Mapping the genetic landscape of suicide risk: Insights from genomic SEM

Abstract

The genetic basis of traits associated with suicide risk remains poorly understood. We applied genomic structural equation modeling and integrated multiple post-genome-wide association study (GWAS) analysis strategies to identify potential causal single-nucleotide polymorphisms independent of known high-risk suicide syndrome GWAS variants, identifying four genome-wide–significant, putatively causal loci and six putatively causal genes. Additionally, we applied multiple transcriptome-wide association study (TWAS) methods at the tissue and cellular levels to fine-map susceptibility genes. Subsequently, we analyzed the key regulatory elements driving their expression. Next, we assessed genetic pleiotropy by evaluating genetic correlations between high-risk suicide syndrome and over one hundred common diseases. Furthermore, we constructed a polygenic risk score (PRS) from the summary statistics to quantify each chromosome's contribution to the associated genetic risk. Our study systematically delineated the overall genetic architecture of high-risk suicide syndrome through a GWAS of this previously unmeasured latent phenotype.

Introduction

Albert Camus noted that “suicide is the only genuinely serious philosophical problem”. This observation underscores the central role of suicide in psychiatry. Globally, approximately 720,000 people die by suicide annually (https://www.who.int), resulting not only in the loss of life but also medical and economic burdens on families and society. Despite recent advances in suicide prevention and intervention [1], our understanding of the genetic and biological bases of suicidal behavior remains limited (https://www.hhs.gov).

Suicide involves biological, psychological, and social factors. Evidence suggests that dysregulation in the neurotransmitter, neuroendocrine, and immune systems may contribute to suicidal behavior [2,3]. Additionally, mental disorders, financial stress, and physical illnesses are established external risk factors for suicidal behavior [4,5]. However, these broad factors do not fully explain the substantial interindividual differences observed in the risk of and genetic susceptibility to suicidal behavior [[6], [7], [8]]. Therefore, additional exploration of the genetic underpinnings of suicidal behavior is warranted to better understand differences and identify high-risk individuals.

Phenotypic heterogeneity and measurement error complicate large-scale genome-wide association studies (GWAS) directly evaluating suicidal behavior [9,10]. For example, the “suicide attempt” phenotype often amalgamates clinically distinct behaviors, such as impulsive acts versus highly premeditated attempts, or behaviors driven by different underlying psychopathologies, which may have different genetic underpinnings. Furthermore, measurement error arises from reliance on broad data sources like electronic health records [11,12]. To overcome these obstacles, we built a GWAS framework focused on the latent construct of suicidal behavior. The core of this framework is genomic structural equation modeling (SEM), an advanced statistical genetics approach [13]. Genomic SEM enables researchers to integrate multiple GWAS summary datasets and test specific genetic structural models within a multivariate framework. Indeed, genomic SEM has been successfully applied to numerous complex psychiatric phenotypes. For instance, recent studies have used it to reveal the multivariate genetic architecture of bipolar disorder [14] and to demonstrate that genetic liability to smoking causally increases the risk of major depression and other mental disorders [15]. These findings highlight the power of genomic SEM to break down complex psychiatric traits, motivating our use to explore the shared genetic basis of suicidal behavior. Another key advantage of genomic SEM is its ability to effectively correct for biases due to sample overlap and size differences. This method also distinguishes loci with pleiotropic effects on multiple traits from those that influence only specific traits. We deconstructed the genetic signals from the new GWAS, isolated the components not explained by known biomarkers and designated them potential genetic markers, and conducted a post-GWAS functional analysis. To map the clinical pleiotropy of this genetic construct, we also assessed its genetic correlations with approximately one thousand clinical phenotypes.

This approach is intended to inform the rapid identification of high-risk individuals and guide the development of personalized intervention strategies in clinical practice. Suicidal behavior is shaped by genetic, environmental, and random factors; thus, our approach does not fully capture the complexity of gene–environment interactions. However, the strength of this method lies in its ability to statistically elucidate the confounding effects of each input trait. This method enables an independent genetic investigation into the core latent construct underlying the risk of suicidal behavior.

In our study, we investigate the biological basis of suicidal behavior and reveal candidate genes for new intervention strategies. Moreover, our findings present a translational pathway from statistical genetic discovery and functional characterization to clinical evaluation. In this study, we define high-risk suicide syndrome as a research construct, a latent genetic liability to suicidal behavior, estimated using genomic SEM of seven suicide-relevant GWAS traits (suicide/self-harm, childhood trauma, major depressive disorder, loneliness, suicidal ideation during depression, financial difficulties, multisite chronic pain). This latent factor represents the common genetic variance shared across these indicators and is not intended as a clinical diagnosis or Diagnostic and Statistical Manual of Mental Disorders entity.

Materials and Methods

The technical workflow of the study is shown in Fig. 1. All analyses were performed on publicly available summary statistics. Ethical approval and participant consent documentation were obtained in the original studies.

Fig. 1.

Multivariate genomic framework to map the latent high-risk suicide syndrome. Seven independent GWAS summary datasets spanning psychological, social, and biological domains (suicide/self-harm/ideation, childhood trauma, loneliness, depression, financial difficulties, and multisite chronic pain; n_traits = 7) were integrated using genomic SEM. Cross-trait genetic covariances were estimated with multivariable LD score regression and fit to a single common-factor model after SNP-level quality control (MAF ≥0.01, allele harmonization, MHC exclusion). The resulting latent factor (high-risk suicide syndrome) was subjected to post-GWAS analyses: Bayesian fine-mapping (SuSiE, FINEMAP), TWAS (FUSION) with gene fine-mapping (FOCUS), functional annotation/gene-set testing (FUMA/MAGMA), and chromosome-wise polygenic risk scoring (PRS–CS). These analyses yielded four putatively causal SNPs, six high-confidence causal genes, and enrichments across pathways, cell types, and clinical phenotypes. Abbreviations: GWAS, genome-wide association study; MAF, minor allele frequency; MHC, major histocompatibility complex; QC, quality control; SNP, single-nucleotide polymorphism; TWAS, transcriptome-wide association study.

Data sources for individual GWAS inputs

We selected seven suicide-relevant GWAS traits to capture complementary domains spanning distal exposures, socioeconomic stress, core psychiatric liability, social/affective state, a somatic comorbidity strongly linked to suicidality, a proximal suicidality indicator within depression, and a direct clinical proxy for suicidal behavior. Selection was guided by prior longitudinal and meta-analytic evidence linking these domains to suicidal thoughts, attempts, and/or death as well as by pragmatic requirements for multivariate genomic modeling (public availability of well-powered summary statistics, standard quality control (QC), harmonizable SNPs, and adequate SNP-heritability). The full dataset details are provided in Supplementary Table 1 and a brief overview of each data source is given below.

• (1)Suicide or other intentional self-harm: Data obtained from the FinnGen R12 database, comprising 11,538 cases and 488,810 controls [16]. (2) Childhood trauma: Data from Chen et al.‘s study, comprising 129,017 individuals of European ancestry [17]. (3) MDD: Data from a large-scale GWAS by the Psychiatric Genomics Consortium (PGC), including 59,851 cases and 113,154 controls [18]. (4) Loneliness: Data from Nagel et al.‘s study, with a sample size of 376,352 individuals [19]. (5) Suicidal ideation during depression: Data from a Neale Lab study, comprising 32,630 cases and 30,018 controls. (6) Financial difficulties: Summary statistics from the MRC Integrative Epidemiology Unit (IEU) OpenGWAS database [20]. (7) Multisite chronic pain (MCP): Data from Tang et al.‘s study, including summary statistics for 387,649 individuals [21].

Quality control (QC) of individual GWAS inputs

We performed rigorous QC on all input GWAS summary data. This process was conducted per the default standards recommended by the genomic-SEM framework [13]. First, all seven source GWASs had already undergone rigorous individual-level QC by the original study consortia (filtering of samples with high genotype missingness >5 %). Next, the following filters were applied at the single nucleotide polymorphism (SNP) level: 1. Range restriction: We retained only autosomal SNPs and excluded the major histocompatibility complex (MHC) region (chr6: 28.5–33.5 Mb, hg19) because its complex linkage disequilibrium structure can interfere with the precise localization of association signals [[22], [23], [24]]. 2. Low-frequency variant filter: We removed SNPs with a minor allele frequency (MAF) < 0.01 because the effect estimates of low-frequency variants are prone to random error, and their LD score regression standard errors tend to be high. 3. Ensuring data compatibility and matrix stability: We removed SNPs with non-matching or inconsistent alleles according to the 1000 Genomes Phase 3 European reference panel. We also checked for SNPs with an effect estimate of zero to maintain the stability of the SEM covariance matrix.

Sample overlap in individual GWAS inputs

The seven GWAS datasets were obtained from different international consortia, databases, or independent studies, each representing a distinct participant cohort. When integrating these GWAS data, we evaluated the potential impact of cohort or sample overlap to ensure that our results were accurate and broadly representative.

Genomic structural equation model

We used the GenomicSEM R package (v0.0.5) [13] to model a single common factor capturing the shared genetic susceptibility across the seven suicide-risk–related traits; the latent high-risk suicide syndrome factor is specified as the common source of genetic variation for these observed GWAS indicators. Genomic SEM analysis was conducted in two stages. In stage 1, we estimated the genetic covariance matrix. Specifically, we applied a multivariable extension of cross-trait LD score regression to the quality-controlled summary statistics of the seven GWASs, generating an empirical genetic covariance matrix for input into a common-factor SEM model.

In stage 2, we constructed and fitted a single-factor model. Given that this study aimed to identify the shared genetic susceptibility driving the seven risk traits, we constructed a single-factor (common factor) model. This model assumes that a latent high-risk suicide syndrome factor serves as the common source of genetic variation for the seven observed traits (the input GWAS). The model was then fitted by minimizing the difference between the model-implied covariance matrix and the empirical covariance matrix from stage 1. We evaluated model fit using criteria of CFI >0.90 and SRMR <0.10. Screening was based on the criteria described above; only SNPs meeting these criteria were retained for subsequent GWAS analysis of high-risk suicide syndrome.

Genomic SEM SNP heterogeneity

To test whether each SNP's effect conformed to the single-factor model, we performed a Q–SNP heterogeneity test. This test identifies SNPs with significant heterogeneity; such SNPs may influence certain traits through other biological pathways and thus deviate from the single-factor assumption. A Q-SNP p value < 0.05 was considered evidence of significant heterogeneity for a SNP.

Genomic SEM evaluation

We used another LD Score regression-based genomic control parameter to evaluate the stability of our genomic SEM [25,26]. The specific steps as follows: retaining SNPs with missing values, retaining SNPs with an INFO score <0.9, retaining SNPs with a MAF <0.01, and excluding SNPs with p-values outside the valid range or with ambiguous strand orientation. The two-step estimator was used with a cutoff set to 30.

Defining genomic loci and identifying novel variants

To interpret the biological significance of the new GWAS results, we performed a comprehensive post-GWAS analysis using the FUMA platform [27,28]. First, we input the GWAS summary statistics into FUMA to identify independent risk loci and lead SNPs. Independent significant SNPs were defined as those reaching genome-wide significance (p < 5 × 10⁻⁸) with low LD with one another (r² < 0.6). Using these independent SNPs, we further filtered for lead SNPs meeting a stricter independence criterion (r² < 0.1). This step involved clustering closely linked independent SNPs to extract representative signals. FUMA defines a risk locus by merging lead SNPs that are close to each other or in LD. Specifically, lead SNPs with r² ≥ 0.1 were merged, and loci with boundaries <250 kb apart were aggregated into broader risk regions. Ultimately, each identified risk locus was represented by the lead SNP with the lowest p value. Next, to assess the novelty of the identified lead SNPs, we compared them with the results of the seven original GWAS datasets. We also queried the GWAS catalog to identify other trait associations (p < 5 × 10⁻⁸) reported for these lead SNPs in prior studies to explore their potential pleiotropy.

In addition to the SNP-level analysis, we conducted gene-level association tests to identify candidate genes significantly associated with high-risk suicide syndrome. Multi-marker Analysis of GenoMic Annotation (MAGMA) is a post-GWAS tool designed to evaluate the association between genes and phenotypes [29]. This tool aggregates multiple genetic markers into gene-level signals and calculates gene–phenotype associations, thereby linking genetic signals to gene function. After multiple-testing correction, genes with FDR <0.05 were considered significantly associated with the trait.

Fine-mapping of causal variants

To locate potential causal variants within the identified genomic risk loci, we used two advanced Bayesian fine-mapping methods, SuSIE [30] and FINEMAP [31]. Both methods were implemented via the R package echolocatoR (v2.0.3) [32].

Our analysis workflow was as follows: for each lead SNP, we defined a ±250 kb genomic window as the fine-mapping region. We then applied SuSIE and FINEMAP within this window to estimate the posterior causal probability of each SNP. We defined SNPs identified as causal by both methods as “consensus SNPs,” representing high-confidence candidate causal variants. We quantified the strength of evidence for each consensus SNP by calculating the mean posterior probability (mean PP) across the two methods. We then categorized candidate causal variants into two tiers according to this mean PP: Tier 1 included SNPs with a mean PP > 0.95 (the most likely causal variants at that locus), and Tier 2 included SNPs with a mean PP < 0.95 and PP > 0.95 in either SuSIE or FINEMAP. This workflow filtered out the variants with the highest functional potential from the statistical associations, providing precise targets for subsequent functional validation.

Transcriptome-wide association study (TWAS)

At the functional genomics level, we first conducted a TWAS using FUSION to identify genes related to our GWAS. FUSION integrates GWAS summary data with precomputed expression quantitative trait locus (eQTL) weights to test for associations between gene expression and the trait. The eQTL reference panel was derived from GTEx v8 and comprised 37,920 expression features across 49 tissues. Genes with p < 0.05 in the FUSION results were considered preliminary candidates and were carried forward to subsequent analyses.

Subsequently, we used FOCUS, a method designed specifically for TWAS, to fine-map the significant genes identified. FOCUS is a Bayesian approach that evaluates which gene(s) in each TWAS association region are most likely to be causal [33] by computing the posterior inclusion probability (PIP) for each candidate gene. We integrated the TWAS and FOCUS results, designating any gene that met both of the following criteria as a high-confidence causal gene: (i) Bonferroni-corrected TWAS p < 0.05, and (ii) FOCUS PIP >0.8.

Gene set and disease ontology enrichment analysis

To investigate the biological roles of the genes identified by our GWAS, we performed gene-set and pathway enrichment analyses using MAGMA and the GSEA module of FUMA. The objective was to determine whether these genes are enriched in gene sets or pathways related to Mendelian diseases. Subsequently, we used the MendelVar database [34] to test whether these genes are associated with Mendelian diseases.

Cell-type annotation analysis

We used the CELLECT analysis framework to identify key cell types. CELLECT integrates GWAS data with single-cell transcriptomic data to assess whether trait heritability is enriched in genes specifically expressed by certain cell types [35]. The single-cell data were derived from the Tabula Muris dataset, covering 20 mouse organs and tissues [36]. CELLECT first normalizes the data and then calculates an expression specificity score for the genes in each cell type. These specificity scores are then input into the LD score regression (LDSC) framework to test for heritability enrichment in those cell types. For multiple comparisons, LDSC coefficient P values across all tested cell types were adjusted using the Benjamini–Hochberg false discovery rate procedure (R, p.adjust, method = “fdr”), and cell types with FDR <0.05 were considered associated with the trait.

Heritability contribution across genomic regions

To further dissect the genetic architecture of the trait, we performed a partitioned heritability analysis using LD score regression. This approach partitions the total heritability of the trait into predefined functional genomic categories (e.g., gene regions, enhancers, and repressors) to assess the contribution of each annotated category. Specifically, this method used a weighted LD matrix, allele frequency data, and our GWAS summary statistics (the filtered set in 4.6 Genomic SEM Evaluation LD score regression) to estimate the extent to which each functional annotation contributes to the trait's heritability. This analysis helps identify which functional elements contribute to high-risk suicide genome genetic susceptibility to the greatest extent and provides insight into potential regulatory mechanisms.

Risk factor annotation analysis

Next, we conducted a large-scale association analysis to systematically explore the potential causal relationships between the trait and numerous diseases and biomarkers. In this analysis, we treated our study trait as the outcome, using our full 6,272,353-SNP GWAS summary statistic file (described in Section 2.1) to ensure maximal instrument overlap. We selected hundreds of diseases and phenotypes from the IEU OpenGWAS database [37] as exposures. We used the MendelianRandomization R package (v0.7.0) to perform two-sample MR, primarily using the inverse-variance weighted (IVW) model to estimate causal effects [38,39]. This was conducted as a hypothesis-generation analysis. Therefore, we considered associations with p < 0.05 to be nominally significant, as formal multiple-testing correction was not applied. The inclusion of multiple, similar phenotypes was deliberate, and only signals that were consistent across related traits were prioritized for discussion.

Construction of polygenic risk scores (PRSs) based on summary data

To assess the genetic contribution of each chromosome to the trait, we calculated PRSs for each chromosome using genome-wide summary statistics. We first constructed these scores using the PRS–continuous shrinkage method [40]. This method uses a continuous shrinkage Bayesian model, integrating GWAS summary data with an external LD panel to infer posterior SNP effect sizes. We then partitioned these posterior effect sizes by chromosome. Finally, to quantify the pronounced genetic contribution of each chromosome, we calculated 22 aggregate scores by summing the posterior effect sizes of all SNPs on each respective autosome. These 22 sums were then compared with zero to determine risk or protective tendency.

Results

SEM metric construction

Linkage disequilibrium (LD) score regression demonstrated the heritability contributions of the seven phenotypes (Table 1). Heritability and genetic correlation matrices are provided in Supplementary Table 2 and Fig. 2. The genetic covariance matrix of these seven input traits was well fit by a common factor model (CFI = 0.97, SRMR = 0.05). Latent factor (F1) details and univariate SEM parameters are presented in Supplementary Table 3. These seven phenotypes share a broad genetic correlation primarily driven by a single common latent genetic factor.

Table 1.

LD score regression metrics for the seven suicide-related GWAS traits.

Phenotype	Z-score	λ_GC
SS	15.6	1.23
CT	13.5	1.14
MDD	17.2	1.22
LS	16.9	1.25
FD	18.9	1.31
DS	3.25	1.03
MCP	26.6	1.49

Abbreviations: SS (suicide or other intentional self-harm), CT (childhood trauma), MDD (major depressive disorder), LS (loneliness), FD (financial difficulties), DS (suicidal ideation during depression), MCP (multisite chronic pain), and λGC (genomic control lambda).

Fig. 2.

Cross-trait genetic correlations among seven suicide-related phenotypes. Pairwise genetic correlations (r_g) among the seven input GWAS traits were estimated with multivariable LD score regression and are displayed as a lower-triangle bubble plot. The circle color encodes the signed r_g (scale −1 to 1); the circle area is proportional to |r_g|. Diagonal cells show trait self-correlation. Abbreviations: CT, childhood trauma; MDD, major depressive disorder; LS, loneliness; FD, financial difficulties; MCP, multisite chronic pain; SS, suicide or self-harm; DS, suicidal ideation during depression.

From this well-fitting genomic SEM framework, we derived a new GWAS. Following cross-trait harmonization and stringent QC across all input GWAS datasets, 6,272,353 single-nucleotide polymorphisms (SNPs) were retained for downstream analysis of this factor.

Genomic SEM genomic control evaluation

We conducted a series of evaluations to investigate the source of inflation in test statistics in the model. After applying the filtering parameters, 5,195,372 SNPs were excluded, leaving 1,076,981 SNPs for modeling. The LD score regression intercept was 0.9881 (SE = 0.0095), suggesting that the inflation was not due to uncontrolled confounders such as population stratification. Given the total heritability (h²) of 0.0129 (SE = 0.0005), the genomic control coefficient (λ_GC) of 1.431, and the mean χ² of 1.525, we concluded that the inflation was driven by a true cross-trait polygenic signal rather than by systemic bias.

Structural equation model evaluation based on FUMA

We identified 31 risk loci in the FUMA evaluation of the genomic SEM results (Table 2); Fig. 3 illustrates a subset of the visualizations, with the full set provided in Supplementary Fig. 1. Applying dual criteria of genome-wide significance (p < 5 × 10⁻⁸) and FDR <0.05, we ultimately identified 24 potential risk genes (Table 3; Fig. 4). In parallel, we annotated 32 independent lead SNPs, most of which were located in intronic gene regions (Table 4).

Table 2.

GenomicRiskLoci identified by FUMA. Thirty-one risk loci with lead SNP, chromosomal coordinates, locus boundaries, and SNP counts.

GenomicLocus	rsID	chr	pos	P	LeadSNPs
1	rs490647	1	37242743	2.59 × 10−8	rs490647
2	rs4635553	2	22440788	1.44 × 10−8	rs4635553
3	rs359234	2	60471409	9.31 × 10−11	rs1025128; rs359234
4	rs58117425	2	145681570	8.78 × 10−9	rs58117425
5	rs12992995	2	175197545	1.75 × 10−8	rs12992995
6	rs9837341	3	49664767	5.22 × 10−9	rs9837341
7	rs12635614	3	173113041	2.15 × 10−8	rs12635614
8	rs34811474	4	25408838	3.28 × 10−8	rs34811474
9	rs10027492	4	28647811	4.88 × 10−9	rs10027492
10	rs10518373	4	121574701	3.25 × 10−8	rs10518373
11	rs13151315	4	140933304	9.80 × 10−9	rs13151315
12	rs171697	5	103956516	1.45 × 10−9	rs171697
13	rs7724996	5	117247501	3.36 × 10−8	rs7724996
14	rs72802853	5	152218058	2.05 × 10−8	rs72802853
15	rs55785035	7	3709491	1.09 × 10−8	rs55785035
16	rs1020006	7	12255664	1.72 × 10−9	rs1020006
17	rs73125518	7	39425117	1.49 × 10−9	rs73125518
18	rs7783012	7	114116881	2.64 × 10−12	rs7783012
19	rs1552284	8	21962607	3.00 × 10−10	rs1552284
20	rs11775287	8	30864339	4.70 × 10−8	rs11775287
21	rs13291079	9	96360650	8.75 × 10−12	rs13291079
22	rs9775587	9	122662720	2.66 × 10−8	rs9775587
23	rs7125588	11	113436072	2.86 × 10−9	rs7125588
24	rs2899991	14	47273964	1.78 × 10−10	rs2899991
25	rs4906365	14	104229230	4.60 × 10−10	rs4906365
26	rs8034783	15	47763726	2.71 × 10−8	rs8034783
27	rs8026678	15	91562005	3.56 × 10−9	rs8026678
28	rs184148321	17	60873255	2.15 × 10−29	rs184148321
29	rs17501820	18	50778737	8.14 × 10−11	rs17501820
30	rs596668	18	53264343	6.88 × 10−12	rs596668
31	rs1555133	20	31048382	1.32 × 10−9	rs1555133

Fig. 3.

Part LocusZoom plots for lead SNPs. For each lead SNP (panels A–F), –log₁₀ P values (left y-axis) of surrounding variants are plotted against genomic position (x-axis). Point color denotes linkage disequilibrium (LD, r²) with the lead SNP. Gene models are displayed below each plot. Abbreviations: LD, linkage disequilibrium.

Table 3.

Significant genes in the MAGMA gene-level analysis.

GENE	CHR	ZSTAT	P_FDR
ENSG00000155052	2	5.537	1.40 × 10−5
ENSG00000233276	3	5.4516	2.17 × 10−5
ENSG00000145022	3	6.0461	1.51 × 10−6
ENSG00000145029	3	5.549	1.40 × 10−5
ENSG00000164061	3	5.7185	5.78 × 10−6
ENSG00000173531	3	5.9002	2.94 × 10−6
ENSG00000196782	4	7.2578	1.80 × 10−9
ENSG00000146555	7	6.7282	3.14 × 10−8
ENSG00000106460	7	5.8148	3.96 × 10−6
ENSG00000106536	7	7.0834	4.28 × 10−9
ENSG00000128573	7	6.616	5.61 × 10−8
ENSG00000173566	8	5.9739	2.11 × 10−6
ENSG00000048828	9	6.1817	8.27 × 10−7
ENSG00000197724	9	6.7805	2.73 × 10−8
ENSG00000149295	11	5.5405	1.40 × 10−5
ENSG00000256053	14	5.3966	2.70 × 10−5
ENSG00000126215	14	5.7274	5.78 × 10−6
ENSG00000224997	14	5.7183	5.78 × 10−6
ENSG00000100711	14	5.8897	2.94 × 10−6
ENSG00000088808	14	6.126	1.03 × 10−6
ENSG00000184056	15	5.8509	3.43 × 10−6
ENSG00000141431	18	5.4008	2.70 × 10−5
ENSG00000187323	18	7.5736	3.31 × 10−10
ENSG00000197183	20	5.3692	3.01 × 10−5

Fig. 4.

Manhattan plot of GWAS. Each point represents a SNP tested in the genomic SEM–derived GWAS. The x-axis denotes chromosomal position; the y-axis shows –log10(P). The red horizontal line marks the genome-wide significance threshold (p < 5 × 10⁻⁸); the blue line indicates a suggestive threshold (p < 1 × 10⁻⁵).

Table 4.

Functional annotation of the 32 independent lead SNPs.

rsID	CHR	BP	A2	A1	P	beta	se	Nearest Gene	Func	CADD
rs490647	1	37242743	A	G	2.59 × 10−8	−0.00333966	0.000599857	RP4-614N24.1	intergenic	11.82
rs4635553	2	22440788	C	T	1.44 × 10−8	−0.002922194	0.00051552	AC068490.2	ncRNA_intronic	3.364
rs1025128	2	60175475	C	G	4.86 × 10−9	−0.003026225	0.000517124	RP11-444A22.1	ncRNA_intronic	16.82
rs359234	2	60471409	G	C	9.31 × 10−11	−0.003464993	0.000534903	AC007381.3	intergenic	5.45
rs58117425	2	145681570	A	G	8.78 × 10−9	0.003447626	0.00059931	TEX41	ncRNA_intronic	0.82
rs12992995	2	175197545	A	C	1.75 × 10−8	0.003204395	0.00056867	SP9	intergenic	17.13
rs9837341	3	49664767	G	A	5.22 × 10−9	0.003254966	0.000557367	BSN	intronic	4.36
rs12635614	3	173113041	A	G	2.15 × 10−8	−0.002828163	0.000505062	NLGN1	intergenic	12.73
rs34811474	4	25408838	A	G	3.28 × 10−8	0.003318478	0.00060057	ANAPC4	exonic	23.7
rs10027492	4	28647811	T	A	4.88 × 10−9	0.003007106	0.00051392	RP11-123O22.1	intergenic	3.55
rs10518373	4	121574701	A	G	3.25 × 10−8	0.003216893	0.000581972	RP11-501E14.1	intergenic	6.39
rs13151315	4	140933304	G	C	9.80 × 10−9	0.003011506	0.00052519	MAML3	intronic	1.07
rs171697	5	103956516	G	C	1.45 × 10−9	−0.003340353	0.000552133	RP11-6N13.1	ncRNA_intronic	0.01
rs7724996	5	117247501	T	G	3.36 × 10−8	0.002842591	0.000514796	CTD-3179P9.1	intergenic	2.82
rs72802853	5	152218058	T	C	2.05 × 10−8	0.003098852	0.000552602	AC091969.1	ncRNA_intronic	0.02
rs55785035	7	3709491	G	T	1.09 × 10−8	0.003280588	0.000573945	SDK1	intronic	0.28
rs1020006	7	12255664	A	G	1.72 × 10−9	−0.003144648	0.000522158	TMEM106B	intronic	0.89
rs73125518	7	39425117	A	G	1.49 × 10−9	0.004261784	0.000705003	POU6F2	intronic	4.32
rs7783012	7	114116881	A	G	2.64 × 10−12	−0.003803046	0.00054363	FOXP2	intronic	1.99
rs1552284	8	21962607	G	A	3.00 × 10−10	0.00327183	0.000519431	FAM160B2	downstream	2.30
rs11775287	8	30864339	T	C	4.70 × 10−8	0.002766865	0.000506544	PURG	intronic	1.30
rs13291079	9	96360650	C	T	8.75 × 10−12	0.003667752	0.00053734	PHF2	intronic	6.20
rs9775587	9	122662720	T	G	2.66 × 10−8	−0.002873879	0.000516671	RP11-360A18.2	intergenic	5.12
rs7125588	11	113436072	G	A	2.86 × 10−9	0.00308079	0.000518725	DRD2	intergenic	1.64
rs2899991	14	47273964	C	T	1.78 × 10−10	0.003324606	0.000521155	MDGA2	intergenic	0.35
rs4906365	14	104229230	C	G	4.60 × 10−10	−0.003259725	0.000523045	PPP1R13B:RP11-73M18.11	ncRNA_exonic	9.73
rs8034783	15	47763726	T	C	2.71 × 10−8	−0.004689307	0.000843566	SEMA6D	intronic	1.031
rs8026678	15	91562005	G	A	3.56 × 10−9	−0.003121958	0.000528852	VPS33B	intronic	0.876
rs184148321	17	60873255	G	A	2.15 × 10−29	0.029921527	0.002658095	MARCH10	intronic	2.071
rs17501820	18	50778737	T	C	8.14 × 10−11	−0.00346687	0.000533524	DCC	intronic	0.195
rs596668	18	53264343	G	A	6.88 × 10−12	−0.00515728	0.00075177	TCF4	intronic	2.972
rs1555133	20	31048382	A	G	1.32 × 10−9	−0.003401263	0.00056079	C20orf112	intronic	1.383

A GWAS locus reduction analysis further identified several risk loci, some of which (e.g., rs359234, rs34811474) have been previously associated with multiple traits (Table 5). Rs359234 has been linked to smoking behavior and metabolic traits, rs34811474 has been associated with body mass index (BMI) and cognitive function, and rs7783012 is linked to insomnia and substance abuse (Supplementary Table 4).

Table 5.

Locus reduction analysis.

Novel	rsID	A2	A1
YES	rs490647	A	G
YES	rs4635553	C	T
YES	rs1025128	C	G
No_FD	rs359234	G	C
YES	rs58117425	A	G
YES	rs12992995	A	C
YES	rs9837341	G	A
YES	rs12635614	A	G
No_MCP	rs34811474	A	G
YES	rs10027492	T	A
YES	rs10518373	A	G
No_MDD	rs13151315	G	C
No_MCP	rs171697	G	C
YES	rs7724996	T	G
YES	rs72802853	T	C
YES	rs55785035	G	T
YES	rs1020006	A	G
No_FD	rs73125518	A	G
No_CT	rs7783012	A	G
YES	rs1552284	G	A
YES	rs11775287	T	C
No_LS	rs13291079	C	T
YES	rs9775587	T	G
YES	rs7125588	G	A
YES	rs2899991	C	T
No_FD	rs4906365	C	G
YES	rs8034783	T	C
No_MCP	rs8026678	G	A
YES	rs184148321	G	A
No_MCP	rs17501820	T	C
No_LS, No_FD	rs596668	G	A
YES	rs1555133	A	G

Fine-mapping

Fine-mapping analysis revealed strong associations at multiple genomic locations, identifying four loci with significant colocalization evidence across all tested regions. Among these, rs34811474 (in ANAPC4) and rs184148321 (in MARCH10) had a mean PP > 0.95, suggesting that they may be driven by the same causal variant, thus representing the highest-confidence candidate SNPs. Furthermore, rs186300744 (SuSiE PP = 0.99) and rs28714250 (FINEMAP PP = 0.98) were identified as secondary credible colocalization signals. The fine-mapping results for these four SNPs are illustrated in Fig. 5, and details are provided in Supplementary Table 5.

Fig. 5.

Bayesian fine-mapping results. A, ANAPC4 locus. Manhattan-style regional plot showing GWAS p values (top panel) and PIP from FINEMAP and SuSiE (middle panels), with the mean PIP (bottom). The dashed gray lines mark the genome-wide significance threshold (p < 5 × 10⁻⁸) and PP ≥ 0.95. The missense variant rs34811474 exceeded both thresholds and is highlighted (diamond). B, MARCH10 locus. As in A, rs184148321 (Tier 1; mean PP > 0.95), rs186300744 (Tier 2; SuSiE PP = 0.99), and rs28714250 (Tier 2; FINEMAP.PP = 0.98) are prioritized. Gene models (top) indicate transcript structures within each locus. Abbreviations: PIP, posterior inclusion probability.

Transcriptome-wide association study (TWAS)

We performed a TWAS using FUSION to capture expression-level association signals. After Bonferroni correction, 39 genes reached the significance threshold (Supplementary Table 6; Fig. 6). To pinpoint causal genes, we applied FOCUS to fine-map the TWAS associations. This analysis prioritized 11 genes as high-confidence candidate causal genes (Supplementary Table 7). We then cross-referenced the TWAS and FOCUS results to identify genes supported by multiple lines of evidence. Six genes satisfied both the FOCUS causality criteria and the TWAS significance threshold; thus, these genes were deemed the highest-confidence core risk genes (Table 6). The direction-of-effect analysis showed that TM9SF4, SUSD3, PTPDC1, and VPS33B expression correlated positively with risk (Z > 0), whereas XRCC3 and LSM10 expression correlated negatively (Z < 0). A graphical summary is provided in Fig. 7.

Fig. 6.

TWAS highlights risk and protective genes for the latent high-risk suicide syndrome. Genome-wide TWAS Z-scores (horizontal axis, chromosomal position) are plotted against their signed effect (vertical axis, Z). Red points denote genes exceeding the Bonferroni-corrected significance threshold. |Z| ≈ 4.0; blue dashed lines. Positive Z-scores (above zero) indicate higher genetically predicted expression associated with increased risk; negative Z-scores indicate putative protective effects.

Table 6.

High-confidence causal genes identified by TWAS and FOCUS.

PANEL	ID	CHR	MODEL	TWAS.Z	TWAS.P	pip	P_Bonferroni
sCCA3	TM9SF4	20	lasso	5.0855	3.67 × 10−7	0.997	0.013512206
sCCA3	XRCC3	14	enet	−6.06814	1.29 × 10−9	0.891	4.74952 × 10−5
sCCA2	SUSD3	9	top1	6.03572	1.58 × 10−9	1	5.81724 × 10−5
sCCA3	PTPDC1	9	top1	6.290472	3.17 × 10−10	1	1.16713 × 10−5
sCCA2	LSM10	1	lasso	−5.17816	2.24 × 10−7	0.999	0.008247232
sCCA2	VPS33B	15	lasso	5.19828	2.01 × 10−7	0.999	0.007400418

Fig. 7.

Integrated TWAS + FOCUS analysis prioritizes six high-confidence candidate genes. For each locus, the upper panel shows TWAS –log₁₀p values for all genes in the window; dot size reflects the FOCUS PIP, binned as indicated. Genes exceeding both the Bonferroni threshold and PIP ≥0.8 are boxed in red. The lower triangle displays the correlation matrix of predicted gene expression. A LSM10; B PTPDC1; C SUSD3; D TM9SF4; E VPS33B; and F XRCC3.

Pathway, cell type, and mendelian disease gene enrichment

MAGMA multimarker analysis mapped 56 genes associated with the trait (Supplementary Table 8). When cross-referenced with the TWAS and FOCUS results, XRCC3 and VPS33B remained significant across all three analyses. A gene set enrichment analysis (GSEA) of these 56 genes showed significant enrichment in neural, cognitive, and behavioral function pathways, along with brain structure pathways. These genes were also associated with various physiological and disease phenotypes (e.g., inflammatory bowel disease and gastroesophageal reflux). Biological process enrichment further highlighted key neuronal activities, including synaptic organization, signal transduction, protein localization, postsynaptic structural modifications, and specific processes involving glutamatergic and GABAergic synapses (Supplementary Table 9).

MendelVar enrichment showed significant clustering among multiple Mendelian disease-related genes spanned multiple domains, including neurodevelopmental, hematological, immunological, and musculoskeletal systems (Supplementary Table 10). Cell-type enrichment analysis further localized the association signals specifically to brain tissue. Significant signals were concentrated in several key brain regions, including higher-order cognitive cortices (e.g., the dorsolateral prefrontal cortex, inferior temporal gyrus, and cingulate gyrus), motor-related regions (e.g., the anterior caudate and substantia nigra), and the hippocampus (Supplementary Table 11).

Heritability contribution results at the genomic region level

The partitioned heritability analysis indicated that the trait's genetic signals are unevenly distributed across the genome and significantly enriched in functional noncoding elements. These enriched loci were mostly in regulatory elements characterized by evolutionary conservation, open chromatin, or active enhancer/promoter marks. The Conserved_LindbladToh regions showed the highest heritability enrichment, with signals extending 500 bp into flanking sequences, suggesting important regulatory roles for neighboring DNA. Furthermore, Coding_UCSC regions and Promoter_UCSC regions demonstrated significant enrichment. Details of the gene regulatory network associated with this complex trait are summarized in Supplementary Table 12.

Risk factor annotation analysis

We identified >10 phenotypes showing a potential statistical association with high-risk suicide syndrome, spanning several categories (psychiatric and neurological disorders, immunological and inflammatory diseases, cardiovascular and metabolic conditions, and other clinical phenotypes). In the psychiatric/neurological domain, sleep problems (e.g., insomnia) were associated with increased suicide risk. In the immune/inflammatory category, asthma, rheumatoid arthritis, allergic diseases, irritable bowel syndrome, and gastroesophageal reflux disease were risk factors. In the cardiovascular/metabolic category, obesity, type II diabetes, coronary heart disease, and hypertension were associated with increased risk. A higher basal metabolic rate and breast cancer were found to be potentially associated with increased risk. Detailed results are provided in Supplementary Table 13.

Chromosome-level results

The genetic contribution of each chromosome to the latent factor, quantified by summing the posterior SNP effect sizes per chromosome, showed significant heterogeneity across chromosomes. Chromosomes 2 and 3 exhibited the most pronounced genetic contributions (−0.17 and −0.15, respectively), potentially harboring key genes or core regulatory elements that influence susceptibility to the syndrome.

Discussion

We used genomic SEM and multiple analytical approaches—including summary-statistic PRS, Mendelian randomization using data from the IEU database, fine-mapping, and transcriptomic analyses—to integrate seven risk traits spanning biological, psychological, and social dimensions. This approach enabled the construction of a latent genetic risk factor for increased suicidal behavior. Subsequently, we performed a systematic genome-wide analysis of this latent factor. The constructed single-factor model exhibited a good fit to the data; thus, the seven included risk phenotypes can be largely attributed to a single underlying genetic factor. In other words, these diverse risk factors share a common genetic foundation that drives individual differences in suicide susceptibility. In the GWAS of the latent factor, we identified 24 potential risk genes, four genome-wide significant SNP loci that appear to be high-confidence causal variants, and six putative causal genes. Furthermore, we found that genetic factors not only determine an individual's baseline risk of suicidal behavior but also exert far-reaching effects on suicide propensity. These effects may be mediated through various pathways, including susceptibility genes, physical illnesses, neurodevelopmental processes, and cellular-level mechanisms. These findings provide a novel perspective on the biological basis of suicidal behavior and lay a foundation for future functional studies and eventual clinical translation.

Genomic SEM revealed the genetic covariances among the seven traits, indicating that the seven phenotypes share common genetic factors. A history of suicide or self-harm and major depressive disorder (MDD) emerged as especially influential components of this genetic network. This genome-level finding provides strong biological support for phenomena that have long been observed in clinical and epidemiological studies. Multiple analyses have confirmed that, among all known risk factors, a history of suicidal behavior or self-harm is the single strongest predictor of future completed suicide [41,42]. Other studies have demonstrated that the risk of repeated self-harm and completed suicide is highest within the first year following a self-harm event [43]. The association of MDD with suicide risk is evident throughout the trajectory from suicidal ideation to suicidal behavior, ultimately leading to a markedly elevated risk of death by suicide. Recent meta-analytic evidence has demonstrated that MDD significantly increases the incidence of suicidal ideation, planning, and attempts [44,45] and confers one of the highest risks of death by suicide among all mental disorders [46].

The complex interplay of numerous other risk factors also plays a key role. Childhood trauma is often regarded as a fundamental long-term risk factor for future suicidal behavior. Such trauma lays the groundwork for later mental disorders by affecting neurodevelopment and emotion regulation and increases an individual's risk of suicide attempts in adulthood [47]. Moreover, the risk of suicidality can arise from intense psychological pain exacerbated by loneliness and is often further heightened by external social pressures, such as financial hardship [[48], [49], [50], [51]]. The structural equation model further confirmed the complex genetic links among these factors, suggesting that these traits are not independent but rather interwoven and jointly influential.

We identified multiple key SNPs in the latent factor GWAS. Notably, most of these association signals were not identified in the original single-trait GWAS, suggesting that our multivariate integration strategy uncovered novel variant loci that traditional single-phenotype approaches failed to capture. However, because the significance of a single SNP does not directly imply causation, we used fine-mapping techniques to narrow down the set of candidate variants. We pinpointed specific, likely causal variants at four genomic loci.

Notably, a missense mutation (rs34811474) in the ANAPC4 gene emerged as an important risk locus. ANAPC4 is a critical subunit of the anaphase-promoting complex/cyclosome (APC/C) within the ubiquitin–proteasome system (UPS). UPS dysfunction impairs synaptic function and neural plasticity and is implicated in the pathology of various psychiatric disorders [[52], [53], [54], [55], [56], [57], [58], [59], [60]]. Rs34811474 may alter ANAPC4 function, triggering subtle dysregulation of the APC/C–UPS pathway. This dysregulation could persistently impair neuronal function in the mature brain and potentially stems from subtle deviations during early neurodevelopment, thereby shaping a “stress-sensitive” brain architecture [61,62]. Notably, this locus exhibits broad pleiotropy; it is associated with multiple traits, such as intelligence, cognitive function, and BMI [[63], [64], [65], [66], [67], [68]]. This finding further underscores the importance of ANAPC4 and the UPS pathway as fundamental biological hubs in mental disorders.

MARCH10 is a novel candidate gene identified in this study; to date, no published research has directly linked any specific MARCH10 variants to psychiatric disorders. However, there is a sound biological basis supporting MARCH10's potential pathogenicity. MARCH10 encodes an E3 ubiquitin ligase in the MARCH family, which serves as a key executor of the UPS [69]. Moreover, the critical role of the UPS pathway in the pathology of neuropsychiatric disorders is well documented. More importantly, convergent evidence from multiple databases indicates that MARCH10 is actively expressed in the human CNS. For example, MARCH10 is expressed in both adult and developing brain tissue [70,71], and its protein is enriched in cerebrospinal fluid [72]. Notably, ENCODE project data show that the MARCH10 promoter is active in neural stem/progenitor cells, strongly suggesting that this gene may play a role during critical neurodevelopment stages. Beyond its function and expression profile, MARCH10's chromosomal location provides indirect yet compelling genetic evidence for its involvement in psychopathological processes. MARCH10 is located on chromosome 17, and a landmark GWAS by Docherty et al. revealed that this chromosomal region is significantly associated with suicide risk [73]. In this context, the three intronic MARCH10 SNPs identified in our study (rs184148321, rs186300744, and rs28714250) are particularly noteworthy. Intronic regions are rich in regulatory elements and play a crucial role in the precise spatiotemporal regulation of gene expression [74,75]. Therefore, we posit that these disease-associated intronic variants may disrupt MARCH10 expression in the brain by affecting its transcription efficiency or splicing patterns. Such variant-induced expression changes—especially those occurring during critical periods, such as neurodevelopment—could disrupt normal UPS function. In summary, fine-mapping prioritized two loci with compelling evidence: rs34811474 (ANAPC4; missense) and rs184148321 (MARCH10; intronic). To our knowledge, these have not been reported at genome-wide significance for suicidal behavior, although both regions show associations with related traits, such as body mass index and cognitive measures, suggesting pleiotropic pathways intersecting the latent suicidal behavior liability. The MARCH10 signal resides within a chromosome 17 region previously implicated in suicide risk, reinforcing its relevance. Accordingly, these loci warrant prioritized, hypothesis-driven follow-up to evaluate causality and biological plausibility.

We identified six high-confidence candidate causal genes using a combination of TWAS and FOCUS analyses. These genes were identified as significant by both methods, suggesting that changes in their expression likely directly impact the latent high-risk suicide syndrome phenotype.

The central nervous system (CNS) exhibits an extremely high oxygen consumption rate and an active metabolism, resulting in substantial oxidative stress. Consequently, neuronal genomes are highly susceptible to damage, placing stringent demands on DNA repair mechanisms [76]. In this context, XRCC3—a core component of the homologous recombination DNA repair pathway—plays an indispensable role in safeguarding genomic integrity [77]. Its importance is evident. XRCC3 is stably expressed in the brain and is involved in forming neuron-specific DNA repair complexes [78]. Moreover, XRCC3 helps maintain the mitochondrial genome (mtDNA) [79], which is notable since mitochondrial dysfunction is a known pathological feature of neuropsychiatric disorders. This biologically plausible hypothesis is directly supported by genetic research; for example, a schizophrenia study revealed that individual differences in DNA repair efficiency modulated by XRCC3 polymorphisms contribute to the disorder's pathophysiology [80]. Accordingly, we posit that reduced XRCC3-mediated repair may diminish neurons' capacity to manage ongoing endogenous DNA damage, potentially contributing to increased suicide risk.

Another key gene, VPS33B, is a core regulator of the intracellular membrane transport system. Its primary function is to regulate vesicle-mediated protein transport and SNARE-dependent membrane fusion events [81]. This process is fundamental for maintaining the orderly exchange of materials, secretion, and endocytosis between organelles. Single-cell and bulk tissue data indicate that VPS33B is expressed in many tissues and at the highest level in the CNS (https://www.proteinatlas.org/). The Mendelian disorder ARC syndrome provides insight into VPS33B function. ARC syndrome results from the complete loss of function of VPS33B and is accompanied by severe CNS abnormalities and intellectual disability, directly linking this gene to nervous system integrity [82]. Mechanistically, loss of VPS33B function may contribute to psychiatric disease pathology by disrupting synaptic vesicle cycling and neurotransmitter release. Future research should include functional experiments to validate this model and elucidate the precise role of VPS33B in psychiatric disorders.

Similarly, TM9SF4, a multi-pass transmembrane protein, is a key cellular stress response regulator. TM9SF4 has been linked to inflammation, endoplasmic reticulum stress, and autophagy [[83], [84], [85], [86]]. Endoplasmic reticulum stress and neuroinflammation are increasingly been recognized as core pathophysiological mechanisms in MDD, bipolar disorder, and schizophrenia [87,88]. Thus, TM9SF4 is a compelling target for further study. Moreover, owing to its enzymatic and signaling roles, TM9SF4 is potentially druggable, which has been partially demonstrated in cancer research [85]. SUSD3 and PTPDC1 are located in adjacent regions on chromosome 9. PTPDC1 expression is significantly downregulated in a subset of oligodendrocytes in Parkinson's disease [89] and is expressed in astrocytes and the brain, indicating potential functions in the CNS. SUSD3 regulates cell–cell and cell–extracellular matrix interactions and is involved in cell migration and cytoskeletal organization [90].

Overall, these findings suggest that neuronal fragility may be critical in genetic susceptibility to suicidal behavior. Collectively, these genes illustrate how cells respond to stress. Chronic psychosocial stress is a major environmental risk factor for depression and suicidal behavior; at the cellular level, this stress manifests as an accumulation of misfolded proteins, organelle damage, and metabolic disruption. A healthy neuron depends on internal protein homeostasis to clear these harmful products. The risk genes identified in this study are at different key nodes of this process; prolonged stress could overwhelm neurons, leading to functional disruption that manifests as psychiatric behaviors.

Furthermore, we identified over a dozen phenotypes potentially associated with the genetically derived high-risk suicide syndrome. The breadth of this finding suggests that suicidal tendency is not an isolated psychiatric entity but a complex syndrome tightly interwoven with an individual's overall health. Our results reveal several distinct pathophysiological clusters. First, the strong association with insomnia supports prior observational studies and meta-analyses indicating that sleep disturbances are key risk factors for suicidal ideation and behavior [91]. Furthermore, interventions targeting insomnia (e.g., cognitive-behavioral therapy for insomnia, CBT-I) may be beneficial for suicide prevention [92]. Second, a notable immune–inflammatory signal emerged, involving conditions such as asthma, rheumatoid arthritis, and irritable bowel syndrome [[93], [94], [95]]. Although these diseases can undermine mental health via psychosocial factors (e.g., chronic pain and functional impairment) [[96], [97], [98]], our genetic evidence points to a biological explanation. Similarly, associations with cardiometabolic diseases (obesity, type II diabetes, and coronary heart disease) and cancer risk align with a large body of epidemiological evidence [[99], [100], [101], [102], [103]], further supporting the role of systemic physiological dysregulation in the pathogenesis of suicidal behavior.

Notably, these findings remain exploratory and should be interpreted with caution. We did not apply a multiple-testing correction; thus, some nominally significant associations may be false positives. Furthermore, although Mendelian randomization strengthens causal inference, it cannot completely rule out horizontal pleiotropy or other biases. Despite these limitations, our study has outlined a potential “hypothesis-generation map” that provides direction for further dissection of the complex biological networks intertwined with suicide risk.

GSEA revealed that the functions of the identified genes converge on synaptic, neurocognitive, and neuroimmune pathways. This finding underscores the intimate connection between neurobiological dysregulation and suicide risk. Most prominently, glutamatergic and GABAergic synapses were implicated. This finding suggests that imbalances in excitatory–inhibitory signaling and impaired synaptic plasticity are key neurobiological mechanisms underlying suicide susceptibility, consistent with findings from prior genetic and postmortem brain studies [[104], [105], [106]]. Another notable result was the enrichment of risk genes in inflammatory disease phenotypes, providing potential genetic support for an immunological dimension of suicide risk and echoing substantial evidence linking systemic inflammation with suicidal behavior. Disruption of synaptic function may interact with—and even exacerbate—neuroimmune dysregulation, jointly shaping an individual's susceptibility to suicidal behavior.

Cell–type–specific enrichment analysis localized the genetic signals of suicide syndrome to specific cell types within the brain tissue. We observed significant heritability enrichment in cell populations from key brain regions involved in emotion regulation and cognitive control, namely, the prefrontal cortex (including the dorsolateral prefrontal cortex), cingulate gyrus, hippocampus, and striatum. These brain areas have long been implicated in depression and suicidal behavior. For example, the dorsolateral prefrontal cortex is responsible for cognitive control and impulse inhibition, and reduced activity in this region has been linked to suicidal impulsivity [107,108]. The hippocampus and cingulate gyrus are involved in emotional memory and regulation [109,110]. Both regions have frequently been reported to show abnormalities in the brains of individuals who died by suicide [111,112]. Our results provide genetic evidence that suicide susceptibility is primarily mediated through effects on neural cell function in specific brain regions. Notably, some peripheral tissues and cell types showed no significant enrichment, suggesting that the genetic risk of suicide is more likely mediated through central (brain) mechanisms than via direct effects on peripheral organs.

Our partitioned heritability analysis revealed a key feature of this trait's genetic architecture: its heritability is not evenly distributed across the genome. Instead, heritability is significantly enriched in a small subset of regions with core biological functions, particularly evolutionarily conserved noncoding regulatory elements. These regions have remained highly stable over long evolutionary timescales—a hallmark of strong negative selection—indicating that they carry functions crucial to life [113]. Notably, this enrichment pattern is not unique to our trait of interest; it closely aligns with the genetic architecture of many classic complex traits (e.g., schizophrenia, height, and educational attainment). This similarity suggests that the enrichment pattern may represent a general rule for polygenic traits [114]. Building on this, our findings further hint at more fine-tuned regulatory mechanisms within these regions. Evidence suggests that the driver of trait variation may not be a single isolated conserved element but a functionally synergistic “regulatory module” comprising core conserved sequences and their surrounding auxiliary regulatory regions—a view supported by modern functional genomics studies [[115], [116], [117]]. Additionally, we observed changes in noncoding RNA expression closely associated with disease susceptibility in some high-risk genomic regions.

Our findings suggest several potential translational pathways. First, for risk prediction, a PRS derived from our “high-risk suicide syndrome” factor may provide a more refined genetic signal than a PRS based on a single, heterogeneous “suicide attempt” GWAS. Second, our study provides novel, high-priority targets for therapeutic development. The newly identified loci (ANAPC4, MARCH10) and high-confidence causal genes (XRCC3, VPS33B, TM9SF4) represent a set of potentially druggable targets. Finally, at the pathway level, our GSEA results, strongly implicating glutamatergic and GABAergic synaptic pathways, provide crucial genetic validation for the ongoing development of rapid-acting antidepressants and related neurotherapeutics. We caution, however, that these applications are hypothesis-generating. Any potential for clinical utility would require extensive, prospective validation before any clinical use.

Our study design and analyses resulted in several important findings. However, our study has some limitations. First, regarding the construction of the latent phenotype, the high-risk suicide syndrome factor we extracted via genomic SEM is a statistical abstraction whose validity depends on the comprehensiveness and relevance of the input phenotypes chosen. Although we attempted to cover multiple psychological, social, and biological factors, genetic influences from risk factors that we did not include (e.g., impulsivity, personality traits, and substance abuse) may exist. Furthermore, other potential gene–environment interactions may exist that were not captured. Second, although we identified multiple genetic loci associated with high suicide risk through fine-mapping and transcriptomic analyses, linking these genes to specific biological mechanisms remains a challenge. Third, population applicability is an important consideration. Our study exclusively used GWAS data predominantly from individuals of European ancestry, and the samples in each database were mainly from European-ancestry populations. Therefore, it is uncertain whether our findings can be generalized to other ethnic groups or populations.

In conclusion, this study characterized the latent genetic architecture of high-risk suicide populations using a whole-genome approach. Integrating genomic SEM, fine-mapping, and transcriptomic analyses, we identified several novel loci and described how they mediate genetic links between gene expression and complex traits. Although genetic predisposition does not equate to destiny, these pathways and nodes provide actionable targets for medical and social interventions that can alter the trajectories of high-risk individuals. Genetic analyses reveal modifiable biological pathways underlying suicide susceptibility. As these key nodes are evaluated and translated into clinical tools, genetic information will shift from fate to a foundation for precision prevention.

Data availability

The data supporting the results of this study are available from the corresponding author upon reasonable request.

Contributions

CZ designed, conceptualized the study, supervised this work, wrote and reviewed the paper, KY conducted data analysis and interpretation, and wrote the paper. XC, CL, HS, JZ, HT conducted data analysis and interpretation. All authors contributed to and approved the final version of the paper.

Funding

This work was supported by grants from the National Natural Science Foundation of China (Nos. 81871052 and 82171503), and Tianjin Health Research Project (No. TJWJ2025ZK009). The funders had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary data

The following are the Supplementary data to this article.