A latent genetic factor underlies multiple component phenotypes of cerebral small vessel disease

Abstract

Background

Cerebral small vessel disease (cSVD) encompasses white matter hyperintensities (WMH), enlarged perivascular spaces (PVS), intracerebral hemorrhage (ICH), small vessel stroke (SVS), and cognitive impairment—phenotypes that frequently co-occur yet have been studied as independent genetic endpoints. Whether a shared genetic architecture underlies these co-occurring cSVD-related phenotypes remains unclear.

Methods

We applied Genomic SEM to GWAS summary statistics from six phenotypes (WMH, PVS, ICH, SVS, cognitive performance, systolic blood pressure), followed by multivariate association scanning, Bayesian fine-mapping (SuSiE/FINEMAP), gene-level (MAGMA) and transcriptome-wide (sCCA-TWAS/FOCUS) analyses, pathway enrichment, cell-type specificity (CELLECT) and spatial developmental mapping (gsMap).

Results

A common-factor model fit the data well (CFI = 1.00, SRMR = 0.061). The multivariate scan identified 1,423 genome-wide significant SNPs, resolving into 20 independent lead SNPs mapped to 14 distinct genomic loci, of which 2 were not detected at genome-wide significance in any constituent single-trait GWAS. Fine-mapping prioritized three high-confidence causal variants at EFEMP1, ICOSLG, and AC098824.6. Gene-level and transcriptome-wide analyses converged on effector genes at 17q25.1 (TRIM47/TRIM65), 10q24.33 (SH3PXD2A), and 2q33.1 (NBEAL1). MAGMA additionally identified APOE at 19q13.32. Pathway analysis highlighted epigenetic dysregulation, cellular senescence, and amyloid fiber formation. CELLECT identified vascular endothelial cells and pericytes as enriched cell types. gsMap revealed developmental enrichment in lung, kidney, and meninges exceeding brain tissue.

Conclusions

These findings delineate a coherent latent genetic structure linking multiple cSVD-related phenotypes, rooted in vascular wall cell biology, epigenetic regulation, and cellular senescence, with developmental origins spanning multiple organ microvascular beds.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12868-026-01031-4.

Introduction

Cerebral small vessel disease (cSVD) refers to a group of pathological processes affecting the small arteries, arterioles, venules, and capillaries of the brain [1–3]. Its manifestations include characteristic neuroimaging markers such as white matter hyperintensities (WMH), enlarged perivascular spaces (PVS), cerebral microbleeds, and lacunes [4], with clinical consequences ranging from spontaneous intracerebral hemorrhage (ICH) and small vessel stroke (SVS) to cognitive impairment and dementia, accounting for approximately 25% of ischemic strokes and representing the most important vascular contributor to dementia [5]. Among vascular risk factors, elevated systolic blood pressure (SBP) is the most important modifiable driver of cSVD onset and progression [6], and Mendelian randomization studies have confirmed that SBP causally increases the risk of PVS, lacunar stroke, and WMH volume [7]. These neuroimaging and clinical phenotypes frequently co-occur, and multi-trait genetic analyses have confirmed significant genetic correlations among them [8]. Integrating the co-occurring cSVD markers into a single measure is an established practice—the individual neuroimaging markers have long been aggregated into a total SVD burden score that predicts clinical outcomes more completely than any single marker [9, 10], and cSVD has recently been modelled as a latent construct integrating multiple imaging endophenotypes [11]. However, such integrative efforts have remained at the phenotypic level, and the shared genetic basis of these co-occurring phenotypes remains largely uncharacterized [12].

Large-scale MRI-based clinical cohort studies have repeatedly demonstrated that cSVD markers, including lacunes, WMH, cerebral microbleeds, and PVS, frequently co-occur within the same individual [13], and that a composite total SVD burden score derived from these markers is independently associated with post-stroke functional outcomes, with predictive power exceeding that of any single neuroimaging marker [10]. At the level of vascular risk factors, both cross-sectional and longitudinal cohort studies have confirmed that hypertension is independently associated with the total SVD burden score [9]. Hypertension compromises the structural and functional integrity of the cerebral microcirculation, promoting microvascular rarefaction, endothelial dysfunction, and neurovascular uncoupling. These pathological changes in turn drive the development of multiple neuroimaging markers including WMH, lacunar infarcts, and microhemorrhages [14]. However, these observational studies are inherently limited by incomplete control of confounding and reverse causation. They cannot distinguish whether the co-occurrence of cSVD phenotypes arises from a shared etiological basis or merely reflects comorbidity under common risk factor exposure, nor can they reveal the causal mechanisms at the molecular genetic level.

At the genetic level, large-scale genome-wide association studies (GWAS) over the past decade have identified risk loci for individual cSVD component phenotypes, implicating inflammatory, glial, extracellular-matrix, pericyte, and APOE-related pathways [15–19], and have confirmed genetic overlap between cognitive performance and vascular risk factors [20]. However, single-trait GWAS face a persistent statistical power bottleneck—all identified WMH loci collectively explain only approximately 6% of the narrow-sense heritability [19], and ICH GWAS have yielded a single genome-wide significant locus despite heritability estimates reaching 73% [15]. More fundamentally, analyzing each phenotype as an independent endpoint cannot capture the genetic synergy and latent hierarchical structure among the component traits, limiting a systematic understanding of the shared genetic architecture across cSVD-related phenotypes. Cross-trait methods have begun to address this gap. Linkage disequilibrium score regression (LDSC) has confirmed significant genetic correlations between periventricular WMH and small vessel ischemic stroke while revealing both distinct and shared genetic architectures among cSVD neuroimaging markers [21], and multi-trait analysis of GWAS (MTAG) has jointly analyzed correlated cSVD phenotypes to identify risk loci undetectable by single-trait GWAS [8]. Yet LDSC provides only pairwise correlations, and although MTAG enhances discovery power, neither method models the latent factor structure among multiple phenotypes within a unified framework, nor separates shared from phenotype-specific genetic effects.

Building on these foundations, this study employs Genomic Structural Equation Modeling (Genomic SEM) [22] to construct and analyze the shared genetic architecture across cSVD-related phenotypes by integrating GWAS summary statistics from six core phenotypes (WMH, SVS, ICH, PVS, cognitive performance, and SBP). An overview of this analytical framework is presented in Fig. 1. We first validate the latent hierarchical structure through confirmatory factor modeling, then leverage multivariate GWAS to identify small-effect pleiotropic variants. The resulting signals are further resolved through Bayesian fine-mapping, gene-level and transcriptome-wide association analyses (TWAS), pathway enrichment, cell-type specificity testing, and spatial developmental mapping, to identify candidate genes, pathways, and cell types underlying the shared genetic susceptibility across cSVD-related phenotypes.

Fig. 1.

Schematic overview of the shared genetic architecture underlying the cSVD. Notes: Six component phenotypes (WMH, PVS, ICH, SVS, cognitive performance, and SBP) converge on a latent genetic factor identified through Genomic SEM. Multi-layer downstream analyses reveal key effector genes at five genomic loci (17q25.1, 10q24.33, 2q33.1, 6p22.1, and 19q13.32), three enriched biological pathways (epigenetic dysregulation, cellular senescence, and amyloid fiber formation), two principal cell types (vascular endothelial cells and pericytes), and a developmental enrichment gradient across multiple organ microvascular beds, with lung and kidney exceeding brain tissue

Methods

GWAS summary data and trait selection

This study drew on summary-level statistics from published large-scale GWAS to delineate the shared genetic structure among the constituent cSVD-related phenotypes. Six phenotypes were included, each representing a distinct pathological dimension of the syndrome. WMH volume (N = 37,355; GCST90472808) and PVS (N = 38,903, comprising 8,950 cases and 29,953 controls; GCST90244153) capture parenchymal neuroimaging burden, SVS (N = 198,048, comprising 5,386 cases and 192,662 controls; ebi-a-GCST006909) and ICH (N = 473,513, comprising 1,935 cases and 471,578 controls; ebi-a-GCST90018870) capture acute cerebrovascular events, cognitive performance (N = 2,266,733; GCST90446168) captures the neurofunctional outcome, and SBP (N = 757,601; ieu-b-38) captures upstream vascular risk exposure. SBP was included as an upstream component of the shared genetic liability rather than as an external covariate. To confirm that our findings do not depend on this choice, a four-trait sensitivity analysis restricted to WMH, PVS, ICH and SVS was performed. The six-trait model was retained as the primary analysis as it represents the full liability spectrum of cSVD from upstream risk to downstream outcome. To maintain consistency of linkage disequilibrium reference structures, all input GWAS were restricted to cohorts of European descent. Cohort sources and sample sizes for each dataset are detailed in Table S1.

Variant harmonization and filtering

All summary statistics were first aligned to a common genomic coordinate and allele orientation framework using the European-ancestry reference haplotypes of the 1000 Genomes Project Phase 3 [23]. A sequential variant exclusion procedure was then applied to ensure analytical robustness. Variants with minor allele frequency (MAF) below 0.01 were removed owing to imprecise effect-size estimation at low frequency. A/T and C/G palindromic variants were discarded to eliminate unresolvable strand ambiguity. The 25–35 Mb interval on chromosome 6, encompassing the major histocompatibility complex (MHC), was also excluded because its extended haplotype structure inflates genetic covariance estimates and introduces artefactual correlations into factor models.

Latent factor estimation through genomic SEM

The GenomicSEM R package (version 0.0.5) [22] was used to evaluate whether a single latent genetic factor can account for the heritable covariance among the six cSVD-related phenotypes. The six sets of QC-passed GWAS summary statistics were converted into an empirical 6 × 6 genetic covariance matrix, together with sampling variance estimates, through multivariate LDSC [24]. Trait-specific SNP heritabilities (h²SNP) were estimated concurrently (Table S2). A one-factor confirmatory model was then fitted to this matrix, constraining the model-implied covariance structure to reproduce the empirical observations. Goodness of fit was assessed through four complementary statistics (Tables S3 and S4), including a χ² test (with P value), the comparative fit index (CFI), the standardized root mean square residual (SRMR), and the Akaike information criterion (AIC). After confirming adequate fit, each autosomal SNP was projected onto the latent factor to yield multivariate association P values. Consistency of effect directions was safeguarded by a per-SNP Cochran Q test, with variants returning Q-statistic P below 0.05 removed.

Genome-wide locus discovery and novelty assessment

Genome-wide significant signals (P < 5 × 10⁻⁸) from the multivariate association scan were mapped to genomic regions and functionally annotated through the Functional Mapping and Annotation of Genetic Associations (FUMA) platform [25]. Lead variants were consolidated into independent loci by applying a pairwise r² threshold of 0.1 with respect to the 1000 Genomes European reference panel. To evaluate whether these lead SNPs represented signals not captured by the individual input GWAS, a physical distance criterion was applied whereby the coordinates of each lead SNP were compared against every variant that had reached genome-wide significance in any of the six constituent single-trait GWAS. A lead SNP was designated as not detected at the single-trait level only when the nearest such variant lay more than 1 Mb away. This criterion reflects detection status within the specific input datasets rather than the absence of any prior report in the literature. As a separate step, previously reported associations for each lead SNP were independently retrieved from the GWAS Catalog to provide external biological context. Gene-level and gene-set tests were conducted with MAGMA [26], retaining results at a false discovery rate (FDR) below 0.05.

Bayesian fine-mapping of candidate causal variants

To distinguish causal variants from those in linkage disequilibrium, Bayesian fine-mapping was performed at each genome-wide significant locus. A 250 kb window flanking each lead SNP was defined using the echolocatoR R package (v2.0.3), within which two independent algorithms, SuSiE [27] and FINEMAP [28], each estimated per-variant posterior inclusion probabilities (PIPs) and constructed a 95% credible set. High-confidence causal status required two simultaneous conditions. A mean PIP exceeding 0.95 when averaged over the SuSiE and FINEMAP outputs, and membership in the 95% credible set returned by each algorithm independently.

Expression-mediated association analysis

To link variant-level genetic associations with gene expression, TWAS was conducted under the sparse canonical correlation analysis TWAS (sCCA-TWAS) framework [29]. sCCA-TWAS integrates the cis-eQTL weights of each gene across the 49 GTEx v8 tissues into a small number of composite features (sCCA1–sCCA3), each representing a cross-tissue co-regulation pattern rather than a single tissue. Because this aggregation does not resolve the tissue in which regulation occurs, we additionally performed single-tissue TWAS across GTEx v8 tissues for the prioritized effector genes, and evaluated colocalization between the latent-factor GWAS and cis-eQTL signals (a posterior probability PP4 > 0.75 taken as evidence of a shared causal variant). Each panel (sCCA1–sCCA3) was independently tested for association between genetically predicted expression and the cSVD latent factor, and genes reaching FDR < 0.05 in any panel were retained. These genes then entered the Bayesian causal gene fine-mapping pipeline of the FOCUS tool [30], which computes a PIP for each gene while accounting for local linkage disequilibrium and colocalization information. Genes with PIP above 0.8 were designated as high-confidence causal mediators.

Biological pathway annotation

To characterize the broader biological context of genes associated with the cSVD latent factor, gene set enrichment analysis (GSEA) was carried out using the Molecular Signatures Database (MSigDB) as the reference catalogue [31]. These analyses targeted genes prioritized by MAGMA from the Genomic SEM output, with all protein-coding genes serving as the background set for enrichment evaluation. Statistical significance was defined at FDR < 0.05.

Cell-type enrichment and regulatory element partitioning

The contribution of individual cell populations to the heritability of the cSVD latent factor was assessed with the CELLECT pipeline [32]. In brief, CELLEX first quantifies the expression specificity of each gene across cell types from single-cell RNA-sequencing input. These specificity scores are then combined with GWAS summary statistics in a stratified LDSC (S-LDSC) framework to determine whether genes preferentially expressed in a given cell type account for a disproportionate share of trait heritability. The cSVD latent factor summary statistics were paired with transcriptomic data from the Tabula Muris mouse single-cell atlas [33]. In a parallel analysis, S-LDSC under the baseline-LD model [34] partitioned SNP heritability across 53 functional genomic annotation categories, spanning histone modification marks, DNase I hypersensitive sites, and evolutionarily conserved elements, to quantify the relative contribution of each category to the genetic architecture of the cSVD latent factor.

Spatial developmental mapping

To map the anatomical distribution of cSVD-associated expression programs during embryonic development, the genetically informed spatial mapping algorithm (gsMap) was applied to project the latent factor’s genetic risk profile onto anatomically defined cell populations within intact embryonic tissue sections. A mouse embryonic tissue atlas at embryonic day 16.5 (E16.5_E1S1.MOSTA) served as the spatial reference, and cross-species inference was facilitated through a murine-to-human orthologue mapping file.

Results

Genetic covariance landscape of cSVD component traits

After quality control, between 1.10 and 1.18 million autosomal SNPs per trait were available for LDSC analysis. SNP heritability (h²SNP) estimates varied considerably across the six phenotypes (Table S2). WMH exhibited the highest value (0.25, SE = 0.024, Z = 10.6), with SBP close behind (0.143, SE = 0.006, Z = 24.2) and ICH at 0.163 (SE = 0.053, Z = 3.07). SVS and PVS both fell within the 0.06 to 0.07 interval, yet their LDSC intercepts diverged markedly. SVS registered an intercept of 1.073, suggesting mild confounding inflation in its source GWAS, whereas the PVS intercept of 0.980 sat below unity. Across all six phenotypes, LDSC intercepts fell between 0.98 and 1.07, indicating that the observed genomic inflation reflected genuine polygenicity rather than residual population structure. Pairwise genetic correlation analysis uncovered an asymmetric landscape of shared heritability (Fig. 2; Table S3). The correlations between SVS and ICH (rg = 0.835, SE = 0.276) and between SVS and PVS (rg = 0.872, SE = 0.290) approached saturation, suggesting that these cerebrovascular event phenotypes share genetic determinants to a substantial extent. WMH shared a moderate genetic correlation with SVS (rg = 0.519, SE = 0.129), while its correlations with ICH (rg = 0.297) and PVS (rg = 0.277) were comparatively modest. At the genetic level, SBP was most tightly linked to SVS (rg = 0.484), with positive but weaker correlations with PVS and ICH. This SVS-centered pattern of high genetic sharing provided the rationale for fitting a common-factor model in the subsequent analysis. To verify that the latent factor is driven by phenotypes directly relevant to cSVD, we refitted the common-factor model using only the four core neuroimaging and cerebrovascular phenotypes (WMH, PVS, ICH, SVS), excluding SBP and cognitive performance. This four-trait model showed excellent fit (CFI = 1.00, SRMR = 0.046; χ² = 1.25, df = 2, P = 0.53), with all four phenotypes loading significantly on a single common factor in the same direction and relative order as in the six-trait model, confirming that the shared genetic structure is robust to trait selection.

Fig. 2.

Pairwise genetic correlations among six cSVD-related phenotypes estimated by cross-trait LDSC. Notes: Each cell shows the genetic correlation (rg) between two phenotypes; colour intensity indicates the magnitude and direction of the correlation, with warmer colours denoting stronger positive correlation, and the numeric value in each cell is the rg estimate

Latent factor model and phenotype contributions

Fitting a confirmatory one-factor model to the genetic covariance matrix produced satisfactory results (Table S4). The χ² P value of 0.479 (χ² = 8.56, df = 9) indicated no perceptible departure of the model-implied covariance structure from the empirical observations. CFI reached the ideal benchmark of 1.00, SRMR was 0.061, and AIC was 32.56. Standardized factor loadings revealed a tiered contribution hierarchy (Table S5). SVS, PVS, and ICH served as the core indicators of the common factor, all with loadings reaching statistical significance (P < 10⁻⁶). WMH and SBP contributed moderate loadings (P < 10⁻⁶), while cognitive performance showed a modest but significant loading (P = 0.0095).

Genome-wide multivariate scan and novel locus discovery

When the validated factor structure was extended to a genome-wide multivariate scan, 1,423 SNPs exceeded the P < 5 × 10⁻⁸ threshold (Fig. 3; Table S6). After clumping at r² < 0.1, these signals resolved into 20 independent lead SNPs distributed across 14 distinct genomic loci (Table S7). The association of each lead SNP with all six constituent phenotypes is also reported in Table S7. Signal strength varied substantially across loci. rs34974290 at 17q25.1 ranked first (P = 3.81 × 10⁻³³), a genomic interval repeatedly linked to WMH volume, SBP, coronary artery disease, and white matter microstructure in prior studies (Table S8). rs7596872 at 2p16.1 constituted the second strongest signal (P = 2.79 × 10⁻²⁴), a locus reported across multiple independent WMH GWAS. rs769449 at 19q13.32 (P = 2.52 × 10⁻¹¹) mapped to the APOE genomic region, a locus with extensive prior associations spanning Alzheimer’s disease, lipid metabolism, inflammatory markers, and cerebrovascular disease. The remaining lead signals distributed across 6q25.1, 10q24.33, 2q33.1, 13q34, and 3q27.1 among other regions. Of the 20 independent lead SNPs, 2 (rs11838776 and rs8127114) did not reach genome-wide significance in any of the six constituent single-trait GWAS after applying the 1 Mb distance criterion (Table S9). This gain in single-trait-undetected signals confirms that multivariate integration improves discovery power even when individual GWAS coverage is already extensive.

Fig. 3.

Manhattan plot of the genome-wide multivariate association scan for the cSVD latent factor. Notes: Each point represents a SNP, plotted by chromosomal position (x-axis) against − log₁₀(P) for its association with the latent factor (y-axis). The horizontal dashed line denotes the genome-wide significance threshold (P = 5 × 10⁻⁸)

Causal variant prioritization through Bayesian fine-mapping

Joint Bayesian fine-mapping with SuSiE and FINEMAP refined the association signals from locus-level to single-variant resolution (Fig. 4; Table S10). Six variants satisfied the dual fine-mapping criterion of mean PIP ≥ 0.95 and membership in both credible sets. Among these, three lead variants additionally attained genome-wide significance in the multivariate scan and were designated as high-confidence causal variants, whereas the remaining three represent additional members of the same credible sets whose posterior probabilities reflect shared support within the local haplotype rather than independent association signals. Functional annotation through FUMA further distinguished the three causal variants. rs7596872 (2p16.1, PIP = 1.00) is located within an intron of EFEMP1 (CADD = 9.27) in a genomic segment with an active-transcription-associated chromatin state. rs8127114 (21q22.3, PIP = 1.00), the variant detectable only at the latent-factor level, maps to the 3′ untranslated region of ICOSLG (RegulomeDB = 5), indicating regulatory potential and providing a plausible post-transcriptional link between the fine-mapped signal and ICOSLG expression. rs4953671 (2p22.1, PIP = 1.00) lies in an intergenic region approximately 18.8 kb from AC098824.6 (RegulomeDB = 7), consistent with a regulatory role that awaits functional characterization.

Fig. 4.

Bayesian fine-mapping of three high-confidence causal loci (EFEMP1, ICOSLG, and AC098824.6). Notes: For each locus, the tracks show, from top to bottom, the multivariate GWAS association (− log₁₀(P)), the FINEMAP posterior inclusion probability (PIP), the SuSiE PIP, and the mean PIP across the two methods. Points are coloured by linkage disequilibrium (r²) with the lead variant; the horizontal dashed lines denote the genome-wide significance threshold (P = 5 × 10⁻⁸) for the association track and the high-confidence threshold (PIP = 0.95) for the fine-mapping tracks

Positional gene-level mapping via MAGMA

Aggregating SNP-level association signals to the gene level through MAGMA brought 101 genes past the FDR < 0.05 significance threshold (Fig. 5; Table S11). The three top-ranked genes mapped to the two genomic regions with the strongest genetic signals. SH3PXD2A (10q24.33, Z = 8.15, P = 1.82 × 10⁻¹⁶) encodes a scaffold protein involved in invadopodia formation and extracellular matrix remodeling. TRIM47 (17q25.1, Z = 7.75, P = 4.44 × 10⁻¹⁵) and WBP2 (17q25.1, Z = 7.69, P = 7.11 × 10⁻¹⁵) co-localized to the same genomic interval as the strongest multivariate GWAS lead signal (rs34974290), further consolidating the centrality of 17q25.1 in the cSVD genetic architecture. To evaluate the gene-detection gain afforded by multivariate integration, we additionally performed gene-based MAGMA analysis on each of the six constituent single-trait GWAS using identical parameters (Table S11). Of the 101 genes significant at the latent-factor level, 6 (PGBD1, PPP1R18, HIST1H2AA, DEGS2, HIST1H4A, TUBB) were not significant in any single constituent trait, whereas genes such as SH3PXD2A were significant in both the multivariate and single-trait analyses, confirming consistency between the two approaches.

Fig. 5.

Manhattan plot of MAGMA gene-level associations for the cSVD latent factor. Notes: Each point represents a gene, plotted by chromosomal position against − log₁₀(P); the horizontal dashed line denotes the FDR < 0.05 significance threshold, and the top-ranked genes are labelled

Expression-mediated gene prioritization

The sCCA-TWAS linked the cSVD latent factor to 114 genes at FDR < 0.05 (Fig. 6; Table 1), with a modest excess of downregulated expression signals (71 negative versus 43 positive TWAS Z-scores). Bayesian causal gene fine-mapping via FOCUS prioritized 13 genes at the high-confidence level (PIP > 0.8). The 17q25.1 region contributed the most prominent expression-mediated signals. TRIM65 exhibited the strongest TWAS association among all causal genes (Z = -9.40, P = 5.67 × 10⁻²¹, PIP = 1.00) and belongs to the same tripartite motif protein family as the MAGMA-captured TRIM47. UNK (Z = -7.10, PIP = 1.00) and GALK1 (Z = 10.99, PIP = 0.991) also mapped to this region. At 10q24.33, three genes formed a tight expression association cluster. SH3PXD2A (Z = 6.51, PIP = 1.00) provided independent cross-validation with the MAGMA results, NEURL1 (Z = 5.66, PIP = 1.00) encodes an E3 ubiquitin ligase involved in Notch signaling, and STN1 (Z = -5.70, PIP = 0.995) encodes a component of the CST telomere maintenance complex, suggesting that telomere biology may also contribute to the pleiotropic mechanisms at this locus. Additionally, BTN3A2 (6p22.1) achieved PIP above 0.99 in two independent panels, and HLA-C (6p21.33, PIP = 0.887) likewise cleared the causal gene threshold. KLHL24 (3q27.1) reached PIP = 1.00 in two panels but with opposing Z-score directions, a bidirectional expression pattern hinting at tissue-specific regulatory mechanisms. NBEAL1 (2q33.1, PIP = 0.857 and 1.00 across panels), RPL12P16 (2q33.1, PIP = 0.990), and ZSCAN9 (6p22.1, PIP = 0.942) also met the high-confidence causal gene threshold. Colocalization analysis between the latent-factor GWAS and cis-eQTL signals supported a shared causal variant (PP4 > 0.75) for the majority of prioritized effector genes, including NBEAL1, RPL12P16, KLHL24, ZSCAN9, BTN3A2, UNK, and SH3PXD2A. For a minority of genes at the highly polygenic 17q25.1 and MHC regions (e.g. TRIM65, GALK1, HLA-C), the analysis favoured two distinct causal variants (PP3 > 0.9), with a negligible probability of a GWAS-only signal, indicating genuine eQTL signal whose single-variant resolution was limited by local LD complexity and cross-tissue aggregation. To identify the tissues in which regulation occurs, we further performed single-tissue TWAS across GTEx v8 tissues for the effector genes (Table S12). The strongest associations localized to vascular and brain tissues, and the colocalization support for the PP3-dominant genes was recovered at single-tissue resolution.

Fig. 6.

Manhattan plot of sCCA-TWAS associations for the cSVD latent factor. Notes: Each point represents a gene, plotted by chromosomal position against − log₁₀(P); positive and negative TWAS Z-scores indicate genetically predicted increased and decreased expression, respectively, and the horizontal dashed line denotes the FDR < 0.05 significance threshold

Table 1.

High-confidence causal genes identified by sCCA-TWAS with FOCUS for the cSVD latent factor

sCCA Feature	Gene	CHR	Heritability Squared	TWAS Z	TWAS FDR	COLOC.PP3	COLOC.PP4	Focus PIP
sCCA1	NBEAL1	2	0.5012	-6.592552	4.32e-11	0.01	0.99	0.857
sCCA3	NBEAL1	2	0.4972	-6.84531	7.63e-12	0.01	0.99	1
sCCA2	RPL12P16	2	0.159	-6.633967	3.27e-11	0.005	0.994	0.99
sCCA1	KLHL24	3	0.2107	-5.72	1.04e-08	0.026	0.974	1
sCCA3	KLHL24	3	0.1366	5.64165	1.68e-08	0.022	0.978	1
sCCA1	HLA-C	6	0.7071	4.99022	6.03e-07	0.969	0.014	0.887
sCCA1	BTN3A2	6	0.792	-5.49452	3.92e-08	0.054	0.945	0.993
sCCA2	BTN3A2	6	0.739	5.3872	7.15e-08	0.054	0.946	0.999
sCCA2	ZSCAN9	6	0.1345	-6.5024	7.91e-11	0.024	0.976	0.942
sCCA1	TRIM27	6	0.4796	4.88902	1.01e-06	0.022	0.725	0.811
sCCA1	NEURL1	10	0.1477	5.6637	1.48e-08	0.996	0.003	1
sCCA2	STN1	10	0.0867	-5.69712	1.22e-08	0.024	0.684	0.995
sCCA3	SH3PXD2A	10	0.131	6.5149	7.27e-11	0.063	0.847	1
sCCA2	UNK	17	0.1006	-7.10093	1.24e-12	0.065	0.905	1
sCCA1	GALK1	17	0.12	10.986255	4.45e-28	0.899	0.017	0.991
sCCA3	TRIM65	17	0.1888	-9.39591	5.67e-21	0.97	0.03	1

Pathway enrichment landscape

Gene-set enrichment analysis characterized the functional profile of cSVD latent factor-associated genes (Table S13). Among Reactome canonical pathways, the most prominent signals involved epigenetic regulation, including HDAC-mediated histone deacetylation (FDR = 0.023), HAT-mediated histone acetylation (FDR = 0.039), DNA methylation (FDR = 0.023), and PRC2-mediated histone methylation (FDR = 0.039). The cellular senescence dimension was equally prominent, with SASP (FDR = 0.023) and cellular senescence (FDR = 0.023) both reaching significance. Enrichment of the amyloid fiber formation pathway (FDR = 0.023) pointed to a potential intersection with neurodegenerative pathology. Among GO molecular functions and cellular components, the structural constituent of chromatin (FDR = 4.63 × 10⁻⁴), DNA packaging complex (FDR = 7.04 × 10⁻⁴), and nucleosome (FDR = 1.75 × 10⁻³) were consistent with the epigenetic pathway cluster. GWAS Catalog gene-set analysis confirmed substantial overlap between the cSVD latent factor and prior WMH genetic findings (WMH, FDR = 2.43 × 10⁻²⁷), and additionally revealed significant intersections with brain morphology (MOSTest, FDR = 2.80 × 10⁻¹⁵), lacunar stroke (FDR = 6.05 × 10⁻⁶), Alzheimer’s disease (FDR = 9.09 × 10⁻⁴), coronary artery disease (FDR = 0.024), and Parkinson’s disease (FDR = 2.52 × 10⁻⁵). Positional analysis confirmed 6p22 (FDR = 2.14 × 10⁻²⁴) and 17q25 (FDR = 6.11 × 10⁻⁶) as the two core genomic bands concentrating cSVD genetic signals.

Cellular origins and regulatory element architecture of cSVD heritability

After CELLECT partitioned the heritability of the cSVD latent factor across cell types in the Tabula Muris atlas, 7 cell types showed significantly enrichment trends (P < 0.05; Table S14). Among these, 4 were vascular endothelial cells spanning multiple organs: lung endothelial cells (P = 0.0016), liver sinusoidal endothelial cells (P = 0.0082), heart endothelial cells (P = 0.015), and kidney endothelial cells (P = 0.044). Two were mural cells, namely brain pericytes (P = 0.0068) and heart smooth muscle cells (P = 0.016). An unclassified heart cell type also reached significance (P = 0.046). Partitioning the latent factor heritability across 53 baseline-LD functional annotations (Table S15) revealed that cSVD-associated genetic variation was concentrated in regulatory and constrained genomic elements rather than in broadly defined non-coding space. Fetal DNase I hypersensitive sites reached significance (fold enrichment = 3.66, FDR = 0.047), suggesting that a fraction of the cSVD genetic risk originates in regulatory programs active during early development. Among adult-tissue annotations, evolutionarily conserved sequences carried the largest heritability concentration (4.36-fold, FDR = 5.47 × 10⁻⁵), followed by super-enhancers (4.12-fold, FDR = 1.28 × 10⁻⁵), H3K9ac-marked active promoters (8.88-fold, FDR = 1.28 × 10⁻³), H3K4me1 enhancer peaks (5.94-fold, FDR = 0.015), and protein-coding exons (19.30-fold, FDR = 0.030).

Developmental tissue localization of cSVD genetic signals

The gsMap projected the cSVD latent factor genetic signals onto the mouse E16.5 embryonic spatial transcriptomic atlas (Table S16). The most significant spatial enrichment was observed in lung tissue (P = 5.49 × 10⁻⁶) and kidney (P = 8.28 × 10⁻⁶), followed by connective tissue (P = 1.90 × 10⁻⁵), mucosal epithelium (P = 2.01 × 10⁻⁵), meninges (P = 2.49 × 10⁻⁵), and muscle (P = 2.76 × 10⁻⁵). Brain tissue showed moderate enrichment (P = 0.018).

Discussion

In this study, we applied Genomic SEM to integrate GWAS summary statistics from six core phenotypes of cSVD and constructed a confirmatory one-factor model that adequately captured the shared genetic covariance among them. The model fit indices (CFI = 1.00, SRMR = 0.061) supported the existence of a coherent latent genetic factor, with SVS, PVS, and ICH serving as core components and WMH and SBP contributing moderate loadings. A genome-wide multivariate association scan based on this latent factor identified 1,423 significant SNPs corresponding to 20 independent lead SNPs across 14 loci, including 2 that were not detected at genome-wide significance in any constituent single-trait GWAS. Subsequent analyses integrating Bayesian fine-mapping, gene-level and transcriptome-wide association mapping, pathway enrichment, cell-type specificity, and spatial developmental profiling converged on vascular wall cell biology, epigenetic dysregulation, and cellular senescence as the central biological themes underlying the shared genetic susceptibility across cSVD-related phenotypes. Below, we discuss the mechanistic implications of these findings at the variant, gene, pathway, cell-type, and developmental tissue levels.

Leveraging the validated cSVD common-factor model [22], we performed Bayesian fine-mapping with SuSiE and FINEMAP to resolve the multivariate association signals from locus-level to single-variant resolution. Among the identified high-confidence loci, the EFEMP1 locus was particularly prominent, harboring two potential causal variants with near-perfect posterior support (rs7596872 and rs80110259). EFEMP1 encodes the extracellular matrix protein fibulin-3, and a prior whole-exome association study of extreme cSVD phenotypes had already identified common variants in the 5′ UTR region of this gene as significantly associated with cSVD severity, with Mendelian randomization supporting a causal effect [12]. At the population cohort level, the Framingham Heart Study confirmed that elevated circulating plasma EFEMP1 levels are independently associated with increased WMH burden, smaller hippocampal volumes, and elevated risk of all-cause dementia [35]. A recent genome-wide Mendelian randomization study further positioned EFEMP1 as a druggable gene for cSVD, exclusively detected in brain tissue, with cSVD risk factors mediating 5.5% to 18.5% of the total effects through this gene [36]. The second high-confidence variant, rs8127114, was detectable only at the latent-factor level and maps to the 3′ UTR of ICOSLG. ICOSL expressed on CNS endothelial cells promotes firm T cell adhesion to the BBB, and blocking it markedly impairs this adhesion [37], while multi-omics analysis of carotid endarterectomy specimens showed that ICOSLG expression at plaque sites independently predicts future adverse cerebrocardiovascular events [38]. Given that the core pathological features of cSVD include cerebral microvascular endothelial dysfunction and perivascular inflammatory cell infiltration, the causal signal at the ICOSLG locus suggests that costimulatory molecule-mediated endothelial-lymphocyte interactions may contribute to the shared genetic susceptibility. The third high-confidence variant, rs4953671, lies in an intergenic region near AC098824.6 whose function remains uncharacterized, though its PIP and prior associations of nearby variants with periventricular WMH and SBP suggest a vascular regulatory role awaiting functional elucidation. The fine-mapping analysis thus resolved the multivariate signals to single-variant resolution, with the three causal loci pointing to extracellular matrix remodeling and vascular wall integrity (EFEMP1), endothelial-lymphocyte immune costimulation and BBB permeability regulation (ICOSLG), and an uncharacterized vascular regulatory program (AC098824.6).

Gene-level analyses extended these variant-level findings to effector gene identification. The strongest cross-validated cluster mapped to 17q25.1, coinciding with the top multivariate lead signal. TRIM47 ranked among the top MAGMA hits, and TRIM65 exhibited the strongest TWAS association with perfect FOCUS causal probability. These two TRIM-family genes operate through complementary pathways. Endothelial-specific Trim47 deletion in mice increases BBB permeability via disrupted KEAP1/NRF2-dependent antioxidant signaling, with NRF2 activator treatment reversing both the BBB damage and cognitive deficits [39], while TRIM65 limits endothelial inflammation by targeting VCAM-1 for ubiquitin-mediated degradation [40]. At 10q24.33, SH3PXD2A occupied the top rank in both MAGMA and TWAS. This gene was identified as a lacunar-stroke risk locus classified under vascular extracellular matrix disruption alongside COL4A2 and HTRA1 [16], and its DNA methylation status tracks WMH burden with functional enrichment in BBB and immune response pathways [41]. At 2q33.1, NBEAL1 received dual MAGMA/TWAS validation and was previously associated with non-lobar ICH combined with SVS as a significant arterial cis-eQTL [42], with an independent WMH meta-analysis confirming this locus [43]. BTN3A2 (6p22.1) reached significance in both methods, supporting innate immune involvement, and MAGMA additionally identified APOE and APOC1 at 19q13.32, confirming a genetic intersection with CAA and lipid metabolism consistent with the Alzheimer’s disease gene-set enrichment. Collectively, these cross-validated genes implicate BBB integrity (TRIM47), endothelial inflammation (TRIM65), extracellular matrix remodeling (SH3PXD2A), vesicle trafficking (NBEAL1), innate immunity (BTN3A2), and lipid metabolism (APOE) as key dimensions of the cSVD genetic architecture.

The pathway enrichment results placed these gene-level findings within a broader biological framework. The most prominent cluster involved epigenetic regulation. HDAC9-mediated histone deacetylation promotes endothelial-mesenchymal transition in atherosclerotic plaques [44], and microglial HDAC3 loss promotes post-stroke oligodendrogenesis and white matter integrity [45], providing systems-genetics support for HDAC involvement in cSVD. The second cluster, cellular senescence, is equally relevant. Senescent endothelial cells accumulate under hypertensive and diabetic stress, exhibiting altered tight junctions, increased non-selective transcytosis, and enhanced inflammatory responses that lead to BBB disruption [46]. A single-cell transcriptomic study in a hypertensive mouse model showed that endothelial senescence emerged as early as 3 days after angiotensin II exposure, well before the cognitive impairment apparent at 42 days [47], positioning endothelial senescence as one of the earliest events in hypertension-driven small vessel injury. Enrichment of the amyloid fiber formation pathway converged with the APOE locus and the Alzheimer’s disease gene-set enrichment, implicating a genetic intersection between cSVD and CAA. GWAS Catalog analysis further revealed overlap with lacunar stroke, Parkinson’s disease, and coronary artery disease, indicating that cSVD genetic risk extends across the cardiocerebrovascular and neurodegenerative spectrum. These three clusters are not independent but may form a self-reinforcing pathological circuit in which epigenetic dysregulation accelerates endothelial senescence, and senescent cells release pro-inflammatory factors and matrix metalloproteinases through the SASP, which in turn promote vascular inflammation and amyloid deposition.

The cell-type enrichment results anchored these molecular findings to specific vascular populations. Of the 7 significant cell types in CELLECT, 4 were vascular endothelial cells spanning lung, liver, heart, and kidney, and 2 were mural cells (brain pericytes and heart smooth muscle cells). This multi-organ endothelial enrichment aligns with evidence that cSVD may reflect a systemic microvascular endothelial disorder. A Maastricht Study cohort demonstrated that an extracerebral microvascular dysfunction composite score was independently associated with WMH volume [48], and a systematic review of 23,056 participants confirmed that albuminuria is associated with all four neuroimaging markers of cSVD, explicitly concluding that the kidney and brain share microvascular pathology [49]. The brain pericyte enrichment is equally informative. Using sPDGFRβ as a pericyte injury marker, a landmark study showed that individuals with early cognitive dysfunction exhibit hippocampal capillary damage and BBB breakdown independent of amyloid-beta and tau pathology [50]. A direct functional link connects the MAGMA-identified APOE locus to this cell type. In a human iPSC-derived BBB model, calcineurin-NFAT signaling and APOE expression are selectively dysregulated in pericytes of APOE4 carriers, directly driving CAA pathology [51]. The convergence of gene-level evidence (TRIM47 for endothelial BBB maintenance, APOE for pericyte-driven CAA), pathway-level evidence (endothelial senescence, epigenetic regulation), and cell-type enrichment (endothelial cells and pericytes) points consistently to vascular wall cells as the core cellular substrate of the cSVD genetic architecture.

The gsMap analysis projected these findings onto a developmental spatial atlas, uncovering a multi-organ enrichment pattern led by lung and kidney, followed by connective tissue, mucosal epithelium, meninges, and muscle, while brain tissue showed only moderate enrichment. The top ranking of lung tissue converges with that of lung endothelial cells in CELLECT. A longitudinal UK Biobank cohort confirmed that low lung function is associated with greater WMH volume and predicts progressive WMH increases over 2.5 years [52], and microvascular dysfunction has been proposed as a common etiological basis placing cSVD and pulmonary arterial hypertension within the same spectrum of endothelial dysfunction [53]. Kidney ranked second. In a cohort of 875 ICH patients, CKD was independently associated with total cSVD burden, and CKD patients were more likely to exhibit a mixed SVD phenotype [54]. At the mechanistic level, an animal model demonstrated that CKD promotes cerebral microhemorrhages without elevating blood pressure, through microglial activation and uremic toxin-mediated impairment of cerebral microvascular endothelial barrier function [55]. The enrichment of meninges is pathophysiologically meaningful. Brain fluid clearance depends on vascular dynamics, perivascular space maintenance, and meningeal lymphatic drainage [56], and in cSVD patients the DTI-ALPS glymphatic index was independently and negatively associated with all four neuroimaging markers [57]. Brain tissue ranked below these mesenchyme-derived tissues, aligning with the pericyte enrichment in CELLECT and suggesting that the developmental genetic susceptibility to cSVD is embedded within vascular wall cell regulatory programs rather than neuron-specific pathways. This cross-layer convergence suggests that the genetic risk for cSVD may be pre-programmed during embryonic development through effects on vascular wall cell differentiation across multi-organ microvascular beds.

Several limitations of the present study should be noted. First, all input GWAS were restricted to European-ancestry cohorts, and the generalizability of our findings to other populations remains to be established. Second, the six constituent GWAS differ substantially in sample size and statistical power, which may influence the relative contributions of individual phenotypes to the latent factor. Third, the independent lead SNPs were defined using linkage-disequilibrium clumping (r² < 0.1), and a formal stepwise conditional analysis (e.g. GCTA-COJO) was not performed, which would provide a more rigorous confirmation of signal independence at loci harboring multiple lead SNPs. Fourth, other systemic risk factors relevant to cSVD, such as type II diabetes and chronic kidney disease, were not modelled here, and their integration into future multivariate frameworks represents a valuable direction for extending the present analysis. Finally, the gsMap analysis used a mouse embryonic spatial transcriptomic atlas at E16.5, and cross-species inference from murine to human developmental programs should be interpreted with caution.

Conclusion

The central finding of this study is that the multiple neuroimaging and clinical phenotypes of cSVD share a coherent latent genetic structure, with cerebrovascular event traits (SVS, ICH) and perivascular space burden (PVS) forming the core of this shared architecture. This genetic coherence is not merely statistical but biologically grounded. Convergent evidence across gene-level, pathway, cell-type, and spatial developmental analyses points to vascular wall cell biology as the unifying substrate, with epigenetic dysregulation and cellular senescence as central pathological mechanisms. The cross-validated effector genes at 17q25.1 (TRIM47/TRIM65), 10q24.33 (SH3PXD2A), 2q33.1 (NBEAL1), 6p22.1 (BTN3A2), and 19q13.32 (APOE) collectively implicate BBB integrity, vascular extracellular matrix remodeling, innate immunity, and lipid metabolism as key biological dimensions of the cSVD. The spatial developmental mapping further reveals that the genetic susceptibility to cSVD extends across multiple organ microvascular beds, with enrichment in lung, kidney, and meninges exceeding that of brain tissue itself, suggesting that the developmental origins of this syndrome are embedded within mesenchyme and neural crest-derived vascular programs rather than neuron-specific pathways. These results reframe cSVD from a collection of co-occurring neuroimaging markers to a set of phenotypes linked by a coherent, genetically grounded systemic microvascular liability, and nominate specific genes, pathways, and cell types as priorities for functional validation and mechanism-oriented therapeutic development.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Yan Zeng: Conceptualization, Methodology, Formal analysis, Data curation, Writing – original draft. Xiaoqian Huang: Data curation, Visualization, Writing – review & editing. YuBiao Yang: Supervision, Writing – review & editing. Chun Guo: Conceptualization, Supervision, Funding acquisition, Writing – review & editing.

Funding

This work was supported by the Hunan Provincial Natural Science Foundation (Grant No. 2026JJ81752), the Changsha Municipal Guidance Science and Technology Plan Project (Grant No. kzd2501107), and the Joint Foundation of Hunan University of Chinese Medicine (Grant No. 2025XYLH020).

Data availability

All GWAS summary statistics used in this study are publicly available from their respective repositories. Specific accession numbers for each dataset are provided in Table S1.

Declarations

Ethical approval

This study was conducted exclusively using publicly available GWAS summary statistics. No individual-level data were accessed or analyzed. All original GWAS included in this study had received ethical approval and informed consent from their respective institutional review boards. No additional ethical approval was required for the present analyses.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.