Disorder-specific and shared genetic architecture underlying schizophrenia and bipolar disorder

Abstract

Schizophrenia (SCZ) and bipolar disorder (BIP) share substantial common-variant liability but differ in cognitive, comorbidity, and treatment response. Here we decomposed these disorders into schizophrenia-predominant, bipolar-predominant, and shared psychosis dimensions to test whether these components show distinct pleiotropic and biological profiles than the original disorder. Using the largest available SCZ and BIP GWAS, we applied bidirectional mtCOJO and Genomic SEM to derive SCZcondBIP, BIPcondSCZ, and PSY-shared and validated them using inter-component genetic correlations, FinnGen psychiatric endpoints, and Genomic SEM latent factors. We then characterized each component across cognitive, cardiometabolic, and immune traits, followed by genomic risk-locus discovery, pathway analysis, developmental expression profiling, and drug-target enrichment. The three components showed marked divergence. SCZcondBIP was negatively correlated with cognition, education, metabolic syndrome, C-reactive protein, and neutrophil percentage, whereas BIPcondSCZ showed the opposite cognitive profile and shifted toward positive cardiometabolic and immune correlations. PSY-shared remained positively correlated with education and negatively correlated with cognitive task performance, immune and metabolic traits. We identified 248 consensus genomic risk loci, including 81 not detected in the input disorder GWAS. Biologically, PSY-shared was enriched for synaptic signalling, ion-channel, and neurodevelopmental pathways; SCZcondBIP primarily implicated synaptic-signalling and cellular-homeostasis pathways; and BIPcondSCZ showed weaker but distinct enrichment for synaptic-vesicular biology. Drug-target enrichment further separated the components, with strong antipsychotic enrichment for PSY-shared and distinct non-antipsychotic signals for the conditional factors. These findings show that SCZ and BIP genetic risk is best understood as biologically distinguishable shared and disorder-predominant dimensions that differentially map onto cognitive, cardiometabolic, immune, and molecular architecture. These findings provide a framework for evaluating whether component-specific polygenic scores improve stratification of cognitive, cardiometabolic, and inflammatory heterogeneity across severe psychiatric illness.

Introduction

Schizophrenia (SCZ) and bipolar disorder (BIP) are severe psychiatric disorders that impose a substantial global burden worldwide^1–3, typically emerging during late adolescence or early adulthood that is a critical period for educational and social development^4,5. Both disorders are highly heritable, with twin-study estimates ranging from 60–80% for SCZ^6,7 and 60–85% for BIP^8,9. Although the two disorders are strongly genetically correlated (rg ~ 0.67)¹⁰, reflecting an extensive shared biology^11–14, they retain highly distinct symptom profiles, longitudinal trajectories, and treatment responses^15–19. This clinical heterogeneity, combined with their incomplete genetic overlap, underscores that a shared transdiagnostic liability cannot fully account for the unique pathophysiology of either disorder²⁰.

Genome-wide association studies (GWAS) have identified a growing number of common-variant risk loci^8,21–23. The largest schizophrenia GWAS identified 263 loci²¹, while bipolar GWAS have expanded from 64²³ to as many as 298 loci in multi-ancestry analyses⁸. Despite these advances, these loci nonetheless explain only a fraction of heritable liability and largely implicate broad neuronal and synaptic processes, offering limited resolution into mechanisms that distinguish disorder-specific biology^20,24.

To disentangle shared and disorder-specific genetic risk, several complementary approaches have been developed. Multi-trait conditional and joint analysis (mtCOJO)²⁵, conditions one disorder on another to recover independent genetic signals, although early applications were constrained by limited GWAS power²⁶. Genomic Structural Equation Modeling (Genomic SEM)²⁷ models genetic covariance among disorders as latent factors. Landmark analyses of 11 and 14 disorders identified the shared architecture (p-factor) of the psychiatric spectrum across a broad range of psychiatric disorder but it inherently fuses SCZ and BIP onto a single shared dimension (F2_SchizophreniaBipolar), rather than delineating separable specific liabilities^13,14; Case–Case GWAS (CC-GWAS) directly contrasts clinical populations to identify divergent allele frequencies, but remains underpowered for the SCZ–BIP comparison and cannot yield a propagatable disorder-specific liability for downstream directional pleiotropy²⁸.

Recently, a Genomic SEM–based GWAS-by-subtraction approach partitioned psychotic-disorder liability into a schizophrenia-specific component and a component shared with bipolar disorder²⁹. By removing the genetic variance shared between SCZ and BIP, the residual schizophrenia-specific component revealed a negative genetic correlation with educational attainment (rg = −0.06), whereas the shared component showed a positive correlation (rg = 0.11), suggesting that disorder-level genetic correlations may obscure opposing relationships embedded within shared and specific liability components. However, the biological pathways underlying the schizophrenia-specific liability remain unresolved, bipolar-specific liability was not explicitly modelled, and phenotypic characterisation was restricted to educational attainment — leaving open how these components relate to the broader endophenotypic landscape of psychosis, including cognitive, cardiometabolic, and immune traits.

Patients with schizophrenia and bipolar disorder exhibit elevated rates of cardiometabolic comorbidity^30–32 and immune dysregulation^33–35, yet whether these reflect shared or disorder-specific genetic liability remains unclear. Prior GWAS-by-subtraction analysis²⁹ offered an initial window into this question, demonstrating that the schizophrenia-specific and shared psychosis components diverge sharply on educational attainment, consistent with our observation that schizophrenia genetic risk is composed of at least two biologically distinct groups of variants: one that simultaneously increases disease risk and reduces cognitive performance, and another that paradoxically increases both disease risk and educational attainment^36,37. Furthermore, our recent comprehensive mendelian randomisation analysis of circulating proteins across psychiatric phenotypes also demonstrated that immune and inflammatory proteins exhibit the strongest cross-phenotype associations³⁸, and separately, neutrophil ratio of leukocyte has since been identified as a potential predictor of antipsychotic treatment response^39,40, implicating immune-inflammatory biology as a further axis along which shared and disorder-specific liability may diverge. Taken together, we hypothesise that the shared and disorder-specific components of psychosis liability are biologically dissociable along cognitive/neurodevelopmental, cardiometabolic, and immune-inflammatory dimensions, a structure that remains obscured when disorders are analysed at the diagnostic level.

Previous conditional GWAS²⁶, Genomic SEM^13,14, and GWAS-by-subtraction²⁹ approaches demonstrated the feasibility of, and laid the conceptual groundwork for, resolving shared and disorder-specific components of psychosis liability. Inspired by these findings, here we applied mtCOJO and Genomic SEM to the latest large-scale GWAS of SCZ and BIP to resolve those into three genetically distinguishable components: SCZ conditioned on BIP (SCZcondBIP), BIP conditioned on SCZ (BIPcondSCZ), and a shared psychosis factor (PSY-shared).

Characterizing each component against cognitive (Cognitive task performance, educational attainment), cardiometabolic (Metabolic syndrome), and immune phenotypes (CRP, Neutrophil percentage), and through gene-set and drug-target enrichment analyses, we reveal divergent genetic profiles and distinct biological architectures that are obscured when disorders are analyzed in isolation.

Methods:

Overview of Methodological Framework:

To systematically disentangle the shared and distinct genetic architectures of schizophrenia and bipolar disorder, we implemented a comprehensive, multi-stage analytical pipeline. First, we leveraged the largest available GWAS for schizophrenia (PGC3; 76,755 cases and 243,649 controls)²¹ and bipolar disorder (57,607 cases and 371,549 controls)⁸ to isolate disorder-specific genetic variance using multi-trait conditional and joint analysis (mtCOJO)²⁵ and captured the shared transdiagnostic liability using Genomic Structural Equation Modelling (Genomic SEM)²⁷ yielding three genetically distinguishable components: SCZcondBIP, BIPcondSCZ, and PSY-shared.

We validated the decomposition of the shared and disease specific components from the original disorder by estimating their inter-component genetic correlations in addition to genomic correlation with independent psychiatric endpoints from FinnGen F5 spanning the schizophrenia and bipolar spectra (schizophrenia: 7,312 cases, 484,297 controls; bipolar disorders: 9,221 cases, 432,672 controls; schizoaffective disorder: 3,168 cases, 484,297 controls; other and unspecified psychotic disorders: 9,216 cases, 484,297 controls)⁴¹, and with the schizophrenia–bipolar (SB) and general psychopathology (p) latent factors from a Genomic SEM analysis of 14 psychiatric disorders¹⁴ using LDSC⁴² with pre-computed LD scores from the 1000 Genomes Project Phase 3 European-ancestry panel.

We subsequently characterised these derived components by estimating their genetic correlations with cognitive task performance (N = 260,576)⁴³, educational attainment (N = 766,345)⁴⁴, metabolic syndrome (N = 4,947,860)⁴⁵, C-reactive protein (N = 575,531)⁴⁶ and neutrophil percentage of leukocytes⁴⁷ (N = 394,642).

Finally, we functionally characterised each component through genomic risk-locus definition and candidate-gene mapping (FUMA, version 2.1.5)⁴⁸, identification of novel loci against published GWAS (BEDTools)⁴⁹, competitive gene-set enrichment (MAGMA)⁵⁰, and drug-target and druggability assessment (DGIdb version 5.0/Enrichr)^51–54 (Figure 1). The analytical workflow is detailed in the Supplementary Methods section.

Figure 1 |. Study design: decomposition of schizophrenia and bipolar disorder GWAS into disorder-specific and psychosis-shared components and their downstream characterisation.

Summary-statistic-level GWAS of schizophrenia (SCZ; PGC3) and bipolar disorder (BIP; 2024 release) were jointly decomposed into three components. Bidirectional multi-trait conditional and joint analysis (mtCOJO) was used to derive SCZcondBIP (SCZ conditioned on BIP) and BIPcondSCZ (BIP conditioned on SCZ), isolating schizophrenia-specific and bipolar-specific common-variant liability. In parallel, Genomic Structural Equation Modelling (Genomic SEM) fitted a common factor on the two disorders to derive the psychosis-shared component (PSY-shared), indexing liability shared across SCZ and BIP. The three components were then characterised through six complementary downstream analyses: (i) LDSC genetic correlations with related psychiatric disorders and cognitive, cardiometabolic and inflammatory endo-phenotypes; (ii) identification of novel genomic risk loci relative to the original SCZ and BIP GWAS; (iii) developmental-timing analysis of implicated genes across BrainSpan periods; (iv) MAGMA gene-set pathway enrichment; (v) Enrichr drug-target enrichment; and (vi) DGIdb-based druggability assessment of FUMA-mapped genes.

Result:

Genetic correlations

Relationships among the derived components

The three derived components showed distinct but interpretable patterns of genetic correlation. The schizophrenia-conditional (SCZcondBIP) and bipolar-conditional (BIPcondSCZ) components were significantly negatively correlated (rg = −0.18, SE = 0.028, p = 2.28 × 10⁻¹⁰), indicating partial divergence of the disorder-specific liability isolated by mtCOJO. Both conditional components were positively correlated with the PSY-shared factor (SCZcondBIP: rg = 0.69, SE = 0.016; BIPcondSCZ: rg = 0.59, SE = 0.016), consistent with a common underlying liability captured by the Genomic SEM common factor. Each conditional component aligned strongly with its index disorder and only modestly with the alternate disorder. The schizophrenia-conditional component was highly correlated with the input PGC3 schizophrenia GWAS (rg = 0.92) and only modestly with the bipolar 2024 GWAS (rg = 0.32). Conversely, the bipolar-conditional component correlated strongly with bipolar disorder (rg = 0.88) and weakly with schizophrenia (rg = 0.23). The PSY-shared factor was highly correlated with both schizophrenia (rg = 0.92) and bipolar disorder (rg = 0.91), supporting its interpretation as a shared psychosis liability dimension (Figure 2a; Supplementary Table 1).

Figure 2 |. Correlational profiles of SCZcondBIP, BIPcondSCZ and the PSY-shared factor relative to the original SCZ and BIP GWAS.

LDSC genetic correlations (rg) between the two original disorder GWAS (SCZ, BIP), the two mtCOJO-derived conditional components (SCZcondBIP, BIPcondSCZ) and the Genomic-SEM common factor (PSY-shared, “Shared factor”) and a set of psychiatric, cognitive, cardiometabolic and inflammatory reference phenotypes. Error bars in b–edenote ±1 standard error.

a, Pairwise rg among the five profiles (SCZ raw, BIP raw, SCZcondBIP, BIPcondSCZ, PSY-shared). Colour scale from blue (rg = −1) to dark red (rg = +1); diagonal fixed to 1.0. The two conditional components are modestly negatively correlated (rg = −0.18), while both remain positively correlated with the psychosis-shared factor (rg = 0.69 for SCZcondBIP, rg = 0.59 for BIPcondSCZ), supporting the interpretation of three biologically distinct but partially overlapping axes.

b, rg with FinnGen psychosis-spectrum endpoints (F5_Schizophrenia, F5_Schizoaffective disorder, F5_Bipolar disorders, F5_Other psychotic disorders). PSY-shared and SCZcondBIP track FinnGen schizophrenia and other psychotic disorders most strongly, whereas BIPcondSCZ tracks the FinnGen bipolar endpoint; schizoaffective disorder correlates with all five profiles.

c, rg with higher-order Genomic-SEM factors from the 14-disorder model of Grotzinger et al. (2025): the schizophrenia–bipolar (SB, F2) factor and the general psychopathology (p) factor. PSY-shared aligns most closely with the SB factor (rg ≈ 1) and shows a positive but weaker correlation with the p factor; both conditional components show attenuated correlations with each higher-order factor.

d, rg with cognitive task performance (CTP; Lam et al. 2021) and educational attainment (EDU; Lee et al. 2018). SCZ, BIP and PSY-shared show the familiar cross-disorder paradox (negative rg with CTP, mixed or positive rg with EDU); decomposition resolves this into an SCZcondBIP profile that is negatively correlated with both traits and a BIPcondSCZ profile that is positively correlated with both.

e, rg with cardiometabolic and inflammatory phenotypes: metabolic syndrome (MetS; GCST90444487), C-reactive protein (MTAG; GCST90179146) and neutrophil percentage (GCST9046809). SCZcondBIP is significantly negatively correlated with all three phenotypes, BIPcondSCZ shifts toward positive correlations, and PSY-shared lies between the two conditional components for every phenotype (rg = −0.06, −0.04 and −0.03 for MetS, CRP and neutrophil percentage, respectively).

Correlations with FinnGen psychiatric endpoints

External validation using FinnGen R13 endpoints further distinguished the components. The schizophrenia-conditional component showed strongest correlations with psychotic-spectrum phenotypes, including schizophrenia (F5_SCHZPHR, rg = 0.92), other and unspecified nonorganic psychosis (F5_PSYCHOTH, rg = 0.74), and schizoaffective disorder (F5_SCHIZOAFF, rg = 0.67), with a weak correlation with bipolar disorder (F5_BIPO, rg = 0.19) (Figure 2b; Supplementary Table 1).

The bipolar-conditional component showed the opposite pattern, correlating most strongly with bipolar disorder (F5_BIPO, rg = 0.78) and more modestly with schizoaffective disorder (F5_SCHIZOAFF, rg = 0.40), other psychosis (F5_PSYCHOTH, rg = 0.26), and schizophrenia (F5_SCHZPHR, rg = 0.10).

The PSY-shared factor showed consistently strong correlations across both psychotic and affective endpoints, including schizoaffective disorder (rg = 0.84), schizophrenia (rg = 0.83), other psychosis (rg = 0.80), and bipolar disorder (rg = 0.71) (Figure 2b; Supplementary Table 1).

Correlations with GenomicSEM psychiatric factors

We next examined correlations with the shared p-factor and schizophrenia-bipolar (SB) factor derived using GenomicSEM on 14-disorder¹⁴. All three components were positively correlated with the SB factor, most strongly for the PSY-shared factor (rg = 0.99), followed by the schizophrenia-conditional (rg = 0.73) and bipolar-conditional (rg = 0.51) components. All components also correlated with the higher-order general psychopathology (p) factor, again with the strongest association for the PSY-shared factor (rg = 0.64), followed by the bipolar-conditional (rg = 0.47) and schizophrenia-conditional (rg = 0.35) components (Figure 2c; Supplementary Table 1).

Correlations with cognitive and educational traits

The components exhibited divergent relationships with cognitive task performance (CTP)⁴³ and educational attainment (EA3)⁴⁴. The schizophrenia-conditional component was negatively correlated with both cognitive performance (rg = −0.28, SE = 0.023) and educational attainment (rg = −0.037, SE = 0.023). In contrast, the bipolar-conditional component showed positive correlations with cognitive performance (rg = 0.077, SE = 0.025) and educational attainment (rg = 0.14, SE = 0.026). The PSY-shared factor showed a mixed pattern resembling that of original schizophrenia, with a negative correlation with cognitive performance (rg = −0.17, SE = 0.019) and a positive correlation with educational attainment (rg = 0.074, SE = 0.018). Thus the shared axis inherits the cognition-negative signature of schizophrenia, whereas the sign inversion between SCZcondBIP and BIPcondSCZ reveals opposing residual cognitive relationships that are masked in the original disorders (Figure 2d; Supplementary Table 2).

Correlations with cardiometabolic and inflammatory profiles

Genetic correlations with cardiometabolic and immune traits paralleled the cognitive pattern. The schizophrenia-conditional component was negatively correlated with metabolic syndrome (MetS: rg = −0.10, SE = 0.019, p = 7.35 × 10⁻⁸), C-reactive protein (CRP: rg = −0.078, SE = 0.019, p = 3.46 × 10⁻⁵) and neutrophil percentage (rg = −0.060, SE = 0.019, p = 1.57 × 10⁻³). By contrast, the bipolar-conditional component showed positive but non-significant correlations with metabolic syndrome (rg = 0.033, SE = 0.020, p = 0.109), C-reactive protein (rg = 0.040, SE = 0.022, p = 0.068) and neutrophil percentage (rg = 0.028, SE = 0.022, p = 0.198). The original input disorders showed the same directional asymmetry in attenuated form: PGC3 schizophrenia was negatively correlated with metabolic syndrome (rg = −0.085), CRP (rg = −0.065) and neutrophil percentage (rg = −0.042), whereas the bipolar disorder GWAS was essentially unrelated to these traits (MetS rg = −0.021, CRP rg = −0.003, neutrophil rg ≈ 0.000, all p > 0.23). Sign asymmetry between SCZcondBIP and BIPcondSCZ was therefore substantially larger than between the raw disorder GWASs, indicating that the conditional decomposition unmasks a cardiometabolic and inflammatory dimension that is diluted at the disorder level (Figure 2e, Supplementary Table 2). PSY-shared showed intermediate correlations with the same phenotypes, sitting between SCZcondBIP and BIPcondSCZ in magnitude and direction. It was significantly negatively correlated with metabolic syndrome (rg = −0.063, SE = 0.014, p = 1.32 × 10⁻⁵) and with C-reactive protein levels (rg = −0.043, SE = 0.015, p = 4.47 × 10⁻³), while its correlation with neutrophil percentage did not reach significance (rg = −0.026, SE = 0.016, p = 0.104).

Conditional decomposition amplifies phenotypic differences that are diluted between the unpartitioned disorders

To determine whether the divergent phenotypic profiles of the two conditional components (SCZcondBIP vs. BIPcondSCZ) and, in parallel, of the two unpartitioned disorders (SCZ vs. BIP) reflected statistically significant differences in their genetic correlations with the same third phenotype, we applied a correlated-overlapping test of the difference between two genetic correlations that share a common trait (see Methods). The test was applied to five candidate discriminating phenotypes, cognitive task performance (CTP), educational attainment (EDU), metabolic syndrome (MetS), C-reactive protein (CRP) and neutrophil percentage and, using LDSC as implemented in Genomic SEM to jointly estimate the underlying genetic-covariance and sampling-covariance matrices, was evaluated under two complementary estimators of the sampling covariance between the two correlations: the analytic Olkin/Steiger covariance and the LDSC block-jackknife covariance read directly from the sampling-covariance matrix V returned by ‘ldsc()’.

Between the two conditional components, all five phenotypes showed statistically significant differences in their genetic correlations under both estimators. Cognitive task performance showed the largest difference (Δrg = rg_SCZcondBIP − rg_BIPcondSCZ = −0.36; analytic Z = −9.63, p = 5.7 × 10⁻²²; LDSC-jackknife Z = −8.64, p = 5.8 × 10⁻¹⁸), reflecting the opposing signs of the two components (−0.28 vs. +0.077). Educational attainment showed an almost identical pattern (Δrg = −0.35; analytic p = 1.4 × 10⁻²¹; LDSC-jackknife p = 1.3 × 10⁻¹⁷). Metabolic syndrome (Δrg = −0.13; analytic p = 8.4 × 10⁻⁶; LDSC-jackknife p = 5.6 × 10⁻⁵), C-reactive protein (Δrg = −0.12; analytic p = 1.6 × 10⁻⁴; LDSC-jackknife p = 7.7 × 10⁻⁴) and neutrophil percentage (Δrg = −0.089; analytic p = 5.0 × 10⁻³; LDSC-jackknife p = 1.2 × 10⁻²) also differed significantly, and remained so under the more conservative LDSC-jackknife covariance. The sign inversion observed between SCZcondBIP and BIPcondSCZ across cognitive, cardiometabolic and immune domains is therefore not attributable to sampling variance.

When the same test was applied to the unpartitioned SCZ and BIP GWASs — as the natural reference against which the conditional-component difference must be judged — significant differences were again recovered, but of markedly smaller magnitude across every phenotype. For cognitive task performance the difference between SCZ and BIP was less than half that observed between the two conditional components (Δrg = −0.19; LDSC-jackknife p = 7.2 × 10⁻¹⁸), and the same attenuation was observed for educational attainment (Δrg = −0.19; LDSC-jackknife p = 8.9 × 10⁻¹⁸), metabolic syndrome (Δrg = −0.064; LDSC-jackknife p = 1.6 × 10⁻⁴), C-reactive protein (Δrg = −0.062; LDSC-jackknife p = 1.1 × 10⁻³) and neutrophil percentage (Δrg = −0.042; LDSC-jackknife p = 4.0 × 10⁻²). Across all five phenotypes the effect-size of the between-component difference was approximately twice that of the between-disorder difference (Δrg ≈ −0.36 vs. −0.19 for cognition and education; Δrg ≈ −0.13 vs. −0.064 for MetS; Δrg ≈ −0.12 vs. −0.062 for CRP; Δrg ≈ −0.089 vs. −0.042 for neutrophil percentage), and the analytic Z-statistic was correspondingly larger for the conditional pairing in the cognitive and cardiometabolic domains. Because the shared axis of liability is removed from each conditional component but retained in each unpartitioned disorder, this systematic amplification demonstrates that the decomposition surfaces a disorder-specific phenotypic axis that is masked at the raw disorder level. Full test statistics for both anchor pairings across all five phenotypes and both covariance estimators are provided in Supplementary Table 2.

Genomic risk loci and cross-component overlap

The schizophrenia-conditional component (SCZcondBIP) mapped to 66 independent genomic risk loci, the bipolar-conditional component (BIPcondSCZ) to 30 loci, and the PSY-shared factor to 197 loci. To characterise the relationships among these signals, we merged loci across the three components at a minimum overlap of one base pair on the same chromosome, yielding 248 distinct consensus loci (Figure 3a; Supplementary Table 3). The majority of consensus loci were component-specific: 159 were unique to the PSY-shared factor, 31 to SCZcondBIP, and 16 to BIPcondSCZ. Overlap between components was modest and dominated by the PSY-shared factor: 28 consensus loci were shared between SCZcondBIP and the PSY-shared factor, 8 between BIPcondSCZ and the PSY-shared factor, and only 4 between the two conditional components. Two loci were common to all three components. Consequently, the two conditional components overlapped at just 6 consensus loci in total (4 shared exclusively between SCZcondBIP and BIPcondSCZ, plus 2 shared by all three), underscoring their partial genetic divergence (Figure 3b; Supplementary Table 3).

Figure 3 — Pleiotropic locus landscape across components and partitions.

(a) Miami plot of −log₁₀(P) values from SCZcondBIP concordant/discordant, BIPcondSCZ and PSY-shared

(b) Venn diagram showing overlap of loci between SCZcondBIP, BIPcondSCZ and PSY-shared

Identification of novel loci

Intersection of the three component locus sets with the published PGC3 schizophrenia and 2024 bipolar-disorder GWAS loci identified 10 novel loci for the schizophrenia-conditional component, 7 for the bipolar-conditional component, and 64 for the PSY-shared factor, totalling 81 loci absent from both input GWASs and revealed only after the mtCOJO / Genomic SEM decomposition (Supplementary Table 4).

Gene-set enrichment (MAGMA)

To investigate the shared and distinct biological pathophysiologies of schizophrenia and bipolar disorder, we performed competitive gene-set analysis against the Gene Ontology (GO) collection for each of the three components. Enriched gene sets were grouped into three biological categories — synaptic structure and signalling; neurodevelopmental processes (including transcriptional, epigenetic and RNA-processing regulation); and cellular homeostasis and metabolic support — and each component was characterised by the category in which it was predominantly enriched.

The PSY-shared factor produced the strongest and most extensive enrichment, with 92 gene sets significant at FDR < 0.05, principally in synaptic structure and receptor signalling, and voltage-gated ion-channel and electrophysiological activity (80 gene sets). The synaptic category encompassed the synapse (FDR-p = 5.33 × 10), postsynaptic specialization (FDR-p = 1.12 × 10), glutamatergic synapse (FDR-p = 9.35 × 10), GABAergic synapse (FDR-p = 5.14 × 10³), and presynapse (FDR-p = 1.24 × 10³); high voltage-gated calcium channel activity (FDR-p = 1.58 × 10); the L-type voltage-gated calcium channel complex (FDR-p = 2.12 × 10); voltage-gated potassium channel activity (FDR-p = 1.04 × 10²); and calcium-ion transmembrane transport (FDR-p = 1.07 × 10). A smaller neurodevelopmental component (10 gene sets) included generation of neurons (FDR-p = 4.18 × 10), neurogenesis (FDR-p = 1.41 × 10), and subpallium development (FDR-p = 3.65 × 10) (Figure 4; Supplementary Table 5).

Figure 4 — Biological pathway enrichment across SCZcondBIP, BIPcondSCZ and PSY-shared.

(a) MAGMA gene-set enrichment results across the SCZcondBIP, BIPcondSCZ, PSY-shared. Arrows link each partition to broad pathway categories: solid lines denote major enriched pathway groups, dashed lines indicate additional pathway groups with fewer enrichments, and thin dotted lines denote pathway groups that were enriched but did not reach FDR significance.

The schizophrenia-conditional component (SCZcondBIP) was enriched for 31 gene sets at FDR < 0.05, likewise predominantly in synaptic structure and receptor signalling, including the somatodendritic compartment (FDR-p = 1.17 × 10³), dendritic tree (FDR-p = 1.17 × 10³), postsynapse (FDR-p = 4.32 × 10³), and neuron-to-neuron synapse (FDR-p = 2.75 × 10²), together with the L-type voltage-gated calcium channel complex (FDR-p = 4.32 × 10³) and the monoatomic ion channel complex (FDR-p = 3.68 × 10²). It additionally showed enrichment in neurodevelopmental processes (protein–DNA complex, FDR-p = 5.85 × 10³; GABAergic neuron differentiation, FDR-p = 1.95 × 10²; subpallium development, FDR-p = 3.68 × 10²) and in cellular homeostasis and metabolic support (pyrimidine deoxyribonucleoside metabolic process, FDR-p = 1.21 × 10³; proteolysis involved in protein catabolic process, FDR-p = 4.00 × 10²) (Figure 4; Supplementary Table 5).

The bipolar-conditional component (BIPcondSCZ) yielded no gene set at FDR < 0.05, indicating a weaker enrichment signal than the other two components. Its leading nominal associations fell within synaptic structure and receptor signalling but, in contrast to the postsynaptic and ion-channel emphasis of the other two components, were directed toward vesicular compartments, comprising the neuronal dense-core vesicle membrane (P = 7.85 × 10), synapse (P = 1.61 × 10), dense-core granule membrane (P = 1.67 × 10), and postsynapse (P = 8.21 × 10) (Figure 4; Supplementary Table 5).

Developmental Age (BrainSpan)

We next examined the developmental timing of the genes implicated by each component using MAGMA gene-property analysis across BrainSpan developmental periods. The schizophrenia-conditional component (SCZcondBIP) was enriched at FDR correction in postnatal periods: late infancy (FDR-P = 4.10 × 10³) and young adulthood (FDR-P = 7.70 × 10³), with a nominal trend in adolescence (P = 0.051, FDR-P = 0.186). The bipolar-conditional component (BIPcondSCZ) showed FDR-significant enrichment in the same two postnatal periods, late infancy (FDR-P = 2.89 × 10²) and young adulthood (FDR-P = 2.22 × 10²), with a nominal trend in adolescence (P = 0.018, FDR-P = 0.067). The PSY-shared factor showed the broadest developmental profile, with FDR-significant enrichment in the early-mid-prenatal (FDR-P = 1.35 × 10³), late-mid-prenatal (FDR-P = 1.42 × 10³), late infancy (FDR-P = 2.20 × 10) and middle-adulthood periods (FDR-P = 1.87 × 10³), with an additional nominal trend in early prenatal life (P = 0.040, FDR-P = 0.089) (Supplementary Table 6).

Drug-target enrichment (Enrichr / DGIdb)

As a positive control for the biological relevance of the implicated gene sets, we tested each component for enrichment of drug–gene targets in DGIdb, evaluated at nominal significance (P < 0.05) (Supplementary Table 7).

Schizophrenia-conditional component (SCZcondBIP): The genome-wide schizophrenia-conditional component yielded no antipsychotic enrichment among nominally significant hits. Eight compounds reached P < 0.05: the selective estrogen receptor modulator raloxifene (P = 5.22 × 10³), the nicotinic partial agonist entry IUPHAR.LIGAND:2434 (corresponding to varenicline, P = 1.82 × 10²), menadione (P = 1.96 × 10²), the anti-manic purinergic agent allopurinol (P = 2.28 × 10²), eptifibatide (P = 3.11 × 10²), lusutrombopag (P = 3.11 × 10²), fenretinide (P = 4.41 × 10²), and dantrolene (P = 4.49 × 10²). The three nominally enriched compounds with prior clinical-trial histories in psychiatric illness — raloxifene, allopurinol and varenicline — represent candidate cognitive-adjunct / repurposing agents, whereas the remaining hits comprised non-psychiatric compounds (Supplementary Table 7).

Bipolar-conditional component (BIPcondSCZ). The genome-wide bipolar-conditional component similarly yielded no antipsychotic enrichment but recovered six nominal hits: the proteasome inhibitors carfilzomib (P = 2.15 × 10²) and bortezomib (P = 2.83 × 10²), the GABA-B agonist baclofen (P = 2.56 × 10²), the nicotinic entry IUPHAR.LIGAND:2434 (varenicline; P = 3.15 × 10²), the central antihypertensive methyldopa (P = 3.34 × 10²), and the selective estrogen receptor modulator afimoxifene (P = 4.44 × 10²) (Supplementary Table 7).

PSY-shared factor: The genome-wide shared component produced the strongest antipsychotic enrichment, comprising the phenothiazine antipsychotics promazine (P = 4.73 × 10), triflupromazine (P = 2.69 × 10), thiethylperazine (P = 3.61 × 10), prochlorperazine (P = 1.52 × 10³), propiomazine (P = 2.46 × 10³), and fluphenazine (P = 1.29 × 10²); the butyrophenone haloperidol (P = 4.27 × 10³); and the atypical antipsychotics remoxipride (P = 6.30 × 10³), quetiapine (P = 1.19 × 10²), and sulpiride (P = 2.28 × 10²). Genes in this group were also enriched for the mood-stabilising gabapentinoids gabapentin (P = 2.33 × 10³) and pregabalin (P = 8.98 × 10³); the dopamine agonist bromocriptine (P = 4.01 × 10³); the system-xc agent sulfasalazine (P = 2.30 × 10³); the anti-inflammatory celecoxib (P = 2.51 × 10²); the neurosteroid progesterone (P = 4.66 × 10²); and the tricyclic antidepressants imipramine (P = 1.07 × 10²), desipramine (P = 3.27 × 10²), and doxepin (P = 1.92 × 10²) (Supplementary Table 7).

Druggability of the implicated genes

Across the three components, a comparable and substantial fraction of the FUMA-mapped genes were targets of at least one approved drug, ranging from 6.2% in the schizophrenia-conditional component to 16.8% in the bipolar-conditional component and 17.3% in the shared component. Beyond the established antipsychotic and mood-stabilising agents expected for these disorders, the mapped genes also interacted with approved drugs spanning several additional neuropharmacological mechanisms. These included dopaminergic agents (the dopamine agonists apomorphine, rotigotine, and cabergoline and the vesicular monoamine transporter inhibitors reserpine and tetrabenazine); cholinergic agents (the acetylcholinesterase inhibitors donepezil and galantamine and the nicotinic agents mecamylamine and varenicline); α2-adrenergic agonists (clonidine and guanfacine); monoaminergic antidepressants (mirtazapine, duloxetine, venlafaxine and tricyclic agents); glutamatergic and neuroprotective agents (riluzole and ibudilast); L-type calcium channel blockers (nimodipine, nicardipine, and isradipine); and anti-inflammatory and immunomodulatory agents (the cyclooxygenase inhibitors naproxen and ketorolac, and interferon beta-1a and beta-1b) (Supplementary Table 8a–c).

Discussion

Schizophrenia and bipolar disorder share a substantial fraction of their common-variant genetic risk (rg ≈ 0.67), yet differ markedly in symptom profile, course, treatment response, and in their relationship to cognition, educational attainment and cardiometabolic and inflammatory comorbidity. While Genomic SEM^13,14 and GWAS-by-subtraction²⁹ have successfully characterised transdiagnostic and schizophrenia-specific liability to some extent, bipolar-specific liability has remained unmodelled, leaving its architecture and downstream biology uncharacterized. The present analyses resolve this gap by deriving three components (an SCZspecific factor, SCZcondBIP; a BIP-specific factor, BIPcondSCZ; and a shared psychosis factor, PSY-shared) from the largest available GWAS of SCZ and BIP using bidirectional mtCOJO and Genomic SEM, and by functionally annotating each component through genomic risk-locus definition, competitive gene-set enrichment, developmental-timing analysis, and drug-target enrichment. Together these analyses map an explicit disease-specific and shared axis onto genetic-correlation profiles, novel locus discovery, biological pathways, developmental age and thereby illuminate heterogeneity in endophenotypes (cardiometabolic, cognitive and immune) even within the same disorder.

The three components behave as biologically distinct dimensions rather than statistical artefacts. SCZcondBIP and BIPcondSCZ remained modestly and negatively correlated, indicating that conditioning recovered disorder-predominant liability rather than merely redistributing shared signal, and implying that some risk alleles may increase liability to one disorder while reducing it for the other. Both conditional components remained positively correlated with PSY-shared, consistent with a common substrate that still contributes to each disorder-specific axis. Alignment with the SB and higher-order p factors from the recent Genomic SEM study¹⁴ and with the psychotic-spectrum FinnGen endpoints supports the view that PSY-shared indexes broad psychosis-spectrum liability while the two conditional components retain diagnostically specific variance.

The most clinically informative divergence emerged in cognition and educational attainment. At the disorder level, both SCZ and BIP show the familiar paradox of a negative genetic correlation with cognition alongside a positive correlation with education, leaving unresolved whether this reflects one mechanism or several. Decomposition splits this paradox cleanly: SCZcondBIP is negatively correlated with both cognition and education, consistent with schizophrenia-specific liability being tied to early neurocognitive burden and impaired premorbid functioning, whereas BIPcondSCZ is positively correlated with both, suggesting that bipolar-specific risk is less coupled to early cognitive compromise once shared psychosis liability is removed. PSY-shared retains the mixed disorder-level profile and therefore appears to drive the cross-disorder paradox. A correlated-correlation Steiger test further showed that the between-component difference in genetic correlations with cognition and education was approximately twice the SCZ-vs-BIP difference and remained highly significant under the more conservative LDSC block-jackknife covariance estimator, indicating that the decomposition surfaces a disorder-specific cognitive axis diluted between SCZ and BIP.

The same asymmetry and post-decomposition amplification was recovered in the cardiometabolic and inflammatory domains. SCZ shows modest negative genetic correlations with metabolic syndrome, C-reactive protein and neutrophil percentage, whereas BIP is essentially unrelated to these traits; decomposition produces a mirror-image split in which SCZcondBIP is significantly negatively correlated with all three and BIPcondSCZ shifts toward positive correlations. The Steiger test confirmed that the between-component contrast exceeded the SCZ-vs-BIP reference for every cardiometabolic and inflammatory phenotype tested, and remained significant under jackknife covariance. To our knowledge, this is the first evidence that cardiometabolic and inflammatory liability differentiates schizophrenia-specific from bipolar-specific risk at the level of common-variant genetic correlation, and that this differentiation is systematically amplified once the shared psychosis axis is removed. The pattern is consistent with the view that schizophrenia-specific liability is intrinsically coupled to metabolic dysregulation, low-grade systemic inflammation and altered neutrophil biology, as reported in first-episode and antipsychotic-naïve cohorts, whereas bipolar-specific liability is largely orthogonal to these peripheral indices. PSY-shared genomic correlation is almost at the mid-point between the two conditional components across all three phenotypes, retaining a modest negative direction with metabolic syndrome and C-reactive protein and effectively no correlation with neutrophil percentage, consistent with a shared psychosis axis that partially averages the divergent cardiometabolic and inflammatory profiles of SCZcondBIP and BIPcondSCZ rather than aligning with either.

Locus discovery reinforces this decomposition. Most consensus loci were component-specific, with only a handful shared between SCZcondBIP and BIPcondSCZ, mirroring the correlational divergence at the locus level. Intersection with PGC3 SCZ and 2024 BIP identified novel loci dominated by PSY-shared. Because conditional and common-factor GWAS can resolve associations attenuated in marginal analyses, for example when disease-specific and shared components act in opposing directions at the same locus, these novel loci reflect biological decomposition rather than increased test count.

The PSY-shared component was enriched for synaptic signalling pathways (glutamatergic and GABAergic synapses, postsynaptic specialization, presynapse, L-type voltage-gated calcium channel complexes, voltage-gated potassium channel activity, and calcium-ion transmembrane transport), together with a few neurodevelopmental genesets (generation of neurons, neurogenesis, subpallium development; Supplementary Table 4). Expression of PSY-shared genes was significant across prenatal and infancy windows and again in middle adulthood. This pattern aligns with the neurodevelopmental model of psychosis, in which early disruptions create abnormal adolescent synaptic pruning that set the stage for hallucinations, delusions, and cognitive impairment observed later in adolescence^55,56. This is consistent with prior evidence anchoring shared schizophrenia–bipolar liability in excitatory/GABAergic networks, calcium channel biology, and synaptic signalling^{26,28,29,57–62}; our analyses extend that work by resolving the shared factor into an explicit synapticelectrophysiological programme co-anchored in glutamatergic, GABAergic, and Ca²/K channel machinery, with a prenatal-to-late-infancy developmental profile extending into middle-adulthood synaptic pruning. Prior efforts to characterize disorder-specific components of schizophrenia and bipolar disorder have largely been underpowered to detect pathway-level signal. We show, for the first time, that SCZcondBIP is enriched primarily for synaptic-signalling and signalling-regulation pathways, with only a small number of cellular-homeostasis and neurodevelopmental terms. Specifically, the schizophrenia-specific component, independent of bipolar liability, showed near-exclusive enrichment for synaptic pathways, including voltage-gated channels and GABAergic neuron differentiation. These findings are consistent with previous GWAS studies showing that schizophrenia risk variants converge on pathways regulating synaptic elimination, formation, and plasticity. Together, our results support version III of the synaptic hypothesis, which proposes that genetic and/or environmental risk factors render synapses vulnerable to excessive glia-mediated elimination triggered by stress during later neurodevelopment^63,64. SCZcondBIP genes were also significantly expressed in late infancy and young adulthood, supporting a model in which prenatally established synaptic vulnerability manifests as disrupted synaptic function during adolescence and early adulthood, the typical age of schizophrenia onset. The BIPcondSCZ gene set did not survive FDR correction, but the leading nominal enrichments trended toward synaptic-structure pathways, departing from the postsynaptic/ion-channel emphasis of the other components and pointing instead to vesicular compartments (neuronal dense-core vesicle membrane, dense-core granule membrane, synapse, and postsynapse). Prior PET/MRI work has shown reduced fronto-limbic Synaptic vesicle glycoprotein 2A, a marker of synaptic density in individuals with bipolar disorder⁶⁵. Notably, lithium and valproate, two established bipolar treatments, both act on large dense-core vesicle biology, suggesting that altered dense-core vesicle dynamics may underlie symptom improvement and bipolar-specific liability⁶⁶. BIPcondSCZ genes were also significantly expressed during late infancy and young adulthood, mirroring SCZcondBIP and suggesting a similar multi-hit developmental timing operating through distinct pathways.

Drug-target enrichment provided orthogonal support for the biological coherence of the three components. PSY-shared produced the strongest antipsychotic signal, spanning phenothiazine, butyrophenone and atypical antipsychotics alongside mood-stabilising gabapentinoids, dopaminergic and anti-inflammatory agents and tricyclic antidepressants, recapitulating the classical pharmacology of psychotic illness. The two conditional components yielded no antipsychotic enrichment among nominal hits and instead showed enrichment for non-antipsychotic candidates: such as hormonal, purinergic and cholinergic agents for SCZcondBIP, and GABAergic, cholinergic, hormonal and proteasome-inhibitor agents for BIPcondSCZ. This might suggest that PSY-shared engages classical antipsychotic and mood-stabilizing machinery whereas the disorder-specific components highlight potential adjunctive strategies. FUMA-mapped genes for each of the three components was also showed to be the target of at least one approved drug, indicating that disorder-specific signals are no less druggable than the shared factor.

Conclusion

The genetic architecture of schizophrenia and bipolar disorder is not biologically unitary but resolves into schizophrenia-specific, bipolar-specific and psychosis-shared components with distinct correlational profiles, pathway enrichments, developmental signatures and drug-target relationships. Schizophrenia-specific liability aligns with cognitive disadvantage, a cardiometabolic and inflammatory signature and a synaptic-signaling substrate; bipolar-specific liability centers on synaptic-vesicular biology; and shared psychosis component was mostly enriched for synaptic structure and receptor signaling, voltage-gated ion-channel and electrophysiological activity. Between-component contrasts in cognitive, cardiometabolic and inflammatory phenotypes are approximately twice as large as the corresponding SCZ-vs-BIP contrasts, showing that the decomposition unmasks disorder-specific phenotypic axes diluted at the diagnostic level. Pleiotropic psychiatric risk is therefore best understood as a set of partially overlapping, biologically coherent dimensions rather than a single diffuse liability.

Future directions

The three components identified here may be more clinically informative than aggregate disorder-level risk when interpreting heterogeneity in illness course, treatment response, functional impairment and medical comorbidity. Component-specific polygenic scores could be evaluated in deeply phenotyped cohorts to test whether they stratify individuals by cognitive trajectory and cardiometabolic and inflammatory burden more effectively than diagnostic labels alone. Integrating the three components with cell-type- and developmental-stage-resolved transcriptomic resources, and extending the same decomposition to additional intermediate phenotypes, would further test whether this substructure generalises.

Limitations

All analyses rely on European-ancestry GWAS, limiting generalisability across ancestry groups. BIPcondSCZ inherits the smaller effective sample size of the BIP GWAS, which limits power for locus discovery and gene-set enrichment and likely explains the absence of FDR-significant MAGMA gene sets for this component despite biologically coherent nominal enrichments. Drug-target enrichments depend on gene-to-drug mappings from DGIdb, which are biased by prior biological characterisation of individual gene families, and should be interpreted as hypothesis-generating rather than confirmatory.

Funding Statement

U.B., J.J., M.P., M.L., and T.L. were supported by the National Institute of Mental Health of the National Institutes of Health (NIH) under award no. R01MH117646 (T.L., principal investigator). The funding agency had no direct 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. The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH.