP and ADNI datasets. For the HCP young participants, we identified that BMI was positively associated with rs7937515 in the first step of mediation analysis, demonstrating a risk effect. rs7937515 located in the regions of FAM86C1/FOLR3 was previously discussed in literatures (Gao et al., 2015; Gao, 2017) and positively linked to waist circumference in the meta-analysis based on the Insulin Resistance Atherosclerosis Family Study (IRASFS) (Palmer et al., 2015), which was designed to investigate the genetic and environmental basis of insulin resistance and adiposity. FAM86C1 (Family With Sequence Similarity 86 Member C1) and FOLR3 (Folate Receptor Gamma) had been reported for their associations with various weight-related phenotypes such as bone mineral density (Li et al., 2019) and BMI (Hair, 2014; Mrozikiewicz et al., 2019), which closely related to osteoporosis (Li et al., 2019; Mrozikiewicz et al., 2019) and obesity (Gómez-Ambrosi et al., 2004). In the second and third steps of mediation analysis, we illustrated that BMI was indirectly influenced by rs7937515 (Figures 4, 5), and iQTs such as clustering coefficient and local efficiency measured at the left middle temporal gyrus respectively played a mediating role. We also examined the genetic association with MMSE, but no evidence indicated any genetic associations to MMSE. In contrary, for the ADNI elderly participants, neither significant associations between rs7937515 and BMI nor MMSE were identified in the first step of mediation analysis, so there was not a necessary to examine mediated effect in this dataset. Our results demonstrated a disappearance of outcome-relevant genetic effect in the elderly participants, this discrepancy from young to elderly participants might due to the dominated influences from life style, environment or other non-genetic factors.

For the second aspect, recent studies (Lowe et al., 2019; Azevedo et al., 2019) showed that structural changes in brain tissues could affect food consumption behaviors and mediate BMI, which implied connectome alteration could be a causal agent and a promising imaging biomarker in this study. Thus, our goal was set to reveal the mapping between connectivity alterations and phenotypic outcome, and discover outcome-relevant imaging biomarkers. For young adult participants, segregation measures (e.g., clustering coefficient or local efficiency measured at left middle temporal gyrus) previously demonstrated their potential to play a mediating role in genetic association discoveries, in this step, we focused on examining their direct associations to the outcomes. Thus, we performed a targeted association analysis between the mentioned segregation measures and the common outcomes (e.g., BMI or MMSE) evaluated in both HCP and ADNI studies (Table 2) by employing linear regression models. For the young participants, clustering coefficient and local efficiency measured at left middle temporal gyrus were negatively associated with BMI. Similar observation was obtained in (Chen et al., 2018) which linked lower structural network segregation to obesity (higher BMI). Our findings suggested that there was a mapping between brain network segregation attributes and human physical conditions, and segregation features of the left middle temporal gyrus showed their potential as neuroimaging biomarkers to detect BMI-associated complex diseases such as dementias (Emmerzaal et al., 2015), cardiovascular disease, cancer, respiratory disease and diabetes (Stenholm et al., 2017). For elderly adult participants, no significant associations were identified between segregation measures and any outcomes, which suggested an interesting topic for further explorations.

Multiple regression analyses demonstrated that middle temporal gyrus was linked to weight-related issues. For example, Veit et al. (2014) and Gómez-Apo et al. (2018) revealed that BMI, visceral fat and age were negatively associated with cortical thickness of the left middle temporal gyrus, Ou et al. (2015) indicated that greater adiposity was associated with lower gray matter (GM) volumes in the middle temporal gyrus, Yokum et al. (2012) found positive correlation between BMI and white matter (WM) volume in the middle temporal gyrus, Rapuano et al. (2016) illustrated left middle temporal gyrus was detected significantly greater activation in response to food commercials than to non-food commercials, Salzwedel et al. (2019) concluded that maternal adiposity influenced neonatal brain functional connectivity in middle temporal gyrus, and Peven et al. (2019) identified that cardiorespiratory fitness was negatively associated with functional connectivity in the right middle temporal gyrus. To the best of our knowledge, our investigations for the association between structural connectivity in the middle temporal gyrus and BMI was among the first weight-related studies with high-level imaging features measured from structural network connectivity, and our results confirmed several previous findings analyzed from thickness data, T1-weighted MRI data, and fMRI data.

For the third aspect, since there was an emerging interest in understanding the segregation and the integration of brain networks (Cohen and D’Esposito, 2016; Mohr et al., 2016) as well as other important network attributes such as centrality (Zuo et al., 2012) and resilience (Karwowski et al., 2019), our goal was to expand our focus on comprehensively discussed segregation attributes to a more complete set of network attributes including segregation, integration, centrality and resilience. For both node level network attributes measured at left and right middle temporal gyrus and global network attributes, we applied targeted genetic association analyses on global efficiency and density (integration, network level), betweeness and eigenvector centrality (centrality, node level) and assortativity coefficient (resilience, network level) of the structural connectivity. We identified several pairwise associations between rs7937515 and these network attributes in both HCP and ADNI datasets (Table 3), and noticed a significant association between rs7937515 and global efficiency in both datasets, which suggested that rs7937515 was involved into the dynamic fluctuations of segregation and integration of neural information. This finding partially answered an elusive question of revealing genetic basis for brain mechanisms of balancing network segregation and integration. Another worth noting finding was that rs7937515 was associationed density and eigenvector centrality respectively in our targeted analyses, while such associations were vanished in elderly participants, which suggested inconsistent genetic influences across different life stages.

With the awareness of the hemispheric asymmetry of network organization, a genetic basis to explain the differences in connectome between two hemispheres were under discovered. In this work, we identified an obvious inconsistency of genetic influences on human connectome in different brain hemispheres (Table 3). As reported in several recent studies (Tian et al., 2011; Shu et al., 2015; Jiang et al., 2019), the topological organizations of structural networks were not uniformly affected across brain hemispheres, which lead to a non-uniformly distributed destruction on neural network of the left and right hemispheres. Our finding gave an explanation from the point-view of genetics, with the potential for further investigations as many of the destruction on neural network (as iQT) were linked to cognitive and behavioral functions and dysfunctions, and their genetic mechanisms were still under discovered.

In this work, we constructed the structural network connectivity, extracted complex-network attributes and examined the heritability of network segregation measures. Then, we revealed a novel association between the minor allele (G) of rs7937515 and decreased network segregation measures of the left middle temporal gyrus across HCP young participants and ADNI elderly participants, which demonstrated a consistent genetic risk effect on brain network connectivity across lifespan. We elucidated the neurobiological pathway from SNP rs7937515 and genes FAM86C1/FOLR3 to brain network segregation, and to BMI. In such pathway, we concluded a genetic risk effect on BMI due to their positive association, examined the mediated effect of network segregation measures, and discovered network segregation of the left middle temporal gyrus as BMI-related neuroimaging biomarkers by identifying a negative association between them. We also examined genetic associations of a more complete set of important network attributes including integration, centrality and resilience, and demonstrated the consistency and discrepancy in genetic associations in brain aging. At last, we illustrated hemispheric asymmetry of network organization in both datasets in the aspect of genetic effect.

In sum, with the awareness that BMI is directly and indirectly associated to multiple complex diseases, this study performed a systematic analysis that integrated genetics, connectomics and phenotypic outcome data, with the goal of revealing biological mechanisms from the genetic determinant to intermediate brain connectomic traits and to the BMI phenotype at two different life stages. We identified the genetic effect of rs7937515 on human brain network segregation in both young and elderly participants and on BMI in young adult cohort. Our findings confirmed several previous genetic and imaging biomarker discoveries. Moreover, we provided outcome-relevant genetic insights in the aspect of complex-network attributes of human brain connectome. Similar analytical strategies can be applied to future integrative studies linking genomics with connectomics, including the analyses of structural and functional connectivity measures, additional network attributes, longitudinal or dynamic connectomic features, as well as other types of brain imaging genomic data.

Data used in preparation of this article were obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database (adni.loni.ucla.edu). As such, the investigators within the ADNI contributed to the design and implementation of ADNI and/or provided data, but did not participate in analysis or writing of this report. A complete listing of ADNI investigators can be found at: http://adni.loni.usc.edu.