
==== Front
9215515
20498
Neuroimage
Neuroimage
NeuroImage
1053-8119
1095-9572

39009250
10.1016/j.neuroimage.2024.120739
nihpa2016110
Article
Joint modeling of human cortical structure: Genetic correlation network and composite-trait genetic correlation
Shen Jiangnan a
Zhang Yiliang a
Zhu Zhaohan a
Cheng Youshu a
Cai Biao ab
Zhao Yize a
Zhao Hongyu acd*
a Department of Biostatistics, Yale School of Public Health, New Haven, CT, USA
b Department of Management Sciences, City University of Hong Kong, Hong Kong S.A.R, China
c Department of Genetics, Yale University School of Medicine, New Haven, CT, USA
d Program of Computational Biology and Biomedical Informatics, Yale University, New Haven, CT, USA
* Corresponding author at: Department of Biostatistics, Yale School of Public Health, Ste 503, 300 George Street, New Haven, CT 06511, USA. hongyu.zhao@yale.edu (H. Zhao).
26 8 2024
15 8 2024
14 7 2024
02 9 2024
297 120739120739
https://creativecommons.org/licenses/by-nc/4.0/ This is an open access article under the CC BY-NC license (http://creativecommons.org/licenses/by-nc/4.0/).
Heritability and genetic covariance/correlation quantify the marginal and shared genetic effects across traits. They offer insights on the genetic architecture of complex traits and diseases. To explore how genetic variations contribute to brain function variations, we estimated heritability and genetic correlation across cortical thickness, surface area, and volume of 33 anatomically predefined regions in left and right hemispheres, using summary statistics of genome-wide association analyses of 31,968 participants in the UK Biobank. To characterize the relationships between these regions of interest, we constructed a genetic network for these regions using recursive two-way cut-offs in similarity matrices defined by genetic correlations. The inferred genetic network matches the brain lobe mapping more closely than the network inferred from phenotypic similarities. We further studied the associations between the genetic network for brain regions and 30 complex traits through a novel composite-linkage disequilibrium score regression method. We identified seven significant pairs, which offer insights on the genetic basis for regions of interest mediated by cortical measures.

Imaging genetics
Cortical structure
Genetic correlation
Gaussian graphical model
Spectral clustering
==== Body
pmc1. Introduction

The human cerebral cortex is critical for higher-order brain functions, including cognition, information processing, perception, thought, language, and consciousness. Abnormalities in the cortex are linked to several cognitive disorders, such as Alzheimer’s disease (Yao et al., 2010) and Parkinson’s disease (Pan et al., 2017). Cortical thickness (CTh), cortical surface area (CSA), and cortical volume (CV) are three key measures of morphology in brain integrity and neuroimaging studies. Significant increases or decrease in these measures can signal abnormal brain structure. For example, psychogenic nonepileptic seizures have been associated with a distinct profile of changes in CTh within the insula (Ristić et al., 2015); CSA has been associated with leftward asymmetry in auditory-related cortical areas (Meyer et al., 2014), and CTh has been associated with rightward asymmetry in the primary and secondary auditory cortex (Meyer et al., 2014). Recent research has investigated shared genetic factors and enhanced the understanding of regional associations across different measures. Specifically, several studies have consistently illuminated high genetic correlations between homologous regions of the opposing cerebral hemispheres (Alexander-Bloch et al., 2013; Biton et al., 2020; Fürtjes et al., 2023; Alexander-Bloch et al., 2019; Wen et al., 2016). Additionally, CTh and CSA exhibit substantial genetic overlap within cortical regions (Grotzinger et al., 2023). The variability of brain asymmetry was found to be significantly influenced by environmental factors, emphasizing the importance of considering both genetic and environmental factors in understanding brain morphology (Biton et al., 2020).

A vital tool for unraveling genetic architecture of complex human traits is genetic correlation. It quantifies shared genetic contributions between traits or disorders. Among different methods to estimate genetic correlation, cross-trait linkage disequilibrium score regression (LDSC) is commonly used to estimate genetic correlation based on summary statistics from genome wide association studies (GWAS) (Bulik-Sullivan et al., 2015a,b). Motivated by LDSC, several methods have been developed to move beyond global genetic correlation, such as stratified by functional annotations (Finucane et al., 2015) and chromosomal regions (Shi et al., 2017; Guo et al., 2021). Instead of a single, genome-wide measure of genetic overlap, these partitioned and localized correlations offer a more granular, region-specific understanding of shared genetic factors among traits. For example, as an extension of an annotation-stratified genetic correlation estimation method GNOVA (Lu et al., 2017), SUPERGNOVA (Zhang et al., 2021) can elucidate both global and local genetic correlations.

Several published studies have applied various dimension reduction and regression techniques to further explore genetic correlation patterns of cortical regions. For instance, Grotzinger et al. (2023, 2019) employed Genomic Structural Equation Modeling (Genomic SEM) to model genetic overlap across brain regions; while Fürtjes et al. (2023) used genomic principal components (PCs) to reveal shared genetic etiology of genetic covariance and brain-wide morphometry. Similarly, Oblong et al. (2024) applied principal component analysis (PCA) and independent component analysis (ICA) to break the multiple brain trait GWAS statistics into latent components to better interpret genetic effects. Several other methods built on GWAS summary data and correlations have been proposed to identify shared genetic factors between traits (Warrier et al., 2023; Soheili-Nezhad et al., 2021; de Leeuw et al., 2015). Notably, Alexander-Bloch et al. (2019) identified the hierarchical structure of cortical measures, showing modular architecture and functional and genetic asymmetry between hemispheres. This study also found an overall decrease in correlation with regional distance, consistent with the correlation pattern observed in postmortem gene expression.

To address the brain’s modular structure, we extend prior analyses using modularity to select and determine the clusters that align most with the correlation structure. More specifically, we constructed a Gaussian graphical model through constrained ℓ1-minimization for inverse matrix estimation (CLIME) (Cai et al., 2011). We then applied recursive two-way cut (Shi and Malik, 2000) to the genetic and phenotype similarity matrices to identify the optimal number of clusters and generate genetic and phenotype clusters. Our approach enables a more objective division of the brain regions that is coherent with genetic patterns. Additionally, we compared the brain maps for CSA, CTh, and CV, and demonstrated concordance between anatomical and genetic analyses. Our results confirmed and extended the findings from earlier studies (Alexander-Bloch et al., 2019; Hofer et al., 2020) and offered a novel perspective through the application of network analysis on correlation matrices. By comparing the genetic and phenotypic networks, we could capture and interpret the genetic effects on phenotype correlations. Our overall analysis framework facilitates the characterization of the genetic basis of cortical measures associated with ROIs.

To understand the interplay between regions of interest (ROIs) and complex traits, we proposed composite-trait Linkage Disequilibrium Score Regression (composite-trait LDSC), a new method that treats the traits within a cluster as a composite phenotype for genetic correlation estimation. As an extension of LDSC, composite-trait LDSC integrates information across genetically correlated traits and jointly analyze the clustered ROIs. By amalgamating ROIs within the same cluster, composite-trait LDSC reduces the number of imaging traits considered, enhances statistical robustness, and mitigates the impact of some outliers. We applied composite-trait LDSC on the clustering results and identified seven significant pairs with genetic correlations between each component of these composite-ROIs and 30 complex traits. Our findings suggest that composite-trait LDSC may reveal genetic pathways shared between brain structures and complex diseases.

2. Methods

2.1. Imaging genetics summary data for cortical structure and complex traits

We performed a genetic analysis of brain imaging data downloaded from the Oxford Brain Imaging Genetics Server-Big40 (Smith et al., 2021), a large-scale population-based cohort study with over 4000 imaging-derived phenotypes (IDPs) from multimodal brain imaging in the UK Biobank (UKBB) (Fig. 1). Based on a high-resolution T1-weighted magnetic resonance imaging (MRI) scan, the cerebral cortex is segmented into 33 contralateral homologous regions with unique anatomical and functional properties. The 66 ROIs in the left and right hemisphere were generated with Freesurfer by parcellation of the white surface using Desikan-Killiany parcellation. The CSA, CTh, and CV of each brain region were derived and quantified from 31,968 subjects, which were shown to be highly heritable and associated with various brain structure and functional processes (Hofer et al., 2020). CSA was calculated as the sum of the areas of each tessellation falling within a given ROI (Palaniyappan et al., 2011). CTh was computed as the average distance between the parcellated portions of white and pial surfaces (Mensen et al., 2017). Lastly, CV integrated the information of CSA and CTh and was calculated by multiplying area by thickness at each vertex (Winkler et al., 2018). Each of these measures offers a unique perspective on the genetic influence on brain structure, capturing complex interaction between genetic factors and brain morphology. For genetic association analysis, we removed single nucleotide polymorphisms (SNPs) with poor imputation quality (INFO <= 0.9) and low minor allele frequencies (MAFs) (MAF <= 0.01). Moreover, non-SNP variants, strand-ambiguous SNPs, and SNPs exhibiting duplicate rs numbers were also removed.

To study genetic correlation between brain structure and complex traits, we downloaded GWAS summary statistics of 30 complex traits, with detailed information provided in Supplementary Table 1. After imputation quality control, we retained SNPs shared between GWAS and the 1000 Genomes Project, while excluding those with MAF < 0.05 and having strand-ambiguity. SNPs situated in the Major Histocompatibility Complex (MHC) region were removed and only autosomal chromosomes were considered in our analyses.

2.2. Clustering ROIs

To effectively discern the patterns from the large number of connections in the inferred genetic network, we cluster ROIs based on the estimated genetic correlations to give a more actionable and insightful map of genetic correlations and to enable the prioritization of important relationships. In this study, we constructed a non-negative similarity matrix Wij = rg(i,j) + 1, where rg(i, j) represents the genetic correlation between ROIs i and j. To segment the graph, we used a global criterion Normalized Cut (Ncut) to compute the similarity within the groups and dissimilarity between the groups. The minimized Ncut value determines the splitting point for each optimal bipartition of the graph. Furthermore, we used the threshold of Ncut to decide the recursively repartition of the segmented parts (Supplementary Note). To evaluate the strength of division of a network into clusters and ascertain the Ncut threshold, we employed modularity (Newman, 2006) given by: (5) Q=12m∑i,jWij−kikj2mδCi,Cj,δCi,Cj=1Ci=Cj0Ci≠Cj

where W = (Wij)1≤i,j≤N is the weight matrix of the graph; m=12∑ijWij is the sum of all the connections, ki=∑jWij denotes the total weights from node i to all other nodes, kikj2m is the expected weight of an edge between node i and j, and Ci is the community of the ith node. Since the modularity quantifies the alignment between the clusters and the network, we chose the best-fitting cluster to explain the genetic correlation structure with the maximal modularity value across all the partitions with different Ncut thresholds.

2.3. Gaussian graphical models for ROI network

We employed Gaussian graphical models (GGM) to infer the conditional dependence among different ROIs (Lauritzen, 1996). The inverse of the covariance matrix, also known as the precision matrix in GGM, describes the conditional independence in that the zero element in the precision matrix implies conditional independence of the two corresponding variables given the other variables, representing no edges in the graphical network. Based on the covariance matrix derived from LDSC, we set the covariances with p-values exceeding 0.05 to zero and performed an eigenvalue decomposition of covariance matrix Σ to ascertain the positive semidefiniteness. We then applied the constrained l1-minimization for inverse matrix estimation (CLIME) (Cai et al., 2011) method, which is less computationally costly, more sparse and less noisy than the other methods like SCAD (Fan et al., 2009) and Glasso (Banerjee et al., 2008). Following the estimation of the precision matrix, we used the elementwise indicator to represent the adjacency matrix of the graphical model (Supplementary note).

2.4. Composite-trait LDSC

We may improve our understanding of genetic impacts on complex phenotypes by considering composite traits simultaneously instead of a single trait. To facilitate such composite-trait analysis, we generalize the LDSC framework through integrating the trait cluster information for composite-trait correlation estimation. For the precision matrix, conventional test statistics may fail to maintain asymptotic normality due to the presence of large nuisance parameters. Thus, we used de-biased CLIME estimator proposed by Neykov et al. (2018) to mitigate the impact of nuisance parameters and generate the revised genetic precision matrix: Ez0ja1z1j/N1+a2z2j/N2+…+apzpj/Np=N0Mlj∑i=1paiρgi+∑i=1paiρiNsiNiN0

(7) ω^ij0d=ω^ij0−ω^i.0TΣω^.j0T−ejω^i.0TΣj

where ω^i.0 and ω^.j0 are the ith row and jth column of the CLIME estimator Ω^0=ω^ij0, is the jth canonical basis in RN, and Σj is the jth row of the covariance matrix. Given the marginal association statistics from GWASs (i.e., z scores for jth SNP z1j, …, zpj), we consider a synthetic trait which is the linear combination of the traits with weights a = a1, a2, …, ap). The synthetic trait z score can be calculated by the marginal linear regression between the composite-trait and SNP j. In particular, the sample size weighted z score can be expressed as a1z1j/N1+a2z2j/N2+…+apzpj/Np, where Npi denotes the sample size of the ith study, where i = 1…p. Then the heritability of the composite-trait can be estimated as (8) h^multi2=aTPgaaTPa

where P and Pg are the p × p phenotypic and genetic covariance matrices, derived from the de-biased inverse precision matrix. To increase the statistical power of the trait pair association, we maximize the heritability and rewrite (8) as: P−12PgP−12y=λy,

where y=P12a. Notably, P−12PgP−12 is symmetric positive semidefinite as the Pg is positive semidefinite. By Rayleigh quotient (Lauritzen, 1996), the largest eigenvector of P−12PgP−12 corresponds to the real valued solution of y, from which the weights a can be derived. In general cases, the composite-trait covariance could be written in a similar form as LDSC: Ea1z1j/N1+a2z2j/N2+…+apzpj/Np2=ljM∑i,k=1paiakρgik+∑i,k=1paiakNiNkρikNsik

where Ni is the sample size of study i, ρgi, ρgik is the genetic covariance, ρi, ρik is the phenotypic correlation among the Nsi and Nsik overlapping samples. In our composite-trait analysis, z0j is the z score of the common trait, and z1j, …, zpj are the z scores for p ROIs within one cluster. Given that all the traits are from one study, we have N1 = N2 = … = Np = N. Consequently, the sample size z score is (a1z1j + a2z2j + … + apzpj) /N, and the composite-trait genetic correlation can be obtained by ∑i=1paiρgi.

3. Results

3.1. Heritability and genetic correlations for cortical structure and complex traits

We estimated the SNP heritability and genetic correlation across three measurements (Fig. 2, Supplementary Tables 2–5, Supplementary Figs. 1–5). All cortical regions across three measurements showed significant heritability after adjusting for multiple comparisons, with the highest heritability in the pericalcarine region in CSA and CV. Besides, we identified 742 pairs of significant genetic correlations among ROIs (120 pairs for CV, 193 pairs for CSA, 430 pairs for CTh; p < 0.05/(33×65) = 2.33e–05). Genetic correlations were observed more robustly between contralateral homologous regions (mean = 91.93 %, SD = 7.76 %) than other inter-regional pairs (mean = 31.43 %, SD = 23.02 %). Specifically, a total of nine regions for CTh and one region for CSA showed a correlation coefficient of 1 with their contralateral homologs. Correlations between CSA and CV had a higher degree of similarity than those with CTh both genetically and phenotypically (Supplementary Table 6), and regional connectivity was more pronounced in CTh than CSA and CV. The GWAS analysis between cortical ROIs and 30 complex traits demonstrated eight pairs of significant genetic correlations (two pairs for CV, five pairs for CSA, pair for CTh; p < 0.05/(33×30) = 5.05e–05; Supplementary Tables 7–9, Supplementary Figs. 6–8). To measure the degree of similarity between how indices of genetic and phenotypic influences resemble each other, we calculated linear associations between phenotypic and genetic correlations. Our analysis revealed a positive association across 2145 pairs of genetic and phenotypic interregional correlations (rCSA2=0.71,rCTh2=0.57,rCV2=0.69), and the magnitudes of genetic correlations tended to be marginally higher than phenotypic correlations (Supplementary Fig. 9). These findings were in line with what was observed in Fürtjes et al. (2023). We also reported the correlation, mean difference, and p value for the paired t test of phenotypic and genetic correlation (Supplementary Table 10). These results suggest the concordance between phenotypically and genetically implied cortical architecture.

Since global analysis may mask some significant local genetic effects, we applied SUPERGNOVA to estimate local genetic correlations and calculated the proportion of correlated regions for each pair of ROIs for CSA, CTh and CV. With LD estimated from the 1000 Genomes Project phase 3 samples of European ancestry, the genome was partitioned into 2325 approximately independent regions by LDetect. Among a total of 1980 pairs, 459 for CV, 485 for CSA, and 489 for CTh were locally correlated after Bonferroni correction for 4,596,049, 4,596,108 and 4,595,819 regions for CV, CSA and CTh, respectively (pvolume < 0.05/4, 596,049, parea < 0.05/4,596,108, pthickness < 0.05/4,595,819, Fig. 3A, Supplementary Figs. 10–11, Supplementary Tables 11–16). Genetic correlations for CV and CSA showed similar patterns between the left and right hemispheres. In contrast, CTh had marked disparities, particularly in contralateral homologous regions. Specifically, 1.87 % and 1.95 % of regions demonstrated correlated genetic effects between BMI and CV in the left and right superior parietal regions, respectively, compared to 9.46 % and 9.47 % for CSA. However, CTh presented a stark discrepancy with significantly correlated proportions at 23.21 % and 5.39 % in the left and right hemispheres, reflecting substantial differential genetics in homologous regions. Besides, we also applied LAVA for local genetic correlation. Among 1980 trait pairs, 67 for CSA, 47 for CTh and 37 for CV had at least one significantly correlated region after Bonferroni correction for 2,349,152, 2,365,050 and 2,336,420 regions for CV, CSA and CTh, respectively (pvolume < 0.05/2, 349, 152, parea < 0.05/2, 365, 050, pthickness < 0.05/2, 336, 420, Supplementary Figs. 12–14, Supplementary Tables 17–19). LAVA identified similar patterns between homologous regions as those by SUPERGNOVA. The genetic covariance of LAVA and SUPERGNOVA showed low similarity with correlations of 0.36, 0.35, and 0.35 for CSA, CTh, and CV, respectively. When excluding the estimates that were not significant in both models, the correlation increased significantly to 0.87, 0.86, and 0.80 for CSA, CTh, and CV, respectively (Supplementary Fig. 15).

SUPERGNOVA identified correlations in specific chromosomal regions that were not present in global analyses. Although there were no significant global correlations between Parkinson’s disease (PD) and cortical measurements, we identified significant local correlation at a locus on chromosome 5 (58,834,852–60,880,207) between CSA, CTh, and CV and PD. This specific locus achieved genome-wide significance in a PD GWAS (Fig. 3B). Similarly, education attainment and smoking initiation exhibited correlations with ROIs at the same chromosomal locus 5 (58,834,852–60,880,207) (Fig. 3B). Alcoholism, education attainment, rheumatoid arthritis, cognitive performance, and Crohn’s disease exhibited correlations with ROIs in the region on chromosomes 16 (27,446,054–29,023,966). Schizophrenia and education attainment exhibited correlations with ROIs in the region on chromosomes 17 (46,828,412–48,027,295). These correlations are further corroborated by locus-specific GWAS analyses (Supplementary Figs. 16 and 17). Besides, we performed enrichment analysis for 30 complex traits in tissue specificity of gene sets with 10 defined tissue and cell types (i.e., Blood, Breast, Brain, Epithelial, ESC, Fetal, GI, Heart, Lung and Muscle) (Supplementary Figs. 18–20), showing different biological processes across different tissue-specific genomes.

3.2. Cluster and network construction

The clusters inferred from genetic correlation network revealed genetic sharing patterns. Specifically, the optimal number of genetic clusters for CSA, CTh and CV was 3, 3 and 5, respectively, compared to 2, 2, and 3 for phenotype clusters for these three measures. Our results suggest that there may be different modular structures across different morphological markers. In contrast to lobe mapping, genetic clusters largely respect cortical segmentation by geometric and neuroanatomical conventions. The occipital lobe displayed a clear delineation across CSA, CTh, and CV maps, setting it apart from the other lobes. Within the CSA framework, the remaining two clusters tended to divide the ROIs into parietal-temporal and frontal-cingulate sections. For CTh, the clusters identified the cingulate from other combined lobes, whereas CV mapping discerned patterns within the parietal and frontal-cingulate regions, further segregating the temporal ROIs into two distinct groups. These clusters closely correspond to traditional lobe mapping, indicating their inherent connection with functional parcellation. All the contralateral homologous regions except for the CTh lateral orbitofrontal cortex were clustered in the same cluster. Even without imposing constraints on anatomical adjacency, the genetic cluster structures displayed both spatial contiguity and symmetry. To assess the modular structure across scales, we then evaluated the modularity of random networks and genetic correlations with clusters selected with varying fractions of strongest connections. CSA and CV showed a strong modular structure with correlations compared to randomly connected networks with the same degree distribution (Supplementary Fig. 21).

The phenotype clusters exhibited both shared and distinct patterns with genetic clusters. For CSA, the phenotype clusters were in close agreement with the genotypic clusters, suggesting a potentially strong genetic basis for CSA. For CV, while one of its clusters matched one genetic cluster, intriguingly, the remaining ROIs merged into one phenotype cluster. For CTh, the phenotype clusters partitioned ROIs along the cerebral hemispheres, distinct from the symmetrical patterns observed in its genetic clusters. The detailed node information of clusters can be found in Supplementary Table 20. By modularity, the correlation trends of CSA and CV clusters draw remarkable parallels to traditional lobe mappings than CTh, potentially indicating the distinct evolutionary or developmental dynamics for CTh (Supplementary Table 21).

The graphical model revealed a complex and heterogenous network characterized by its consistency in connections between contralateral homologous regions. These connections faded more slowly compared to their non-homologous counterparts as the graphical models became increasingly sparse. Moreover, regions housed within the same cluster harbored richer connections as opposed to regions that straddling disparate clusters. This illustrates the strong associations knitting together regions within a given cluster, suggesting perhaps a shared genetic underpinning or co-evolutionary pressures. When the tuning parameter was set to 0.1, 237 pairs emerge as conditionally dependent in the CV graphical network, whereas the CSA and CTh networks had 256 and 315 paired connections, respectively (Fig. 4, Supplementary Figs. 22–26). Regional CSA, CT and CV networks showed higher genetic and phenotypic similarity than the CTh group based on the Jaccard similarity (Haque and Neubert, 2020) (Supplementary Table 22). To check whether our results can accurately capture the correlation pattern, we measure the modularity between the cluster and the network (Supplementary Table 23). For both CSA and CV, the genetic modularity is greater than the phenotypic modularity, suggesting that the genetic cluster is better able to explain the underlying structure of the genetic association. Cortical thickness (CTh) may have a more diffused genetic architecture and less interpretability with smaller modularity than CSA and CV.

3.3. Composite-trait LDSC

Through composite-trait LDSC, we calculated the correlation between each component of composite -trait and 30 complex traits (Supplementary Tables 24–26, Fig. 5, Supplementary Figs. 27 and 28). We found seven significant pairs for three measurements, including cluster 2 of CSA and cognitive performance (CP; ρg = 0.115; p = 3.65E–05), cluster 1 of CSA and educational attainment (EA; ρg = − 0.150; p = 1.70E–06), cluster 1 of CSA and body mass index (BMI; ρg = 0.164; p = 4.21E–06), cluster 3 of CV and education attainment (ρg = 0.118; p = 4.49E–07), cluster 5 of CV and education attainment (ρg = − 0.164; p = 1.30E–05), and cluster 3 of CV and age at menarche (AAM; ρg = 0.109; p = 3.87E–05). Moreover, to discern hemisphere-centric influences, we amalgamated regions across both hemispheres, assigning them equivalent weights (Supplementary Tables 27–29, Fig. 5, Supplementary Figs. 29 and 30). This integration revealed a significant correlation between the CV of the rostral anterior cingulate and educational attainment (ρg = − 0.140; p = 2.73E–06). However, this effect was not detected when the individual hemispheres were studied separately, demonstrating the value of composite trait analysis in capturing overarching patterns.

Our investigation also clarified different genetic effects of complex traits across three cortical measures. Pertaining to PD, the CSA of integrated regions predominantly displayed positive correlations with PD, whereas the CTh of all the regions generally exhibited negative relationships. The CV, however, exhibited a more complex interaction with both positive and negative associations. This result suggests uniformity across all regions for CSA and CTh and diverse correlation patterns that different cortical morphology markers hold with common traits.

4. Discussion

In this study, we used the brain imaging data from the UKBB to estimate heritability of CSA, CTh, and CV of 66 ROIs and their genetic correlations using LDSC, SUPERGNOVA and LAVA. For genetic correlations with 30 common traits, we showed that while global genetic correlations did not reveal many significant relationships, local genetic correlation analyses inferred significant associations in specific genomic regions. For example, educational attainment (EA) with CSA of left bankssts, schizophrenia (SCZ) with CTh of left lateral occipital, cognitive performance (CP) with CV of right superior temporal gyrus were not significant in global analyses but showed significant local correlations by SUPERGNOVA and LAVA. This emphasizes the importance of examining genetic correlations at a more granular level, where global analyses might miss patterns of significance. However, the results of SUPERGNOVA and LAVA were not consistent in local genetic covariance estimation, with only 21 pairs being significant in both methods. For a comprehensive comparison of different local genetic correlation methods, Zhang et al. (2023) showed that SUPERGNOVA produced more stable local genetic correlation estimates using different reference panels compared to LAVA. Specifically, SUPERGNOVA outperformed LAVA when using an external reference panel with matched LD. Therefore, we recommend considering SUPERGNOVA results as more reliable and suggest using LAVA when an in-sample reference panel is available for brain imaging data. To evaluate the performance of two models, the downstream analyses, such as GenomicSEM (Grotzinger et al., 2023; Grotzinger et al., 2019) and GenomicPCA (Fürtjes et al., 2023), could be used to better infer local genetic covariance structure.

The cluster results suggested distinct genetic influences across three measures. Prior studies have reported that CSA and CTh are largely independent (Winkler et al., 2010) and weakly associated (Strike et al., 2019). Such independence could be because of differences in the cellular processes that influence each measure during corticogenesis (Rakic, 2009). As the product of CSA and CT, CV will tag properties of both measures to improve the ability to simultaneously detect influences on thickness and area (Winkler et al., 2018). Our results of CSA and CV being more similar than CTh provides evidence for the regionally distinct genetic factors. This observation largely agrees with GWAS meta-analysis for cortical structure (Hofer et al., 2020) that CSA and CV show a substantial genetic correlation and CSA and CTh have a generally weak correlation.

Moreover, the significant tissue-specific signals and shared genetics in specific chromosomal regions indicate different biological processes involved in brain structures and different traits. For example, our analysis identified local genetic correlation for variants within chromosome 5 (58,834,852–60,880,207) between PD and all three measures of ROIs, consistent with previous multi-trait association analysis showing relations between PD and these variants, localized in the SMIM15 and ERCC8 genes (Tirozzi et al., 2023). The SMIM15 gene was found to be associated with brain volume measurement (Zhao et al., 2019), which can be a potential impact factor of the PD. For ERCC8, its expression level is linked to cognitive profiles (Baez et al., 2013), with cognitive deficits observed in patients with Parkinson’s Disease (PD) (Dubois and Pillon, 1996). Furthermore, the pattern of ERCC8 gene expression aligns with the observed global loss of gray matter volume in both the cerebral and cerebellar regions (Baez et al., 2013). Our finding of local genetic correlations corroborates previous findings and enriches our understanding of the specific gene’s role in association with PD and brain architecture.

Our analysis of ROI clusters and networks has revealed strong correlations and similar clustering patterns among homologous regions located in opposing cerebral hemispheres. This result is in line with the multiple previous studies that found a strong correlation between homologous regions (Alexander-Bloch et al., 2013; Biton et al., 2020; Fürtjes et al., 2023; Alexander-Bloch et al., 2019; Wen et al., 2016; Hofer et al., 2020; Akkose et al., 2021). The observed clusters and their corresponding correlations aligned closely with conventional brain functional lobe mapping, especially for CV and CSA. This consistency allows a coherent understanding of connection between morphological markers and anatomical functionalities. Furthermore, conditional dependence association within clusters was more pronounced within clusters than between clusters. The interconnectedness of these regions could stem from a variety of factors, including evolutionary pressures, susceptibility to similar environmental and developmental influences, shared genetic mechanisms and functions.

Divergences between genetic and phenotypic networks could offer a lens into the effects of non-genetic factors on brain structures. As an additive combination of genetic and environmental covariances, phenotypic covariance allows us to quantify the contributions of genetic and environmental factors between traits. Specifically, genetic factors explain approximately 34.1 % of the phenotypic covariance for cortical surface area (CSA), 27.0 % for cortical thickness (CTh), and 22.7 % for cortical volume (CV), suggesting a significant role for environmental factors. The similarity between phenotypic and genetic networks corroborates previous findings of high congruence between phenotypic and genetic morphometry (Fürtjes et al., 2023), and indicates the potential for similar trend between shared genetic and environmental factors.

The genetic association results help to better understand the functional characteristics of ROIs. The genetic CV clusters suggest similar genetic influences between frontal lobe structures and cingulate gyrus structures, as these two lobes are included in one cluster. Besides, the cluster results support that the temporal gyri are less similar with the parahippocampal gyrus and entorhinal cortex than the frontal lobes and the cingulate cortex in functional and genetic ties. The temporal gyri play a critical role in an array of functions including the processing of sounds, comprehension of language, and audio-visual emotional recognition (Bigler et al., 2007; Buckley et al., 1997). In contrast, the parahippocampal gyrus is chiefly responsible for encoding and recognition of environmental scenes (Luck et al., 2010; Köhler et al., 1998), while the entorhinal cortex acts as a central hub for memory, navigation, and time perception (Maass et al., 2015; Canto et al., 2008). On the other hand, the frontal lobes are instrumental in an array of cognitive functions encompassing reward processing, attention, short-term memory tasks, planning, and motivation (Chayer and Freedman, 2001), and the cingulate cortex is pivotal for emotion formation and processing, as well as learning and memory, playing a significant role in linking motivational outcomes to behavior (Devinsky et al., 1995). Our genetic correlation pattern from clustering results provides an intuitive reference for the functional distinctions of different brain regions, suggesting that temporal gyri bear less genetic resemblance to the parahippocampal gyrus and entorhinal cortex than the frontal lobes and the cingulate cortex. Our findings link genetic associations with biological similarity and address the genetic basis for ROIs mediated by cortical measures.

Based upon ROI network to group ROIs into different clusters, the composite-trait LDSC gains additional insights into the genetic architecture of brain imaging and could potentially reveal novel phenotypes for extensive GWAS analysis. In our study, we found seven significant pairs between the clustered phenotypes, i.e., composite traits, and 30 complex traits across different measurements and discerned hemisphere-centric influences. By integrating regions across hemispheres, we may identify patterns previously undetectable under singular hemispheric scrutiny. This shows the capabilities of composite-trait LDSC in simplifying the intricate genetic web and might uncover phenotypes for comprehensive GWAS study.

There are several limitations of this study. First, our analysis relies on LD Score Regression (LDSC), so our study is subject to the same limitations as the LDSC methodology. As HapMap SNPs are common variants in the human genome, this precludes the consideration of less common and rare variants. If the cumulative effects from less common and rare variants are substantial at the population level, this may bias the estimates and have an impact on how broadly applicable our findings about genetic correlations are across the whole genome. Second, our research had relatively small GWAS sample size and was conducted on the UK Biobank sample, so it could be questionable whether our study could consider the cohort effect and be generalized to the broader population. Our findings can be extended by incorporating larger and more diverse GWAS datasets and examining variations across datasets. Additionally, our results also depend on the selection of brain cortex parcellation, the reference panel of SUPERGNOVA and LAVA, and the methodology of correlation estimation.

In conclusion, our work introduces a general framework for investigating the genetic architecture of the cerebral cortex. With the application of network analysis to genetic correlation, our findings support and extend previous research. Additionally, we have developed composite-trait LDSC to jointly analyze the clustered ROIs by treating the traits within a cluster as a single phenotype. This methodology can be used in other studies to explore a wider range of research questions related to brain imaging and downstream analysis for genetic correlation.

Supplementary Material

1

2

3

Acknowledgments

This work was supported in part by NIH grants R01 GM134005, R01 AG068191, RF1 AG081413 and NSF grant DMS 1902903.

Data availability

The authors do not have permission to share data.

Fig. 1. Workflow of our analysis for genetic and phenotypic covariance. Details on the statistical framework are described in the Methods section.

Fig. 2. Estimated genetic correlations between regional CV. The ROIs are grouped by spectral clustering and are marked with different colors in the diagonal line. Asterisks highlight significant genetic correlations after Bonferroni correction for 2145 pairs. Colored bars on the left indicate the lobe mapping traits belongs to. Traits started with ‘lh_’ and ‘rh_’ denote regions in the left and right hemispheres, respectively.

Fig. 3. (A) Estimated proportions between CV of ROIs and 30 complex traits. Asterisks indicate at least one significantly correlated region between traits. The ROIs are grouped by hierarchical clustering applied to genetic correlations. (B) LocusZoom plots for Parkinson’s disease (upper) and educational attainment (lower) at the locus on chromosome 5 (58,834,852–60,880,207). The correlations between PD, EA and ROIs were significant in the highlighted region.

Fig. 4. Genetic network for regional CV. All the 66 nodes labeled with the trait names and the connecting lines adopt colors based on the specific clusters of the two ROIs they link. Clustering results obtained by Recursive Two-Way Ncut are marked in the outer edge of concentric circles, and the lobe mapping are marked in the center part of the concentric circle with five different colors.

Fig. 5. (A) Correlation between each composite-trait of CV and 30 complex traits. Asterisks highlight significant genetic correlations after Bonferroni correction. (B) Comparison of three cortical measures based on phenotype clusters. (C) Comparison of three cortical measures based on genetic clusters. 66 traits are listed for each measure, with colors corresponding to the clusters they belong to. Edges are drawn between nodes that correspond to the same ROIs.

Ethics statement

The BIG40 Open Data Server is provided by the Wellcome Centre for Integrative Neuroimaging, which is supported by center funding from the Wellcome Trust (203139/Z/16/Z). It follows the UK Biobank approval from the North West Multi-centre Research Ethics Committee (MREC) for obtaining and disseminating data and samples from participants. For more details, see https://www.ukbiobank.ac.uk/.

CRediT authorship contribution statement

Jiangnan Shen: Writing – review & editing, Writing – original draft, Visualization, Project administration, Methodology, Formal analysis, Conceptualization. Yiliang Zhang: Writing – review & editing, Visualization, Methodology, Formal analysis, Conceptualization. Zhaohan Zhu: Visualization, Formal analysis, Conceptualization. Youshu Cheng: Writing – review & editing, Formal analysis, Conceptualization. Biao Cai: Writing – review & editing, Supervision, Conceptualization. Yize Zhao: Writing – review & editing, Supervision, Conceptualization. Hongyu Zhao: Writing – review & editing, Supervision, Resources, Conceptualization.

Declaration of competing interest

The authors declare that they have no conflict of interest.

Supplementary materials

Supplementary material associated with this article can be found, in the online version, at doi:10.1016/j.neuroimage.2024.120739.
==== Refs
References

Akkose U , Kaya VO , Lindsey-Boltz L , , 2021. Comparative analyses of two primate species diverged by more than 60 million years show different rates but similar distribution of genome-wide UV repair events. BMC Genom. 22 , 1–13.
Alexander-Bloch A , Giedd JN , Bullmore E , 2013. Imaging structural co-variance between human brain regions. Nat. Rev. Neurosci 14 (5 ), 322–336.23531697
Alexander-Bloch AF , Mathias SR , Fox PT , , 2019. Human cortical thickness organized into genetically-determined communities across spatial resolutions. Cereb. Cortex 29 (1 ), 106–118. 10.1093/cercor/bhx309.29190330
Baez S , Couto B , Herrera E , , 2013. Tracking the cognitive, social, and neuroanatomical profile in early neurodegeneration: type III Cockayne syndrome. Front. Aging Neurosci 5 , 80.24324434
Banerjee O , El Ghaoui L , d’Aspremont A , 2008. Model selection through sparse maximum likelihood estimation for multivariate Gaussian or binary data. J. Mach. Learn. Res 9 , 485–516.
Bigler ED , Mortensen S , Neeley ES , , 2007. Superior temporal gyrus, language function, and autism. Dev. Neuropsychol 31 (2 ), 217–238.17488217
Biton A , Traut N , Poline JB , , 2020. Polygenic architecture of human neuroanatomical diversity. Cereb. Cortex 30 (4 ), 2307–2320.32109272
Buckley MJ , Gaffan D , Murray EA , 1997. Functional double dissociation between two inferior temporal cortical areas: perirhinal cortex versus middle temporal gyrus. J. Neurophysiol 77 (2 ), 587–598.9065832
Bulik-Sullivan B , Finucane HK , Anttila V , , 2015b. An atlas of genetic correlations across human diseases and traits. Nat. Genet 47 (11 ), 1236–1241. 10.1038/ng.3406.26414676
Bulik-Sullivan BK , Loh PR , Finucane HK , , 2015a. LD Score regression distinguishes confounding from polygenicity in genome-wide association studies. Nat. Genet 47 (3 ), 291–295. 10.1038/ng.3211.25642630
Cai T , Liu W , Luo X , 2011. A constrained ℓ 1 minimization approach to sparse precision matrix estimation. J. Am. Stat. Assoc 106 (494 ), 594–607. 10.1198/jasa.2011.tm10155.
Canto CB , Wouterlood FG , Witter MP , 2008. What does the anatomical organization of the entorhinal cortex tell us? Neural Plast. 2008 .
Chayer C , Freedman M , 2001. Frontal lobe functions. Curr. Neurol. Neurosci. Rep 1 (6 ), 547–552.11898568
de Leeuw CA , Mooij JM , Heskes T , , 2015. MAGMA: generalized gene-set analysis of GWAS data. PLoS Comput. Biol 11 (4 ), e1004219.25885710
Devinsky O , Morrell MJ , Vogt BA , 1995. Contributions of anterior cingulate cortex to behaviour. Brain 118 (1 ), 279–306.7895011
Dubois B , Pillon B , 1996. Cognitive deficits in Parkinson’s disease. J. Neurol 244 , 2–8.
Fan J , Feng Y , Wu Y , 2009. Network exploration via the adaptive LASSO and SCAD penalties. Ann. Appl. Stat 3 (2 ), 521. 10.1214/08-AOAS215SUPP.21643444
Finucane HK , Bulik-Sullivan B , Gusev A , , 2015. Partitioning heritability by functional annotation using genome-wide association summary statistics. Nat. Genet 47 (11 ), 1228–1235. 10.1038/ng.3404.26414678
Fürtjes AE , Arathimos R , Coleman JRI , , 2023. General dimensions of human brain morphometry inferred from genome-wide association data. Hum. Brain Mapp 44 (8 ), 3311–3323.36987996
Grotzinger AD , Mallard TT , Liu Z , , 2023. Multivariate genomic architecture of cortical thickness and surface area at multiple levels of analysis. Nat. Commun 14 (1 ), 946.36806290
Grotzinger AD , Rhemtulla M , de Vlaming R , , 2019. Genomic structural equation modelling provides insights into the multivariate genetic architecture of complex traits. Nat. Hum. Behav 3 (5 ), 513–525.30962613
Guo H , Li JJ , Lu Q , Hou L , 2021. Detecting local genetic correlations with scan statistics. Nat. Commun 12 (1 ), 1–3. 10.1038/s41467-021-22334-6.33397941
Haque IRI , Neubert J , 2020. Deep learning approaches to biomedical image segmentation. Inform. Med. Unlocked 18 , 100297.
Hofer E , Roshchupkin GV , Adams HHH , , 2020. Genetic correlations and genome-wide associations of cortical structure in general population samples of 22,824 adults. Nat. Commun 11 (1 ), 4796. 10.1038/s41467-020-18367-y.32963231
Köhler S , Black SE , Sinden M , , 1998. Memory impairments associated with hippocampal versus parahippocampal-gyrus atrophy: an MR volumetry study in Alzheimer’s disease. Neuropsychologia 36 (9 ), 901–914.9740363
Lauritzen SL , 1996. Graphical Models. Clarendon Press.
Lu Q , Li B , Ou D , Erlendsdottir M , Powles RL , Jiang T , , 2017. A powerful approach to estimating annotation-stratified genetic covariance via GWAS summary statistics. Am. J. Hum. Genet 101 (6 ), 939–964. 10.1101/114561.29220677
Luck D , Danion JM , Marrer C , , 2010. The right parahippocampal gyrus contributes to the formation and maintenance of bound information in working memory. Brain Cogn. 72 (2 ), 255–263.19837500
Maass A , Berron D , Libby LA , , 2015. Functional subregions of the human entorhinal cortex. Elife 4 , e06426.26052749
Mensen VT , Wierenga LM , van Dijk S , , 2017. Development of cortical thickness and surface area in autism spectrum disorder. NeuroImage 13 , 215–222. 10.1016/j.nicl.2016.12.003.28003960
Meyer M , Liem F , Hirsiger S , Jäncke L , Hänggi J , 2014. Cortical surface area and cortical thickness demonstrate differential structural asymmetry in auditory-related areas of the human cortex. Cereb. Cortex 24 (10 ), 2541–2552. 10.1093/cercor/bht094.23645712
Newman MEJ , 2006. Modularity and community structure in networks. Proc. Natl. Acad. Sci 103 (23 ), 8577–8582. 10.1073/pnas.0601602103.16723398
Neykov M , Ning Y , Liu JS , A unified theory of confidence regions and testing for high-dimensional estimating equations. 2018. 10.1214/18-STS661.
Oblong LM , Soheili-Nezhad S , Trevisan N , , 2024. Principal and independent genomic components of brain structure and function. Genes, Brain Behav. 23 (1 ), e12876.38225802
Palaniyappan L , Mallikarjun P , Joseph V , , 2011. Regional contraction of brain surface area involves three large-scale networks in schizophrenia. Schizophr. Res 129 (2–3 ), 163–168. 10.1016/j.schres.2011.03.020.21497489
Pan PL , Zhang Y , Liu Y , , 2017. Abnormalities of regional brain function in Parkinson’s disease: a meta-analysis of resting state functional magnetic resonance imaging studies. Sci. Rep 7 (1 ), 40469. 10.1038/srep40469.28079169
Rakic P , 2009. Evolution of the neocortex: a perspective from developmental biology. Nat. Rev. Neurosci 10 (10 ), 724–735.19763105
Ristić AJ , Daković M , Kerr M , Kovačević M , Parojčić A , Sokić D , 2015. Cortical thickness, surface area and folding in patients with psychogenic nonepileptic seizures. Epilepsy Res. 112 , 84–91. 10.1016/j.eplepsyres.2015.02.015.25847343
Shi HWB , Mancuso N , Spendlove S , Pasaniuc B , 2017. Local genetic correlation gives insights into the shared genetic architecture of complex traits. Am. J. Hum. Genet 101 (5 ), 737–751. 10.1016/j.ajhg.2017.09.022.29100087
Shi J , Malik J , 2000. Normalized cuts and image segmentation. IEEE Trans. Pattern. Anal. Mach. Intell 22 (8 ), 888–905. 10.1109/CVPR.1997.609407.
Smith SM , Douaud G , Chen W , , 2021. An expanded set of genome-wide association studies of brain imaging phenotypes in UK Biobank. Nat. Neurosci 24 , 737–745. 10.1038/s41593-021-00826-4.33875891
Soheili-Nezhad S , Beckmann CF , Sprooten E , 2021. Independent genomic sources of brain structure and function. BioRxiv, 2021.01. 06.425535.
Strike LT , Hansell NK , Couvy-Duchesne B , , 2019. Genetic complexity of cortical structure: differences in genetic and environmental factors influencing cortical surface area and thickness. Cereb. Cortex 29 (3 ), 952–962.29377989
Tirozzi A , Quiccione MS , Cerletti C , , 2023. A multi-trait association analysis of brain disorders and platelet traits identifies novel susceptibility loci for major depression, Alzheimer’s and Parkinson’s disease. Cells 12 (2 ), 245. 10.3390/cells12020245.36672180
Warrier V , Stauffer EM , Huang QQ , , 2023. Genetic insights into human cortical organization and development through genome-wide analyses of 2347 neuroimaging phenotypes. Nat. Genet 55 (9 ), 1483–1493.37592024
Wen W , Thalamuthu A , Mather KA , , 2016. Distinct genetic influences on cortical and subcortical brain structures. Sci. Rep 6 (1 ), 32760.27595976
Winkler AM , Greve DN , Bjuland KJ , , 2018. Joint analysis of cortical area and thickness as a replacement for the analysis of the volume of the cerebral cortex. Cereb. Cortex 28 (2 ), 738–749. 10.1093/cercor/bhx308.29190325
Winkler AM , Kochunov P , Blangero J , , 2010. Cortical thickness or grey matter volume? The importance of selecting the phenotype for imaging genetics studies. Neuroimage 53 (3 ), 1135–1146.20006715
Yao Z , Zhang Y , Lin L , , 2010. Abnormal cortical networks in mild cognitive impairment and Alzheimer’s disease. PLoS Comput. Biol 6 (11 ), e1001006 10.1371/journal.pcbi.1001006.21124954
Zhang C , Zhang Y , Zhang Y , , 2023. Benchmarking of local genetic correlation estimation methods using summary statistics from genome-wide association studies. Brief. Bioinform 24 (6 ), bbad407.37974509
Zhang Y , Lu Q , Ye Y , , 2021. SUPERGNOVA: local genetic correlation analysis reveals heterogeneous etiologic sharing of complex traits. Genome Biol. 22 , 262. 10.1186/s13059-021-02478-w.34493297
Zhao B , Luo T , Li T , , 2019. Genome-wide association analysis of 19,629 individuals identifies variants influencing regional brain volumes and refines their genetic co-architecture with cognitive and mental health traits. Nat. Genet 51 (11 ), 1637–1644.31676860
