Introduction
Alzheimer’s disease (AD) represents the predominant type of dementia, marked by significant neurodegeneration, alongside symptoms such as memory impairment and cognitive degradation. It is characterized by pathological processes involving β-amyloid deposition and tau pathology (Braak, Alafuzoff, Arzberger, Kretzschmar, & Del Tredici, Reference Braak, Alafuzoff, Arzberger, Kretzschmar and Del Tredici2006; Villemagne et al., Reference Villemagne, Burnham, Bourgeat, Brown, Ellis, Salvado and Australian Imaging2013). However, the neurobiological mechanisms of AD have not yet been elucidated.
The foundational organizational structure of the human brain encompasses a hierarchical architecture, facilitating the encoding and integration of information from sensory perception to cognitive processes (M. Mesulam, Reference Mesulam2012; M. M. Mesulam, Reference Mesulam1998). This hierarchical architecture serves as an optical organization for information transferring within the human brain (Wang et al., Reference Wang, Li, Zhao, Chen, Yang, Yao and Liu2024), and uncovering it could provide insight into how the integrated nature of neural processing can give new perspectives on understanding the roots of function and dysfunction (Xue et al., Reference Xue, Guo, Zhu, Liang, Xu, Ma and Liu2023). To study the hierarchical nature of the brain, researchers have proposed the concept of gradients, which are low-dimensional manifold representations of high-dimensional brain network features obtained by diffusion map embedding algorithm (Margulies et al., Reference Margulies, Ghosh, Goulas, Falkiewicz, Huntenburg, Langs and Smallwood2016). Different gradients represent different axes of variance of brain network features, along which cortical regions are arranged in a spatial continuum (Huntenburg, Bazin, & Margulies, Reference Huntenburg, Bazin and Margulies2018; Margulies et al., Reference Margulies, Ghosh, Goulas, Falkiewicz, Huntenburg, Langs and Smallwood2016). The principal gradient explains the maximum variation in the brain network. An increasing array of studies on gradients have sought to examine the hierarchy of brain organization. Notably, the principal functional network gradient delineates a spectrum spanning unimodal (sensory–motor) regions to transmodal (default mode) regions in healthy adults (Huntenburg et al., Reference Huntenburg, Bazin and Margulies2018; Margulies et al., Reference Margulies, Ghosh, Goulas, Falkiewicz, Huntenburg, Langs and Smallwood2016). This spectrum is consistent with the hierarchical architecture of the brain, and many diseases involve abnormal changes in this spectrum (Dong et al., Reference Dong, Yao, Wang, Hong, Genon, Xin and Luo2023; Hong et al., Reference Hong, de Wael, Bethlehem, Lariviere, Paquola, Valk and Bernhardt2019; Xia et al., Reference Xia, Liu, Mechelli, Sun, Ma, Wang and He2022). A previous study has found that AD is associated with an increase in functional network gradients in unimodal regions and a decrease in transmodal regions (Zheng, Zhao, Yang, & Guo, Reference Zheng, Zhao, Yang and Guo2024b).
In parallel, structural network gradients delineate the spatial continuum in morphology or microstructure, offering a pathway to unveil the underlying mechanisms of neurodevelopment (Burt et al., Reference Burt, Demirtas, Eckner, Navejar, Ji, Martin and Murray2018; Huntenburg et al., Reference Huntenburg, Bazin and Margulies2018). To construct the structural network gradients, we first need to construct the structural network. Traditionally, the construction of the structural covariance network based on structural magnetic resonance imaging (sMRI) has been carried out almost exclusively at the group level, ignoring differences at the individual level (Modinos et al., Reference Modinos, Vercammen, Mechelli, Knegtering, McGuire and Aleman2009). A recent study suggests that the structural characteristics of the human cortex can be more accurately assessed by integrating multiple sMRI-derived morphometric indices for each specific region. This could include, for example, combining cortical thickness with gray matter volume (Sabuncu et al., Reference Sabuncu, Ge, Holmes, Smoller, Buckner and Fischl2016). On this foundation, a morphological similarity network (MSN) was proposed, which combined multiple sMRI-derived cortical characteristics, including cortical thickness (CT), cortical volume (CV), cortical area (CA), gaussian curvature (GC), and mean curvature (MC). MSN quantifies the similarity between cortical regions, and instead of measuring inter-regional correlations between individual cortical morphological features of a group of subjects, it measures the correlation between multiple cortical morphological features of each individual, thus allowing to obtain structural brain networks at the individual level (M. J. Cai et al., Reference Cai, Ma, Wang, Zhao, Zhang, Wang and Liu2023; Seidlitz et al., Reference Seidlitz, Vása, Shinn, Romero-Garcia, Whitaker, Vértes and Consortium2018). MSN has a complex topological organization, including modularity. MSN modules recapitulate known classes of cortical cytoarchitecture, and the edges of the MSN strongly correlate with regional gene co-expression in the human brain. In rhesus monkeys, brain regions with higher morphometric similarity were found to be more likely to form axonal tracts with each other (Seidlitz et al., Reference Seidlitz, Vása, Shinn, Romero-Garcia, Whitaker, Vértes and Consortium2018). In addition, MSN could be considered a neuroimaging phenotype correlating structural alterations in the brain with transcriptomic features, thus capturing cellular, molecular, and functional features of the brain. However, MSN itself characterizes the brain from a discrete perspective, while MSN gradient analysis can capture how various brain regions are assembled from a more integrated perspective. Therefore, mapping MSN gradients can reproduce the hierarchical properties of cortical organization (Li et al., Reference Li, Seidlitz, Suckling, Fan, Ji, Meng and Liao2021; Seidlitz et al., Reference Seidlitz, Nadig, Liu, Bethlehem, Vértes, Morgan and Raznahan2020). In healthy subjects, the first MSN gradient manifests in the sensory and motor cortex at both extremes and the first MSN gradient is strongly correlated with cortical organization properties (Yang et al., Reference Yang, Wagstyl, Meng, Zhao, Li, Zhong and Liao2021). In patients with major depression, the first MSN gradient decreases in sensorimotor regions and increases in visual-related regions compared to normal controls (Xue et al., Reference Xue, Guo, Zhu, Liang, Xu, Ma and Liu2023). However, it is still unclear whether and how the first MSN gradient changes in AD.
As a genetically driven disease, AD has a heritability of 50–70 percent (Gatz et al., Reference Gatz, Pedersen, Berg, Johansson, Johansson, Mortimer and Ahlbom1997; Pedersen, Gatz, Berg, & Johansson, Reference Pedersen, Gatz, Berg and Johansson2004). Emerging evidence indicates that genetic factors are crucial in the formation and organization of the human brain connectome (Arnatkeviciute et al., Reference Arnatkeviciute, Fulcher, Oldham, Tiego, Paquola, Gerring and Fornito2021; Thompson et al., Reference Thompson, Cannon, Narr, van Erp, Poutanen, Huttunen and Toga2001). The integration of microarray-based gene expression data from post-mortem adult brains, such as those provided by the Allen Human Brain Atlas (AHBA) (Hawrylycz et al., Reference Hawrylycz, Lein, Guillozet-Bongaarts, Shen, Ng, Miller and Jones2012), with neuroimaging techniques have led to the development of the emerging field of neuroimaging transcriptomics. This approach offers a feasible method for identifying the genetic mechanisms that influence various neuroimaging phenotypes. It also establishes a crucial link between the detailed patterns of gene expression at the micro level and the large-scale organization of brain networks (Fornito, Arnatkevičiūtė, & Fulcher, Reference Fornito, Arnatkevičiūtė and Fulcher2019; F. Liu, Tian, Li, Li, & Zhuo, Reference Liu, Tian, Li, Li and Zhuo2019; J. Liu, Xia, Wang, Liao, & He, Reference Liu, Xia, Wang, Liao and He2020; Richiardi et al., Reference Richiardi, Altmann, Milazzo, Chang, Chakravarty, Banaschewski and Greicius2015; Xia et al., Reference Xia, Liu, Mechelli, Sun, Ma, Wang and He2022; Xue et al., Reference Xue, Guo, Zhu, Liang, Xu, Ma and Liu2023; Zhu et al., Reference Zhu, Yuan, Gao, Xu, Xue, Zhu and Yu2021). Specifically, prior work identified a correlation between alterations in functional connectome gradients in patients with AD and gene expression profiles (Zheng, Zhao, et al., Reference Zheng, Zhao, Yang and Guo2024b). Therefore, elucidating the association between alterations in MSN gradients and genetic expression in AD will contribute to a deeper comprehension of the molecular genetic underpinnings driving changes in MSN gradients in AD.
To fill these gaps, we applied both the sMRI and gene expression datasets to investigate the first MSN gradient in patients with AD, and we analyzed their correlations with transcriptome profiles and cognitive functions. Specifically, our hypotheses were that: (i) the first MSN gradients alter in AD; (ii) the first MSN gradients in AD are associated with cognition; and (iii) the regions with altered first MSN gradient in AD are related to transcriptome profiles enriched in specific biological function.
Methods
Subjects
The dataset employed in our research was obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI, https://adni.loni.usc.edu/). Specifically, the discovery dataset was derived from stage 2 (ADNI-2), while the validation dataset was obtained from stage 3 (ADNI-3). The discovery dataset included 133 AD and 157 normal controls (NCs); the validation dataset included 53 AD and 61 NCs. Baseline sMRI, demographic information, and cognitive measures of all subjects were downloaded in de-identified form. Cognitive measures included the Mini-Mental State Examination (MMSE) and four composite cognitive scores for memory function (ADNI-MEM), executive function (ADNI-EF), visuospatial function (ADNI-VS), and language function (ADNI-LAN) (Choi et al., Reference Choi, Mukherjee, Gibbons, Sanders, Jones, Tommet and Crane2020; Crane et al., Reference Crane, Carle, Gibbons, Insel, Mackin, Gross and Alzheimer’s2012; Gibbons et al., Reference Gibbons, Carle, Mackin, Harvey, Mukherjee and Insel2012). In both the discovery and validation datasets, there were no significant differences in age, sex, and number of years of education between AD and NCs (Table 1). However, to account for the potential confounding effects of these variables of no interest, they were used as covariates in the between-group comparison.
Demographics for all participants
T1-weighted image acquisition and preprocessing
T1-weighted (T1w) images were acquired for all subjects using a 3 Tesla scanner. Detailed scan information of T1w images is available on the ADNI website (http://adni.loni.usc.edu/methods/mri-tool/mri-analysis/). The recon-all pre-processing pipeline in FreeSurfer version 7.4.1 software (https://surfer.nmr.mgh.harvard.edu/) was used for T1w image preprocessing, including skull stripping, tissue segmentation, surface reconstruction, metric reconstruction, and spherical normalization parameter estimation (Xue et al., Reference Xue, Guo, Zhu, Liang, Xu, Ma and Liu2023). After that, individual cortical surfaces were reconstructed. A total of four subjects failed to have their T1w images segmented (discovery dataset: 1 patient with AD and 1 NC; validation dataset: 1 patient with AD and 1 NC) after the T1w data pre-processing was completed. Therefore, only the subjects shown in Table 1 were used in this study.
Construction of MSN gradients
Cortical morphological metrics were extracted using the DK-1533 template (Romero-Garcia, Atienza, Clemmensen, & Cantero, Reference Romero-Garcia, Atienza, Clemmensen and Cantero2012; Yang et al., Reference Yang, Wagstyl, Meng, Zhao, Li, Zhong and Liao2021). The DK-1533 template derives from the Desikan–Killian 68 (DK-68) template (Desikan et al., Reference Desikan, Ségonne, Fischl, Quinn, Dickerson, Blacker and Killiany2006) and consists of 1533 parcellations with an approximate size of 1 cm2 for each parcellation. The DK-1533 template was transformed to each subject’s individual surface space, and five cortical morphological metrics including cortical thickness (CT), cortical volume (CV), cortical area (CA), gaussian curvature (GC), and mean curvature (MC) were extracted for each parcellation. The morphological metrics of each subject were standardized across brain regions (z-score). For each subject, a symmetric matrix of size 1533 × 1533 can be generated by calculating Pearson’s correlation coefficients between the morphological metrics of each pair among the 1533 brain regions. This resulting matrix is referred to as the MSN.
The MSN gradients were generated using the BrainSpace toolbox (Vos de Wael et al., Reference Vos de Wael, Benkarim, Paquola, Lariviere, Royer, Tavakol and Bernhardt2020). The main steps are as follows:
-
1. Referring to previous studies (Hong et al., Reference Hong, de Wael, Bethlehem, Lariviere, Paquola, Valk and Bernhardt2019; Margulies et al., Reference Margulies, Ghosh, Goulas, Falkiewicz, Huntenburg, Langs and Smallwood2016; Vos de Wael et al., Reference Vos de Wael, Benkarim, Paquola, Lariviere, Royer, Tavakol and Bernhardt2020; Xia et al., Reference Xia, Liu, Mechelli, Sun, Ma, Wang and He2022; Xue et al., Reference Xue, Guo, Zhu, Liang, Xu, Ma and Liu2023; Yang et al., Reference Yang, Wagstyl, Meng, Zhao, Li, Zhong and Liao2021), only the top 10% of elements per row in the MSN were retained to compute a cosine similarity matrix that captured similarity in morphological similarity profiles. The similarity matrix underwent an additional transformation into a normalized matrix of angles.
-
2. We adopted the diffusion map embedding algorithm (Coifman et al., Reference Coifman, Lafon, Lee, Maggioni, Nadler, Warner and Zucker2005), which is a nonlinear manifold reduction approach, to generate the descending gradient components that explain the MSN variance.
-
3. The gradient components for each subject were aligned to the template deriving from an average MSN with the usage of the Procrustes rotation (Hong et al., Reference Hong, de Wael, Bethlehem, Lariviere, Paquola, Valk and Bernhardt2019; Xia et al., Reference Xia, Liu, Mechelli, Sun, Ma, Wang and He2022; Xue et al., Reference Xue, Guo, Zhu, Liang, Xu, Ma and Liu2023).
The first MSN gradient captures the maximum variance in MSN and is along a spectrum from the sensory cortex to the motor cortex (Yang et al., Reference Yang, Wagstyl, Meng, Zhao, Li, Zhong and Liao2021), therefore, we focused only on the first MSN gradient.
Comparison of the first MSN gradients in AD and NCs
Before making comparisons between groups, we first harmonized for site effects using ComBat (Fortin et al., Reference Fortin, Cullen, Sheline, Taylor, Aselcioglu, Cook and Shinohara2018; Fortin et al., Reference Fortin, Parker, Tunç, Watanabe, Elliott, Ruparel and Shinohara2017; Johnson, Li, & Rabinovic, Reference Johnson, Li and Rabinovic2007). A general linear model (GLM) was conducted to assess the regional MSN gradient differences between AD and NCs, in which age, sex, and educational attainment were covariates. Furthermore, GLM was leveraged to assess the MSN gradient differences between AD and NCs within two other atlases which were referred to as the Yeo functional network atlas (Yeo et al., Reference Yeo, Krienen, Sepulcre, Sabuncu, Lashkari, Hollinshead and Buckner2011) and the Von Economo cytoarchitectural class atlas (Economo, Koskinas, & Triarhou, Reference Economo, Koskinas and Triarhou2008), respectively. Age, sex, and educational attainment were also served as covariates. Yeo atlas consists of seven networks, namely somatomotor network (SOM), visual network (VIS), dorsal attention network (DAN), salience network (SAL), limbic network (LIM), frontoparietal network (FPN), and default mode network (DMN). The Von Economo atlas consists of seven classes, namely primary sensory cortex (Prim sens), secondary sensory cortex (Sec sens), association cortex 1 (Asso1), association cortex 2 (Asso2), limbic regions (Limbic), primary motor cortex (Prim motor), and insula regions (Insula). Kolmogorov–Smirnov test was performed to investigate the differences in the distribution of MSN gradients based on global and cytoarchitectural classes between AD and NCs. P-values <0.05 were regarded as statistical significance. P-values were adjusted for false discovery rate (FDR) multiple comparisons (Benjamini & Hochberg, Reference Benjamini and Hochberg2018).
Correlation analysis between MSN graph metrics and gradients in AD
We used the Brain Connectivity Toolbox (Rubinov & Sporns, Reference Rubinov and Sporns2010) to calculate two graph metrics, the clustering coefficient, which measures MSN separation, and the average path length, which measures MSN integration (Liao, Vasilakos, & He, Reference Liao, Vasilakos and He2017; Watts & Strogatz, Reference Watts and Strogatz1998) (Supplementary Methods). GLM was conducted to assess the AD-NCs differences in the clustering coefficient and the average path length adjusting age, sex, and educational attainment. Graph metrics with statistical significance were further used incorrelation analysis with MSN gradients.
Correlation analysis between MSN gradients and cognition
The MSN gradients reflect the fundamental properties of cortical organization, and to capture the relationship between the first MSN gradients at the level of Von Economo cytoarchitectural class and cognition in AD, partial correlation analysis was conducted between the first MSN gradients at the level of Von Economo cytoarchitectural class and cognition, with age, sex, and education level as covariates. The gradient within each cytoarchitectural class was computed as the mean of gradients across all brain areas belonging to that class.
Prediction of cognitive scores
We employed support vector regression, a common supervised machine learning technique, to evaluate the predictive ability of the first MSN gradients for cognition in AD. MSN gradients were the predictor variables, and cognition was the response variable, respectively. We built a nested 5-fold cross-validation framework, and Pearson’s correlation coefficient between observed and predicted cognition was used to assess the predictive performance of the model, and its significance was determined by a permutation test (1 000 times). For comparison with the MSN gradients, we also evaluated the predictive power of the regional MSN (the sum of correlation coefficients for each region) (Supplementary Methods).
Transcriptomic data and preprocessing
We used the brain transcriptomic datasets from AHBA (Hawrylycz et al., Reference Hawrylycz, Lein, Guillozet-Bongaarts, Shen, Ng, Miller and Jones2012), derived from brain samples of six adult post-mortem donors (Supplementary Table 1). Following the standard workflows (Arnatkeviciute, Fulcher, & Fornito, Reference Arnatkeviciute, Fulcher and Fornito2019), we mapped the transcriptomic data to the DK-1533 template using abagen toolbox (Markello et al., Reference Markello, Arnatkeviciute, Poline, Fulcher, Fornito and Misic2021), which resulted in a 1533 × 15631 matrix. In brief, the standard workflows include probe re-annotations, filtering, and selection, mapping the transcriptomic data to the DK-1533 template, and data normalization (Supplementary Methods).
Transcription-neuroimaging association analysis
To investigate the relationship between gene expression profiles and AD-NCs differences in principal MSN gradients (M. Cai et al., Reference Cai, Ji, Zhao, Xue, Sun, Wang and Liu2024; Ma et al., Reference Ma, Xue, Wang, Cai, Wang, Li and Liu2023), partial least squares (PLS) regression analysis was conducted where transcriptional profiles were the independent variables and unthresholded t-statistics of the AD-NCs differences in principal MSN gradients were the dependent variables (Supplementary Methods). We focused only on the first PLS component (PLS1), which was a linear combination of transcriptional profiles and was responsible for the most variation in unthresholded t-statistics. Permutation tests (10 000 times) were used to assess the significance of the explanatory variance of PLS1. The significance of the correlation coefficient between PLS1 and unthresholded t-statistics was also established using the permutation test (10 000 times). Bootstrapping (10 000 times) was leveraged to assess the variance of the genes contributing to PLS1, the ratio of the weight for each gene to its standard error (z score) was the corrected weight.
PLS1 genes+ had normalized positive PLS1 weights which indicated high gene expression corresponding to increased alterations in the first MSN gradient in AD compared to NCs, whereas, PLS1 genes- had normalized negative PLS1 weights which indicated low gene expression corresponding to decreased alterations in the first MSN gradient in AD compared to NCs. Gene sets having positive weights (PLS1 genes+) or negative weights (PLS1 genes-) (pFDR < 0.001) (Shen et al., Reference Shen, Rolls, Cheng, Kang, Dong, Xie and Feng2022; Zheng, Xiao, Zhao, Yang, & Guo, Reference Zheng, Xiao, Zhao, Yang and Guo2024a; Zheng, Zhao, et al., Reference Zheng, Zhao, Yang and Guo2024b) were retained for enrichment analyses including pathways and Gene Ontology (GO) terms which include biological process (BP), molecular function (MF), and cellular component (CC). Genes enrichment analyses were performed in the ToppGene portal (https://toppgene.cchmc.org/). P-values <0.05 were regarded as statistically significant for all enrichment analyses (FDR correction).
Validation analysis
To validate our main results, we did the following: (i) validating the reproducibility of dominant MSN gradients in a validation dataset; and (ii) adjusting total intracranial volume (TIV) when comparing the first MSN gradient between AD and NCs.
Results
Between-group differences in the first MSN gradient
MSN variance explained by the first gradient was not significantly different between AD and NCs (AD, 12.8% ± 0.2%; NCs, 12.8% ± 0.3%, p = 0.24, Figure 3A, left). The first MSN gradient exhibited anchoring points at the opposing poles of the motor and sensory cortices, while the association cortex occupied an intermediary position along its continuum (Figure 1A and 1B).
Differences in the first MSN gradient between AD and NCs. A: The first MSN gradient mapping in AD; B: The first MSN gradient mapping in NCs; C: T-statistic map of the differences in the first MSN gradient between AD and NCs; D: brain regions with significant differences in the first MSN gradient between AD and NCs. In C and D, cool colors indicate regions where the gradient is decreased in AD compared to NCs, while warm colors indicate regions where the gradient is increased in AD.
Comparisons across different brain regions indicated that in AD patients, the areas showing a significant increase in the first MSN gradient values were predominantly located in the cingulate gyrus, while the regions exhibiting a significant decrease were primarily situated in the temporal lobe (Figure 1D, Supplementary Table 2). Additionally, we utilized two established frameworks for categorizing cortical areas, to broaden the scope of our investigations into the function and cytoarchitecture of cerebral cortex. In patients with AD, at the level of the Von Economo cytoarchitecture class, the first MSN gradient values demonstrate a significant increase in the primary and secondary sensory cortices, while exhibiting a significant decrease in the association cortex 1 (Figure 2A, Supplementary Table 3). In patients with AD, at the level of the Yeo functional networks, the first MSN gradient values demonstrate a significant increase in VIS and SAL, while exhibiting a significant decrease in DAN, LIM, and DMN (Figure 2B, Supplementary Table 4).
Differences in the first MSN gradient between AD and NCs at the level of Von Economo classes and Yeo functional networks. A: The first MSN gradient differences between AD and NCs at the Von Economo class level; B: The first MSN gradient differences between AD and NCs at the Yeo functional network level. *: p < 0.05, **: p < 0.01, ***: p < 0.001.
The distributions of the first MSN gradient in AD and NCs were significantly different both at the global level (Supplementary Figure 1A, and Supplementary Table 5) and at the level of the Von Economo cytoarchitecture class (Supplementary Figure 1B, and Supplementary Table 5).
Association between the first MSN gradient and MSN topology in AD
In addition to the first gradient explained ratio, we also compared two other global gradient metrics, the first gradient range, and the first gradient variance. As shown in Figure 3A, middle and right, the first gradient range (AD, 0.238 ± 0.014; NCs, 0.241 ± 0.012, p = 0.03) and gradient variance (AD, 0.068 ± 0.004; NCs, 0.069 ± 0.003, p = 0.04) of AD were significantly smaller than those of NCs.
The first MSN gradient and MSN topological properties. A: Differences in global metrics of the first MSN gradient between AD and NCs; B: Differences in MSN topological properties between AD and NCs, and the associations between MSN topological properties and the first MSN gradient. *: p < 0.05, **: p < 0.01, ns: not significant.
The MSN clustering coefficient (AD, 0.558 ± 0.004; NCs, 0.557 ± 0.004, p = 0.003) and average path length (AD, 2.862 ± 0.031; NCs, 2.852 ± 0.028, p = 0.021) for AD were all significantly larger than NCs (Figure 3B, left and middle). In patients with AD, the average path length and the explained ratio of the first gradient were significantly positively correlated (r = 0.46, p < 0.001, Figure 3B, right).
Association between the first MSN gradient and cognition in AD
As illustrated in Figure 4, the first MSN gradient of association cortex 1 was significantly positively correlated with executive function (r = 0.204, p = 0.02), language function (r = 0.194, p = 0.03), memory function (r = 0.220, p = 0.01), and MMSE (r = 0.223, p = 0.01) in AD patients. While the primary sensory cortex was significantly negatively correlated with executive function (r = −0.177, p = 0.04) and language function (r = −0.213, p = 0.02), the secondary sensory cortex was significantly negatively correlated with language function (r = −0.186, p = 0.03).
Correlation between the first MSN gradient and cognitive scores in AD. *: p < 0.05.
The predictive ability of the first MSN gradient on cognition in AD
As shown in Supplementary Table 6, the first MSN gradient was able to significantly predict memory function, executive function, language function, and visual function in AD. Compared with the first MSN gradient, the regional MSN values significantly predicted only executive function, language function, and visual function, and the regional MSN values were less predictive of these cognitive scores than the first MSN gradient was of these cognitive scores.
Gene expression profiles associated with the first MSN gradient alterations in AD
The PLS1 explained 14.8% (p < 0.0001) of the variation in AD-associated alterations in the first MSN gradient. Following this, our attention was exclusively channeled towards analyzing the PLS1 in further assessments. The PLS1 displayed distinct transcriptomic patterns, characterized by high expression predominantly in occipital lobe regions and low expression primarily in temporal lobe regions (Figure 5A, right). We also found a significant positive correlation between the PLS1 and between-group t-statistics of the first MSN gradient (r = 0.385, p < 0.0001) (Figure 5A, middle). Gene set enrichment analysis revealed that PLS1 genes+ were mainly enriched in biological functions and pathways related to gene expression (Figure 5B), whereas PLS1 genes- were mainly enriched in biological processes such as synaptic signaling, molecular functions such as oxidoreductase activity, cellular components such as synapses, and pathways related to metabolism and neural development (Figure 5C).
Relationship between regionally altered first MSN gradient in AD and gene expression, alongside enrichment results of PLS1 genes. A: Regional mapping of PLS1 scores, T-map of the first MSN gradient differences between AD and NCs, and correlation between PLS1 scores and the T-map. B: Top enrichment terms for PLS1 genes+. C: Top enrichment terms for PLS1 genes-.
Validation results
In the validation set, we obtained the first MSN gradient pattern similar to that of the discovery set, where the motor and sensory cortices were situated at opposite ends, with the association cortex positioned in the middle (Supplementary Figure 2A and Supplementary Figure 2B). Between-group comparisons of the first MSN gradient in the validation set revealed significant brain regions similar to those identified in the discovery set (Supplementary Figure 2D). Specifically, Pearson’s correlation coefficient between the t-statistics obtained from region-based AD-NCs comparisons in the validation set and those from the discovery set was 0.5 (Supplementary Figure 2E). Additionally, when conducting between-group comparisons at the regional level, Pearson’s correlation coefficient between the t-statistics obtained with and without TIV as a covariate exceeded 0.9 (Supplementary Figure 2F).
Discussion
In this investigation, we have, for the first time, presented the dysfunction of the morphological similarity network hierarchy in AD, and elucidated its correlation with cognitive function and gene expression profiling. We found that the two extremes of the first MSN gradient axis are the sensory cortex and the motor cortex, with the association cortex positioned in the middle. In patients with AD, the abnormal first MSN gradient pattern is primarily characterized by increases in the primary sensory cortex and secondary sensory cortex and decreases in association cortex 1. This pattern is similar to the functional connectome gradient abnormalities observed in AD (Zheng, Zhao, et al., Reference Zheng, Zhao, Yang and Guo2024b). In patients with AD, the first MSN gradient in association cortex 1 is positively correlated with functions such as executive functioning, language, and memory, whereas the first gradient in the sensory cortex is negatively correlated with these functions. Furthermore, the first MSN gradient is a significant predictor of cognitive function in AD patients. We also discovered that the genes associated with alterations in the first MSN gradient are predominantly involved in synaptic signaling and metabolic activity. These findings provide new insights into the mechanisms of structural brain network impairment in AD patients, enhancing our understanding of the neurobiological processes responsible for impaired cognition in patients with AD, and offering a new dimensional structural biomarker for AD.
The first MSN gradient alterations in AD
Compared to group-level structural covariance networks, one of the advantages of MSN is that it integrates multiple morphological features of the cerebral cortex to construct individual-level structural networks (Seidlitz et al., Reference Seidlitz, Vása, Shinn, Romero-Garcia, Whitaker, Vértes and Consortium2018). We applied a neuroimaging phenotype that reflects the continuous spectrum of cortical structural networks – the MSN gradients – to research the differences between AD and NCs. We found that, in both NCs and AD, the sensory cortex and motor cortex are located at the two ends of the first MSN gradient axis, which is consistent with previous studies. They have similar endpoints (Xue et al., Reference Xue, Guo, Zhu, Liang, Xu, Ma and Liu2023; Yang et al., Reference Yang, Wagstyl, Meng, Zhao, Li, Zhong and Liao2021). This differs from the functional network gradient axis, which has unimodal (sensory–motor) regions and transmodal (default mode) regions at its two ends, respectively (Margulies et al., Reference Margulies, Ghosh, Goulas, Falkiewicz, Huntenburg, Langs and Smallwood2016). This indicates that MSN gradients and functional network gradients characterize different hierarchical aspects of the brain from different perspectives. The first MSN gradient pattern is closely related to various fundamental properties of the cortex, such as gene expression (Yang et al., Reference Yang, Wagstyl, Meng, Zhao, Li, Zhong and Liao2021). Compared to healthy controls, AD patients exhibited smaller variance and a narrower range of the first MSN gradient, indicating a smaller disparity in the morphological connectivity patterns between the sensory and motor cortices in AD patients. The regions exhibiting increased first MSN gradient in patients with AD are primarily localized within the frontal and cingulate gyrus, whereas decreased regions are predominantly situated within the temporal lobe. Research has indicated that the medial temporal lobe is affected even before cognitive decline becomes apparent in AD (Burggren & Brown, Reference Burggren and Brown2014). Our findings further support the crucial role of the temporal lobe in the pathogenesis of AD from a network hierarchy perspective.
On the Yeo functional network, this is reflected by an increase in the first MSN gradient in the VIS and a decrease in the first MSN gradient in networks such as DMN and LIM. In the class of Von Economo, there is a notable increase in the first MSN gradient primarily in the primary and secondary sensory cortices, while there is a decrease in the first MSN gradient in the association cortex 1. The VIS receives external signals and filters relevant information before transmitting the signals to higher-order systems to induce appropriate responses. Functional alterations in the VIS occur even before the onset of cognitive deficits in AD and worsen as the disease progresses (Devanand et al., Reference Devanand, Liu, Tabert, Pradhaban, Cuasay, Bell and Pelton2008; Lin et al., Reference Lin, Metter, O’Brien, Resnick, Zonderman and Ferrucci2011; Verghese, Wang, Lipton, Holtzer, & Xue, Reference Verghese, Wang, Lipton, Holtzer and Xue2007). Disruption of cortical connectivity in AD leads to insufficient bottom-up audiovisual integration (Festa, Katz, Ott, Tremont, & Heindel, Reference Festa, Katz, Ott, Tremont and Heindel2017). Functional impairments in the DMN are also associated with the pathophysiology of memory deficits and various cognitive impairments in AD (Ibrahim et al., Reference Ibrahim, Suppiah, Ibrahim, Mohamad, Hassan, Nasser and Saripan2021; Mandal, Banerjee, Tripathi, & Sharma, Reference Mandal, Banerjee, Tripathi and Sharma2018). The pattern of changes in the first MSN gradient observed in this study parallels the alterations in the functional network gradients in AD (Zheng, Zhao, et al., Reference Zheng, Zhao, Yang and Guo2024b). This finding supports previous observations of deficits in early sensory and visual processing and highlights the importance of the DMN in AD, from the perspective of continuity of cortical spatial organization.
The relationship between the first MSN gradient and cognition
We also found that association cortex 1 is positively correlated with functions such as memory, language, and executive functions, whereas the primary sensory cortex is negatively correlated with these functions. Furthermore, the first MSN gradient can significantly predict memory, language, executive, and visual functions in AD patients. This suggests that alterations in the first MSN gradient may be an endophenotype of AD pathology. The study also discovered that both integration and segregation of MSN are increased in AD patients, and there is a notable association between these topological changes of MSN and the first MSN gradient.
Transcription-neuroimaging association analysis
Our transcription-neuroimaging association analysis revealed a connection between AD-related changes in the first MSN gradient and gene expression enriched in synaptic signaling and metabolic activity. This suggests that synaptic signaling and metabolic activity may be potential mechanisms underlying the alterations in the principal MSN gradient in AD. Synaptic dysfunction plays a crucial role in the pathology of AD (Selkoe, Reference Selkoe2002). Trans-synaptic signaling is one of the most fundamental biological processes, involving a series of critical molecular functions, including synapse formation and the regulation of synaptic plasticity (de Wit & Ghosh, Reference de Wit and Ghosh2016; Fossati et al., Reference Fossati, Assendorp, Gemin, Colasse, Dingli, Arras and Charrier2019). Disruption of synaptic signaling across numerous important pathways affects synapse formation and stability, which is pivotal in the pathology of AD (Selkoe, Reference Selkoe2002). Metabolic abnormalities are key characteristics of the neuronal cellular milieu in AD, including impaired energy metabolism, disrupted lipid metabolism, and decreased overall metabolic capacity. These abnormalities have the potential to cause instability in neural network activity, impair neuroplasticity, and disrupt the hierarchical organization of the brain, thereby undermining the harmonious functioning of the brain (Xu, Liu, Qin, & Wang, Reference Xu, Liu, Qin and Wang2023). Genes associated with changes in functional network gradients are enriched in ion transmembrane transporter activity (Zheng, Zhao, et al., Reference Zheng, Zhao, Yang and Guo2024b), while genes related to changes in the first MSN gradients are not enriched in this biological function. We cautiously infer that the reason for this difference may be that functional network gradients and the first MSN gradients explore the hierarchical nature of the brain from different aspects, respectively. Compared to NCs, the alterations in functional gradients and the first MSN gradients at the region level are distinct in AD. Moreover, neuroimaging-transcriptome association analysis is based on PLS regression between T-values of region-level differences in gradients and gene expression values, which may lead to different results in gene enrichment. Given the current challenges in the measurement of the expression of regional genes in the living brain, the findings of this study offer valuable clues for understanding the relationship between microscopic biological events and the macrostructural changes observed in AD.
Limitations
The current study has several limitations. First, although the data analyzed in this article comes from public databases, the participants’ race or ethnicity information is unavailable, and we are unsure whether there are any racial or ethnic differences, which is a limitation of this study. Second, indicators of AD-related pathology, such as beta-amyloid and tau proteins, were not included. Future research should incorporate these indicators to assess how they affect the MSN gradient associated with AD. Third, gene expression datasets are from donors without a diagnosis of AD. Additional transcriptomic data from patients with AD was essential to link MSN gradients to transcriptomic profiles. To elucidate the genetics underlying AD-associated gradient changes, genome-wide association studies are needed in future studies to examine the synergistic roles of related genes in the pathogenesis of AD, given its complex polygenic nature. Finally, besides genetic factors, environmental influences are also crucial. Subsequent investigations ought to prioritize exploring exposome-neuroimaging relationships to advance our comprehension of the etiological contributors influencing MSN gradients in AD (F. Liu et al., Reference Liu, Xu, Guo, Qin, Liang, Schumann and Yu2023).
In conclusion, the findings of our study support the hypotheses that changes in the first MSN gradient are present in patients with AD, and these changes are correlated with cognitive scores and transcriptomic profiles. Additionally, our study showed that the genes involved in AD showed an enrichment in pathways that were relevant to neurobiology. Collectively, our findings shed new light on the altered structural coordination in AD and had the potential to provide a novel endophenotype for further exploration of complicated mechanisms underlying AD.
Supplementary material
The supplementary material for this article can be found at http://doi.org/10.1017/S0033291725000091.
Data availability statement
The sMRI and cognition data were obtained from ADNI (https://adni.loni.usc.edu/), and gene expression data were obtained from AHBA (http://human.brain-map.org/static/download).
Acknowledgments
We thank all the investigators of ADNI for their contributions to data collection and sharing. Data collection and sharing for this project were funded by the Alzheimer’s Disease Neuroimaging Initiative (ADNI) (National Institutes of Health Grant U01 AG024904) and DOD ADNI (Department of Defense award number W81XWH-12-2-0012). ADNI is funded by the National Institute on Aging, the National Institute of Biomedical Imaging and Bioengineering, and through generous contributions from the following: AbbVie, Alzheimer’s Association; Alzheimer’s Drug Discovery Foundation; Araclon Biotech; BioClinica, Inc.; Biogen; Bristol-Myers Squibb Company; CereSpir, Inc.; Cogstate; Eisai Inc.; Elan Pharmaceuticals, Inc.; Eli Lilly and Company; EuroImmun; F. Hoffmann-La Roche Ltd and its affiliated company Genentech, Inc.; Fujirebio; GE Healthcare; IXICO Ltd.; Janssen Alzheimer Immunotherapy Research & Development, LLC.; Johnson & Johnson Pharmaceutical Research & Development LLC.; Lumosity; Lundbeck; Merck & Co., Inc.; Meso Scale Diagnostics, LLC.; NeuroRx Research; Neurotrack Technologies; Novartis Pharmaceuticals Corporation; Pfizer Inc.; Piramal Imaging; Servier; Takeda Pharmaceutical Company; and Transition Therapeutics. The Canadian Institutes of Health Research is providing funding to support ADNI clinical sites in Canada. Private sector contributions are facilitated by the Foundation for the National Institutes of Health (www.fnih.org). The grantee organization is the Northern California Institute for Research and Education, and the study is coordinated by the Alzheimer’s Therapeutic Research Institute at the University of Southern California. ADNI data are disseminated by the Laboratory for Neuro Imaging at the University of Southern California. A complete listing of ADNI investigators can be found at: https://adni.loni.usc.edu/wp-content/uploads/how_to_apply/ADNI_Acknowledgement_List.pdf.
This work was supported by the National Natural Science Foundation of China (Grant No. 12071124 [to Shuixia Guo]); the National Natural Science Foundation of China (Grant No. 12471273 [to Shuixia Guo]); and Hunan Provincial Natural Science Foundation of China (Grant No. 2023JJ40461 [to Wei Zhao]).
Author contribution
Data used in preparation of this article were obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database (adni.loni.usc.edu). As such, the investigators within the ADNI contributed to the design and implementation of the ADNI and/or provided data but did not participate in the analysis or writing of this report. A complete listing of ADNI investigators can be found at: https://adni.loni.usc.edu/wp-content/uploads/how_to_apply/ADNI_Acknowledgement_List.pdf.
Competing interest
The authors report no biomedical financial interests or potential conflicts of interest.
Open access
