Regional juxtacortical microstructural changes and gene expression provide insights into neurodegeneration in Alzheimer’s disease Authors: Junfang Zhang, Lipeng Sun, Xinyuan Yang, Lin Kang, Yingting Zheng, Qi Huang, Wei Xu, Yihui Guan, Jun Liu, Yulei Deng, Fang Xie, Xiaomeng Xu, Binyin Li Published: 2026 ABSTRACT AbstractWhite matter microstructural changes play a crucial role in cognitive decline in aging and neurodegenerative disorders including Alzheimer’s disease (AD). However, the processes underlying white matter microstructural changes and the molecular pathways leading to these changes in AD remain largely unknown. AD involves cortical and juxtacortical microstructural changes, with free water fraction (FWF) as a potential imaging marker. We measured FWF using diffusion magnetic resonance imaging in 68 juxtacortical regions of 153 cognitively normal controls and 194 patients with AD as evidenced by elevated amyloid PET. We estimated the expression of 15,633 genes in the same regions using transcriptomic data from the Allen Human Brain Atlas. The biological processes and cell types associated with the linked genes were evaluated. Mediation analysis was used to examine whether FWF mediates the association between APOE ε4 status and cognitive performance. Gene ontological analyses revealed that these genes were enriched for biological processes relating to lipid metabolic process, ensheathment of neurons, and synaptic signaling and were predominantly expressed in oligodendrocytes, GABAergic neurons, and pyramidal neurons from the hippocampus CA region. These ontological enrichment results were replicated in two additional datasets. Furthermore, mediation analyses revealed a domain-specific role of FWF in the association between APOE ε4 status and cognitive performance. Our findings provide mechanistic insights into regional juxtacortical microstructural changes in AD, particularly the processes involving lipid metabolism, offering potential therapeutic targets.Subject terms: Molecular neuroscience, Clinical genetics Introduction Alzheimer’s disease (AD) is a progressive neurodegenerative disorder and the main cause of dementia, characterized by senile plaques and neurofibrillary changes composed of amyloid beta (Aβ) and tau proteins [1]. In addition to the gray matter pathology, emerging evidence highlights the distinct white matter abnormalities in AD [2]. White matter microstructure is important in brain function, comprising myelinated axons that enable efficient neuronal signal transmission [3]. White matter microstructural changes play a crucial role in cognitive decline in aging and AD-related cognitive impairment [4]. Diffusion magnetic resonance imaging (dMRI) measures the diffusion of water molecules in the brain, enabling in vivo evaluation of alterations in white matter microstructure [5]. Conventional diffusion tensor imaging metrics, including fractional anisotropy and mean diffusivity, are limited by partial volume effects. The free water fraction (FWF) derived from a multicompartment diffusion model [6] serves as a sensitive microstructural marker that precedes morphological atrophy, and correlates with cerebrospinal fluid Aβ-42 and cognitive decline [7, 8]. Our previous study demonstrated a positive association between juxtacortical FWF change and cortical Aβ and tau deposition on positron emission tomography (PET) [9]. However, the processes underlying white matter microstructural changes and the molecular pathways leading to these changes remain largely unknown. Studies integrating free water measures with genetic data could help identify the underlying biological mechanisms associated with these metrics [4]. Recently, the analytical approach of combining neuroimaging with transcriptomic data has been methodologically established, providing insight into how disease-related alterations at the microscale architecture drive macroscale brain abnormalities in various neuropsychiatric disorders, including major depressive disorder [10], schizophrenia [11], Huntington’s disease [12], and Parkinson’s disease [13]. In this work, we used the spatial information about juxtacortical FWF changes and combined it with regional gene expression profiles from the Allen Human Brain Atlas (AHBA) transcriptomic data [14] to identify possible molecular mechanisms for regional microstructural abnormalities in AD participants. First, we described the juxtacortical FWF changes in AD using surface-based analysis. Second, we used partial least squares (PLS) regression to test whether juxtacortical FWF changes in AD correlates with specific patterns of gene expression. Third, we performed an enrichment analysis on the identified gene expression patterns linked to FWF changes to determine the biological and cell processes linked with the process of microstructural changes in AD. Finally, we investigated whether FWF mediates the association between APOE ε4 status and cognitive performance. These observations help us understand the pathophysiological mechanisms underlying white matter microstructural changes, and ultimately facilitate the development of interventions targeting these changes in AD. Methods Study population We included participants from the RJNeuroBank-Dementia (RJNB-D) cohort for the analyses. The RJNB-D was an observational cohort launched in 2021 to explore the diagnostic capabilities of various tools such as MRI, PET, blood biomarkers, and neuropsychological assessments in different causes of dementia, as well as their specific roles in AD development [15]. The cohort was registered on ClinicalTrials.gov (NCT05623124, Registration Date: 2022-10-23). The RJNB-D cohort was approved by the ethics committee at Ruijin Hospital, Shanghai Jiao Tong University School of Medicine, China (Approval No. 2021-337). All participants or their caregivers provided written informed consent. The researchers affirm that the study adhered to the ethical principles set forth in the 1964 Declaration of Helsinki and its subsequent revisions. In the discovery dataset, we enrolled 347 participants who had at least one visit with T1, dMRI and Aβ PET scans, and had been diagnosed as either cognitively normal (CN, n = 153), or with mild cognitive impairment (MCI) and dementia due to AD (AD continuum, n = 194) by evidence of brain Aβ pathologic abnormalities on PET imaging. Specifically, subjects diagnosed with MCI due to AD or AD dementia were grouped as AD continuum group (AD group for short). Participants with a primary diagnosis other than cognitively unimpaired controls, MCI or dementia due to AD were excluded from the study. The diagnostic criteria for dementia and MCI were based on the guidelines set by the National Institute on Aging-Alzheimer’s Association (NIA-AA, 2016) workgroups [16]. Cognitively normal participants with negative Aβ on PET were defined as controls. The analysis pipeline of this study is shown in Fig. 1. Study schema. This schematic illustrates the experimental design and sample distribution across the discovery and replication cohorts. A total of 194 AD participants with Aβ positivity on PET (AD continuum group) and 153 cognitively normal controls (CN group) from the RJNB-D cohort were enrolled in the discovery dataset. Participants from the discovery dataset with 1-year follow-up T1 and dMRI scans (n = 190) were included in the RJNB-D retest dataset (AD continuum group n = 107; CN n = 83). Additionally, 104 controls and 74 participants with MCI or mild AD from ADNI cohort were included in another replication dataset. (A) Regional juxtacortical FWF difference analysis. Regional FWF values were extracted using the Desikan-Killiany Atlas. AD-CN group differences across regions were then analyzed. (B) Gene expression profiles. Gene expression profiles from the Allen Human Brain Atlas in 34 regions (left hemisphere only) were averaged across six postmortem brains. PLS regression was then used to identify imaging-transcriptomic associations. Next, gene ontological analyses and expression-weighted cell-type enrichment analyses were performed on the gene list associated with the first component of PLS (PLS1). A subset of participants from the discovery dataset with MS-based plasma proteomic data available were enrolled to analyze the shared pathways between transcriptomic and proteomic data. AD Alzheimer’s disease, ADNI Alzheimer’s Disease Neuroimaging Initiative, Aβ amyloid beta, dMRI diffusion magnetic resonance imaging, FWF free water fraction, MCI mild cognitive impairment, MS mass spectrometry, PLS partial least squares, RJNB-D RJNeuroBank-Dementia cohort. Participants underwent comprehensive neuropsychological assessments including Mini-Mental State Examination (MMSE, Chinese Version) for global cognition [17], Auditory Verbal Learning Test (AVLT) 5-min for memory, verbal fluency test (VFT) and Boston naming test (BNT) for language, and shape-trail test (STT) parts A and B for executive function. Venous blood samples were used for apolipoprotein E (APOE) genotyping. MRI image acquisition and preprocessing MRI data were acquired on a 3T scanner (uMR 890, United Imaging Healthcare, Shanghai, China) with a dedicated 64-channel head coil. The protocol of three-dimensional (3D) T1, T2 fluid-attenuated inversion recovery (FLAIR) and dMRI scans has been reported in our previous studies [15]. The preprocessing pipelines for 3D T1, 3D T2 FLAIR and dMRI generally followed the protocols of an image-processing pipeline developed by UK Biobank [18]. The 3D T1-weighted images were acquired with a 3D fast spoiled gradient-echo sequence, with 0.75 mm isotropic voxels (field of view read: 240 mm, output images of 360 × 439× 480 slices, repetition time (TR) / inversion time (TI) = 7.5 ms /1100 ms, echo time (TE) = 3.4 ms, flip angle = 7 degree). The 3D FLAIR acquisition parameters were: TR = 6000 ms, TE = 408 ms, and TI = 1841ms, with 0.75 mm isotropic voxels. Diffusion images were acquired with an echo-planar imaging (EPI) sequence, and the parameters were: TR/TE = 3500/77.3 ms, voxel size = 1.5 × 1.5 × 1.5 mm, 100 slices, multiple band = 10, diffusion direction = 32/64, b = 1500/3000, phase encoding direction = PA. The original diffusion MRI data were corrected for susceptibility-induced distortion by top-up method [19], and then corrected for head motion and eddy currents using the Eddy tool in FSL (version 6.0) [20]. For structural MRI, cortical surface modelling was performed using the Freesurfer (version 7.1.1, https://surfer.nmr.mgh.harvard.edu/) software package through standard preprocessing steps, as previously described [15]. The T1-weighted images were processed through the standard ‘recon all’ pipeline, generating grey matter/white matter and grey matter/cerebrospinal fluid boundaries (white and pial surface, respectively). Surface reconstructions were visually inspected in all subjects and manual corrections were applied to resolve topological inaccuracies. Cortical thickness was measured as the Euclidean distance between the corresponding vertices of the white and pial surfaces. The FWF was measured quantitatively by neurite orientation dispersion and density imaging (NODDI) model derived from the multi-shell dMRI. NODDI estimation for FWF was analyzed using Accelerated Microstructure Imaging via Convex Optimization [6]. After co-registration with the reconstructed structural surface, the juxtacortical white matter FWF values were sampled at a depth of 1 mm beneath the border of the gray/white matter interface. FWF values were smoothed with a Gaussian kernel of 5 mm in full width at half maximum (FWHM) utilizing a surface-based smoothing procedure that averaged data across neighboring cortical locations. We extracted FWF values using Desikan-Killiany atlas (Freesurfer) which includes 68 regions of interest (ROIs). PET image acquisition and preprocessing The tracers of 18F Florbetapir and 18F-MK-6240 were used to visualize Aβ and tau accumulation separately with a PET/CT scanner (Biograph mCT Flow, Siemens, Erlangen, Germany) using previously described parameters [21]. Twenty-minute scans were conducted 50 min after intravenous injection of ~370 MBq (±10%) 18F-florbetapir and 90 min after intravenous injection of 185 MBq (±10%) 18F-MK-6240. The tracer of 18F-SynVesT-1 was used to visualize synaptic vesicle glycoprotein 2 A (SV2A) binding. 18F-SynVesT-1 scans were collected on a 3 T whole-body PET/MR scanner (uPMR 790, United Imaging, China). A 30-min static PET scan started at 60-min post-injection of 18F-SynVesT-1 (~3.7 MBq/kg body weight) [22]. After the acquisition, the PET images were reconstructed by a filtered back-projection algorithm with corrections for decay, normalization, dead time, photon attenuation, scatter and random coincidences [23]. The Aβ positivity of each participant was determined by visual assessment from two senior PET specialists (F. X. and Y. G.). The Aβ-PET, tau-PET, and SV2A-PET images were further processed using PETSurfer within FreeSurfer (version 7.1.1) to derive cortical standardized uptake value ratios (SUVR), with the whole cerebellum cortex used as the reference region, as it is relatively preserved from pathological involvement and provides a stable signal in Alzheimer’s disease PET studies. Structural images were used as the high-resolution template to run the partial volume correction (PVC) methods. The SUVRs of the cortex were finally sampled onto the brain surface and smoothed on the two-dimensional surface by a Gaussian kernel of 5 mm in FWHM for subsequent analyses. Estimation of regional gene expressions The neuroimaging transcriptomic analysis was processed according to the pipeline reported by Arnatkevic et al. [24, 25]. Briefly, regional microarray expression data were obtained from six post-mortem brains (1 female, aged 24–57 years) provided by the AHBA dataset (http://human.brain-map.org) [14]. The Abagen Toolbox was utilized to process and map the transcriptomic data to 34 parcellated cortical regions in the left hemisphere from the Desikan-Killiany atlas [26]. The processing details are in the Supplementary materials. Scaled expression profiles were then averaged across all donors to generate a single matrix, with rows representing brain regions and columns corresponding to the 15,633 retained genes. Because the AHBA data were available from six donors for the left hemisphere but only two donors for the right hemisphere, we limited our transcriptomic-imaging association analyses to the left hemisphere samples only, as these were considered more robust [13]. Regional changes in FWF and gene expression Partial least squares (PLS) regression was used to identify genes significantly associated with the FWF value differences between AD and CN group [27]. Gene expression data were used as predictor variables of regional changes in FWF in the PLS regression. The first component of the PLS (PLS1) was the linear combination of gene expression values that was most strongly correlated with regional changes in FWF. To account for spatial autocorrelation, permutation testing (10,000 times) based on spherical rotations of FWF map was used to test the null hypothesis that PLS1 explained no more covariance between FWF map and whole-genome expression than expected by chance [28]. Boot strapping (n = 1000) was used to estimate the variability of each gene’s PLS1, and the ratio of the weight of each gene to its bootstrap standard error was used to calculate the Z scores and rank the gene according to their contribution to PLS1. The set of genes significantly associated with regional changes in FWF (false discovery rate [FDR]-adjusted p < 0.05), either positive (PLS1+) or negative (PLS1–), constituted the FWF-related gene list. These two sets of genes (PLS1+ and PLS1-) were used for further enrichment analyses, as described below. Analysis of AD-related genes from GeneCards gene database We selected genes with a relevance score greater than 7 for Alzheimer’s disease from the GeneCards database (https://www.genecards.org/) as AD-related candidate genes. To explore the contribution of the AD-related genes in the PLS analysis, we first obtained the overlapped genes from the AD-related candidate gene list and 15,633 background genes. We then estimated the associations between overlapped gene expressions and group differences in FWF in the left hemisphere. We also used permutation testing (10,000 times) based on spherical rotations of the FWF map to account for spatial autocorrelation. Enrichment analysis We put all genes in PLS1+ and PLS1- into Metascape, a bioinformatics analysis platform (https://metascape.org) [29], to conduct enrichment analysis, investigating whether these genes map to common or relevant biological pathways. Gene ontology (GO) analysis and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis were conducted, focusing on biological processes (BPs), cellular components (CCs), and molecular functions (MFs). The threshold for significance was set at 5%, corrected for FDR. We also performed gene set enrichment analysis (GSEA) on the entire PLS1 gene set (ranked by Z-scores) using the fgsea R package, focusing on Gene Ontology (GO) Biological Processes. Pathways with an adjusted p < 0.05 and |normalized enrichment score (NES)| > 1 were considered significantly enriched. Brain cell-type enrichment analysis We performed expression-weighted cell-type enrichment (EWCE) analysis to determine whether PLS1+ and PLS1- gene sets have higher expression within a particular cell type [30]. We used human-derived cell type-specific gene sets from two independent sources to validate our results, including AHBA-seq dataset and DroNC-seq dataset [31, 32]. Briefly, AHBA-seq dataset containing single-cell transcription data from the middle temporal gyrus; human-derived DroNC-seq dataset containing data from five donors and 19,550 cells from both the hippocampus and the prefrontal cortex. Plasma proteomic measurement Plasma samples were processed to isolate the nanoparticle-associated protein corona, followed by denaturation, reduction, alkylation, and overnight digestion with trypsin. The resulting peptides were purified and reconstituted, following by the analysis on an Orbitrap Astral mass spectrometer (Thermo Fisher Scientific, USA). Chromatographic separation was performed on a Vanquish Neo UHPLC system. Acquisition parameters included high-resolution full mass spectrometry (MS) scans and Data-Independent Acquisition (DIA)-MS2 acquisition. DIA data were analyzed using DIA-NN (v1.8.1) and an unpublished plasma spectral library containing 126,661 peptide precursors mapping to 10,109 protein groups [33]. Full experimental details are provided in the Supplementary materials. Integrated enrichment analysis of transcriptomic and proteomic results For the proteomic data, enrichment was conducted using the “gsva” method within the gene set variation analysis (GSVA) framework [34]. The activation score of each pathway was calculated using GSVA, considering the identified proteins associated with the respective pathway. A higher score indicates greater activity of the corresponding pathway in that participant. To explore the shared biological mechanisms between transcriptomics and proteomics, we first identified significantly enriched pathways separately from the transcriptomic and proteomic datasets, and then obtained a set of overlapped pathways between two enriched pathways lists. Replication of transcriptional signatures To replicate the findings of transcriptional signatures from discovery dataset, we leveraged two datasets, namely the RJNB-D retest dataset and the Alzheimer’s Disease Neuroimaging Initiative cohort (ADNI). Specifically, we included a subset of participants (n = 190) from the discovery dataset who had one-year follow-up dMRI as a retest dataset (CN, n = 83; AD, n = 107; Table 1). We also included 178 participants from the ADNI cohort (CN n = 104; mild cognitive impairment [MCI] n = 64; mild AD n = 10; Supplementary Table S1). In the ADNI dataset, the participants with MCI and mild AD were grouped as the cognitive impairment group. Details of the ADNI dataset, including participant characteristics, imaging acquisition parameters and preprocessing procedures, are provided in the Supplementary Materials and Table S1. Table 1Participant demographics in the RJNB-D cohort.Discovery set (n = 347)Retest set (n = 190)CN(n = 153)AD(n = 194)pCN(n = 83)AD(n = 107)pAge67.2 ± 8.071.0 ± 7.6<0.00166.9 ± 7.373.3 ± 7.3<0.001Sex (M:F)54:9977:1170.4725:5841:660.28Education years12.2 ± 3.411.7 ± 4.00.2513.0 ± 3.311.2 ± 3.9<0.001MMSE28.5 ± 1.320.9 ± 6.1<0.00128.8 ± 1.321.3 ± 5.7<0.001APOE4 (non-carrier: carrier)95:3363:111<0.00159:2039:65<0.001Continuous variables are presented as means ± SDs.AD Alzheimer’s disease, CN cognitively normal controls, MMSE mini-mental state examination, RJNB-D RJNeuroBank-Dementia cohort. Participant demographics in the RJNB-D cohort. Continuous variables are presented as means ± SDs. AD Alzheimer’s disease, CN cognitively normal controls, MMSE mini-mental state examination, RJNB-D RJNeuroBank-Dementia cohort. In these two replication datasets, the group comparisons of FWF between controls and cognitive impairment participants were validated. The group comparison t-maps were obtained using linear regression model, regression out age, sex, and years of education. A multi-gene-list meta-analysis was conducted between the PLS1+ and PLS1- gene lists from discovery and replication datasets. All obtained pathways were thresholded for significance at 5%, corrected for FDR. Statistical analysis Statistical analyses and data visualization were performed using R version 4.2.0 and Python version 3.12. Continuous variables with a normal distribution were compared using a two-sample t-test. For variables with non-normal distributions, the Mann–Whitney test was applied. Categorical variables were compared using the Chi-squared test. Linear regression model was used for group comparison of FWF between AD and CN groups as follows: FWF = intercept + β1 × group + β2 × age + β3 × sex + β4 × education, with age, sex, and years of education as covariates. The primary contrast of interest was assessed using a two-tailed t-test comparing the estimated group parameter. The significance level was set at a two-tailed p-value < 0.05. The Benjamini-Hochberg’s FDR correction was used to control for false positive results caused by multiple comparisons. Pearson’s correlation coefficient (r) was used to test the linear relationship between regional FWF and synaptic density. A one-tailed test was performed with the directional hypothesis that higher free water fraction is associated with lower synaptic density.Mediation analysis was used to investigate whether FWF mediates the relationship between APOE ε4 status and cognitive performance, with age and sex as covariates. Results Sample Characteristics In the present study, a total of 347 participants including 194 participants with AD (mean age 71.0 ± 7.6 years; 117 women) and 153 CN participants (mean age 67.2 ± 8.0 years; 99 women) were recruited from the RJNB-D cohort as the discovery dataset. AD participants were older (p < 0.001) and had worse cognitive performance (p < 0.001) than CN participants. There were no significant differences in sex (p = 0.47) or years of education (p = 0.25) between AD and CN participants. A total of 190 participants (CN n = 83; AD n = 107) from the discovery dataset with one-year follow-up dMRI were included in the retest dataset for replication analyses. The characteristics of the study sample are given in Table 1. AD-related changes in FWF AD participants exhibited increased FWF values in cingulate, middle frontal, parietal, temporal, and lateral occipital lobes (t-value range from 2.06 to 5.86; p-value range from <0.001 to 0.04) than CN participants. The two-sided mean t-statistic values were extracted and the t-map of FWF is shown in Fig. 2A. Gene expression profiles related to FWF changes in AD group in the discovery dataset. A FWF changes in AD compared to controls (unthresholded). B A weighted gene expression map of regional PLS1 scores in the left hemisphere (unthresholded). C A scatterplot of regional PLS1 scores (a weighted sum of 15,633 gene expression scores) and regional changes in FWF demonstrating a negative correlation; each data point represents one of 34 cortical regions. D The distribution of bootstrapped gene weights on PLS1. FDR was used to correct for multiple comparisons. E AD-related genes from GeneCards in the adult human brain negatively (i.e., KLK6) and positively (i.e., CHI3L1) correlated with FWF changes in AD. AD Alzheimer’s disease, CHI3L1 chitinase-3-like protein 1, CN cognitively normal controls, FDR false discovery rate, KLK6 kallikrein-related peptidase 6, PLS partial least squares. Relating FWF changes to variation in gene expression patterns in AD We used PLS regression to identify the pattern of gene expression that correlated with FWF changes in AD. The PLS1 explained the largest proportion of variance in FWF changes (47.51%). The distribution of the PLS1 weighted map reflects an anterior-posterior gradient of gene expression (Fig. 2B). Notably, we found that the PLS1 weighted gene expression map was spatially correlated with the FWF group comparison t-map (Pearson’s r = −0.69, p < 0.001; Fig. 2C). We then used PLS1 to rank and select significantly weighted genes, yielding a set of 189 significantly up-weighted and 690 significantly down-weighted genes (all FDR-adjusted p < 0.005, Fig. 2D). In total, 879 genes constituted the regional changes in FWF gene list in participants with AD (Supplementary Table S2). To further determine the relationship between gene expression profiles of previously reported AD-related genes and regional changes in FWF, we searched the disease term “Alzheimer” and selected genes with a relevance score greater than 7 from the GeneCards database, identifying 377 AD-related genes. Of them, 282 genes overlapped with the 15,633 background genes, and 24 genes further overlapped with significantly weighted genes in PLS1. When assessing the Pearson correlation between these 24 genes and t-value of FWF changes, KLK6 gene showed the strongest negative correlation (r = −0.601, p = 0.00017), and CHI3L1 gene showed the strongest positive correlation (r = 0.631, p < 0.0001; Fig. 2E). Enrichment pathways associated with FWF changes in AD The enrichment results are displayed in Fig. 3. In the PLS1- (down-weighted) genes related GO analyses, several terms associated with “lipid metabolic process” and “ensheathment of neurons” were enriched. Enrichment for biological process (BP) is involved in “membrane lipid metabolic process” (GO:0006643, –logP = 7.22), “supramolecular fiber organization” (GO:0097435, –logP = 5.85), and “ensheathment of neurons” (GO:0007272, –logP = 5.75). Regarding to cellular component (CC), the term “actin cytoskeleton” (GO:0015629, –logP = 6.73), “intermediate filament” (GO:0005882, –logP = 6.29), “myelin sheath” (GO:0043209, –logP = 6.05), and “apical plasma membrane” (GO:0016324, –logP = 5.52) showed significant enrichment. Compared to BP and CC, molecular function (MF) terms showed less statistically significant results, with the term “endopeptidase activity” (GO:0004175, –logP = 4.26) enriched. There were several significant KEGG pathways involving lipid metabolism, such as “Ether lipid metabolism” (hsa00565, –logP = 4.83). Functional enrichment of gene transcripts in the discovery dataset. A Ontology terms for PLS1- genes. Top 5 terms were shown for each category. B Metascape enrichment network visualization showing the intra-cluster and inter-cluster similarities of enriched terms for PLS1- genes. C Ontology terms for PLS1+ genes. Top 5 terms were shown for each category. D Metascape enrichment network visualization showing the intra-cluster and inter-cluster similarities of enriched terms for PLS1+ genes. BP biological process, CCs cellular components, KEGG Kyoto Encyclopedia of Genes and Genomes, MFs molecular functions, PLS1 first partial least squares component. In the PLS1+ (up-weighted) genes related results, several terms associated with “synaptic signaling” were enriched. Enrichment for BP is involved in “synaptic signaling” (GO:0099536, –logP = 6.53), “behavior” (GO:0007610, –logP = 6.26), “modulation of chemical synaptic transmission” (GO:0050804, –logP = 5.99), “regulation of trans-synaptic signaling” (GO:0099177, –logP = 5.98), and “trans-synaptic signaling” (GO:0099537, –logP = 5.42). Terms such as “postsynaptic density” (GO:0014069, –logP = 8.93) and “dendrite” (GO:0030425, –logP = 8.44) were enriched for CC. For MF, terms such as “phosphotransferase activity” (GO:0016773, –logP = 7.30), “actin binding” (GO:0003779, –logP = 6.84), and “kinase activity” (GO:0016301, –logP = 6.78) were enriched. There were no significant results for KEGG pathways. Additionally, we performed GSEA on the entire PLS1 gene set and identified 44 significant deregulatory pathways (Supplementary Table S3). The downregulated biological processes included lipid metabolism related pathways, whereas upregulated pathways were mainly involved in synapse related processes (Supplementary Fig. S1). Considering that dysfunction in synaptic signaling may be accompanied by synaptic structural abnormalities, we performed Pearson correlation analyses between regional FWF and synaptic density measured using SV2A PET in participants with AD (n = 94). As shown in Supplementary Fig. S2A, higher FWF values were significantly correlated with lower SV2A SUVRs in bankssts, entorhinal and temporal cortex (e.g., right bankssts: r = −0.53, FDR-adjusted p < 0.001; left bankssts: r = −0.24, FDR-adjusted p = 0.04). These associations remained consistent after additional adjustment for cortical thickness (see Supplementary Fig. S2B). Transcriptional signatures for canonical cell types We evaluated whether the top up- and down-weighted genes associated with FWF changes in AD showed enriched expression in specific cell types (Fig. 4). We found that down-weighted genes were significantly associated with oligodendrocytes (ODC) based on both AHBA and DroNC-seq gene lists (FDR-adjusted p < 0.01), while the up-weighted genes were associated with GABAergic neurons (GABA) based on AHBA gene list (FDR-adjusted p < 0.01) and associated with pyramidal neurons from the hippocampus CA region (exCA) based on DroNC-seq gene list (FDR-adjusted p < 0.01). Expression-weighted cell-type enrichment analyses in the discovery dataset using the AHBA and DroNc single-cell transcription dataset. A Top 30% down-weighted and up-weighted genes using AHBA-seq gene lists. B Top 30% down-weighted and up-weighted genes using DroNc-seq gene lists. Cell types in which the target gene lists are significantly enriched are marked with asterisks. ** FDR-adjusted p < 0.01. ABHA Allen Human Brain Atlas, ASC astrocytes, DG granule neurons from the hippocampus dentate gyrus region, END endothelial cells, exCA pyramidal neurons from the hippocampus CA region, GABA GABAergic neurons, Glu Glutamatergic, GluPF glutamatergic neurons from the prefrontal cortex, MG microglia, NSC neuronal stem cells, ODC oligodendrocytes, OPC oligodendrocyte precursor cells. Reproducibility of transcriptomic profile in AD related changes in FWF We replicated the FWF changes associated with transcriptomic profiles in the RJNB-D retest dataset and in an independent external cohort (the ADNI dataset). In the RJNB-D retest dataset (CN n = 83; AD n = 107), the participants in AD group were older (p < 0.001), had less years of education (p < 0.001), and had worse cognitive performance (p < 0.001) than the participants in CN group. There was no significant difference in sex between the two groups (Table 1). In the ADNI dataset (CN n = 104; MCI n = 64; mild AD n = 10; Table S1), the cognitive impairment group (including MCI and mild AD) had a lower proportion of females (p < 0.001) and exhibited poorer cognitive performance (p < 0.001) compared with the CN group. There were no significant differences in age or years of education between the cognitive impairment group and the CN group. In the RJNB-D retest dataset, we found that 85 PLS1+ genes, and 307 PLS1− genes were significantly overexpressed, yielding a total of 392 genes associated with regional FWF differences. The gene lists in the RJNB-D discovery and retest datasets were overlapped (Supplementary Fig. S3; odds ratio [OR] = 39.376, p < 0.0001, Fisher’s exact test). After correcting for enrichment pathways, several ontological terms including “lipid biosynthetic process” and “ensheathment” survived. The overlapping ontology terms between discovery and replication cohorts concentrated on “ensheathment of neurons”, “axon ensheathment”, “myelin sheath”, “apical plasma membrane”, “Ether lipid metabolism”, “dendrite”, “phosphotransferase activity”, and “kinase activity” (Supplementary Figs. S4 and S5). Significantly overlapping ontology support the generalized relationship between gene expression and the AD-related changes in FWF. Highly overlapped cell types including ODC are shown in the Supplementary Fig. S6. In the ADNI dataset, we identified 300 genes associated with regional FWF differences that overlapped with those in the discovery dataset (OR = 3.524, p < 0.0001, Fisher’s exact test; Supplementary Fig. S7). After correcting for enrichment pathways, we found the overlapping ontological terms between the discovery dataset and the ADNI cohort including “ensheathment of neurons”, “axon ensheathment”, and “postsynaptic density” (Supplementary Fig. S8). Overlapped cell types between discovery dataset and ADNI dataset are shown in the Supplementary Fig. S9. Proteomic signatures associated with entorhinal subcortical FWF changes in AD To examine whether transcriptomics-derived pathways were recapitulated at the proteomic level, we included a subset of participants with MS-based proteomics data available from the discovery dataset (total n = 113; AD n = 81, CN n = 32). As shown in Fig. 2A, the entorhinal subcortical FWF values showed significant differences between AD and CN groups (left, t = 5.65, p < 0.001; right, t = 5.66, p < 0.001). We therefore performed Pearson correlation analyses between the mean of left and right entorhinal subcortical FWF value and the abundance of each protein. The proteins significantly correlated with entorhinal subcortical FWF (p < 0.05) were identified (n = 302), and these proteins were further analyzed by functional enrichment analysis. As shown in Supplementary Fig. S10, the enrichment terms involved “mitochondrial membrane” (GO:0031966, –logP = 22.1), “endomembrane system organization” (GO:0010256, –logP = 12.74), “Golgi membrane” (GO:0000139, –logP = 11.75), “early endosome” (GO:0005769, –logP = 11.75), and “focal adhesion” (GO:0005925, –logP = 11.45). Integrated enrichment analysis of transcriptomic and proteomic results Among the significantly enriched pathways common to both the transcriptomics and proteomics datasets, the top 25 pathways ranked by enrichment significance are selected for visualization (Supplementary Fig. S11). Several pathways related to intracellular trafficking (e.g., early endosome and lysosome) and lipid metabolism (including glycerolipid, lipid, and phospholipid metabolic processes) were enriched. The Spearman correlation of the significantly enriched pathways with clinical features are shown in Supplementary Fig. S11. Notably, several neuropsychological scores (e.g., VFT, STT-A and STT-B) were significantly correlated with the activation scores of pathways related to glycerolipid and phospholipid metabolic process (FDR-adjusted p values: 0.006–0.027). FWF mediates the relationship between APOE ε4 status and cognitive performance in a domain-specific manner In the discovery dataset, a total of 302 participants (AD n = 174; CN n = 128) with APOE genotyping information were included for the mediation analysis. The APOE ε4 carriers showed significantly higher entorhinal subcortical FWF than the APOE ε4 non-carriers (APOE ε4 carriers: 0.0792 ± 0.0706; APOE ε4 non-carriers: 0.056 ± 0.059; t = −2.91, p = 0.0039). Mediation analysis revealed a domain-specific role of FWF in the association between APOE ε4 status and cognitive performance (Fig. 5). Specifically, the entorhinal subcortical FWF completely mediated the effects of APOE ε4 on language (VFT) and executive measures (STT-A), accounting for approximately 66.0–67.8% of the total effect. Partial mediation was observed for global cognition (MMSE), naming performance (BNT), and executive function (STT-B), with mediated proportions ranging from 26.6 to 42.7%. In contrast, only a modest mediating effect was detected for episodic memory performance (AVLT-5, mediated effect = 8.4%). Domain-specific mediation effect of APOE ε4 on cognition by FWF. Mediation models illustrating the role of FWF in the association between APOE ε4 status and cognitive performance across six cognitive measures. Entorhinal subcortical FWF completely mediated APOE ε4-related effects on language (VFT) and executive measures (STT-A), partially mediated global cognition (MMSE), naming performance (BNT), and STT-B, and showed minimal mediation for episodic memory (AVLT-5). Standardized regression coefficients (β) and p values are shown for direct and indirect effects. All models were adjusted for age and sex. AVLT Auditory Verbal Learning Test, BNT Boston naming test, FWF free water fraction, MMSE Mini-Mental State Examination, STT-A shape-trail test part-A, STT-B shape-trail test part-B, VFT verbal fluency test. Discussion Using diffusion MRI to define the maps of AD-related changes in microstructural integrity in humans, we found that juxtacortical FWF changes were related to gene transcripts in AD. Specifically, FWF changes were partially overlapped with prior defined AD-related genes and were ontologically enriched for “membrane lipid metabolic process”, “ensheathment of neurons”, and “synaptic signaling” related terms that were validated in two replication datasets. In addition, we mapped FWF changes related genes in AD to biological processes associated with oligodendrocytes, GABAergic neurons, and pyramidal neurons from the hippocampus CA region. Furthermore, we revealed that FWF mediates the relationship between APOE ε4 status and cognitive performance in a domain-specific manner. These findings reveal the pathophysiological processes underlying FWF changes in AD, and bridge the gap between transcriptomics and neuroimaging, promoting an integrative understanding of AD. Recently, human imaging genetics has emerged as a powerful strategy for understanding the molecular basis of brain organization [11, 35]. In the present work, we first validated our previous study that AD participants had increased juxtacortical FWF compared to controls [9]. Then, to combine the regional pattern of FWF changes with potential biological processes, we used the PLS method and found that significantly weighted genes in the first PLS component may have roles in the pathogenesis of AD. For example, in the present study, the KLK6 gene showed significant negative correlation with FWF changes in AD, while CHI3L1 gene showed positive correlation with FWF changes. Human kallikrein-related peptidase 6 (KLK6) is highly expressed in the central nervous system [36], which has been involved in the regulation of myelin sheath volume and degradation of a β-amyloid peptide [37]. In patients with elevated phosphorylated tau in cerebrospinal fluid (CSF), CSF KLK6 levels were significantly increased, suggesting its potential role in tau pathology in AD [38]. Similar to KLK6, elevated levels of chitinase-3-like protein 1 (CHI3L1, also known as YKL-40) in CSF are detected in AD participants, and can predict the rate of cognitive decline in AD [39]. As a neuroinflammatory marker, CHI3L1 is mainly expressed by astrocytes and significantly elevated in human cortical tissues in AD [40]. CHI3L1 was reported as a modulator of glial phagocytic activation and AD pathogenesis [40, 41]. In the GO enrichment analysis, we found those PLS1− (down-weighted) genes were significantly enriched for lipid metabolic process and ensheathment of neurons. Dysregulation in membrane lipids was considered as an important pathological mechanism in AD [42]. In the list of top risk genes for the sporadic late onset AD, APOE, TREM2, APOJ, PICALM, ABCA1, and ABCA7 are directly involved in lipid trafficking or metabolism [43]. Among them, APOE ε4 is the strongest genetic risk factor for late-onset AD [44]. The APOE plays an important role in cholesterol transport and lipid metabolism, which also has a role in myelin maintenance [45]. Cholesterol is essential for membrane repair, synapse formation, and myelination. Previous studies have reported that free water was higher in APOE ε4 carriers on white matter microstructure [3, 46]. APOE ε4 isoform is less efficient at transporting lipids, which can impair myelination and the maintenance of white matter microstructure [47, 48]. Lipid metabolism represents a potential therapeutic target for AD which requires further investigation. On the other hand, those PLS1+ (up-weighted) genes were significantly enriched for synaptic signaling. Synapse loss and damage are central features of AD, which contribute to the onset and progression of its clinical and pathophysiological features [49]. Pathogenic Aβ and tau species can drive synaptic deficits in AD [49]. Our previous study using 18F-SynVesT-1 PET have demonstrated that synaptic density loss correlated with impaired functional and structural connectivity in AD [22]. In this study, we identified the significant associations between higher FWF values and lower SV2A SUVRs in bankssts, entorhinal and temporal cortex. These findings suggested that dysfunction in synaptic signaling might be accompanied by synaptic structural abnormalities, which warrants further investigation. We found that the distribution of gene expression could be partly explained by variation in the distribution of cell types. Specifically, the down-weighted genes were significantly associated with oligodendrocytes, while the up-weighted genes were linked to GABAergic neurons and pyramidal neurons from the hippocampus CA region. Oligodendrocytes are responsible for the production and maintenance of the myelin sheath. Myelin abnormalities have been detected in several different mouse models of amyloidosis [50, 51]. A previous study reported that cholesterol gene expression and localization were dysregulated in oligodendrocytes, leading to impaired myelination [48]. Pharmacologically promoting cholesterol efflux could increase myelination in vitro and in vivo, and improve cognition in APOE ε4/ε4 targeted replacement mice [48]. GABAergic neurons are one of the major types of neurons responsible for synaptic transmission in the brain. Emerging evidence suggested that the alterations of GABAergic circuits might contribute to AD by disrupting the overall network function [52]. In addition, hippocampal CA1 pyramidal neurons play a key role in the circuits mediating the connection between the entorhinal cortex and the hippocampus [53]. AD-linked synaptic loss was reported in hippocampal CA1 pyramidal neurons in mouse and human [54]. Given that FWF changes were ontologically enriched for lipid metabolism-related pathways, and considering APOE ε4 as a major genetic determinant of lipid dysregulation in AD [47], we further revealed the domain-specific mediation effects of APOE ε4 on cognition by FWF. Specifically, the executive and language-related functions were largely or entirely mediated by FWF, whereas episodic memory showed minimal mediation. These findings demonstrated FWF plays a significant role in the relationship between APOE ε4 and cognition in AD, offering potential therapeutic targets. Our study has limitations. First, our analysis mainly used gene expression data from AHBA derived from donors unaffected by neuropsychiatric disease and are not Alzheimer’s disease specific. Therefore, spatial expression profiles are based on unaffected individuals and conclusions can only apply to this intrinsic variability, not to changes in gene expression that occur in AD. Future studies using brain-wide gene expression data of postmortem brains from age- and sex-matched aging individuals and patients with AD are required to replicate our findings. Second, we restricted our analysis to left-hemisphere data from the AHBA because right-hemisphere data were only available from two participants. Hence, the relationship between genes and AD-related changes in FWF does not represent the entire brain’s condition. Future work with gene expression data from both hemispheres in a larger number of donors could help to investigate the differences between hemispheres. Third, there were significant differences in age between controls and AD participants in our study. Therefore, age was included as a covariate in all the relevant analyses. Fourth, our cell-type specificity analysis was inferred from bulk transcriptomic data, which may obscure cell-type specific signals. In future studies, integration with single-cell data would allow more precise attribution. Fifth, although PLS analysis identifies latent components that maximize covariance between imaging and transcriptomic data, these components do not imply causal relationships. In summary, we used juxtacortical FWF data and transcriptomic analyses to examine the relationship between microstructural changes and gene expression in AD. Our findings provide insights into the connections between genes, biological pathways, cell classes, and in vivo microstructural changes, supporting the idea that the lipid metabolic process and synaptic dysfunction are implicated in the pathophysiology of AD. These findings may facilitate the identification of pathogenetic biomarkers and therapeutic targets for AD in the future. Supplementary information Supplementary_information_ADNI_Acknowledgement_List Supplementary_materials Supplementary Table S2 Supplementary Table S3