Cell type resolved MR based on brain single cell eQTLs corroborated by single cell RNA sequencing uncovers neuroimmune and vascular programs in intracerebral hemorrhage

Abstract

Background

Intracerebral hemorrhage (ICH) lacks effective neuroprotective therapies. We integrated cell type–resolved genetic inference with single-cell profiling to map putative causal programs and multicellular circuitry relevant to ICH.

Methods

Cis-eQTLs from eight human brain cell types were used as instruments for two-sample Mendelian randomization (MR), with an ICH meta-analysis from large biobanks and a stroke consortium as the outcome. Instruments were LD-pruned and restricted to strong variants (F > 10). Inverse-variance weighting (IVW) was the primary estimator, supported by robustness methods, heterogeneity/pleiotropy diagnostics, and false discovery rate control. Experimental validation used mouse collagenase ICH single-cell RNA-seq at 24 h (n = 3 sham; n = 3 ICH) with Seurat integration, composition testing, Slingshot pseudotime, and CellChat. An independent mouse cohort underwent qRT–PCR for selected genes.

Results

The ICH meta-analysis showed acceptable genomic control, supporting downstream MR. We identified 524 nominal gene–cell type associations, with a glia-weighted signal landscape. Enrichment implicated autophagy/mitophagy, antigen processing, cytoskeletal and vesicular trafficking, endothelial matrix–adhesion programs, ferroptosis, and myelin stress pathways. In mouse scRNA-seq, disease-associated microglia expanded with reciprocal loss of homeostatic microglia and increased neutrophils and T cells. Prioritized genes showed directional concordance; qRT–PCR confirmed ARPC3 and EIF2AK2 upregulation and TBCK and SPECC1 downregulation in ICH versus sham. Pseudotime supported a shift toward disease-associated microglial states, and CellChat indicated increased network interaction strength with microglia and endothelium as hubs.

Conclusions

Cell type–specific MR combined with single-cell validation highlights neuroimmune and neurovascular programs in ICH and links genetic signals to state transitions and inferred intercellular communication.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12967-026-07842-7.

Introduction

Intracerebral hemorrhage (ICH) carries a substantial global burden, with limited therapeutic options and poor outcomes, and its molecular pathobiology and druggable targets remain insufficiently defined [1]. As a major cause of stroke-related death and disability worldwide, ICH is characterized by a high incidence and high case-fatality and disability rates, persistent risks of recurrence and long-term dependency, and considerable medical and caregiving costs, with a particularly heavy burden reported in East Asian populations [2]. Current surgical and medical treatments focus on blood pressure control, hemostasis, and hematoma management, but they do not effectively halt secondary injury cascades that include hematoma toxicity, sterile inflammatory responses, disruption of the blood–brain barrier (BBB), and white matter injury [3]. The lesion milieu in ICH is highly heterogeneous, and bulk tissue omics often dilute cell type–specific and cross-cellular signals, obscuring actionable drivers and interactions. There is therefore an urgent need to reconstruct a cell-resolved mechanistic map of ICH at the scale of the neurovascular unit, with single-cell resolution and an explicit view of intercellular communication, to inform target discovery and biomarker development.

Prior studies indicate that microglial activation, including disease-associated microglia, BBB dysfunction, astrocytic reactivity, and stress within oligodendrocytes and oligodendrocyte precursor cells jointly contribute to the onset and progression of intracerebral hemorrhage, yet causal localization at the cell-type level and an intercellular communication map remain lacking [4–7]. In the microglial and disease-associated microglial compartment, upregulation of phagocytic and adhesion programs may amplify inflammation, participating in clearance of hematoma and necrotic tissue while also potentially driving secondary injury; however, direct anchoring of these executional pathways to genetic susceptibility has not been established [8]. In endothelial cells and pericytes, signaling through ANGPT–TIE2, VEGF pathways, and axes involving ICAM, VCAM, and selectins is closely related to BBB disruption and remodeling, but the links connecting these pathways to genetic risk signals and to specific cell states have not been systematically defined [9]. Astrocytic support via ApoE–LRP1 and via metabolic and autophagy programs is impaired under hemorrhagic or ischemic conditions, and the potential inflammatory “bridging” role of astrocytes in intracerebral hemorrhage has been underappreciated [10]. Methodologically, conventional genome-wide association and transcriptomic studies are largely conducted in bulk tissue and thus cannot resolve cell-specific effects; single-cell approaches reveal state heterogeneity but have rarely been integrated with causal inference and network-level communication analysis. Consequently, although the likely players have been identified, a framework that sequentially links causality, cell states, and intercellular communication is still needed.

To address critical gaps across causality, cell states, and intercellular networks in ICH, this study aims to build a testable mechanistic chain and reconstruct a disease map of the neurovascular unit. At the level of genetic causality, we leverage genome-wide association meta-analysis together with cell type–resolved single-cell eQTL–based Mendelian randomization. At the level of cell states, we analyze mouse brain single-cell transcriptomic composition and differential expression. At the level of intercellular communication, we apply CellChat to quantify information flow and sender–receiver relationships. Our innovation lies in anchoring genetic effects at the resolution of gene-by–cell type and situating the corresponding execution pathways within a systems context centered on disease-associated microglia. This approach not only clarifies which cell compartments and programs are driving pathology and how they interact, but also nominates druggable axes such as SPP1–integrin signaling, ANGPT–TIE2 signaling, and C3–C3AR1 signaling, together with translational biomarkers including a disease-associated microglia score, circulating adhesion and vascular factors, and cerebrospinal fluid chemokines.

Methods

Single-cell / single-nucleus brain eQTL data (exposure), instrument selection, LD pruning, and metabrain replication filter

We used the published single-cell/single-nucleus brain eQTL compendium as the exposure dataset [11]. The resource aggregates data from 192 donors and reports cis-eQTLs across eight major brain cell types—excitatory neurons, inhibitory neurons, oligodendrocytes, oligodendrocyte precursor cells (OPCs), astrocytes, microglia, endothelial cells, and pericytes. Cis-eQTL mapping was performed with fastQTL in a ± 1 Mb window around each gene transcription start site, with study-level and expression/genotype covariates as described in the original work; multiple testing was controlled at FDR = 0.05 to define significant eGenes. From the full summary statistics accompanying that study, we extracted, within each cell type, SNP–gene pairs (effect allele, non-effect allele, effect size, standard error, P value, and gene symbol) as candidate instrumental variables. Following our primary workflow, candidate SNP–gene pairs were first screened by nominal association (P < 0.05) within each gene × cell-type stratum, and then pruned for linkage disequilibrium using PLINK clumping (window = 100 kb, r² < 0.10, clump-P1 = 1) with a matched-ancestry reference panel constructed locally (MAF > 0.01) [12]. To strengthen instrument reliability and address concerns regarding cross-donor and cross-region stability, we applied an external replication filter using MetaBrain, a large meta-analyzed bulk brain eQTL resource [13]. Specifically, for each SNP–gene pair we aligned on chr: position: alleles + gene and retained variants only if they replicated in MetaBrain at nominal P < 0.05 with concordant effect direction. Instrument strength was subsequently assessed using the F statistic derived from the eQTL effect and its standard error (F = β²/SE²), and only instruments with F > 10 were retained to limit weak-instrument bias. The final instrument set therefore consisted of LD-independent, MetaBrain-replicated, direction-consistent, and strong variants carried forward to MR. Replication rates and sign concordance by cell type are summarized in Supplementary Table S1.

Allele harmonization and exclusion of strand-ambiguous palindromic variants

Exposure and outcome summary statistics were harmonized to a common effect-allele convention using the TwoSampleMR framework under a stringent action setting. Effect and non-effect alleles were aligned between the single-cell eQTL instruments and the ICH GWAS, and variants that could not be unambiguously matched were removed. Strand-ambiguous palindromic variants (A/T or C/G) were excluded when allele-frequency information was insufficient for strand disambiguation, and unresolved allele mismatches were filtered out. All coordinates and allele codes followed the conventions of the source datasets.

Outcome GWAS for intracerebral hemorrhage (ICH)

We conducted a genome-wide association meta-analysis of intracerebral hemorrhage (ICH) by aggregating summary statistics from FinnGen R12 (5,112 cases, 450,016 controls; 19,172,494 SNPs), UK Biobank (1,027 cases, 407,633 controls; 5,334,979 SNPs), and the International Stroke Genetics Consortium (ISGC; 1,545 cases, 1,481 controls) [14–17]. Per-study results were harmonized to a common genome build; effect and non-effect alleles were aligned to the forward strand; strand-ambiguous palindromic variants with allele frequency near 0.5 were removed; and variants failing study-level QC (e.g., low imputation quality or extreme standard errors) were excluded. After harmonization and filtering to MAF ≥ 0.01, a total of 9,482,104 SNPs were retained. Mixed-effects inverse-variance–weighted meta-analysis was performed in METAL, and between-study heterogeneity was assessed using Cochran’s Q and I². Genome-wide significance was defined at P ≤ 5 × 10⁻⁸. Quantile–quantile plots were generated with FUMA, and linkage disequilibrium score regression (LDSC) was used to estimate the LDSC intercept and SNP-based heritability on the liability scale. Regional association plots for lead loci were produced with the R package gassocplot using ancestry-matched LD.

Mendelian randomization, sensitivity estimators, robustness diagnostics, and multiple-testing control

Two-sample Mendelian randomization (MR) was conducted at the gene × cell-type level, treating cell-type–specific eQTL effects as exposures and the meta-analyzed ICH summary statistics as the outcome. The primary estimator was inverse-variance weighting (IVW) [18], complemented by MR-Egger regression, the weighted median, and simple and weighted mode-based methods as robustness estimators [19–22]. We reported effect sizes as odds ratios (OR = exp[β]) from IVW together with 95% confidence intervals and P values, and provided standard visualizations including MR scatter plots, single-instrument forest plots, and funnel plots. Between-instrument heterogeneity was quantified using Cochran’s Q under both IVW and MR-Egger models [23], and directional horizontal pleiotropy was assessed using the MR-Egger intercept. MR-PRESSO with 1,000 simulations was used to detect and, where appropriate, correct outlier-driven pleiotropy [24], and leave-one-out analyses were performed to evaluate the influence of any single instrument on the causal estimate [25]. To account for multiple parallel tests across all gene × cell-type MR analyses, IVW P values were adjusted using the Benjamini–Hochberg FDR procedure.

Sensitivity analysis: replication-filtered instruments across three metabrain brain regions

As a pre-specified sensitivity analysis, we repeated the MR pipeline after restricting instruments to SNP–gene pairs that replicated in MetaBrain cortex, basal ganglia, or hippocampus (P < 0.05) with concordant directions [13]. After allele harmonization, variants with F > 10 and LD independence (PLINK clumping: 100-kb window, r² < 0.10; ancestry-matched reference panel [12]) were retained. We re-estimated IVW as the primary estimator together with MR-Egger, weighted median, and mode-based methods, re-assessed heterogeneity (Q) and pleiotropy (Egger intercept), re-applied BH-FDR (global and within-cell-type), and repeated downstream enrichment analyses using the retained gene sets. Cell-type-level replication rate, sign concordance, and concordance summaries are reported in the Supplement.

Enrichment analysis (GO/KEGG)

Gene set enrichment analyses were performed in a cell-type–stratified manner using clusterProfiler. For each cell type, we constructed gene lists from the primary MR results (IVW/Wald ratio) and carried out Gene Ontology (GO, Biological Process) and KEGG pathway enrichment with Benjamini–Hochberg adjustment of enrichment P values. In response to multiple-testing concerns, enrichment results used for biological interpretation were based on the FDR-significant gene–cell-type associations (BH-FDR < 0.05) from the primary MR analyses, and we reported FDR-adjusted enrichment statistics (q values; default qvalueCutoff = 0.25 unless otherwise specified).

Because the number of FDR-significant genes within some cell types was small, enrichment was performed only when a minimum gene-set size was available (≥ 10 mapped Entrez IDs). When the FDR-significant set for a given cell type did not meet this minimum, we did not over-interpret pathway enrichment for that cell type; instead, we provided an exploratory enrichment based on nominal MR associations (P < 0.05) as hypothesis-generating results and explicitly labeled them as such, with all outputs cross-referenced to the corresponding FDR status.

Animals

All procedures were approved by the Institutional Animal Care and Use Committee of Guizhou Provincial People’s Hospital (Guizhou, China) and complied with NIH guidelines and ARRIVE recommendations. Male C57BL/6J wild-type mice (8 weeks; 22–25 g; Tengxin Biotechnology Co., Chongqing, China) were housed under controlled conditions (12-h light/dark cycle, ad libitum food and water) and acclimated for 7 days prior to experimentation. For the single-cell RNA-seq cohort, mice were randomly assigned to sham or ICH groups (n = 3 per group); additional animals from the same colony were reserved for parallel assays. All surgical and downstream analyses were performed blinded to group allocation.

ICH induction and tissue processing for single-cell RNA-seq

Mice were anesthetized with sevoflurane (induction 162 µM/L; maintenance 81 µM/L) and secured in a stereotaxic frame. After scalp preparation, the periosteum was cleared with 3% H₂O₂, and bregma and lambda were exposed and aligned. Type IV collagenase, dissolved in sterile PBS (total infusion volume 0.5 µL), was delivered into the right striatum at 0.25 µL/min using the following coordinates relative to bregma: anteroposterior + 0.8 mm, mediolateral + 2.5 mm (right), and dorsoventral − 3.5 mm; the needle was retained in place for 5 min to minimize reflux. Sham mice underwent identical procedures without collagenase infusion. Animals recovered in a warmed environment and were euthanized 24 h after surgery. To minimize blood-borne contamination, mice were transcardially perfused with ice-cold PBS under light anesthesia. The right cerebral hemisphere was rapidly dissected, meninges removed, and tissue gently minced on ice for enzymatic and mechanical dissociation into a single-cell suspension. After filtration (70 μm then 40 μm), red blood cell lysis, and debris/myelin removal, viable cells (≥ 85%) were counted and loaded onto the 10x Genomics Chromium platform (Single Cell 3′ v3.1 chemistry) targeting ~ 6,000–10,000 cells per sample, followed by Illumina NovaSeq sequencing according to the manufacturer’s recommendations. Raw data were processed with Cell Ranger (mm10 reference) to generate gene–barcode matrices for downstream single-cell analyses.

Single-cell transcriptomic validation (ICH versus sham)

To validate Mendelian randomization–prioritized genes at the transcriptional level, we analyzed in-house mouse brain single-cell RNA sequencing data from intracerebral hemorrhage and sham groups in R version 4.3.0 Seurat. Raw UMI count matrices were converted to Seurat objects; cells with fewer than 200 detected genes were removed. A second quality-control filter retained cells with more than 300 detected genes and a mitochondrial gene fraction no greater than 20%, and only genes detected in at least three cells were carried forward. To mitigate batch effects, we applied an anchor-based integration workflow with consistent normalization and variance stabilization using LogNormalize, selected the top 2,500 highly variable genes with SelectIntegrationFeatures, identified anchors with FindIntegrationAnchors, and generated an integrated expression matrix with IntegrateData. Principal component analysis was performed on the integrated matrix, retaining 30 components; the first 15 components were used to construct the shared nearest neighbor graph with FindNeighbors and to cluster cells with FindClusters at a resolution of 1.0. UMAP and t-SNE visualizations were computed using the same 15 components. Clusters were annotated by canonical markers, and clusters coexpressing markers from divergent lineages were considered potential doublets and excluded from downstream comparisons. Within each annotated cell type, differential expression between intracerebral hemorrhage and sham was tested primarily with Seurat’s FindMarkers using the Wilcoxon method, with a sample-level pseudobulk approach analyzed by DESeq2 as a robustness check. Statistical significance was defined as a Benjamini–Hochberg–adjusted P value less than 0.05 combined with an absolute log2 fold change of at least 0.25. For gene–cell type pairs prioritized by Mendelian randomization, we specifically evaluated concordance in effect direction between differential expression and causal inference, and visualized group differences with violin, box, and dot plots.

For microglia-focused analyses, we subset microglial cells and reclustered them, then scored each cell against disease-associated microglia and homeostatic gene sets using AUCell to derive orthogonal activity metrics. Pseudotime trajectories were inferred with Slingshot on the low-dimensional embedding to position cells along a continuum from homeostatic through transitional to disease-associated states; group-wise enrichment along pseudotime was assessed with Wilcoxon testing. Changes in cellular composition across states were quantified at the sample level with propeller and corroborated using robust t testing. These procedures enabled joint assessment of state transitions, composition shifts, and gene-level concordance with Mendelian randomization signals.

Intercellular communication networks were inferred from the single-cell data using CellChat version 1.6.1, which estimates the number and relative strength of signaling interactions among defined cell types.

TMT-based proteomics for ICH

We downloaded and reanalyzed the ProteomeXchange dataset PXD033791, which generated TMT-based quantitative proteomics profiling of perihematomal brain tissue collected 24 h after collagenase IV–induced ICH in young and aged mice [26]. Raw mass spectrometry data were processed in Proteome Discoverer (Thermo Fisher Scientific) using the SEQUEST HT search engine against the UniProt mouse proteome database. Peptide-spectrum matches and proteins were filtered at a FDR < 1% at both the peptide and protein levels, and protein identification required ≥ 1 unique peptide. TMT reporter ion intensities were normalized to the total signal to account for between-channel differences in loading and labeling efficiency, followed by log2 transformation. Differential protein abundance was assessed in R using limma, fitting linear models with empirical Bayes variance moderation to improve statistical power. Multiple testing was controlled using the Benjamini–Hochberg FDR procedure.

qRT–PCR

Total RNA was extracted from perihematomal brain tissue (24 h post–collagenase IV–induced ICH) using TRIzol reagent (Beyotime, Shanghai, China) according to the manufacturer’s instructions. RNA quantity and purity were assessed by spectrophotometry, and equal amounts of RNA were reverse-transcribed into cDNA using NovoScript^® Plus 1st Strand cDNA Synthesis SuperMix (Novoprotein Scientific Inc., Shanghai, China). Quantitative real-time PCR was performed on a real-time PCR system using SYBR High-Sensitivity qPCR SuperMix (Novoprotein Scientific Inc., Shanghai, China). Gene expression levels of ARPC3, EIF2AK2, TBCK, and SPECC1 were normalized to the internal reference gene GAPDH. Each sample was run in technical triplicates, and no-template controls were included to monitor potential contamination. Relative mRNA expression was calculated using the 2^−ΔΔCt method. Primer sequences are provided in Supplementary Table S2.

Results

High-quality meta-GWAS enables cell type–resolved MR, revealing a glia-weighted risk landscape in ICH

The intracerebral hemorrhage meta–genome-wide association analysis showed minimal inflation, with a genomic inflation factor equal to 1.061, no evidence of residual confounding by LD score regression as indicated by an intercept equal to 1.0021, and low SNP-based heritability estimated at 0.002; this dataset was therefore used for downstream Mendelian randomization analyses. In two-sample Mendelian randomization with inverse-variance weighting as the primary estimator and other robust methods as comparators, we identified 524 nominally significant gene–cell type associations with ICH risk at an unadjusted P value less than 0.05. The distribution across cell types was as follows: excitatory neurons 182 associations (34.7%), oligodendrocytes 134 (25.6%), inhibitory neurons 66 (12.6%), astrocytes 59 (11.3%), oligodendrocyte precursor cells 39 (7.4%), microglia 28 (5.3%), endothelial cells 11 (2.1%), and pericytes 5 (1.0%). Aggregated by lineage, neuron-related signals accounted for 47.3% (248 of 524), glial signals for 49.6% (260 of 524), and cerebrovascular cell types including endothelium and pericytes for 3.1% (16 of 524).(Supplementary Table S3-S45).

Pathway enrichment suggests convergent biological axes across cell types in ICH

Functional enrichment analyses indicated that proteins with putative causal links to intracerebral hemorrhage were organized along convergent biological axes that recurred across cell types. In astrocytes, signals concentrated on autophagy and mitophagy and nutrient responses (for example, ATG7, CTSK, CHKA), regulation of cell division and cytokinesis (CDC14B, PKN2, NUSAP1), fatty acid and xenobiotic metabolism (CYP4F12, CYP2D6), and peroxisomal and lysosomal transport (PEX6, CDIP1, CCZ1B), together with terms related to purinergic receptors and ion channels (GPR171, LRRC8A) and chromatin and centrosome assembly (CENPP). Nominal KEGG enrichment pointed to the cell cycle, tight junctions, motor protein–related processes, the pentose phosphate pathway, and categories related to autophagy. In endothelial cells, enrichment centered on the basement membrane and the extracellular matrix to receptor adhesion axis, including LAMC1 and ANGPTL1, accompanied by focal adhesion and blood–brain barrier homeostasis terms, consistent with early barrier injury and adhesion remodeling after hemorrhage. In excitatory neurons, Gene Ontology categories emphasized innate immunity and antigen processing and presentation (CTSS, HLA-DMA, HLA-F, UNC93B1, IRGM, AP3B1, AP4M1), together with endosomal and lysosomal trafficking and cytoskeletal coupling, which included ARPC3; KEGG trends suggested ferroptosis, lysosome, and antigen presentation. Inhibitory neurons showed nominal signals related to dopaminergic synapse, lysosome, apelin signaling, endocytosis and autophagy, and microtubule and ciliary structures. Microglia were dominated by immune receptor signaling and actin dynamics (BTN3A2, PTPRJ, CD2AP, ARPC3), with cellular component and molecular function terms indicating lamellipodium and ruffle structures and regulation by small GTPases and PI3K; KEGG pathways related to infection-like stimuli, DNA repair, and Hippo and circadian programs, in keeping with a disease-associated microglia–like activation state characterized by cytoskeletal remodeling and migratory phagocytosis. In oligodendrocytes, nominal enrichment centered on axon–myelin biology and stress signaling, including integrin, VEGFR, and MAPK routes, translation elongation, adherens junctions, and lysosomal membrane processes, suggesting myelin homeostasis under stress reprogramming. Oligodendrocyte precursor cells showed KEGG trends for complement and coagulation, ferroptosis, autophagy, and glycosphingolipid and folate metabolism, with Gene Ontology terms implicating class I major histocompatibility complex antigen processing, including ERAP2, regulation of coagulation via PROCR and KLKB1, and microtubule and spindle assembly involving CKAP2, SMC6, and SENP6. In pericytes, enrichment pointed to ERAP2-driven peptide antigen processing and metalloaminopeptidase activity, endoplasmic reticulum lumen localization, and kinetochore and centromere assembly with DNA damage responses exemplified by XRRA1 (Figs. 1 and 2; Supplementary Table S6-S7).

Fig. 1.

GO enrichment of MR-implicated genes across brain cell types in ICH. (A–H) Cnetplots views for astrocytes, endothelial cells, excitatory neurons, inhibitory neurons, microglia, oligodendrocytes, oligodendrocyte precursor cells (OPCs), and pericytes. Square nodes denote genes prioritized by two-sample Mendelian randomization; circular nodes denote enriched GO Biological Process terms (node size ∝ gene Count). Edges connect genes to their annotated GO terms. Colors group terms into manually curated functional themes (e.g., autophagy/lysosome, cytoskeleton/cytokinesis, ECM/adhesion, immune/antigen processing and presentation, metabolism/transport, chromatin/centrosome); mini-legends on the right indicate category membership. Only representative terms per cell type are displayed for visual clarity

Fig. 2.

KEGG pathway enrichment of MR-prioritized genes across brain cell types in intracerebral hemorrhage (ICH). (A–F) Cnetplots summarize KEGG pathways significantly enriched among genes with putative causal effects on ICH, resolved by cell type (A, astrocytes; B, endothelial cells; C, excitatory neurons; D, microglia; E, oligodendrocytes; F, oligodendrocyte precursor cells, OPC). Edges connect MR-prioritized genes (gray nodes) to enriched KEGG terms (colored spokes). Line colors distinguish individual pathways. Node size denotes the number of pathway memberships for a given gene (degree), and the accompanying dot scale encodes the multiple-testing–adjusted enrichment PPP value (larger dots indicate stronger enrichment). Only representative terms per cell type are displayed for visual clarity

Taken together, these cross-cell enrichment patterns align with core pathologies of intracerebral hemorrhage, including endothelial barrier disruption and extracellular matrix remodeling, neuroimmune activation and antigen presentation, cytoskeletal and vesicular trafficking reconfiguration, oligodendrocyte and myelin stress, coupling between coagulation and inflammation, and metabolic reprogramming. It should be emphasized, however, that most signals were nominal and did not broadly surpass false discovery rate thresholds; they should therefore be interpreted as directional leads pending validation and consolidation in larger samples, independent cohorts, and functional experiments.

sceQTL–MR FDR-significant findings and functional enrichment

After global multiple-testing correction across all gene × cell-type MR tests (IVW; Benjamini–Hochberg FDR), 91 gene–cell-type associations remained significant at FDR < 0.05 (Supplementary Table S8). Significant signals were enriched in neuronal and glial lineages, comprising 27 associations in excitatory neurons, 15 in oligodendrocytes, 14 in inhibitory neurons, 9 in oligodendrocyte precursor cells (OPCs), 8 in astrocytes, 8 in microglia, and 10 in endothelial cells; no FDR-significant associations were observed in pericytes under this stringent threshold. Several genes showed concordant signals across cell types. Notably, ICA1L reached global FDR significance in three cell types (excitatory neurons, astrocytes, and OPCs). IGHMBP2 was significant in both endothelial cells and microglia and exhibited effects in a consistent direction across these cell types.

Functional annotation of the globally FDR-significant gene sets (reported using q-values) revealed the most coherent enrichment patterns in vascular/immune-related cell types. GO terms were primarily driven by antigen processing and presentation (including MHC-related processes) and core programs related to protein synthesis and targeting. KEGG analyses further supported enrichment of immune–phagocytic/lysosomal pathways, including Antigen processing and presentation, Phagosome, and Lysosome, with the strongest signals observed in endothelial cells and microglia and related signals also detectable in OPCs. In contrast, enrichment results for astrocytes and neuronal lineages were comparatively sparse, although terms related to autophagy/lysosomal trafficking and cellular homeostasis remained detectable (Supplementary Tables S9–S10).

MetaBrain (cortex / basal ganglia / hippocampus)

• Using MetaBrain bulk brain eQTL as an external replication filter, we evaluated the cross-dataset stability of sc-eQTL instruments and re-estimated MR effects across three brain regions. Among 64,020 candidate sc-eQTL SNP–gene pairs tested per region, the nominal replication rates (P < 0.05) were 42.6% in cortex (27,265/64,020), 20.9% in basal ganglia (13,354/64,020), and 37.5% in hippocampus (24,029/64,020), with broadly comparable patterns across the eight cell types within each region. After enforcing effect-direction concordance and instrument QC (including strength and LD-independence), the final replication-filtered instrument sets retained 16,488 (cortex), 8,689 (basal ganglia), and 10,397 (hippocampus) independent instruments, supporting 6,357, 3,997, and 4,517 gene×cell-type MR tests, respectively. In the replication-filtered MR analyses (IVW; global BH-FDR), 39, 287, and 65 associations remained significant in cortex, basal ganglia, and hippocampus, respectively, with signals enriched in neuronal and oligodendroglial lineages; IGHMBP2 showed consistent evidence across all three regions, while ICA1L remained significant in multiple regions/cell types. Functional enrichment of FDR-significant gene sets revealed convergent immune–lysosomal axes, including antigen processing/presentation and phagosome/lysosome pathways, most prominently in vascular/immune-related cell types and extending to select glial/neuron strata (Supplementary Tables S11-S14).

Single-cell atlas of ICH versus sham brains and concordance with MR-prioritized genes

In our in-house mouse brain scRNA-seq dataset, stringent quality control and anchor-based integration generated a stable cellular atlas with broadly comparable sequencing depth and gene detection between the intracerebral hemorrhage (ICH) and sham groups (Supplementary Fig. 1). Total UMI counts correlated strongly with detected gene numbers, and mitochondrial read fractions did not show systematic group-level shifts. Clusters were annotated using canonical markers, including microglia (Aif1, P2ry12), disease-associated microglia (Itgax, Lgals3, Spp1), astrocytes (Gfap), oligodendrocytes (Mbp), endothelial cells (Kdr, Pecam1), ependymal cells (Foxj1), and T cells (Cd3d). Global clustering patterns were consistent with marker expression and UMAP separations (Fig. 3A).

Fig. 3.

Single-cell atlas of ICH versus sham brains and concordance with MR-prioritized genes (10x scRNA-seq, 24 h post-ICH; n = 3/group). (A) UMAP of all cells, colored by curated cell types (Astrocytes, Endothelial, Ependymal, Excitatory/Inhibitory neurons, Microglia incl. DAM, Neutrophils, Oligodendrocytes, OPC, T cells). Clusters were annotated by canonical markers (for example, Aif1/P2ry12 for microglia and Itgax/Lgals3/Spp1 for DAM; Gfap for astrocytes; Mbp for oligodendrocytes; Kdr/Pecam1 for endothelium; Foxj1 for ependyma; Cd3d for T cells). (B) Stacked bar charts showing sample-level cell-type composition in sham and ICH. (C) Differential abundance analysis (propeller with robust t-tests) reporting the proportion difference (ICH – sham) with 95% CIs for each cell type; bars in red denote FDR < 0.05. (D) Bubble summary of MR effects across cell types for prioritized gene–cell pairs: fill encodes MR log(OR), bubble size encodes –log10(FDR). (E) Concordance plot relating MR effect size (y-axis, beta/log[OR]) to single-cell differential expression (x-axis, log2FC ICH vs. sham). Points are colored by cell type and sized by expression percentile; representative genes are labeled. Together, the data show expansion of disease-associated microglia and infiltrating immune cells, with gene-level expression changes that broadly align with MR-inferred causal directions

Sample-level composition analyses (propeller with robust t testing) indicated marked immune-state remodeling accompanied by inflammatory infiltration after ICH. The disease-associated microglia compartment increased from ~ 0.20% in sham to ~ 12.85% in ICH (FDR ≈ 2.2 × 10⁻³), whereas homeostatic microglia decreased from 48.53% to 29.69% (FDR ≈ 1.6 × 10⁻⁴). Neutrophils rose from near-zero to 2.25% (FDR ≈ 1.1 × 10⁻³), and T cells increased from 0.35% to 6.46% (FDR ≈ 3.2 × 10⁻³). Endothelial and ependymal fractions were also higher in ICH, while astrocytes showed a modest decrease; these latter shifts were largely nominal (Fig. 3B, C).

Within each annotated cell type, we examined expression changes of MR-prioritized gene–cell pairs and evaluated directional concordance between the human genetic estimates and the acute (24 h) transcriptional response. In microglia, TBCK was downregulated in ICH, consistent in direction with MR estimates suggesting lower ICH risk with higher expression (OR ~ 0.85), whereas ARPC3 was upregulated in disease-associated microglia, directionally consistent with an MR risk-increasing estimate (OR ~ 1.08). In oligodendrocytes, EIF2AK2 was increased with an MR risk-increasing direction of similar magnitude, while SPECC1 was decreased, consistent with an MR protective direction (OR ~ 0.88). In astrocytes, CSDC2, SMIM8, ZKSCAN8, and CDIP1 were reduced, aligning in direction with MR estimates indicating lower ICH risk with higher expression and suggesting attenuation of astroglial homeostatic programs after hemorrhage. A subset of genes, exemplified by RASSF4, showed discordant directionality (MR risk-increasing estimate but lower expression in ICH), which is compatible with time- and state-dependent regulation as microglia transition from homeostatic to disease-associated phenotypes, and with the possibility that acute injury responses may not directly mirror lifelong genetically proxied effects (Fig. 3D, E; Supplementary Table S15).

Taken together, these scRNA-seq results indicate substantial post-ICH remodeling of cellular composition and gene expression, including expansion of disease-associated microglia, induction of cytoskeletal/phagocytic programs, concurrent vascular-stress signatures, and stress-associated regulation within the oligodendrocyte lineage, providing expression-level context for MR-prioritized candidates in the acute phase (Fig. 3; Supplementary Table S15).

Microglial trajectory toward a disease-associated state with concordant gene-level signals

After subsetting microglia and reclustering, we quantified homeostatic and disease-associated microglial (DAM-like) programs using both AddModuleScore and AUCell. Compared with sham, ICH microglia displayed a consistent shift toward a disease-associated transcriptional profile (Fig. 4A). Slingshot pseudotime inference positioned microglia along a dominant continuum spanning homeostatic, transitional, and disease-associated states (Fig. 4B–E). ICH-derived cells were significantly enriched at later pseudotime positions, with a higher median pseudotime than sham (Wilcoxon test, P < 0.001), whereas most sham microglia remained concentrated at the earliest state. State proportion estimates further supported this redistribution in ICH (~ 53% disease-associated, ~ 30% transitional, and ~ 17% early), consistent with a pronounced activation shift following hemorrhage. Concordantly, propeller-based composition analysis indicated an expansion of the disease-associated compartment with a reciprocal reduction of homeostatic microglia (Fig. 4F), and ICH cells again occupied later pseudotime positions than sham (Fig. 4G).

Fig. 4.

Microglial trajectory toward a disease-associated state with concordant gene-level signals (10x scRNA-seq, 24 h post-ICH; n = 3/group). (A) Violin plots of disease-associated microglia (DAM) module activity scored by AddModuleScore and AUCell show a shift toward DAM programs in ICH versus sham. (B–E) Slingshot pseudotime inferred on reclustered microglia orders cells along a principal path from homeostatic (early) through transitional to DAM (late); panels show pseudotime, discrete states, group labels, and cell-type labels, respectively. (F) Stacked composition of pseudotime states by group indicating enrichment of DAM (≈ 53%) and transitional (≈ 30%) states and depletion of early/homeostatic cells (≈ 17%) in ICH. (G) Violin plot of pseudotime by group showing later positions for ICH microglia (Wilcoxon P < 0.001). (H) Dot-plot summary of marker expression across states highlights preferential activation of DAM programs in ICH. (I) Pseudotime heatmaps for representative MR-prioritized genes: ARPC3 (risk-increasing in MR) is upregulated toward the DAM end of the trajectory; TBCK (protective in MR) shows low overall expression with downregulation in ICH; RASSF4 is relatively higher in non-DAM/early states and lower in late DAM, consistent with time- and state-dependent dynamics

At the gene level, ARPC3, which showed a risk-increasing direction in MR, exhibited higher expression in the disease-associated microglial subset. In contrast, TBCK, with a protective MR direction, showed generally low microglial expression and was reduced in ICH. RASSF4 was comparatively lower in the late disease-associated state but higher in non–disease-associated microglia, suggesting potential state dependence across the trajectory. Together, these analyses indicate that acute ICH is associated with a marked microglial transition toward disease-associated programs and with gene-level patterns that are broadly directionally concordant with MR-prioritized candidates, while also highlighting possible time- and state-dependent regulation for selected signals such as RASSF4 (Fig. 4H, I).

Remodeling of cell–cell communication networks in ICH

CellChat analysis suggested a marked reconfiguration of intercellular communication in ICH relative to sham, with increases in both the number of inferred interactions (ICH 1,217 vs. sham 757) and overall interaction strength (ICH ~ 25.3 vs. sham ~ 20.7) (Fig. 5A). Network centrality metrics indicated a redistribution of sender–receiver roles: disease-associated microglia showed higher outgoing and incoming signaling strength and ranked among the most connected populations, while endothelial cells maintained prominent sender activity. Astrocytes remained major receivers with detectable outgoing contributions, and several parenchymal populations exhibited generally elevated outgoing strength in ICH (Fig. 5B). Consistent with these shifts, network topology analyses pointed to DAM-like microglia as key hubs and suggested increased endothelial participation, including reinforced self-loop signaling compatible with a barrier-stress context (Fig. 5C).

Fig. 5.

Remodeling of cell–cell communication networks in ICH (CellChat on 10× scRNA-seq, 24 h post-ICH; n = 3/group). (A) Global communication summary showing more inferred interactions (ICH 1,217 vs. sham 757) and greater overall interaction strength (ICH ≈ 25.3 vs. sham ≈ 20.7) in ICH. (B) Network centrality. Disease-associated microglia (DAM) become a major hub with increased outgoing and incoming strength; endothelial cells remain strong senders; astrocytes are prominent receivers with appreciable output. (C) Outgoing (sender) pathway heatmaps by cell type in sham and ICH highlight amplification of inflammatory/adhesion/vascular axes in ICH (e.g., TNF–TNFR, CCL/CXCL–CCR/CXCR, ICAM/VCAM–integrins, SELP–SELPLG, SPP1–integrins, ANGPT–TEK/VEGF). (D) Incoming (receiver) pathway heatmaps by cell type show heightened endothelial and microglial receipt of pro-inflammatory and barrier-stress signals in ICH, with reduced homeostatic trophic inputs. (E) Differential pathway dot plots (ICH vs. sham). Left: pathways increased in ICH (examples above); Right: pathways decreased in ICH, including ApoE–LRP1, laminin/ECM-supportive modules, and other neurotrophic/metabolic signals. Together, the networks indicate a shift toward DAM-driven pro-inflammatory emission, reinforced endothelial crosstalk, and loss of astroglial homeostatic signaling

At the pathway level, differential signaling patterns favored inflammatory, adhesion, and vascular-remodeling programs—TNF–TNFR, CCL/CXCL–CCR/CXCR, ICAM/VCAM–integrins, SELP–SELPLG, SPP1–integrins, and ANGPT–TIE2/VEGF—whereas several neurotrophic/metabolic and homeostatic modules (e.g., ApoE–LRP1 and laminin/ECM-supportive signals) showed relative attenuation. Astrocyte-centered signaling displayed a similar shift, with reduced ApoE–LRP1 and increased bridging along C3–C3AR1 (astro → microglia) and chemokine axes, broadly concordant with the transcriptomic trends in DAM-like microglia (e.g., higher ARPC3) and reduced astroglial protective signatures (e.g., lower TBCK-related programs) (Fig. 5D–F). Overall, these CellChat-inferred network changes provide a systems-level context for the observed cell-state transitions and MR-prioritized candidates, while remaining consistent with the inference nature of expression-based communication modeling.

TMT protein signals

A total of six proteins showed directionally consistent changes in the ICH group relative to sham: Itgam, Itgav, Itgb2, and Arpc3 were upregulated, whereas Tbck and Specc1 were downregulated. Overall effect sizes were modest (log2FC approximately − 0.041 to 0.137), and none of the proteins in this panel reached the FDR < 0.05 significance threshold. Notably, Itgam exhibited the largest increase (log2FC = 0.137, nominal P = 0.068), suggesting a potential early post-ICH trend toward protein-level changes involving immune/adhesion-related molecules; however, given the current sample size and multiple-testing adjustment, these findings are best interpreted as directionally supportive / trend-consistent evidence rather than definitive significance (Supplementary Fig. 2).

qRT–PCR validation

In an independent mouse cohort, qRT–PCR indicated that, relative to sham, the ICH group showed higher ARPC3 and EIF2AK2 mRNA levels and lower TBCK and SPECC1 expression. These trends were consistent in direction with the patterns observed in our single-cell analyses and provide additional, gene-level support for these candidates at 24 h post-ICH (Supplementary Fig. 3).

Discussion

Under minimal genomic inflation (genomic inflation factor 1.061) and an LD score regression intercept close to unity (1.0021), we integrated genome-wide association meta-analysis with cell type–resolved single-cell eQTL–based Mendelian randomization, mouse brain single-cell RNA sequencing profiling, and CellChat-based inference of intercellular communication. Collectively, these layers are compatible with a coherent and testable framework in which genetically anchored, cell-type–specific expression proxies nominate candidate pathways that show broad consistency with injury-associated cell-state remodeling and altered cross-cell signaling after intracerebral hemorrhage. Rather than establishing definitive causality for individual genes or ligand–receptor pairs, the results provide partially convergent patterns across modalities that help prioritize candidate axes for follow-up validation. Across this triangulation, three themes are suggested: (i) genetic signals are distributed across both neuronal and glial compartments, with some biologically interpretable enrichment in glia and vascular-related cells; (ii) expansion of disease-associated microglia and infiltration of peripheral immune cells in the acute model occur alongside endothelial stress signatures and extracellular matrix remodeling; and (iii) astrocytic homeostatic metabolic/autophagy-related programs tend to be reduced while inflammatory and immune-bridging transcriptional features increase. Together, these observations support a tentative narrative relating genetic prioritization to cell-state changes and inferred network rewiring in ICH.

Convergent multimodal patterns are consistent with the possibility that ICH involves a coordinated shift across microglia, endothelium, astrocytes, and the oligodendrocyte lineage, including oligodendrocyte precursor cells. Along the microglial/disease-associated microglia axis, CellChat indicates higher inferred information flow in adhesion and inflammatory programs, including SPP1–integrin, TNF–TNFR, and ICAM/VCAM-related routes, a pattern compatible with enhanced motility, phagocytosis, and inflammatory amplification. In parallel, ARPC3 was prioritized by MR with a risk-increasing direction and showed higher expression in disease-associated microglia in the acute dataset, implicating actin assembly and phagocytic execution as plausible downstream components of genetically anchored programs [27]. In endothelium and pericytes, strengthened ANGPT–TIE2 and VEGF-related signaling and increased selectin/integrin-associated routes—together with more pronounced inferred endothelial autocrine activity and receptor-side input—are consistent with blood–brain barrier stress and extracellular matrix remodeling within the neurovascular unit [28]. Astrocytes exhibit a bidirectional reconfiguration in which ApoE–LRP1 and metabolic/autophagy-support programs are reduced whereas C3–C3AR1 and CCL/CXCL chemokine programs increase, suggesting a shift away from homeostatic support toward a more inflammatory bridging role; this directionality aligns with downregulation of TBCK in the acute model, which may indicate attenuation of protective buffering capacity [29]. In the oligodendrocyte/OPC axis, EIF2AK2 (PKR) showed a risk-increasing MR direction while SPECC1 showed a protective direction; together with enrichment trends involving translational stress and complement/coagulation/ferroptosis-related programs, these findings are compatible with dual constraints on myelin stability driven by stress translation and inflammation–coagulation coupling [30]. At the network level, CellChat further suggests redistribution of sender/receiver roles, with stronger inferred emission from disease-associated microglia, enhanced endothelial receiving and autocrine loops, and reduced protective astroglial output alongside increased inflammatory bridging. Importantly, these network-level patterns should be viewed as inferred and hypothesis-generating, yet they are broadly concordant with the MR prioritization and cell-state shifts observed in the single-cell atlas.

Our findings extend current understanding of ICH by adding cell type resolution and an intercellular network context to genetic prioritization. First, classic observations of microglial activation, enhanced phagocytic and migratory capacity, and blood–brain barrier injury are here contextualized by cell-type–resolved MR that anchors candidate genes to specific cellular compartments and by network inference that highlights coordinated pathway-level changes. Specifically, cell type–specific MR prioritizes cytoskeletal/adhesion-related candidates in disease-associated microglia, exemplified by ARPC3, while CellChat suggests that pathways such as SPP1–integrin, ICAM/VCAM–integrin, and TNF–TNFR may be amplified during ICH, providing testable hypotheses for cross-cell propagation [31]. Second, we highlight an underappreciated axis involving attenuation of astroglial metabolic/autophagy support concurrent with a shift toward inflammatory bridging. Decreases in ApoE–LRP1 and metabolic/autophagy-associated outputs alongside increases in C3–C3AR1 and chemokine signaling, aligned with TBCK downregulation, are consistent with a transition from homeostatic support toward an immune-amplifying astroglial state [32]. Regarding white matter pathology, translational stress mediated by EIF2AK2 together with coupling among complement, coagulation, and ferroptosis pathways may plausibly hinder remyelination and repair, in line with the recognized vulnerability of white matter after hemorrhage [33]. Compared with bulk tissue or peripheral biomarker studies, this integrative approach provides cell-specific candidate programs and a network-level context that may help reconcile otherwise heterogeneous observations; nonetheless, the inferences remain contingent on the strengths and limitations of each modality.

Within a two-sample Mendelian randomization framework, we applied instrument selection using F statistics greater than 10, allele harmonization, and sensitivity diagnostics including Cochran’s Q, the MR-Egger intercept, MR-PRESSO, and leave-one-out analyses. Inverse-variance weighted estimates served as the primary analysis, and P values were controlled using the Benjamini–Hochberg false discovery rate procedure. Under FDR control, a smaller subset of gene–cell type associations remained statistically supported relative to the broader nominal set; accordingly, we interpret FDR-significant signals as higher-confidence leads and treat nominal associations as hypothesis-generating. In addition, external replication–filtered sensitivity analyses using bulk eQTL resources across brain regions served as a supplementary check on instrument stability, while also emphasizing that instrument availability and effect estimates can vary by tissue context and technical platform. Taken together, these considerations support an emphasis on coherent directionality and cross-modal consistency rather than strong claims about any single molecular target.

Multimodal evidence offers generally concordant support for several candidate axes. ARPC3 shows a risk-increasing direction in MR and higher expression within disease-associated microglia, and aligns with increased inferred activity in adhesion and cytoskeletal communication programs such as SPP1–integrin, ICAM/VCAM–integrin, and TNF–TNFR signaling. TBCK shows a protective MR direction and is downregulated in ICH, concordant with decreased astroglial ApoE–LRP1 signaling and reduced metabolic/autophagy-support signatures. Endothelium-directed signaling, including ANGPT–TIE2, VEGF-related routes, and selectin/integrin-associated programs, tends to be higher overall, consistent with neurovascular unit stress and extracellular matrix remodeling. At the same time, divergences for individual genes are compatible with time- and state-dependent regulation along trajectories from homeostatic to transitional to disease-associated microglia. Such patterns highlight the need for time-resolved profiling and targeted perturbation studies to clarify whether candidates primarily influence susceptibility/thresholds versus secondary responses during acute injury.

The neurovascular communication map suggests a set of testable intervention hypotheses and candidate biomarker directions. Among pathways that may warrant prioritization for mechanistic follow-up are SPP1–integrin, TNF–TNFR, chemokine (CCL/CXCL) axes, ANGPT–TIE2 signaling, TAM receptor pathways including AXL and MERTK, and C3–C3AR1 signaling; these programs span cytoskeletal/adhesion execution, inflammatory amplification, and regulation of vascular barrier integrity and clearance. For restoration of protective programs, ApoE–LRP1 and astroglial metabolic/autophagy modules linked to TBCK- and mTOR/autophagy-related biology represent plausible buffering mechanisms that could be explored experimentally. Candidate biomarkers suggested by these analyses include disease-associated microglia–aligned transcriptional scores, circulating adhesion/vascular factors, and cerebrospinal fluid chemokines; however, their translational value will depend on independent replication and orthogonal validation.

This work is strengthened by methodological breadth and triangulation across orthogonal modalities, but several limitations constrain inference and motivate future work. First, single-cell/single-nucleus eQTL resources remain limited in sample size and may vary across donors, brain regions, and technical platforms; instrument availability and estimated effects can therefore be sensitive to dataset-specific features. Second, although cell-type–resolved eQTLs reduce confounding by cellular composition, external replication using bulk brain eQTLs cannot fully disentangle cell-composition effects from true cell-intrinsic regulation, and residual confounding cannot be excluded. Third, while FDR control identifies a subset of higher-confidence MR associations, many signals remain nominal; pathway enrichment based on nominal gene sets should therefore be interpreted as exploratory, whereas enrichment based on FDR-significant sets is more conservative but may be underpowered in cell types with fewer discoveries. Fourth, human MR reflects lifelong genetically proxied expression differences, whereas the mouse scRNA-seq captures an acute (24 h) injury response; directional concordance should not be over-interpreted as a direct causal chain, as some MR-prioritized genes may act on susceptibility or early injury thresholds while the mouse data reflect downstream reactive programs. Fifth, the scRNA-seq analysis is based on n = 3 mice per group, which limits power for composition, differential expression, and communication inference; we therefore emphasize consistency of trends and prioritize candidates showing reproducible directions across replicates and analytic strategies. Finally, CellChat infers putative signaling changes from expression patterns and curated ligand–receptor priors and does not establish functional signaling; protein-level and spatial validation will be required to determine whether highlighted interactions correspond to true intercellular signaling changes. These limitations also underscore the importance of expanded single-cell eQTL resources across ancestries and brain contexts and targeted experimental validation in independent cohorts.

Taken together, the data are consistent with a model in which genetic liability in intracerebral hemorrhage is plausibly linked to cell-state remodeling and inferred intercellular communication changes that involve disease-associated microglia, neurovascular stress programs, and shifts in astroglial support biology, alongside contributions from neuronal and oligodendrocyte-lineage candidates. This integrative framework provides a prioritized and testable set of hypotheses for early barrier protection, microglial modulation, and promotion of repair, and it highlights clear directions for replication, time-resolved profiling, and orthogonal functional validation in future studies [34].

Supplementary Information

Below is the link to the electronic supplementary material.

Author contribution

Conceptualization: SY, JQZ, YT, and YJH. Formal analysis: SY, XQR, MD, and YJH. Methodology: SY, JCL, and GQH. Investigation: SY, JQZ, YT, and YJH. Visualization: SY, TX, TL, and YJH. Funding acquisition: JQZ, YT, and GQH. Project administration: YT and YJH. Supervision: JQZ, YT, and YJH. Writing—original draft: SY. Writing—review & editing: All authors. All authors read and approved the final version of the manuscript, SY, JQZ, YT, and YJH have accessed and verified the underlying data.

Funding

This work was supported by the National Natural Science Foundation of China (82360376 and 82360482), Guizhou Provincial Science and Technology Projects (ZK[2024]483), Young Elite Scientist Sponsorship Program By GAST (GASTYESS202414), Guizhou Province’s funding for the cultivation of high-level innovative talents through the Thousand Talents Program (gzwjrs2023-001 and gzwjrs2024-007), and Guizhou Provincial People’s Hospital Youth Fund (GZSYQN202202).

Data availability

Summary statistics from the intracerebral hemorrhage meta analysis and our in house mouse single cell RNA sequencing dataset will be made available by the corresponding author upon reasonable request and subject to institutional approvals. The exposure data for Mendelian randomization comprise cell type specific cis eQTLs from eight human brain cell types reported by Bryois J, et al. Analysis code and processed MR outputs will be shared with requesters and released at publication.

Declarations

Ethics approval and consent to participate

The ethics approval was obtained from the Ethics Committee of Guizhou Provincial People’s Hospital. All animal experiments were conducted in accordance with relevant ethical guidelines and regulations.

Consent for publication

Not applicable.

Conflict of interest

Authors declare that they have no competing interests.