Transcriptomic signatures of synaptic loss in Alzheimer’s disease Authors: Lipeng Sun, Xinyuan Yang, Xiaomeng Xu, Lin Kang, Yingting Zheng, Qi Huang, Wei Xu, Yihui Guan, Yiwen Shen, Jun Liu, Yulei Deng, Shu Liu, Junfang Zhang, Fang Xie, Binyin Li Published: 2026 ABSTRACT SummarySynaptic loss is a major pathological cause of cognitive impairment in Alzheimer’s disease (AD). We integrated in vivo synaptic density imaging using synaptic vesicle glycoprotein 2A positron emission tomography (PET) from a prospective AD cohort with brain transcriptomic data. Partial least squares analysis identified 1,233 genes associated with synaptic loss, enriched for synaptic organization, Tau phosphorylation, cytoskeletal integrity, and ubiquitin-mediated protein degradation. Cell-type enrichment showed downregulation in glutamatergic and GABAergic neurons and upregulation in oligodendrocytes and endothelial cells. Stratification by Tau Braak staging suggested stage-dependent transcriptional programs involved in myelination and inflammation. These associations were supported by retest PET data and longitudinal analyses, which further linked progressive synaptic decline to DNA repair and telomere pathways. Proteomic profiling further highlighted mitochondrial dysfunction. Together, these findings delineate transcriptional signatures underlying synaptic decline in AD with consistency evidence across cross-sectional, retest, longitudinal, and proteomic analyses underscoring their mechanistic relevance and biomarker potential.Subject areas: biological sciences Introduction Alzheimer’s disease (AD) is a progressive neurodegenerative disorder marked by cognitive decline and memory impairment, underpinned by hallmark pathologies such as β-amyloid (Aβ) plaques and Tau hyperphosphorylation.1 While the Aβ cascade and Tau-mediated neurotoxicity remain central to AD theories, synaptic loss has emerged as a more proximal correlate of clinical symptoms and a direct pathological hallmark of neurodegeneration.2,3,4,5,6 Evidence suggests that synaptic degeneration aligns more closely with regions of cognitive dysfunction than either Aβ or Tau pathology.2,7 A growing body of work implicates multiple contributors—including oligomeric Aβ and Tau species localized at synapses, as well as glial dysregulation—in the breakdown of synaptic integrity.8,9,10,11,12 However, the molecular mechanisms that orchestrate synaptic loss in AD, and the interplay among these diverse pathological drivers, remain incompletely understood. In clinical research, positron emission tomography (PET) imaging targeting synaptic vesicle glycoprotein 2A (SV2A) has been established as a reliable tool for measuring in vivo synaptic density.13 SV2A, a transmembrane protein localized to synaptic vesicles, is broadly expressed at both glutamatergic and GABAergic terminals and serves as a reliable proxy for synaptic integrity.14,15,16,17,18 Radioligands such as 11C-UCB-J and 18F-SynVesT-1 have demonstrated robust imaging performance, with the latter used in multicenter studies due to its favorable imaging properties and practical half-life.19,20,21,22 In AD, SV2A PET consistently reveals reduced synaptic density, particularly in the hippocampus, aligning with postmortem evidence.23,24,25,26 However, while SV2A PET captures the spatial footprint of synaptic loss and its relationship with clinical phenotype, it offers limited insight into the transcriptional programs or molecular pathways underlying synaptic vulnerability in AD.13,27 Recent advances in imaging transcriptomics have enabled the integration of neuroimaging phenotypes with spatially resolved gene expression data, offering a powerful framework to uncover the molecular correlates of brain-wide imaging signals. By linking regional imaging features—such as structural atrophy or synaptic density—to transcriptomic profiles, this approach can identify gene networks and biological pathways underlying disease-related imaging changes.28,29 The Allen Human Brain Atlas (AHBA) provides high-resolution whole-brain transcriptomic data from over 3,700 brain samples across six adult donors, covering expression levels of more than 20,000 genes.30 This resource has been increasingly leveraged to dissect the molecular basis of neuropsychiatric and neurodegenerative imaging phenotypes.28,29,31,32 Such integrative approaches offer a systems-level understanding of disease pathology, bridging macro-scale imaging phenotypes and molecular signatures. In this study, we integrated SV2A PET imaging with transcriptomic data of left hemisphere from the AHBA to uncover the molecular architecture underlying synaptic loss in AD. We first mapped regional SV2A PET standardized uptake value ratio (SUVR) differences between individuals with AD and cognitively normal (CN) controls, and applied partial least squares (PLS) regression to identify genes whose spatial expression patterns correlate with SUVR alterations.33 These genes were further examined using postmortem transcriptomic datasets to validate disease relevance, followed by gene set enrichment analysis (GSEA), functional enrichment, and protein-protein interaction (PPI) analyses to elucidate the underlying biological pathways. To assess the specificity of these gene signatures, we conducted cross-referencing with cell-type, hippocampal tissue, and developmental stage expression profiles. In addition, we performed stratified analyses based on Tau Braak staging to explore the relationship between SV2A decline neurofibrillary tangles (NFTs). Validation analyses replicated the transcriptomic associations in a retest subset and extended them to paired longitudinal imaging and proteomic profiling, demonstrating robustness across imaging time points and biological modalities. Results Demographics The discovery subset included 111 individuals in the AD continuum (aged 49–(86) and 54 controls (aged 49–82, Table 1; Figure 1). Groups were comparable in sex and education, while AD participants showed significantly lower MMSE scores. The retest subset comprised 31 follow-up visits from AD 26 individuals and 19 visits from 14 C N participants, conducted at either 12- or 24-month time points (Table S1). The AD 26 individuals had paired longitudinal data at the 12-month follow-up (Table S2). Table 1Participant characteristics of the discovery subsetCNAD continuump valuen54111–Age (years), mean (SD)66.02 (7.33)71.49 (7.17)< 0.001Sex, % females62.96%64.86%0.947Education (years), mean (SD)12.2 (3.43)11.26 (3.54)0.1573MMSE, mean (SD)28.37 (1.26)21.28 (6.23)< 0.001AVLT, mean (SD)4.15 (2.46)0.65 (1.55)< 0.001BNT, mean (SD)22.09 (4.68)17.38 (6.19)< 0.001VFT, mean (SD)19.39 (22.72)10.36 (5.00)< 0.001STT-A, mean (SD)63.78 (19.51)188.36 (265.37)< 0.001STT-B, mean (SD)152.69 (126.33)391.16 (351.40)< 0.001Braak, n–––Braak 0/16/Braak Ⅰ-Ⅱ/3/Braak Ⅲ -Ⅳ/11/Braak Ⅴ-Ⅵ/72/Statistical comparisons between AD and CN groups were performed using two-sided independent t tests for age, Mann-Whitney U tests for years of education, MMSE, AVLT, BNT, VFT, STT-A, and STT-B scores, and chi-square tests for sex. Participant characteristics of the discovery subset Statistical comparisons between AD and CN groups were performed using two-sided independent t tests for age, Mann-Whitney U tests for years of education, MMSE, AVLT, BNT, VFT, STT-A, and STT-B scores, and chi-square tests for sex. Study flowchart. Participants with 18F-SynVesT-1 PET-MRI scans from the RJNB-D cohort formed the basis of this study The discovery subset comprised 165 participants (111 in the AD continuum, 54 C N), with only the first (baseline) available scan from each participant included. Both AD vs. CN model and stratified Tau model were derived from this baseline discovery subset. From this cohort, several subsets were also defined: the retest subset (with repeated 18F-SynVesT-1 PET-MRI scans), the paired longitudinal subset (AD continuum participants with paired baseline and 12-month follow-up scans), and the plasma proteomic subset. Gene expression data were derived from the Allen Human Brain Atlas. All subset were analyzed using PLS regression to identify differentially genes/proteins, with X and Y differing across models. Genes associated with SV2A alterations in AD Compared to CN, AD participants exhibited significantly lower SV2A in 24 of the 83 brain regions (pFDR < 0.05, Table S3), with the most pronounced reductions in the lateral hippocampus, inferior parietal lobule, and right middle temporal cortex. No regions showed increased SV2A in AD relative to CN (Figure 2). Differences in SV2A between AD/CN groups The left image shows the regional distribution of mean SUVR values across all AD and CN individuals. The right image displays t-value representing group differences between AD and CN, with only regions showing statistically significant differences retained (pFDR < 0.05). Regional SUVR values were adjusted for age, sex, and years of education. Color in medial wall area is for visual integrity and has no statistical meaning. We derived a transcriptomic matrix covering 41 left-hemisphere brain regions and 15,633 genes based on the AHBA dataset (Figure S1). The first PLS component (PLS1) explained 36.2% of the variance in SV2A SUVR differences between AD and CN groups, higher than PLS2 (13.3%), and was therefore used in subsequent analyses. The distribution of weighted gene expression values from PLS1 (PLS1 scores) across brain regions reflected regional transcriptional differences and showed a strong spatial correlation with SV2A differences (Pearson’s r = 0.60, p < 0.001; spin permutations test; Figure 3A). PLS captures a reliable set of genes correlated with SV2A alterations (A) The accompanying brain surface map visualizes the spatial distribution of PLS1 scores across the left hemisphere, with warmer colors indicating regions more strongly weighted in the PLS1 component. Color in brain medial area is for visual integrity and has no statistical meaning. (B) Significance of gene weights evaluated using bootstrap resampling. A total of 1,000 bootstrap iterations were performed. Gene significance was determined based on Z scores (weight divided by standard error), with a threshold of pFDR < 0.05. (C) Expression levels of AD-related genes from GeneCards that were significant in PLS1 showed robust spatial associations with SV2A alterations, as illustrated by surface maps of two representative genes in the left hemisphere. (D) The direction of PLS1 gene weights was highly consistent with the direction of differential gene expression observed in postmortem AD brains. Red represents genes with up weighted and upregulated, blue represents down weighted and downregulated, and gray represents genes with inconsistent weighting and regulated. To identify robust associations, we performed 1,000 bootstrap resamplings and computed Z scores of each gene. A total of 1,233 genes met the significance threshold (pFDR < 0.05), comprising 749 negatively weighted (PLS1−) and 484 positively weighted (PLS1+, Table S4). The PLS1− genes were those associated with SV2A reduction, and the most strongly down-weighted gene was PPFIBP1 (Z = −6.08, pFDR < 0.001, Figure 3B). To further evaluate the AD relevance of SV2A-associated genes, we first intersected the significant PLS1 gene set with AD-related genes retrieved from GeneCards (n = 377). Among 282 AD-related genes in AHBA, 25 genes were significant PLS1 genes. Of these, 21 genes (84%) showed significant correlations between their individual expression levels and the SV2A alterations, respectively (p < 0.05, spin permutations test). Among them, SGCE showed the strongest positive correlation (r = 0.553, p < 0.001, rank 38th by Z scores in PLS1), while CDK5R1, a known tau-related gene, showed the strongest negative correlation (r = −0.549, p < 0.001, rank 9th; Figures 3C; Figure S2; Table S5). We also obtained 7,996 (differentially expressed genes) DEGs from postmortem AD hippocampal tissue (GSE173955). 513 of these were included in the set of our 1,233 PLS1 significant genes: 277 genes were both down-weighted and downregulated, while 225 were up-weighted and upregulated, resulting in a total of 502 genes with consistent directional changes. Furthermore, PLS1 gene weights were significantly correlated with differential expression levels (p < 0.001 for upregulated genes; p = 0.018 for downregulated genes; Figure 3D). The alterations in SV2A reflect diverse underlying molecular mechanisms To characterize the biological functions associated with SV2A alterations, we performed GSEA on the entire PLS1 gene set, and identified 204 significant deregulatory pathways (p.adj < 0.05, |NES| > 1, Table S6). The most downregulated biological processes included synapse related pathways, cognition, axonal and dendritic development (Figure 4A), whereas upregulated pathways were mainly involved in cytoplasmic translation, ribosomal and RNA metabolism, and antigen-related process (Figure 4B). In addition to these dominant signals, GSEA also identified enrichment of positive regulation of inflammatory response (p = 0.0073, p.adj = 0.096; Table S6), indicating the presence of inflammatory-related signatures. GO and KEGG pathway enrichment analyses were performed on the significant genes from PLS1− and PLS1+, resulting in 1,189 enriched terms (pFDR < 0.05, Table S7). SV2A alterations associated molecular mechanisms (A–B) GSEA of the entire PLS1 gene sets, showing significantly enriched pathways for downregulation (A) and upregulation (B). (C and D) Functional enrichment analysis of PLS1−/+ genes. Left: Top 5 most significant terms in each module. Right: enriched terms related to potential AD-related pathological processes. Both color and point size represent p values. (E–G) Cell-, tissue-, and developmental stage-specific analyses. The color intensity indicates the significance level (p value), with more intense red/blue representing smaller p values. Up-regulated represent PLS1+ genes, while downregulated represent PLS1− genes. Z score, the number of standard deviations away from the mean. ∗: pFDR < 0.05 (permutation-based EWCE test). NRC: Neurotransmitter release cycle; BDNF: Brain-derived neurotrophic factor. ASC, astrocytes; GABA, GABAergic neurons; Glu, glutamatergic neurons; MG, microglia; OPC, oligodendrocyte precursor cells; ODC, oligodendrocytes; END, endothelial cells; exCA, pyramidal neurons from the hippocampus CA region; DG, granule neurons from the hippocampus dentate gyrus region; GluPF, glutamatergic neurons from the prefrontal cortex; NSC, neuronal stem cells. EPr, early prenatal; MPr, middle prenatal; LPr, late prenatal. NX, entorhinal cortex. PLS1− genes were significantly enriched in terms related to synaptic organization, glutamatergic signaling, axonal structure, phosphotransferase activity, as well as numerous AD-related pathways including cognition, phosphorylation, cellular response to amyloid-beta, tau-protein kinase activity, and microtubule cytoskeleton organization (Figure 4C; Table S7). Semantic clustering revealed overarching biological themes involving synaptic function, protein phosphorylation, and cognition (Figure S3A). In contrast, PLS1+ significant genes were predominantly enriched in ribosomal processes, RNA metabolism, and pathways associated with glial regulation and neurodevelopment (Figures 4D; Figure S3B; Table S7). They were also enriched in processes of negative regulation of cell differentiation, ubiquitin ligase inhibitor activity, and MHC class II protein complex (Figure 4D). Additionally, PPI network analysis of PLS1− significant genes identified 7 distinct functional modules (Figure S4; Table S7). Notably, MCODE1 was enriched for genes regulating cytoskeletal organization, while MCODE3 highlighted genes involved in vesicle-mediated transport in synapse. SV2A related genes in cell types, hippocampal tissues, and developmental stages Cell-type specificity: We incorporated neural cell-type datasets from two independent studies of the human cerebral cortex and subcortical structures (AHBA: https://celltypes.brain-map.org/ and DroNC-seq). PLS1− significant genes were enriched in excitatory (glutamatergic) and inhibitory (GABAergic) neurons, including CA pyramidal neurons and granule neurons of the dentate gyrus, whereas PLS1+ significant genes were enriched in non-neuronal cell types, particularly oligodendrocytes and endothelial cells (Figure 4E). Hippocampal tissue specificity: using a transcriptomic dataset that includes hippocampal subfield and entorhinal cortex, we found that PLS1− significant genes were preferentially expressed in CA2-CA4, whereas PLS1+ were enriched in CA1 and CA2 (Figure 4F). Developmental stage specificity: developmental transcriptomic profiling (https://brainspan.org/) revealed that PLS1− significant genes were predominantly expressed during postnatal stages—infancy, adolescence, and adulthood—while PLS1+ were mainly expressed during prenatal development (Figure 4G). To confirm robustness, specificity analyses were repeated using the top 80%, 50%, 30%, and 10% of genes ranked by absolute PLS1 weight. Results remained consistent across thresholds (Table S8). Results remain consistent in the age-matched subset analysis Based on age matching to the control group, an age-matched subset was constructed (54 C N and AD 54; Table S9), and the same analysis pipeline was applied. The main conclusions were highly consistent with the results in the discovery subset. The spatial differences of SV2A in the age-matched subset closely mirrored that of the discovery subset (Pearson’s r = 0.857, p < 0.001). Using PLS regression, we identified 1,491 significant genes in the age-matched subset (pFDR < 0.05; Table S10), of which 883 genes overlapped with those identified in the discovery subset (OR = 57.229, p < 0.001; Figure S5A), including PPFIBP1, SGCE, and CDK5R1. Enrichment analysis further highlighted key pathways shared across subsets, including presynapse, axon, tau protein binding, MHC protein complex, and cytosolic ribosome (Figure S5B; Table S10). The specificity of these gene signatures was also consistent with the primary analysis, showing enrichment in neurons, oligodendrocytes, and CA pyramidal neurons (Figure S5C; Table S11). Stratified analysis based on Tau Braak stages Compared with the early-stage group (B0–Ⅱ), the middle-to-late stage group (BⅢ–Ⅵ) showed decreased SV2A in bilateral hippocampus (left: t = −5.135, p < 0.001; right: t = −5.658, p < 0.001; Table S12). In the stratified Tau PLS model, GSEA of PLS1 gene sets revealed strong downregulation of vesicle-mediated synaptic transport, whereas cytoplasmic translation remained predominantly upregulated (Figure S6; Table S13). This model identified 2,834 significant genes (pFDR < 0.05), of which 1,127 overlapped with the significant PLS1 significant genes from the primary AD/CN comparison (OR = 79.058, p < 0.001, Figure 5A). Functional enrichment analysis of these genes shared 867 terms with AD/CN comparison, including presynapse, glutamatergic synapse and Tau-protein kinase activity (Table S14). In addition, unique enrichments were observed in pathways related to myelination, inflammatory response, lipid-related metabolism pathways and others (Figure 5B). Cell-type specificity analysis further demonstrated that these genes were broadly enriched across diverse brain cell types (Figures 5C; Table S15). SV2A alterations and transcriptomic associations across Tau Braak stages and time (A) Overlap of PLS1 significant genes across three PLS models. A total of 105 significant genes were shared across all models. The bars aforementioned indicate the number of overlapping genes per category, while the bars on the left show the number of PLS1 significant genes identified in each model. (B) Enrichment analysis of PLS1 significant genes from the stratified Tau PLS model, highlighting AD-related pathological processes. Both dot color and size indicate p values. (C) Cell-type specificity analysis of PLS1 significant genes across models. Red indicates PLS1+ genes, and blue indicates PLS1− genes. Color intensity reflects the absolute Z-value. Dot size represents the p value and only pFDR < 0.05 remained. (D) Enrichment analysis of PLS1 significant genes from the paired longitudinal PLS model, highlighting AD-related pathological processes. Retest subset replicated the primary findings The spatial differences of SV2A in the retest subset closely mirrored that of the discovery subset (Pearson’s r = 0.764, p < 0.001; Figure S7A), with the hippocampus remaining the most affected region (left: t = –3.067, p = 0.003; right: t = –2.639, p = 0.011). Using the same analysis pipeline, we identified 2,028 PLS1+ and 1,507 PLS1− significant genes in the retest subset (pFDR < 0.05, Table S16). Of these, 920 genes were also found in the discovery subset (OR = 13.247, p < 0.001), including 362 down-weighted and 558 up-weighted genes showing consistent directionality (Figure 6A). This set included previously highlighted genes such as PPFIBP1, SGCE, and CDK5R1. Furthermore, GSEA revealed highly consistent biological themes with the most significant enrichments again observed in the upregulated pathway of cytoplasmic translation and the downregulated pathway of vesicle-mediated transport in synapse, consistent with findings from the discovery subset (Figure S7B; Table S17). Enrichment analysis of the significant genes further highlighted key pathways shared across subsets, including presynapse, regulation of cytoskeletal organization, phosphorylation, cellular response to amyloid-beta, and immune-related functions such as MHC protein complex and ubiquitin-dependent protein catabolic process. (Figures 6B; Figure S7C; Table S17). The retest subset replicated the main findings from the discovery subset (A) Overlap of PLS1 genes and enrichment results between the retest and discovery subset. The outer circle indicates differentially weighted PLS1 significant genes, while the dark segments of the inner circle and purple edges denote overlapping genes; blue edges denote overlapping enrichment pathways. (B) Top 20 most significant pathways shared by the PLS1− significant genes of two subsets. Color represents the significance level. Longitudinal SV2A changes associated with AD pathology Paired longitudinal PLS model was constructed to characterize spatiotemporal transcriptional signatures of SV2A decline and identified 2,409 significant genes, of which 154 overlapped with the PLS1 significant genes from the cross-sectional AD/CN comparison (Figure 5A). GSEA indicated strong downregulation in postsynaptic density organization, whereas cytoplasmic translation remained predominantly upregulated (Figure S8; Table S13). Functional enrichment analysis yielded 1,799 enriched terms (pFDR < 0.05), 457 of which overlapped with the AD/CN comparison (Table S14). Overlapping pathways were predominantly related to synaptic function and AD-relevant processes such as amyloid-beta binding, Tau signaling, microtubule cytoskeleton organization. Additionally, unique enrichments were observed in pathways related to DNA repair, telomere maintenance, and mitochondrial metabolism (Figure 5D; Table S14). Cell-type specificity analysis further showed enrichment in oligodendrocytes, CA pyramidal neurons, and dentate granule cells (Figure 5C; Table S15). Plasma proteomics aligns with transcriptomic mechanisms of SV2A loss The demographic information of plasma proteomic subset is provided in Table S18. Each individual had an index of AD-related synaptic loss (CN = −1.055, AD = −1.678, p < 0.001, one-sided t test. Figure S9A). Using these indices, we constructed a plasma proteomics-PLS model and identified 893 proteins in the PLS1 component as significant proteins (pFDR < 0.05). Among these proteins, 62 overlapped with the PLS1 significant genes identified in the discovery subset (Table S19). GSEA revealed strong deregulation of inner mitochondrial membrane organization and vesicle mediated transport in synapse (Figure S9B; Table S20). Functional enrichment analysis revealed that these proteins were primarily involved in protein transport, mitochondrial function, cytoskeleton (Figure 7A), as well as several synaptic- and AD-related pathways (Figures 7B; Table S20). To integrate multi-omics evidence, we applied Stouffer’s method to combine the enrichment results from PLS1 significant proteins and PLS1 significant genes identified in the discovery subset. This analysis revealed 543 shared enriched terms, including glutamatergic synapse, presynapse, actin cytoskeleton, and Tau protein binding (Figures 7C; Table S21). Functional enrichment of plasma proteomics (A) The top 5 enrichment terms of differentially proteins in each module. (B) AD-related terms in enrichment of differentially proteins. (C) The x axis denotes the significance of differential gene enrichment, and the y axis denotes that of differential protein enrichment. Points deviating from the zero axes indicate shared enrichment between the two entities. Log-transformed Stuffer’s p value is the color bar. Each point represents an enriched term. Discussion In this study, we integrated in vivo SV2A PET imaging with spatial transcriptomic analysis to delineate the molecular basis of synaptic loss in the AD continuum. We identified SV2A reduction closely aligned with region-specific gene expression signatures. These transcriptional associations were validated through repeated imaging in a retest subset, and a paired longitudinal subset. A subset of participants with available plasma proteomics data were used to validate SV2A-related findings at the protein level. Enrichment analyses revealed involvement of pathways related to synaptic organization, Tau phosphorylation, and cytoskeletal dynamics. Using PLS regression, we identified 1,233 genes significantly associated with SV2A alterations. Among the negatively weighted genes, PPFIBP1—which encodes liprin-β1, a protein involved in presynaptic assembly and synapse formation—emerged as a key feature.34,35,36 Biallelic loss-of-function mutations in PPFIBP1 have been linked to severe neurodevelopmental disorders, such as profound intellectual disability.37 Among the positively weighted genes, CDK5R1 showed the strongest negative correlation with SV2A alterations. CDK5R1 encodes a critical regulator of CDK5, a kinase involved in Tau hyperphosphorylation and synaptic plasticity in AD.38,39,40,41 Several studies have reported that miRNAs regulating CDK5R1 were downregulated in the hippocampus and temporal cortex of AD brains, with increased CDK5R1 expression observed in AD hippocampal tissue.42,43 We further found that the synapse-related genes identified by PLS were closely associated with transcriptional changes in postmortem AD brains. Of the 1,233 significant genes, 513 overlapped with DEGs in the AD hippocampus, and 502 of them showed consistent directions between expression changes and PLS weights.44 The strong correlation between PLS1 weights and differential expression fold changes suggests that transcriptional alterations associated with synaptic loss are detectable in the AD brain and may precede overt structural degeneration. Functionally, PLS1-weighted genes were significantly enriched in pathways related to synapse organization, synaptic vesicle cycling, endocytosis, and neurotransmission, aligning with the biological relevance of SV2A. Terms associated with learning and memory were also overrepresented, supporting the link between synaptic dysfunction and cognitive decline.7,45 More importantly, we also observed strong enrichment in AD-related pathological processes, including protein phosphorylation, microtubule remodeling, Tau-kinase activity, and responses to Aβ. Aβ and tau synergistically contribute to synaptic dysfunction and AD progression, notably by promoting dendritic spine degeneration, especially at inhibitory synapses.10,46,47 Hyperphosphorylated Tau dissociates from microtubules, forming NFTs and impairing axonal transport. Tau may also accumulate at the postsynaptic membrane, disrupting synaptic transmission and plasticity.10,47,48 Our previous work also demonstrated a close association between SV2A decline and Tau pathology based on functional connectivity network.27,49 In addition, our stratified analyses based on Tau Braak stage highlighted the critical involvement of the myelination and inflammation in pathological progression of NFTs. The findings were in line with evidence which suggests NFTs act as a trigger for inflammatory responses and the Tau propagation shows strong inverse association with regional myelination.50,51,52 Enrichment of ubiquitin-related terms—such as ubiquitin ligase inhibition and protein transferase regulation—further implicates dysregulation of the ubiquitin-proteasome system, a pathway essential for synaptic maintenance and APP turnover.53,54,55,56 Neurotrophic signaling pathways, including BDNF-TrkB, were also significantly enriched, highlighting therapeutic targets relevant to synaptic repair.57,58,59,60 Beyond synaptic functions, these genes also exhibited strong cell type specificity. We observed strong cell-type specificity among PLS1 genes. Consistent to previous studies, negatively weighted genes were enriched in glutamatergic neurons,61 GABAergic neurons,62 and CA pyramidal neurons63—cell types known to exhibit early vulnerability in AD. In contrast, positively weighted genes were enriched in oligodendrocytes, consistent with evidence that age-related oligodendrocyte dysfunction contributes to myelin degeneration and impaired synaptic support.64,65,66 These findings underscore that synaptic vulnerability in AD arises from a complex interplay of neuronal and glial dysfunction. Notably, synapse-related genes were not enriched in astrocytes or microglia—cell types known to mediate synaptic pruning and neuroinflammation.67,68,69 This absence may reflect the limited sensitivity of SV2A PET to glia-mediated processes, or the spatial and temporal heterogeneity of glial activation. Future studies leveraging single-cell or spatially resolved transcriptomic approaches, as demonstrated in recent SV2A-PET work in multiple sclerosis,70 will be essential to refine the cellular attribution of SV2A-PET signals and to disentangle direct neuronal loss from indirect glial effects. Besides validation from overlapped genes and enrichment in the retest subset, paired longitudinal analysis revealed that progressive SV2A decline was accompanied by emerging transcriptional signals involved in DNA repair, telomere regulation, and mitochondrial pathways—alterations that may not show in the cross-sectional analyses. These findings underscore the value of follow-up imaging for capturing time-dependent molecular dynamics in AD. Furthermore, In-house plasma proteomics supported major transcriptomic findings, including synaptic function and tau phosphorylation, and further pointed to prominent mitochondrial dysfunction. This observation aligns with growing evidence from multi-omics studies linking mitochondrial abnormalities to AD pathophysiology.61,62,71 This study relies on a transcriptomic-imaging integration framework that involves inherent assumptions. Specifically, regional gene expression patterns from the AHBA were used as a spatial transcriptomic reference to relate spatial patterns of SV2A decline observed in AD. This approach assumes that normative gene expression gradients reflect regional vulnerability to synaptic loss, which may introduce uncertainty in the interpretation of disease-related molecular associations. Nevertheless, the AHBA provides high-resolution whole-brain transcriptomic data covering more than 20,000 genes, thereby reducing the likelihood that relevant molecular signals are missed. We further employed PLS regression, which identifies multivariate gene expression patterns whose spatial covariance across brain regions maximally aligns with the spatial distribution of SV2A decline. By focusing on shared spatial structure rather than absolute expression levels, PLS enables the use of normative transcriptomic atlases to probe molecular correlates of disease-related brain changes and is well suited for high-dimensional, collinear transcriptomic data. This strategy has been widely applied in prior neuroimaging-transcriptomic studies.28,33,72,73 To further mitigate this limitation, as reported above, we validated our findings using independent postmortem AD hippocampal transcriptomic data (GSE173955),44 demonstrating consistent directional effects and significant correlations with PLS-derived gene weights. Future work using AD-specific multimodal or single-cell atlases could provide a more direct and precise validation of these associations. In summary, our study establishes a comprehensive spatial transcriptomic framework that links in vivo synaptic loss in AD to coordinated molecular programs and specific cellular contributors. By integrating SV2A PET with regional gene expression profiles, we provide mechanistic insight into the molecular architecture underlying regionally selective synaptic vulnerability, disease heterogeneity, and progression. These findings extend current knowledge by contextualizing synaptic degeneration within brain-wide transcriptomic organization, highlight candidate molecular pathways relevant to synaptic resilience and failure, and underscore the value of SV2A PET as a translational bridge between neuroimaging and molecular neuropathology. Limitations of the study Our study has several limitations. First, AHBA data were derived from a small number of healthy donors, with limited right hemisphere sampling, which may affect generalizability and overlook individual heterogeneity, although our preprocessing pipeline aimed to minimize such effects.30 Second, SV2A PET primarily reflects presynaptic density and cannot distinguish excitatory versus inhibitory synapses or infer functional activity.63,74 Third, AHBA transcriptomic data represent bulk RNA and lack single-cell resolution,64 limiting cell-type precision. Finally, longitudinal analyses were based on a modest sample size and warrant replication in larger cohorts. Resource availability Lead contact Further information and requests for resources should be directed to and will be fulfilled by the lead contact, Binyin Li (libinyin@126.com). Materials availability This study did not generate new unique reagents. Data and code availability •The processed SV2A PET data and all public transcriptomic datasets used in this study are publicly available as of the date of publication. Accession numbers are listed in the key resources table. The processed SV2A PET data and all public transcriptomic datasets used in this study are publicly available as of the date of publication. Accession numbers are listed in the key resources table. •All original analysis code has been deposited in GitHub and is publicly available as of the date of publication. Accession numbers are listed in the key resources table. All original analysis code has been deposited in GitHub and is publicly available as of the date of publication. Accession numbers are listed in the key resources table. •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request. Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request. Acknowledgments This study was supported by National Key Research and Development Program of China, Scientific and Technological Innovation 2030—Major Projects (2022ZD0213800), National Natural Science Foundation of China (82271441, 82171473, 81901180, and 82501588). The authors thank the participants and their families for their participation in this study. Author contributions B.L., F.X., and J.Z. led the project and designed the study. B.L., L.S., X.Y., and X.X. analyzed the data, and published the code. B.L., L.S., X.Y., X.X., L.K., Y.G., Y.Z., and Q.H. contributed to the analysis and interpretation of the results. Y.S., F.X., W.X., Y.G., J.L., Y.D., and S.L. performed neuroimaging preprocessing and quality control for the participant cohort. L.S., B.L., and S.L. wrote the manuscript and created the figures. All authors reviewed and commented on the manuscript. Declaration of interests The authors declare no competing interests. STAR★Methods Key resources table REAGENT or RESOURCESOURCEIDENTIFIERBiological samplesRuijin NeuroBank of Alzheimer’s Disease and Dementia (RJNB-D) cohorthttps://clinicaltrials.gov/study/NCT05623124Deposited dataAllen Human Brain AtlasAllen Institute for Brain Science: https://human.brain-map.org/GeneCardshttps://www.genecards.org/National Center for Biotechnology InformationGSE167494Cell types - the Allen Cell Types Databasehttps://celltypes.brain-map.org/Cell types - the DroNC-seqNational Center for Biotechnology InformationGSE278723BrainSpan Atlashttps://www.brainspan.org/static/download.htmlProcessed data in this studyFigshare: https://doi.org/10.6084/m9.figshare.29662010.v1)Software and algorithmsCustom code of this studyThis paper; GitHub: https://github.com/Sahuhu-Sun/Imaging-transcriptomicsPython 3.9https://www.python.orgR version 4.2.0https://www.r-project.org/FreeSurfer v7.1.1https://surfer.nmr.mgh.harvard.edu/ Experimental model and study participant details The discovery subset consisted of 165 participants with completed 18F-SynVesT-1 PET-MRI scans from the Ruijin NeuroBank of Alzheimer’s Disease and Dementia (RJNB-D) cohort between November 2021 and December 2024,65 including 111 individuals diagnosed in the AD continuum (AD group) and 54 C N volunteers (aged 49–86 years; 106 females and 50 males). Specifically, the discovery subset consisted exclusively of each participant’s first (baseline) available 18F-SynVesT-1 PET-MRI scan. An age-matched subset was constructed by matching AD participants to controls, resulting in 54 C N controls and AD 54 participants. All participants underwent 18F-Florbetapir (FBP) PET-CT for β amyloid, 18F-SynVesT-1 PET-MR imaging and neuropsychological assessments. The criteria for AD continuum included positive β-amyloid in FBP PET and clinical assessments that have been described in our previous study.66 A subset of AD participants (n = 102) also received 18F-MK6240 PET-CT scans for Tau Braak staging. All procedures were completed within one month of enrollment. Exclusion criteria included any history of neurological, psychiatric, or systemic illness unrelated to AD, recent major surgery, brain injury, brain tumor, or contraindications to PET-MRI. The retest subset comprised 40 participants (AD 26 and 14 C N) who underwent repeated 18F-SynVesT-1 PET between November 2022 and December 2024. In total, 50 follow-up scans were obtained (31 from the AD group and 19 from the CN group) at either 12- or 24-month timepoints. These repeated scans were not included in the discovery analysis and were used for retest analysis to assess the reproducibility of SV2A PET - transcriptomic associations. For paired longitudinal analysis, AD 26 participants had paired 18F-SynVesT-1 PET scans at both baseline (also included in the discovery subset analysis) and 12-month follow-up, allowing for analysis of progressive SV2A decline and associated transcriptional changes. In addition, plasma samples from 65 participants (a subset of the discovery subset including AD 52 and 13 C N) were analyzed using a data-independent acquisition (DIA) mass spectrometry workflow to evaluate protein-level correlates of SV2A variation. The study was approved by the ethics committee, Ruijin Hospital, Shanghai Jiao Tong University School of Medicine, and was also registered with ClinicalTrials.gov (NCT05623124). All participants or their caregivers provided written informed consent. Method details Imaging acquisition and preprocessing Participants underwent multimodal imaging, including 18F-SynVesT-1 PET, 18F-Florbetapir (FBP) PET and 18F-MK6240 PET, together with high-resolution structural MRI. Structural MRI was preprocessed by FreeSurfer (v7.1.1), including cortical and subcortical segmentation, surface reconstruction, and anatomical parcellation. PET images were co-registered with structural MRI by PETSurfer pipeline followed by partial volume correction. Given prior kinetic analyses demonstrating that late-phase SUVR (60–90 min post-injection) provides a reliable surrogate for tracer uptake estimated using one-tissue compartment modeling, static SUVR images were derived using cerebellar gray matter as the reference region. The extracted SUVR was observed as a reliable and non-invasive measure with minimal variability across groups.13,75 In vivo Braak stage classification for 18F-MK6240 PET followed established definitions.52,76,77 See Methods S1 for details. Identification of genes associated with SV2A The SV2A difference in each Desikan–Killiany atlas region between AD and CN was measured using a general linear model controlling for age, sex, and years of education. Group differences in SV2A were assessed by performing two-tailed t-tests on regression coefficients. Regional gene expression data were obtained from AHBA (http://human.brain-map.org/).30 The dataset comprises whole-brain transcriptomic data from six healthy donors (aged 24–57), with 3,702 tissue samples covering over 20,000 genes. Only left-hemisphere data were used due to limited right-hemisphere sampling (only two donors). Gene expression preprocessing was performed using the abagen toolbox, including probe reannotation, quality filtering, probe selection, spatial matching, and normalization (Figure 1, details in supplemental information).64,78 Expression values were averaged across samples within each Desikan–Killiany atlas region, producing a brain region × gene transcriptomic matrix for subsequent analysis. Next, we employed PLS regression to identify gene expression patterns associated with regional SV2A alterations.33 Specifically, we extracted the first PLS component (PLS1), which captures the maximal covariance between the gene expression matrix (X) and SV2A variation across brain regions (response variable, Y).28 To ensure that the explanatory power of PLS1 was not due to chance or driven by spatial autocorrelation of gene expression, we conducted a spin test with 10,000 spherical permutations of brain region labels.79,80 Additionally, we applied a bootstrap resampling procedure (1,000 iterations) to evaluate the stability and statistical significance of gene weights in PLS1. For each gene, a Z score was calculated as the ratio of its weight to the bootstrap-estimated standard error. We then performed z-tests on the Z scores and applied Benjamini-Hochberg correction to control the false discovery rate (FDR), using a threshold of pFDR < 0.05 for significance. Significant genes were divided into two sets, PLS1+ and PLS1−, based on their weights. AD-related genes in PLS1 significant genes We evaluated the association of AD-related genes with the PLS1 significant genes from two perspectives. First, AD-related genes were retrieved from GeneCards (keyword = ‘Alzheimer’, relevance score >7; https://www.genecards.org/), a comprehensive database integrating extensive human gene information.81,82 Pearson’s correlation analysis was performed between regional SV2A alterations and the expression levels of genes that overlapped between PLS1 significant genes and AD-related gene sets. To account for spatial autocorrelation, 10,000 permutation tests were performed using the spin method. Second, we obtained 7,996 DEGs in the AD hippocampus from a publicly available study (GSE173955)44 and assessed the consistency between gene differential expression and their PLS1 weights. Gene functional enrichment analysis We first performed GSEA on the entire PLS1 gene set (ranked by Z scores) using the fgsea R package, focusing on Gene Ontology (GO) Biological Processes.83 Pathways with an adjusted p < 0.05 and |normalized enrichment score (NES)| > 1 were considered significantly enriched. Then on the PLS1 significant genes, functional enrichment analysis was performed using the Metascape (https://metascape.org).84 Significant genes in PLS1+ and PLS1− were analyzed separately, with “Homo sapiens” selected as the reference species, Analyses included GO biological processes and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment. We also applied semantic summarization using the REVIGO (semantic similarity cutoff = 0.7, similarity metric = SimRel; http://revigo.irb.hr/) to better summarize the overall results of functional enrichments.85 Additionally, PPI network was also be constructed on Metascape platform with the Molecular Complex Detection (MCODE) algorithm to screen for potential core functional subnetworks.86 Specificity analysis To examine whether SV2A changes in AD reflect specific cell types, hippocampal tissues, or developmental stages, we conducted a systematic specificity analysis inspired by the Expression Weighted Cell-type Enrichment (EWCE) framework using significant PLS1+ and PLS1− genes.87 EWCE is a robust approach used to assess whether a given gene set is significantly enriched in particular marker types, providing insights into biological origin and spatiotemporal specificity (details in supplemental information). For cell-type specificity, we integrated datasets from the human cerebral cortex and subcortical structures, including the AHBA single-cell RNA-seq (https://celltypes.brain-map.org/) and DroNC-seq datasets,88 enabling cross-dataset validation. For tissue specificity, we used hippocampal subfields (CA1-(4) and entorhinal cortex expression data from GSE278723.67 For developmental specificity, we used BrainSpan (https://brainspan.org/) transcriptomes spanning seven postnatal stages (early prenatal to adult). Replication analysis The analysis pipeline for the retest subset was identical to gene identification process of the discovery subset. Functional enrichment analysis was conducted using the Metascape, following the same procedures as in the discovery subset. Stratified and paired longitudinal analysis To investigate the association between SV2A reduction and Tau pathology, AD participants were stratified into early-stage (B0–Ⅱ) and middle-to-late-stage (BⅢ–Ⅵ) groups. A linear model was applied to estimate the stratified group-level β values, controlling for age, sex, and years of education. The group β values were compared using a two-tailed t-tests, and the resulting t-values were then used as the Y in the stratified Tau PLS model. Among the AD group, 26 individuals underwent repeated one-year follow-up 18F-SynVesT-1 PET scans. Paired t-tests were performed to compare regional SV2A changes between baseline and follow-up. The resulting t-values captured the spatial profile of synaptic decline across brain regions and were used as Y in the paired longitudinal PLS model. In-house plasma proteomic analysis DIA data were processed with DIA-NN (v1.8.(1) using a plasma spectral library, with standard search parameters (1% FDR, peptide length 7–50 amino acids, variable methionine oxidation, and N-terminal excision).89 Quality control included biological replicates and technical duplicates across batches. After data processing, quantitative measures were obtained for 3,558 proteins, and proteins with >80% missing values were excluded from downstream analyses. Individual-level synaptic loss was estimated by computing one-tailed t test comparing SV2A in AD–CN differential versus non-differential regions, as defined in the discovery subset analysis above. Subsequently, we constructed a proteomic–PLS model linking plasma protein abundances (X) with the synaptic loss (Y) across individuals. PLS1 was selected, and 1,000 bootstrap resamplings were used to calculate Z-scores for each protein. Proteins with pFDR < 0.05 were considered differentially abundant. Functional enrichment analysis and was then performed on the identified PLS1 proteins and results from plasma proteins and differential brain genes were combined using Stouffer’s method to integrate multiple p-values into a composite Z score.90,91 Quantification and statistical analysis All statistical analyses were performed in R (version 4.2.(2) and Python (version 3.12), unless otherwise specified. Two-tailed independent-samples t tests were used for normally distributed variables (e.g., age), whereas Mann–Whitney U tests were applied when normality assumptions were violated, including Mini-Mental State Examination (MMSE) scores, Auditory Verbal Learning Test (AVLT) delayed recall, Boston Naming Test (BNT), Verbal Fluency Test (VFT, semantic fluency), and Shape Trail Test (STT-A and STT-B). Sex distribution was compared using chi-square tests. Age-matched AD participants were selected by matching AD participants to CN controls using one-to-one nearest-neighbor matching based on age. In specificity analyses (cell-type, hippocampal tissue, and developmental stage), EWCE method was performed and repeated using subsets of genes ranked in the top 80%, 50%, 30%, and 10% by absolute PLS1 weight, with statistical significance defined as pFDR <0.05 (permutation-based EWCE test).87 Odds ratios (OR) were calculated to evaluate the consistency of gene identification from 2 significant gene sets, and Fisher’s exact test was used to determine the statistical significance of gene overlap. Details of these analyses are provided in the corresponding Methods sections. Additional resources Processed data generated in this study are publicly accessible on Figshare at https://doi.org/10.6084/m9.figshare.29662010.v1. All custom analysis codes supporting the core workflows of this study are deposited and maintained in a GitHub repository: https://github.com/Sahuhu-Sun/Imaging-transcriptomics. This study was approved by the Ethics Committee of Ruijin Hospital, Shanghai Jiao Tong University School of Medicine, and was prospectively registered at ClinicalTrials.govunder identifier NCT05623124. Footnotes Supplementary data related to this article can be found online at https://doi.org/10.1016/j.isci.2026.116336. Contributor Information Junfang Zhang, Email: tvxq98@sjtu.edu.cn. Fang Xie, Email: fangxie@fudan.edu.cn. Binyin Li, Email: libinyin@126.com. Supplemental information Document S1. Figures S1–S9, Tables S1, S2, S9, and S18, and Methods S1 Data S1. Tables S3–S8, S10–S17, and S19–S21 References Associated Data Supplementary Materials Document S1. Figures S1–S9, Tables S1, S2, S9, and S18, and Methods S1 Data S1. Tables S3–S8, S10–S17, and S19–S21 Data Availability Statement •The processed SV2A PET data and all public transcriptomic datasets used in this study are publicly available as of the date of publication. Accession numbers are listed in the key resources table. The processed SV2A PET data and all public transcriptomic datasets used in this study are publicly available as of the date of publication. Accession numbers are listed in the key resources table. •All original analysis code has been deposited in GitHub and is publicly available as of the date of publication. Accession numbers are listed in the key resources table. All original analysis code has been deposited in GitHub and is publicly available as of the date of publication. Accession numbers are listed in the key resources table. •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request. Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.