Overview of the study

Dysregulation of gene expression in specific brain cell types is a hallmark of psychiatric and neurological disorders [12,13,15,16,29]. To systematically infer cell type–specific causal relationships between gene expression and brain-associated complex phenotypes, we performed two-sample Mendelian randomization (MR) by integrating GWAS summary statistics for 26 disorders and behaviors (DBs) and 123 imaging-derived phenotypes (IDPs) with a recently published cell type–specific cis-eQTL atlas encompassing eight brain cell types [12] (S1–S3 Tables). The eight cell types were astrocytes, endothelial cells, excitatory neurons, inhibitory neurons, microglia, oligodendrocytes, oligodendrocyte precursor cells (OPCs), and pericytes. The 26 DBs comprised three groups: 8 behavioral-cognitive phenotypes, 8 neurological disorders, and 10 psychiatric disorders (mean GWAS sample size: 284,168; S1 Table). The 123 IDPs comprised two groups: 101 brain regional volumes and 22 white matter microstructure phenotypes measured by mean fractional anisotropy (mean GWAS sample size: 22,072; S2 Table). The overall study design is depicted in Fig 1a.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Study workflow and cell type–specific causal eGene counts across brain-associated phenotypes.

(a) Workflow. The cis-eQTL data from eight brain cell types were integrated with GWAS summary statistics for 123 IDPs and 26 DBs. TWAS pre-screening was followed by cell type–specific Mendelian randomization to infer putatively causal eGenes for 149 phenotypes, classified into two classes (IDPs and DBs) and five groups (psychiatric disorders, neurological disorders, behavioral–cognitive phenotypes, brain regional volumes, and white matter tracts). Downstream analyses included shared eGene characterization, single-cell expression profiling, and causal route reconstruction. Brain region schematics adapted from the Desikan-Killiany Atlas [34]; white matter tract schematics from the Atlas Track [35]. (b) Heatmap of cell type–specific causal eGene counts across IDPs (only IDPs with >3 eGenes are shown). (c) Heatmap of cell type–specific causal eGene counts across DBs.

https://doi.org/10.1371/journal.pcbi.1014422.g001

Given evidence of bidirectional causal effects between brain disorders and imaging phenotypes [20,30], we first screened for potential eGene–phenotype associations using transcriptome-wide association study (TWAS) [31], retaining 43,494 eGene–DB pairs and 284,927 eGene–IDP pairs at nominal significance (P < 0.05) for subsequent MR analysis. To ensure instrument validity, we selected as instrumental variables (IVs) only those single nucleotide polymorphisms (SNPs) that achieved genome-wide significance (P < 5 × 10^-8) in the cis-eQTL data. We manually verified the sample descriptions of all datasets and confirmed the absence of participant overlap between the cis-eQTL catalog and the GWAS cohorts.

To enhance the robustness of causal inference, we applied four complementary MR methods: SMR [6], Wald ratio [32], PMR-Egger [8] and GSMR [33]. These methods differ in their IV selection strategies (single-SNP versus multi-SNP; independent versus correlated SNPs), tolerance for horizontal pleiotropy, and analytical assumptions (S1 Fig and Methods), thereby providing convergent evidence for each inferred association. We then assessed the reliability of the predicted cell type–specific causal eGenes using external functional genomic datasets from the human brain. Next, we characterized the pleiotropic landscape of these causal eGenes across phenotype classes and groups, reconstructed putative causality routes linking cell type–specific eGenes to DBs and IDPs, and investigated the spatiotemporal expression dynamics of these genes using independent single-cell data.

Putative causal effects of eGenes on IDPs and DBs

We identified 254 eGenes with putative causal effects on 112 IDPs across eight cell types (FDR < 0.05, corrected across all eGene–cell type–IDP tests), yielding 760 significant eGene–cell type–IDP combinations (Fig 1b and S4 Table). Among these, 184 and 112 eGenes were implicated in 90 brain regional volumes and 22 white matter microstructure phenotypes, respectively. To assess reproducibility, we replicated the MR analyses using cortical eQTLs from BrainMeta [36] and cell type–specific eQTLs from brainSCOPE [37]. Of the eGene–IDP pairs, 64.9% (430/663) and 49.2% (326/663) were replicated in BrainMeta at nominal and FDR < 0.05 levels, respectively (both P < 10^-16, hypergeometric test; S5 Tableand S2 Fig). For eGene–cell type–IDP triples, 41.9% (315/751) and 33.7% (253/751) were replicated in brainSCOPE at nominal and FDR < 0.05 levels, respectively (both P < 10^-16, hypergeometric test; S5 Table and S3 Fig).

These associations were strongly cell type–specific: 91.4% (605/662) of eGene–IDP pairs were significant in only one cell type, consistent with the notion that brain structure is shaped by distinct cellular programs [38–40]. This specificity persisted after accounting for eQTL availability—among genes with valid instrumental cis-eQTLs in two or more cell types, 80.6% (237/294) remained significant in exactly one cell type (S4 Fig). White matter microstructure phenotypes harbored more causal eGenes in oligodendrocytes and astrocytes than brain regional volumes did (Fig 1b), mirroring prior evidence that white matter heritability is enriched in glial cells [22]. The strongest enrichment occurred between white matter volume and oligodendrocytes (12 eGenes), eight of which (ANKRD44, ZCWPW1, SLC16A8, PILRB, RUNX2, LARP6, C7orf61 and SUPT3H) have been linked to neuroimaging phenotypes in the GeneCards database [41]. Collectively, these results nominate candidate cell type–specific regulators of brain structural phenotypes. Nevertheless, because the cis-eQTL data were derived from the prefrontal cortex, temporal cortex, and deep white matter—not all anatomical regions represented by the imaging outcomes—these findings should not be interpreted as pinpointing the precise brain region where each regulatory effect operates [42–44].

Turning to DBs, we identified 217 eGenes with putative causal effects on 26 DBs across eight cell types (FDR < 0.05, corrected across all eGene–cell type–DB tests), comprising 298 eGene–cell type–DB combinations (Fig 1c and S6 Table). Of these, 109, 61, and 85 eGenes were associated with 10 psychiatric disorders, 8 neurological disorders, and 8 behavioral-cognitive phenotypes, respectively. We validated these findings using three independent resources: BrainMeta cortical eQTLs [36], PsychENCODE differential expression data [29], and the GWAS Catalog [1]. Replication rates were 35.7% (87/244) and 23.0% (56/244) in BrainMeta at nominal and FDR < 0.05 levels, respectively; 29.5% (23/78) in PsychENCODE; and 23.8% (57/240) in the GWAS Catalog (all P < 10^-4, hypergeometric test; S7 Table and S2 Fig). For eGene–cell type–DB triples, replication in brainSCOPE [37] was 41.9% (315/751) nominally and 33.7% (253/751) at FDR < 0.05 (both with P < 10^-16, hypergeometric test; S7 Table and S3 Fig). Notably, 12 of the 14 AD-associated cell type–specific causal eGenes identified here were corroborated by Mathys et al. [45].

As with IDPs, the eGene–DB associations were highly cell type–specific: 82.1% (197/240) of pairs were restricted to a single cell type, and among genes with valid instruments in multiple cell types, 69.2% (92/133) were significant in exactly one (S4 Fig). These patterns support the view that most genes influencing brain disorders and behaviors exert their effects through discrete cell types [12,46]. Excitatory neurons harbored the largest number of causal eGenes across nearly all phenotype groups, notably in SCZ, bipolar disorder (BIP), and insomnia (INS). Inhibitory neurons contained fewer causal eGenes overall but remained prominently involved in SCZ. The SCZ–excitatory neuron pair yielded the greatest number of causal eGenes (20), four of which (ANKRD27, LIN28B, UROS and CCHCR1) have been linked to SCZ in the GeneCards database [41]. Moreover, five of the six eGenes predicted to influence AD in excitatory neurons—including MSH3 [47], ICA1L [48], RGS14 [49], C17orf97 [50] and ZSCAN31 [51]—have established roles in AD and other neurodegenerative disorders, suggesting their convergent contributions to neurodegenerative pathogenesis.

Pervasive phenotype pleiotropy of causal eGenes

To characterize the pleiotropic landscape of the identified causal eGenes, we systematically compared their sharing patterns across phenotype classes (IDPs versus DBs) and phenotype groups (brain regional volumes and white matter microstructure phenotypes within IDPs; behavioral-cognitive phenotypes, neurological disorders, and psychiatric disorders within DBs). Significant eGene overlap was observed between every pair of phenotype groups (all with P < 10^-10, hypergeometric test; S5 Fig). Within the IDP class, the strongest overlap occurred between brain regional volumes and white matter microstructure phenotypes (42 shared eGenes, 37.5%), with the inferior fronto-occipital fasciculus (IFO) contributing the largest fraction of white matter phenotypes to this overlap (22.81%; Figs 2a and S6). Notably, this pleiotropy was not fully explained by genetic correlations among IDPs. Across 7,503 IDP pairs, the absolute genetic correlation was only weakly associated with eGene overlap (Spearman’s rank correlation coefficient r = 0.092, P = 1.95 × 10^-15; S7b Fig), and the association remained weak for between-group pairs (Spearman’s rank correlation coefficient r = 0.073, P = 6.10 × 10^-4). Indeed, even the most strongly correlated between-group pair (PLIC_FA and right lateral ventricle, r_g = 0.49) shared no causal eGenes. Within the DB class, four eGenes exhibited pleiotropic effects across all three DB groups—DDHD2 and LIN28B in excitatory neurons, and MAP2K5 and XKR6 in inhibitory neurons (S8 Table)—all of which have been implicated in multiple psychiatric disorders [52–55].

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Causal eGenes shared across IDPs and DBs.

(a) Upset plot of cell type–specific causal eGenes shared across phenotype classes and groups. (b) pLI scores of pairwise shared causal eGenes. (c) Top two enriched gene sets for pairwise shared eGenes. Within-group: shared by ≥2 phenotypes within a group. Within-class: shared across groups of the same class. Between-class: shared between the DB and IDP classes. BRV: Brain regional volume; WMMP: White matter microstructure phenotype; BCP: Behavioral-cognitive phenotype; NEUD: Neurological disorder; PSYD: Psychiatric disorder.

https://doi.org/10.1371/journal.pcbi.1014422.g002

Between-class pleiotropy was similarly extensive. The brain regional volume group overlapped most strongly with the behavioral-cognitive phenotype group (21 shared eGenes, 24.7%), driven primarily by drinks per week (DPW) (10.45%; Figs 2a and S6); and with the psychiatric disorder group (26 shared eGenes, 23.85%), driven primarily by attention-deficit/hyperactivity disorder (ADHD) (15.38%; Figs 2a and S6). More broadly, 17 eGenes were shared between two IDP groups and one DB group, nine between one IDP group and two DB groups, four across four groups (XKR6 in inhibitory neurons, MAPT in astrocytes, ZSCAN31 in astrocytes, and MSH3 in excitatory neurons), and one eGene—XKR6 in inhibitory neurons—was shared across all five phenotype groups (Fig 2a and S8 Table). Collectively, these findings indicate that cell type–specific genetic susceptibilities are shared across diverse categories of brain-associated phenotypes.

We next examined the evolutionary constraint and functional profiles of shared causal eGenes using probability of loss-of-function intolerance (pLI) scores from ExAc [56] and gene-set enrichment analysis [57]. Among genes shared within individual phenotype groups, those in the behavioral-cognitive phenotype group exhibited the highest constraint (μ = 0.302, σ² = 0.203), followed by the brain regional volume group (μ = 0.20, σ² = 0.14; Fig 2b, left panel). Functionally, the behavioral-cognitive shared eGenes were enriched in pathways governing memory, cognition, and neurotransmission (Fig 2c and S9 Table), consistent with their established roles in learning and cognitive performance [58,59]. By contrast, brain regional volume–shared eGenes were enriched in protein-binding functions, notably Hsp90 and MHC class I protein binding (Fig 2c and S9 Table), which regulate synapse formation and [60,61].

For eGenes shared within the same phenotype class but across groups, the neurological disorder–psychiatric disorder overlap exhibited the highest pLI scores (μ = 0.409, σ² = 0.235), whereas the brain regional volume–white matter microstructure overlap showed the lowest (μ = 0.153, σ² = 0.125; Fig 2b, middle panel). Concordantly, the neuropsychiatric shared eGenes were enriched in cell maintenance and signaling pathways such as plasma membrane organization (Fig 2c and S9 Table), which have been linked to the etiology of neuropsychiatric disorders [62,63].

Finally, for between-class shared eGenes, those bridging brain regional volumes and DB groups (behavioral-cognitive, neurological, and psychiatric) exhibited higher pLI scores than those bridging white matter microstructure and DB groups (Fig 2b, right panel). This suggests that volumetric measures may capture more fundamental and evolutionarily conserved genetic determinants of brain disorders than microstructural measures—a pattern consistent with evidence that microstructural changes typically precede volumetric changes during disease progression [64,65], and that the genetic drivers of such early microstructural alterations are correspondingly less constrained. Notably, the brain regional volume–neurological disorder shared eGenes, which displayed the highest between-class pLI scores, were enriched in four pathways related to extracellular matrix dynamics (Fig 2c and S9 Table). Given the central role of the extracellular matrix in brain volume regulation, neural development, and disease pathophysiology [66–68], these results underscore the importance of extracellular matrix biology in the shared genetic architecture of brain structure and neurological disorders.

Shared cell type specificity of causal eGenes between different phenotype classes

Having established the pervasive sharing of causal eGenes across phenotype classes, we next examined the specific cell types that mediate these shared associations between IDPs and DBs. The cell type landscape differed markedly across DB–IDP pairs (Fig 3a). SCZ and drinks per week (DPW) each shared causal eGenes with brain regional volumes across a broad spectrum of cell types—including astrocytes, excitatory neurons, microglia, oligodendrocytes, and OPCs for SCZ, and astrocytes, excitatory neurons, microglia, inhibitory neurons, and OPCs for DPW—consistent with their polygenic architecture [69,70]. By contrast, neurodevelopmental disorders such as autism spectrum disorder (ASD), post-traumatic stress disorder (PTSD), and ADHD showed restricted sharing confined to inhibitory neurons (Fig 3a), supporting the hypothesis that disrupted inhibitory neurotransmission underlies both the structural and functional abnormalities seen in these conditions [71–73]. Notably, several risky behaviors—general risk tolerance (GRT), automobile speeding propensity (ASP), and DPW—also shared inhibitory neuron eGenes with both brain regional volumes and white matter microstructure phenotypes (Fig 3a). Furthermore, OPC-specific sharing with brain regional volumes was observed exclusively for SCZ and the behavioral-cognitive phenotypes DPW, ASP, and educational attainment (Fig 3a), in line with evidence that OPC gene expression shapes neural circuit development and cognitive outcomes [74,75].

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Shared cell type–specific eGenes between IDPs and DBs.

(a) Brain cell types showing significant sharing of causal eGenes between IDP groups and individual DB phenotypes. Cell types were tested by hypergeometric test within each cell type and selected at FDR < 0.05. (b) Causal eGenes shared between brain regional volumes and psychiatric disorders.

https://doi.org/10.1371/journal.pcbi.1014422.g003

These patterns point to a common cell type–specific genetic susceptibility linking brain structure to disorders and behaviors. Among psychiatric disorders, the MAPT gene in astrocytes emerged as a prominent shared signal, implicated in SCZ alongside 26 brain regional volumes and 15 white matter microstructure phenotypes (Figs 3b and S8). Increased astrocytic MAPT expression was associated with elevated SCZ risk, reduced mean fractional anisotropy in 14 white matter tracts, and decreased volume in 17 brain regions. Located at 17q21, MAPT is pivotal for human brain size and patterning, with established links to intracranial volume, white matter microstructure, and cortical architecture [24]. MAPT mutations drive brain atrophy [76], and MAPT dysregulation has been tied to SCZ [77], while astrocytic dysfunction itself is increasingly recognized as a causal contributor to SCZ pathophysiology [78]. The MAPT gene encodes tau, a protein essential for microtubule stabilization and axonal transport [79]. Given tau’s central role across both neuronal and glial compartments, and the established modulation of tau pathology in neurodegenerative disease [80], our findings raise the prospect of targeting astrocytic MAPT expression therapeutically to address SCZ-related structural brain abnormalities.

In the context of neurological disorders, three excitatory neuron eGenes—ZSCAN31 [51,81], SYT14 [82,83] and ICA1L [48,84]—were shared between AD and multiple IDPs (S8 and S9 Figs), converging with prior evidence that variants in these genes influence both brain imaging measurements and AD susceptibility [48,51,81–84].

Finally, among behavioral-cognitive phenotypes, XKR6 in inhibitory neurons was shared across ASP, GRT, 15 brain regional volumes (including right lateral orbitofrontal cortex and left isthmus cingulate), and 5 white matter microstructure phenotypes (including fornix–stria terminalis [FXST], IFO, and uncinate fasciculus [UNC]; S8 and S9 Figs). XKR6 belongs to the Kell blood group complex subunit–related family [85]. Large-scale GWAS of risky behavior have identified XKR6-related variants and implicated inhibitory neurotransmission in shaping individual variation in risk tolerance [86]. Separately, astrocytic SRR expression was causally linked to smoking cessation (SC), DPW, neuroticism (NEU), 5 brain regional volumes, and one white matter microstructure phenotype. Encoding serine racemase, SRR is expressed in excitatory glutamatergic neurons of the human forebrain [87] and plays a key role in glutamatergic synaptic signaling [88]. Its expression has been associated with substance use phenotypes, and SRR represents a promising pharmacological target for smoking cessation [88].

Cell type-informed clustering of phenotypes reveals biologically meaningful patterns

To identify phenotypes with shared cell type–specific regulatory architectures, we computed a phenotype × phenotype Jaccard similarity matrix based on shared causal eGenes and their corresponding cell types (S1 Text and S10 Fig). Oligodendrocytes, excitatory neurons, and astrocytes were the dominant contributors to the global similarity structure (S11 Fig), reflecting their greater overall abundance of phenotype-associated causal eGenes (Fig 1b-1c). Hierarchical clustering identified 12 robust phenotype clusters (S10 Fig and S10 Table), of which eight (C2–C9 and C12) contained both IDP and DB phenotypes (S12 Fig). Clusters C6 and C7 were the most heterogeneous, each spanning all five phenotype groups, with inhibitory neurons as the principal cell type driving shared molecular regulation across these diverse phenotypic categories. Cluster C4 was significantly enriched for white matter microstructure phenotypes (FDR = 0.0014, hypergeometric test) and nominally enriched for psychiatric disorders (P = 0.03), with microglia as the predominant cell type—implicating microglia-mediated mechanisms in the shared genetic basis of white matter integrity and psychiatric vulnerability.

We next asked whether these molecularly defined clusters mirror clinically observed comorbidity patterns. Focusing on disorder–disorder pairs that fell within the same cluster, we surveyed the literature for documented comorbidities and epidemiological associations (S11 Table). All nine psychiatric–psychiatric pairs identified showed established comorbidity support; for example, BIP clustered with both SCZ and anxiety disorder (ANX) [89,90]. Cross-category pairs also displayed compelling clinical parallels. Epilepsy (EPI) and PTSD co-occurred in cluster 6, consistent with documented clinical comorbidity [91]. IS and obsessive-compulsive disorder (OCD) shared cluster 7, supported by longitudinal evidence of elevated stroke risk in OCD patients [92]. PD and alcohol use disorder (AUD) clustered together in C12, aligning with epidemiological findings of increased AUD prevalence among PD patients [93]. Together, these observations indicate that phenotype clusters derived from shared cell type–specific causal eGenes capture not only molecular convergence but also clinically meaningful disease relationships.

Putative causality routes among cell type-specific causal eGenes and brain-associated complex phenotypes

Genetic risk for brain disorders may operate partly through intermediate imaging phenotypes—altering brain structure to influence disease risk, or conversely, manifesting as structural changes downstream of disease processes [20,94,95]. To dissect these causal chains at cellular resolution, we performed bidirectional Mendelian randomization between IDPs and DBs (Methods). This identified six causal effects from IDPs to DBs and three from DBs to IDPs. By overlaying these bidirectional phenotype relationships with the cell type–specific eGene–phenotype associations identified above, we reconstructed four eGene–IDP–DB routes and seven eGene–DB–IDP routes in which the eGene was causal for both the imaging phenotype and the disorder (Fig 4 and S12 Table).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Cell type–specific causal routes linking eGenes to IDPs and DBs.

(a) eGene–IDP–DB routes: the eGene is causal for both the IDP and the DB, and the IDP is causal for the DB. (b) eGene–DB–IDP routes: the eGene is causal for both the IDP and the DB, and the DB is causal for the IDP.

https://doi.org/10.1371/journal.pcbi.1014422.g004

One such route links ICA1L expression in excitatory neurons to ACR microstructure and MS (Fig 4a and S12 Table). ICA1L encodes a BAR-domain protein family member implicated in excitatory synaptic signaling [96,97]. Its dysregulation has been tied to MS pathogenesis [98], and variants in this gene are associated with ACR microstructure [19], a tract critical for executive control [99]. Diffusion tensor imaging (DTI) abnormalities of the corona radiata are also documented in MS [100]. Our results position ICA1L in excitatory neurons as a upstream regulator: it negatively affects ACR microstructure (SMR: β = -0.16, FDR = 5.42 × 10^-4; Wald ratio: β = -0.17, FDR = 1.16 × 10^-36; GSMR: β = -0.15, FDR = 1.91 × 10^-4), which in turn elevates MS risk (PMR-Egger: β = 0.11, FDR = 0.002; GSMR: β = 1.06, FDR = 2.16 × 10^-15). ICA1L also exerts a direct negative causal effect on MS (SMR: β = -0.20, FDR = 0.011; Wald ratio: β = -0.19, FDR = 1.1 × 10^-6; GSMR: β = -0.19, FDR = 9.63 × 10^-3), suggesting both indirect and potentially pleiotropic influences on disease risk.

Conversely, an eGene–DB–IDP route connects ZSCAN31 expression in astrocytes to SCZ and posterior thalamic radiation (PTR) microstructure (Fig 4b and S12 Table). ZSCAN31 encodes a C2H2-type zinc finger protein [101] with established links to both SCZ risk and neuroimaging phenotypes [81]. Thalamic radiation abnormalities are well documented in SCZ [102], and thalamocortical dysconnectivity appears to worsen progressively during psychosis onset [103]. Our results indicate that astrocytic ZSCAN31 expression contributes to the effect of SCZ progression on PTR microstructure (PMR-Egger: β = 0.087, FDR = 1.54 × 10^-5; GSMR: β = 0.033, FDR = 1.45 × 10^-5). ZSCAN31 is itself positively associated with SCZ (Wald ratio: β = 0.17, FDR = 2.38 × 10^-26; GSMR: β = 0.14, FDR = 1.53 × 10^-5; SMR: β = 0.16, FDR = 6.89 × 10^-5) and with PTR microstructure (Wald ratio: β = 0.08, FDR = 2.81 × 10^-8; GSMR: β = 0.08, FDR = 2.57 × 10^-3; SMR: β = 0.08, FDR = 0.03), positioning it as a shared astrocytic regulator of both disease risk and structural brain changes.

Together, these integrated causality routes nominate cell type–specific genes through which brain structure and disorder may be causally coupled, offering testable entry points for mechanistic and therapeutic follow-up.

To identify regulatory variants underlying the 11 causality routes (Fig 4), we further performed SNP-level fine-mapping, colocalization, and regulatory annotation (S2 Text). This yielded 56 SNP–cell type–eGene–phenotype chains, each linking a regulatory variant to its cell type–specific eGene and two associated phenotypes. For example, the chain rs3130455–CCHCR1 (excitatory neurons)–SCZ–PTR (effect direction: + , + , +) places this variant within an active promoter chromatin state (ChromHMM 1_TssA) across neural lineages, where it disrupts neurodevelopmental transcription factor motifs including Klf4 and Klf7. This supports a model in which increased rs3130455 dosage elevates CCHCR1 expression in excitatory neurons, raising SCZ risk and enhancing white matter connectivity in the PTR. Another chain, rs55938136–MAPT (astrocytes)–GCC–SCZ (effect direction: + , − , +), reveals perfect colocalization of the intronic SNP with both astrocytic MAPT expression and genu of corpus callosum (GCC) fractional anisotropy (PP.H4 = 1; PP.H4.snp = 1). This suggests that higher rs55938136 dosage increases astrocytic MAPT expression, diminishes GCC structural connectivity, and thereby elevates SCZ susceptibility [104,105]. All 56 chains are catalogued in S13 Table.

Spatiotemporal expression patterns of causal eGenes for IDPs and DBs

We next examined the expression patterns of causal eGenes in human brain single-cell data (Methods). For DBs, the causal eGenes were markedly enriched in glial cells relative to neurons (P < 2.2 × 10^-16, Wilcoxon test), with the strongest enrichment in astrocytes (S13 Fig). Psychiatric and behavioral-cognitive causal eGenes showed higher neuronal expression during the prenatal period than postnatally, whereas glial cells displayed the inverse pattern (P < 0.001, Wilcoxon test; Figs 5a and S14a). Most neurological disorder eGenes were elevated in neural cells after birth; notably, amyotrophic lateral sclerosis (ALS) eGenes were exclusively neuronal and showed postnatal upregulation (S14b Fig), consistent with prior observations [106–108].

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Spatiotemporal expression for causal eGenes of psychiatric disorders and white matter microstructure phenotypes.

(a) Prenatal versus postnatal expression of psychiatric disorder causal eGenes in glial and neuronal cells. (b-c) Lifespan trajectories of psychiatric disorder causal eGenes in (b) glial and (c) neuronal cells; shaded regions, 95% confidence intervals. (d) Postnatal expression of white matter microstructure causal eGenes in glial and neuronal cells. Developmental stages: P4, 13 PCW ≤ Age < 16 PCW; P8, Birth ≤ Age < 6 Months; P12, 12 ≤ Age < 20 Years. ANX, anxiety disorder; ASD, autism spectrum disorder; BIP, bipolar disorder; MDD, major depressive disorder; SCZ, schizophrenia; TS, Tourette syndrome. Glia: microglia, astrocytes, oligodendrocytes, and OPCs. Neurons: inhibitory and excitatory neurons.

https://doi.org/10.1371/journal.pcbi.1014422.g005

Across the lifespan, psychiatric disorder eGenes in glial cells declined through late mid-fetal development (19–24 PCW) before rising again in early childhood (1–6 years), while their neuronal expression decreased monotonically (Figs 5b-5c and S15). Behavioral-cognitive phenotypes followed a similar trajectory (S16 Fig), suggesting convergent developmental programs for these related phenotype categories.

For IDPs, causal eGenes for white matter microstructure phenotypes were strongly enriched in adult glial cells—particularly astrocytes and microglia (Fig 5d)—mirroring the predominance of glial cells in white matter tissue [109]. These spatiotemporal expression profiles provide an independent biological layer that complements the genetically inferred causal relationships. Whereas the MR framework captures cis-eQTL–driven regulatory architecture, the observed cell type–specific and developmental stage–specific transcriptional dynamics offer convergent evidence of biological plausibility, pointing to coordinated regulatory and transcriptional programs underlying brain-associated complex phenotypes.

Expression of causal eGenes indicates associations of brain-associated complex phenotypes

Brain function relies on the interplay between regional structure and white matter connectivity [110,111], yet how these relationships are manifest at the cellular level remains unclear. To address this, we examined the associations between brain regional volumes and white matter microstructure phenotypes using cell type–specific expression data from the STAB2 [112]. Regional volumes and white matter phenotypes formed largely distinct clusters (Fig 6a and Methods). Notably, causal eGene expression tracked cerebral hemisphere specificity: homologous regions from opposite hemispheres were assigned to different modules (Spearman’s rank correlation coefficient r > 0.7), indicating that genetically driven cell type–specific expression underlies hemispheric structural divergence.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. Coordinated phenotype–phenotype associations revealed by causal eGene expression.

(a) Correlation heatmap among imaging-derived phenotypes (IDPs), with phenotype group membership indicated by color bars (brain regional volumes; white matter microstructure phenotypes). (b) Correlation heatmap between white matter microstructure phenotypes and psychiatric disorders. Values are Spearman’s rank correlation coefficients of mean causal eGene expression across eight cell types (Methods). *, FDR < 0.05; **, FDR < 0.01; ***, FDR < 0.001.

https://doi.org/10.1371/journal.pcbi.1014422.g006

A similar analysis of DBs revealed that causal eGene expression profiles carried functional associations with brain disorders and phenotypes (S17 Fig). For instance, the expression signatures of educational attainment and intelligence correlated strongly with those of AD (Spearman’s rank correlation coefficient r > 0.9), echoing epidemiological evidence that higher cognitive reserve lowers AD risk [113,114].

These expression patterns also corroborate the causal routes linking IDPs to DBs identified above (Fig 4). Fractional anisotropy in the PTR and the inferior fronto-occipital fasciculus (IFO) was significantly associated with SCZ and general risk tolerance (GRT), respectively (Figs 6b and S18), consistent with the bidirectional causal relationships inferred by MR. Beyond these known links, we detected significant expression associations between several regional volumes and psychiatric disorders (FDR < 0.05, Spearman’s rank correlation test). The strongest was between major depressive disorder (MDD) and the left rostral middle frontal cortex (FDR = 4.01 × 10^-45, Spearman’s rank correlation test; S19 Fig), implicating this prefrontal region in the transcriptional architecture of depression.

Regulation by cell type-specific causal eGenes

Previous studies have pointed to an important yet underrecognized role for excitatory neurons in AD etiology [115,116]. In our analysis, six eGenes—SYT14, MSH3, ICA1L, RGS14, C17orf97 and ZSCAN31—were causally associated with AD in excitatory neurons (S6 Table). Among these, five (MSH3, ICA1L, RGS14, C17orf97 and ZSCAN31) showed a consistent positive causal effect. Several of these genes have established links to excitatory neuron or synaptic function: SYT14 mediates calcium-dependent neurotransmitter release and synaptic vesicle trafficking [117,118]; ICA1L and RGS14 modulate synaptic signaling pathways in excitatory neurons [97,119]; and ZSCAN31, a zinc finger protein, regulates neuronal excitability and intrinsic circuitry within this subtype [120,121]. This convergence suggests that these genes may influence AD progression through shared mechanisms affecting neuronal function. Moreover, we observed significant overlap between AD and brain structural phenotypes with respect to cell type–specific causal eGenes in excitatory neurons (Fig 3), including ZSCAN31, SYT14, and ICA1L. These findings collectively strengthen the mechanistic links among genes, excitatory neurons, AD, and brain structure.

ZSCAN31 is among the most pervasively pleiotropic genes in our catalog. In astrocytes, it showed a positive causal effect on all 11 associated phenotypes spanning four groups: MDD and SCZ (psychiatric disorders), GRT (behavioral-cognitive phenotypes), right supramarginal volume (brain regional volumes), and seven white matter microstructure phenotypes—mean fractional anisotropy of the GCC (genu of corpus callosum), PTR (posterior thalamic radiation), SLF (superior longitudinal fasciculus), SCC (splenium of corpus callosum), BCC (body of corpus callosum), ACR (anterior corona radiata), and the average across all tracts (Fig 3b and S4 Table). By contrast, ZSCAN31 in excitatory neurons exerted a negative causal effect on right hippocampal and total white matter volume. This divergence in effect direction across cell types indicates that ZSCAN31 is subject to cell type–specific regulatory control over brain-associated phenotypes, extending prior findings [101]. We accordingly inferred three putative routes linking ZSCAN31 in astrocytes to IDPs and DBs (Fig 4 and S12 Table). For example, the route ZSCAN31 in astrocytes → SCZ → PTR suggests that astrocytic ZSCAN31 expression may mediate the influence of SCZ disease progression on PTR microstructural change. Consistent with this, external single-cell data indicated that causal eGenes in multiple cell types participate in the association between SCZ and PTR (Fig 6b), illuminating how cell type–specific eGene regulation shapes brain-associated phenotype relationships.

Shared genetic signals between risky behaviors and IDPs

Our analysis supports genetically driven links between risky behaviors and multiple IDPs at the cellular level. We observed significant sharing of causal eGenes within inhibitory neurons between risky behaviors (GRT and ASP) and both brain regional volumes and white matter microstructure phenotypes (Fig 3a). Among these shared signals, XKR6 in inhibitory neurons was implicated in both risky behaviors (GRT and ASP) and several IDPs, including cortical volumetric phenotypes and white matter phenotypes such as the FXST, IFO and UNC. Using external expression data across cell types, we also detected significant expression-profile correlations between GRT and IFO, and between GRT and FXST (S18 Fig), pointing to a mediating role for cell type–specific causal eGenes in these associations. These results suggest a shared cell type–specific genetic architecture linking risky behaviors to selected IDPs. They should not, however, be taken as evidence for the precise brain regions in which these effects operate; that anatomical context awaits further validation.

Interpretation of cell type-specific associations

The high cell type specificity observed here (91.4% for eGene–IDP pairs and 82.1% for eGene–DB pairs) should be interpreted cautiously. These rates reflect the specificity of significant eGene–phenotype associations after TWAS and MR filtering, not the inherent specificity of eQTL effects. This definitional distinction may partly explain why our estimates exceed those reported in studies focused on cell type–specific eGenes or eQTLs [13,122]. In addition, methodological factors—such as significance filtering (TWAS + MR) and uneven cis-eQTL availability across cell types (e.g., the comparatively small number in endothelial cells and pericytes)—may inflate the apparent specificity.

It is also important to note that the cis-eQTL data used here were derived from a limited set of brain regions (prefrontal cortex, temporal cortex, and deep white matter) rather than the full complement of anatomical regions represented by the imaging phenotypes. Given the context dependence of molecular eQTL effects and known transcriptomic divergence across brain regions—especially between cortical and subcortical structures [44]—associations involving unmatched regions should be treated with caution. More broadly, MR captures the effect of a genetically proxied expression perturbation and does not identify the exact biological context or site of action [42]. Our results are therefore best regarded as prioritizing putative cell type–specific genes for brain phenotypes, rather than pinpointing the precise brain regions in which each regulatory effect occurs.

Limitations

Several avenues remain open for future work. First, the comparatively small number of cis-eQTLs in endothelial cells and pericytes (S3 Table) [12], likely constrained the number of causal genes identified in these cell types (Fig 1b). Larger single-cell eQTL datasets with greater sample sizes and finer cell type resolution should improve power to detect eGene–phenotype associations in these populations. Second, when performing MR between IDPs and DBs, half of the DB GWASs were excluded because they included participants from UK Biobank [123], limiting the completeness of the inferred causality routes among eGenes, IDPs, and DBs (Fig 4). In addition, shared causal eGenes identified across IDPs and/or DBs may be influenced by unavoidable sample overlap between some GWAS datasets. Future studies should assemble more fully independent GWAS resources for brain-associated phenotypes to strengthen the robustness and coverage of these causal inferences. Third, although our findings highlight putative causal relationships, they do not establish definitive causality and require confirmation by randomized controlled trials, the gold standard for causal inference. Considerable work remains ahead in treatment development, and our results should be interpreted accordingly.

Outlook

In summary, our study provides a systematic investigation of cell type–specific causal genes for brain-associated complex phenotypes, together with evidence that brain structure and disorders/behaviors share genetic foundations operating at the cellular level. We hope that the connections identified among genes, cell types, and complex phenotypes will serve as starting points for developing therapeutic strategies targeting brain structural abnormalities, disorders, and behaviors.

Cell type-specific cis-eQTL summary statistics

We obtained published cell type–specific cis-eQTL summary statistics comprising effect sizes for 67,815,924 SNP–gene pairs across 4,690,822 SNPs and 17,881 gene transcripts [12]. The data cover eight major brain cell types: astrocytes, excitatory neurons, inhibitory neurons, oligodendrocytes, OPCs, microglia, endothelial cells, and pericytes (Fig 1a and S3 Table). The cis-eQTLs were identified by integrating high-quality single-nuclei RNA sequencing data from the prefrontal cortex, temporal cortex, and white matter with corresponding genotype data from 192 individuals. All individuals fell within three standard deviations of the mean for the first and second principal components of European-ancestry populations from the 1000 Genomes Project [12].

GWAS summary statistics for brain disorders and behaviors

We obtained publicly available GWAS summary statistics for 26 brain DBs, including eight behavioral-cognitive phenotypes, eight neurological disorders, and 10 psychiatric disorders (S1 Table). Most were published recently and derived from meta-analyses with large sample sizes (mean ~280,000 individuals). All GWAS summary statistics were of European ancestry, and there was no participant overlap with the cell type–specific cis-eQTL dataset.

The sample size for each GWAS summary statistics dataset is listed below.

• Behavioral-cognitive phenotypes
  1. (1) Automobile speeding propensity (ASP): 404,291 individuals (heritability: 7.9%) [86]
  2. (2) Childhood intelligence (CI): 12,441 individuals (heritability: 22%-46%) [124]
  3. (3) Drinks per week (DPW): 941,280 individuals (heritability: 4%) [125]
  4. (4) Educational attainment (EA): 766,345 individuals (heritability: 12.2%) [126]
  5. (5) General risk tolerance (GRT): 466,571 individuals (heritability: 4.6%) [86]
  6. (6) Intelligence (INT): 269,867 individuals (heritability: 19%) [127]
  7. (7) Neuroticism (NEU): 390,278 individuals (heritability: 10%) [128]
  8. (8) Smoking cessation (SC): 547,219 individuals (heritability: 5%) [125]
• Neurological disorders
  1. (1) Alzheimer’s disease (AD): 71,880 cases and 383,378 controls (heritability: 5.5%) [129]
  2. (2) Amyotrophic lateral sclerosis (ALS): 20,806 cases and 59,804 controls (heritability: not reported in the original GWAS publication) [130]
  3. (3) Epilepsy (EPI): 15,212 cases and 29,677 controls (heritability: 9.7%) [131]
  4. (4) Intracerebral hemorrhage (ICH): 1,545 cases and 1,481 controls (heritability: not reported in the original GWAS publication) [132]
  5. (5) Insomnia (INS): 109,548 cases and 277,440 controls (heritability: 7.2%) [133]
  6. (6) Ischemic stroke (IS): 67,162 cases and 454,450 controls (heritability: not reported in the original GWAS publication) [134]
  7. (7) Multiple sclerosis (MS): 47,429 cases and 68,374 controls (heritability: 19.2%) [135]
  8. (8) Parkinson’s disease (PD): 33,674 cases and 449,056 controls (heritability: 22%-27%) [136]
• Psychiatric disorders
  1. (1) Attention-deficit/hyperactivity disorder (ADHD): 20,183 cases and 35,191 controls (heritability: 21.6%) [137]
  2. (2) Anxiety disorder (ANX): 25,453 cases and 58,113 controls (heritability: 26%) [138]
  3. (3) Autism spectrum disorder (ASD): 18,381 cases and 27,969 controls (heritability: 11.8%) [139]
  4. (4) Alcohol use disorder (AUD): 11,569 cases and 34,999 controls (heritability: 9.0%) [140]
  5. (5) Bipolar disorder (BIP): 41,917 cases and 371,549 controls (heritability: 18.6%) [141]
  6. (6) Major depressive disorder (MDD): 135,458 cases and 344,901 controls (heritability: 8.7%) [142]
  7. (7) Obsessive-compulsive disorder (OCD): 2,688 cases and 7,037 controls (heritability: 28%) [143]
  8. (8) Post-traumatic stress disorder (PTSD): 9,831 cases and 19,225 controls (heritability: 28%) [144]
  9. (9) Schizophrenia (SCZ): 76,755 cases and 243,649 controls (heritability: 24%) [145]
  10. (10) Tourette syndrome (TS): 4,819 cases and 9,488 controls (heritability: 21%) [146]

GWAS summary statistics for imaging-derived phenotypes

We obtained publicly available GWAS summary statistics for 123 brain IDPs, including 101 brain regional (and total) volume phenotypes and 22 white matter microstructure phenotypes (S2 Table). The brain volume GWAS summary statistics were processed by Zhao et al. [147] from 19,629 UK Biobank participants of European ancestry (reported mean heritability ~40%). Volumes included total brain volume, gray matter, white matter, and cerebrospinal fluid, labeled with the Mindboggle-101 atlas [148]. The white matter tract GWAS summary statistics were processed by Zhao et al. [22] from 33,292 UK Biobank participants of European ancestry (reported mean heritability ~46.3%). Tracts were segmented with the JHU ICBM-DTI-81 white-matter atlas [149–151], and microstructure was quantified by mean fractional anisotropy (FA).

Transcriptome-wide association study

To investigate potential links between gene expression and brain-associated complex phenotypes, we conducted a TWAS for the GWAS summary statistics of 123 IDPs and 26 DBs [31]. Prediction weights and covariance were obtained from two resources: (1) PredictDB Data Repository (https://predictdb.org/), where prediction models were built from GTEx V8 expression data [152] across 13 brain regions (amygdala, anterior cingulate cortex [BA24], caudate, cerebellar hemisphere, cerebellum, cortex, frontal cortex [BA9], hippocampus, hypothalamus, nucleus accumbens, putamen, spinal cord [cervical C-1] and substantia nigra); (2) CMC-derived DLPFC prediction models (https://github.com/laurahuckins/CMC_DLPFC_prediXcan), built from CommonMind Consortium expression data [153] from dorsolateral prefrontal cortex (DLPFC) tissue. TWAS was implemented with MetaXcan [31]. eGene–phenotype reaching nominal significance (P < 0.05) in at least one brain tissue were retained for subsequent MR analysis.

Primary Mendelian randomization analysis

To investigate the causal effects of cell type–specific gene expression on brain-associated complex phenotypes, we performed two-sample MR using four complementary methods: two single-SNP methods (SMR [6] and Wald ratio [32]) and two multi-SNP methods (PMR-Egger [8] and GSMR [33]). Details on each method are provided below (S1 Fig).

1. (1) The SMR method tests whether the expression level of a gene is associated with a phenotype under the assumption of either causality or pleiotropy [6]. It is designed specifically for gene expression as the exposure and requires only one SNP as the instrument. The top cis-eQTL associated with each gene expression in a given cell type at genome-wide significance (P < 5 × 10^-8) was selected as the instrument. SMR incorporates the HEIDI test to detect linkage [6]; results with HEIDI P < 0.05 were discarded. SMR and the HEIDI test were implemented with SMR software (version 1.3.1, https://yanglab.westlake.edu.cn/software/smr).
2. (2) The Wald ratio method computes the change in disease risk per standard deviation change in gene expression, using the instrumental cis-eQTL for the target gene. It is applicable to any exposure and requires only one SNP as the instrument under the standard MR assumptions (relevance, independence, and exclusion restriction). For instrument selection, LD clumping (r² < 0.1, window = 250 kb, 1000 Genomes EUR reference panel) was first performed for cis-eQTLs associated with each gene expression at genome-wide significance (P < 5 × 10^-8) using PLINK v1.90 [154]. The independent cis-eQTL(s) for each gene in the given cell type were then selected as instrument(s). After instrument selection, 86.6% of cell type–specific genes had one valid instrumental cis-eQTL, so the Wald ratio could be applied in most analyses. For the remaining 13.4% of genes with more than one valid instrument, the inverse-variance weighted (IVW, fixed-effects) method was used as a supplementary approach [155], which is commonly paired with the Wald ratio in the literature [20,156]. Both methods were implemented with the R package TwoSampleMR (version 0.5.10) [157]. The harmonise_data() function in TwoSampleMR was used to harmonize effect alleles and SNP effects between exposure and outcome. Heterogeneity (function mr_heterogeneity, P < 0.05) and pleiotropy (function mr_pleiotropy_test, P < 0.05) tests were conducted for IVW results.
3. (3) The PMR-Egger method estimates the causal effect of gene expression on a phenotype in the presence of horizontal pleiotropy within an MR likelihood framework [8]. It is designed specifically for gene expression as the exposure, can handle multiple correlated variants as instruments, and accounts for linkage disequilibrium (LD) among them. The cis-eQTLs (clumping r² < 0.9, window = 250 kb, 1000 Genomes EUR reference panel) associated with each gene expression in a given cell type at genome-wide significance (P < 5 × 10^-8) were selected as instruments. PMR-Egger was implemented with the R package PMR (version 1.0). The LD correlation matrix was derived from the 1000 Genomes EUR reference panel, and the maximum number of iterations was set to 5000.
4. (4) The GSMR method extends SMR to estimate the causal effect of an exposure on an outcome, leveraging multiple variants while accounting for LD among them [33]. It is applicable to any exposure under the standard MR assumptions. The cis-eQTLs (clumping r² < 0.9, 1000 Genomes EUR reference panel) associated with each gene expression in a given cell type at genome-wide significance (P < 5 × 10^-8) were selected as instruments. GSMR incorporates the HEIDI-outlier test to identify and exclude instruments with significant pleiotropic effects on the outcome (P < 0.05). Both GSMR and the HEIDI-outlier test were implemented with the R package gsmr (version 1.1.0).

A gene was deemed putatively causal for a phenotype within a given cell type (cell type–specific causal eGene) if it met the following criteria: (1) significant in at least one single-SNP method after FDR correction (FDR < 0.05) [158]; (2) significant in at least one multi-SNP method after FDR correction (FDR < 0.05); (3) consistent causal effect direction across all four methods, reducing the likelihood of false positives.

We also calculated the F-statistic for each instrument using the Cragg-Donald statistic to assess instrument strength [159] by , where represents the proportion of exposure variance explained by the instrument, N is the sample size from the cis-eQTL summary statistics, β is the effect size, se is the standard error, and k is the number of instruments used in the MR estimate (k = 1 for a single instrument). Genes for which the averaged F-statistic across all instruments exceeded 10 were retained. A brief introduction to MR for non-expert readers is provided in S3 Text.

Significance of cell type-specific causal eGenes sharing

We applied a hypergeometric test to assess whether cell type–specific causal eGenes were significantly shared between any two brain-associated complex phenotypes. The background gene set was derived by multiplying the total number of genes (only those with the same biotypes as the causal eGenes) by the total number of cell types. When estimating overlap within a particular cell type, the background comprised the total number of genes with the same biotypes as the causal eGenes. An FDR threshold of 0.05 was used to select significant cell types. Gene annotation was obtained from Ensembl (GRCh37 release 87) [160].

Gene set enrichment analysis

We utilized all gene sets from the Biological Process (BP), Molecular Function (MF), and Cellular Component (CC) sub-ontologies in the Gene Ontology (GO) database for gene set enrichment analysis [161,162]. GO enrichment was implemented with the enrichGO() function in the R package clusterProfiler (version 4.10.0) [57].

Replication of predicted associations

We performed replication analysis on the predicted causal associations between eGenes and brain-associated complex phenotypes identified in our MR analysis. To validate eGene–phenotype pairs, three external datasets were used. (1) The cis-eQTL summary statistics from the BrainMeta portal (https://yanglab.westlake.edu.cn/data/brainmeta/cis_eqtl/), derived from RNA sequencing of 2,865 brain cortex samples from 2,443 unrelated European-ancestry individuals [36]. The same MR analysis procedure was applied to infer causal relationships between cortical gene expression and 149 brain-associated complex phenotypes. An eGene–phenotype pair was considered replicated if it achieved nominal significance in at least one MR method and showed consistent effect directions across all MR methods and with the discovery analysis. (2) The differentially expressed genes and transcripts (at 5% FDR) in ASD, SCZ and BIP from PsychENCODE [29], which used in vivo gene expression profiles to identify transcriptome-wide isoform alterations across the three major psychiatric disorders. (3) Published eGene–phenotype associations from the NHGRI-EBI GWAS Catalog (version 2023-06-09, www.ebi.ac.uk/gwas/) [1].

To validate eGene–cell type–phenotype triples, two external datasets were used: (1) cell type–specific eQTL summary statistics from the brainSCOPE resource (https://brainscope.gersteinlab.org/integrative_files.html), generated from single-cell resolution data of 388 adult DLPFC samples [37]. Six major cell types were considered: astrocytes, excitatory neurons, inhibitory neurons, oligodendrocytes, OPCs, and microglia. The same MR analysis procedure was applied to infer causal relationships between cell type–specific gene expression and 149 brain-associated complex phenotypes. An eGene–cell type–phenotype triple was considered replicated if it achieved nominal significance in at least one MR method and showed consistent effect directions across all MR methods and with the discovery analysis. (2) Differentially expressed genes associated with AD pathology in each cell type from Mathys et al. [45]. For each external dataset, the replication rate was calculated as the number of replicated associations divided by the total number tested, and significance was assessed by hypergeometric test. Replication was evaluated only for phenotypes present in both the discovery and replication studies.

Mendelian randomization between brain IDPs and DBs

To characterize putative causality routes among cell type–specific genes, IDPs, and DBs, we performed bidirectional MR between DBs and IDPs. Because the IDP GWAS was conducted in UK Biobank participants, we selected DB GWAS datasets that did not include UK Biobank samples to minimize sample overlap. Thirteen DB GWASs were retained: CI, EPI, ICH, IS, ALS, MS, OCD, TS, ADHD, AUD, ASD, PTSD, and SCZ. Unlike MR with gene expression as the exposure—where a single independent genome-wide significant cis-eQTL typically serves as the instrument—analyses with DBs or IDPs as exposures often involve multiple independent instruments. We therefore primarily employed multi-SNP MR methods: PMR-Egger, GSMR, and methods from the R package TwoSampleMR (IVW [155], MR-Egger [163], Weighted median [164], Simple mode [165], Weighted mode [166]). or instrument selection, variants were selected at genome-wide significance (P < 5 × 10^-8) with clumping r² < 0.6 for PMR-Egger and r² < 0.1 for GSMR and TwoSampleMR methods. Heterogeneity (function mr_heterogeneity, P < 0.05) and pleiotropy (function mr_pleiotropy_test, P < 0.05) tests were conducted for TwoSampleMR results. A phenotype was deemed putatively causal for another if it met the following criteria: (1) significant in at least two MR methods after FDR correction (FDR < 0.05); (2) consistent causal effect direction across all methods.

Expression analysis of causal eGenes using single-cell data

External normalized single-cell gene expression data were obtained from STAB2 [112]. We extracted cell type–specific expression data of the causal eGenes for the eight cell types used in this study: microglia, astrocytes, oligodendrocytes, OPCs, inhibitory neurons, excitatory neurons, endothelial cells, and pericytes. To assess associations between and within IDPs and DBs at the cell type level, we first compiled gene expression data of the causal eGenes for each IDP or DB in each cell type. For each IDP and DB, we then created an eight-dimensional expression vector, in which each value represented the mean expression of the causal eGenes in a given cell type. Spearman’s rank correlation coefficient was used to evaluate associations between IDPs and DBs based on the expression profiles of the corresponding causal eGenes across cell types. Additionally, hierarchical clustering (complete linkage) was employed to investigate associations of IDPs and DBs based on gene expression correlation coefficients.