Exome sequencing and metabolite profiling identifies genetic variants associated with altered metabolism in early-onset schizophrenia

Summary

Early-onset schizophrenia (EOS) is a severe psychiatric disorder characterized by strong genetic contribution and metabolic alterations, including lipid dysregulation. To investigate the relationship between genetic variation and metabolic changes in EOS, we performed whole-exome sequencing and serum metabolome profiling in 28 patients with EOS and 20 healthy controls. We identified 114 high-risk genes and 117 differentially expressed metabolites. Of the risk genes with variants in multiple patients, 54.35% (25/46) were associated with clinical symptoms, and of the differentially expressed lipids, 34.88% (15/43) were correlated with clinical symptoms. By integrating protein-metabolite interactions and metabolite correlations, we constructed a gene-metabolite network and identified 19 high-risk genes linking to 31 dysregulated lipids. Twenty of these lipids were significantly down-regulated in patients, with 80% (16/20) showing further down-regulation in variant carriers. Our findings provide compelling evidence for a genetic-metabolic interaction in EOS pathogenesis and point to an alternative disease mechanism of schizophrenia.

Graphical abstract

Highlights

• •Integrated multi-omics reveals a genetic-metabolic link in EOS schizophrenia
• •Identified altered lipids as the major metabolites linking to genetic risk variants
• •Identified 19 high-risk genes connecting with 31 dysregulated lipid molecules
• •Detected genetic and metabolic association with PANSS symptom severity

Psychiatry; Human Genetics; Omics; Metabolomics

Introduction

Schizophrenia is a severe neuropsychiatric disorder characterized by substantial genetic and phenotypic heterogeneity.¹ Early-onset schizophrenia (EOS), typically defined by symptom onset before age 18, represents a particularly severe neurodevelopmental subtype associated with poorer clinical outcomes, greater cognitive impairment, and increased familial aggregation compared with adult-onset schizophrenia.^2,3 Previous studies have suggested that EOS exhibits higher heritability and stronger genetic loading than later-onset forms of the disease,^4,5 indicating that genetic and molecular factors may play especially important roles in its pathogenesis. In addition, the early developmental stage at disease onset may involve distinct metabolic and neurobiological alterations that differ from those observed in adult-onset schizophrenia. However, the molecular mechanisms underlying EOS remain poorly understood, limiting the development of targeted therapeutics and objective biomarkers.

A striking feature of schizophrenia is its intrinsic association with metabolic dysfunction. Patients exhibit markedly elevated rates of obesity, insulin resistance, and dyslipidemia, contributing to reduced life expectancy.^6,7,8 Critically, these abnormalities are observed in drug-naïve first-episode patients,^9,10 indicating abnormal metabolism reflects core disease biology, not merely medication side effects. This metabolic-psychiatric comorbidity suggests shared pathogenic mechanisms. Current studies mainly focused on the genetic linkage between schizophrenia and metabolic syndrome,^11,12 while the molecular links between genetic risk and metabolic dysregulation remain largely uncharacterized.

Genome-wide association studies (GWAS) have identified numerous schizophrenia risk loci,^13,14 and whole-exome sequencing has revealed rare, penetrant mutations, attempting to find the root cause of the disease.^15,16 In parallel, metabolomic profiling has consistently detected alterations in lipid and energy-related pathways in patient serum and brain tissue. However, these genetic variations and metabolomic findings have remained largely disconnected. No study has systematically integrated genetic variants with metabolomic profiles to establish functional links between genotypes and metabolic alterations in schizophrenia.

We hypothesized that patients with EOS might harbor damaging mutations in genes critical for metabolism, potentially contributing to the status of the patients’ phenotypes. We performed whole-exome sequencing and untargeted serum metabolomic profiling on 28 patients with EOS and 20 healthy controls, and identified 114 high-risk genes and 117 differentially expressed metabolites (DEMs). To link the risk genes to the dysregulated metabolites, we integrated protein-metabolite interactions and metabolite-metabolite correlations to construct a gene-metabolite network, focusing on a lipid-centered subnetwork containing 27 lipids linked to 19 high-risk genes and 4 lipids correlated with gene-associated lipids. We compared the expression levels of the DEMs between variant-carriers and non-variant carriers of the genes in this network to link the variants to metabolic phenotypes, providing evidence for a direct genetic-metabolic axis in EOS pathogenesis. This work points to a potential alternative disease mechanism.

Results

Integrated genomic analyses identify high-risk genes in early-onset schizophrenia (EOS)

To identify genetic variants contributing to EOS, we performed whole-exome sequencing on a cohort of 28 patients with EOS and 20 healthy controls (Table S1). Variant calling and quality filtering identified 95,445 rare variants (MAF <0.05 in East Asian populations). Using variant effect predictor (VEP), we prioritized 1,292 high-impact variants predicted to substantially alter protein functions, including 899 protein-truncating variants (PTVs) and 393 damaging single amino acid variants (SAVs) across 1,178 genes (Figure 1A). In parallel, we identified 658 copy number variations (CNVs) using XHMM. After filtering against control samples and retaining only CNV regions supported by schizophrenia-associated CNVs in the literature, we identified 175 CNV regions encompassing 537 putative risk genes (Figure 1A). The combined dataset comprised 1,715 putative risk genes (1,178 from SNVs/indels and 537 from CNVs).

Figure 1.

Gene risk scoring prioritizes high-risk genes in early-onset schizophrenia (EOS)

(A) Workflow for identifying high-risk genes in EOS. Variants and copy number variants (CNVs) were detected from whole-exome sequencing data using GATK and XHMM, respectively. Potentially deleterious mutations were filtered based on minor allele frequency (MAF), odds ratio (OR), and predictions of harmful effects. CNVs were filtered according to population information and schizophrenia-related CNVs. Protein-truncating variants (PTVs), single amino acid variants (SAVs), and CNVs were then integrated to evaluate the pathogenic risk of genes in EOS.

(B) Distribution of gene risk score (GRS) among mutated genes in EOS. A cutoff of GRS >4 was applied to define 114 high-risk genes, accounting for 6.84% of all mutated genes. Bars are colored according to support from variants, CNVs, or both.

(C) Parameters contribution to GRS calculation. Three parameters: mutation number, harmful effect prediction, and literature evidence, were considered in GRS estimation. These factors collectively influenced the prioritization of high-risk genes in EOS.

We developed a method to evaluate gene risk score (GRS) by integrating three parameters (see STAR Methods): (1) number and predicted deleteriousness of mutations, (2) number of patients affected, and (3) supporting evidence from published genetic studies (GWAS, WES/WGS, differential expression, differential methylation, and linkage studies). The GRS distribution followed a scale-free pattern, with most genes showing low scores and a small subset exhibiting high scores (Figure 1B). Using GRS >4 as a threshold, we identified 114 high-risk genes representing the top 6.84% of all mutated genes (Figure 1C and Table S2).

High-risk genes associated with symptom severity participate in metabolic pathways

To validate the relevance of the high-risk genes to the disease phenotypes, we compared Positive and Negative Syndrome Scale (PANSS) scores between variant-carrier patients and non-variant-carrier patients. We calculated the association between the high-risk genes and clinical symptom severities. Among these 114 high-risk genes, 46 had mutations in at least two patients with PANSS records. Of these, 16 genes showed significant differences in PANSS negative symptom scores, 4 in positive symptom scores, and 10 in total symptom scores (Figure 2). Variants in CD36, UBAP2, CCDC110, SHOX2, FAXDC2, and MARCKS were associated with higher symptom scores. These findings demonstrate that mutations in specific high-risk genes were directly associated with clinical symptom severity, suggesting they may play functional roles in EOS pathogenesis.

Figure 2.

High-risk gene variants are associated with Positive and Negative Syndrome Scale (PANSS) symptom severity

Of 114 high-risk genes, 46 genes harboring mutations in at least two patients were subjected to the analysis. The genes with a significant symptom score difference between mutation carriers and non-carriers are highlighted in bold. Purple, orange, and gray represent PANSS negative symptom scores, positive symptom scores, and total scores, respectively. p values were calculated using a t test, and the dashed line represents p = 0.05.

To find out if some of these high-risk genes have functions related to metabolic processes, we compared the 114 high-risk genes to metabolic pathway members from the Reactome database.¹⁷ Fifteen high-risk genes overlapped with eight core metabolic pathways: carbohydrate metabolism, inositol phosphate metabolism, lipid metabolism, integration of energy metabolism, TCA cycle and respiratory electron transport, nucleotide metabolism, amino acid metabolism, and biological oxidations (Table S3). Six genes (PTGS1, SREBF1, PITPNM1, CD36, GLB1L2, and INPP5D) are regulatory components of metabolic pathways, with some involved in multiple pathways.

Lipid metabolism is profoundly dysregulated in early-onset schizophrenia

To identify the high-risk genes associated with metabolism dysregulation in this disease, we performed untargeted LC-MS/MS metabolomic profiling on serum from all participants. Principal component analysis demonstrated clear separation between patient and control groups (Figure 3A). To account for potential confounding effects of age and sex, differential metabolite expression between patients with EOS and controls was assessed using a multivariate linear regression model. Applying a stringent threshold (FDR <0.05), we identified 5,071 DEMs, including 2,276 upregulated and 2,795 downregulated metabolites (Figure 3B).

Figure 3.

Metabolomic profiling reveals profound disruption of lipid homeostasis in early-onset schizophrenia

(A) The principal component analysis result is based on metabolite intensities of all identified metabolites in the serum of patients with early-onset schizophrenia and healthy subjects. Sample ages were labeled for the corresponding samples in the image.

(B) Numbers of differentially expressed metabolites (DEMs) under positive and negative ion models.

(C) The DEMs with accurate annotations were identified under positive and negative ion models.

(D) Superclass classification of DEMs based on the HMDB annotation.

(E) Over-representation analysis of DEMs in RaMP pathways. The node sizes represent the ratio of DEMs to all metabolites in the corresponding pathways. The dashed line in purple indicates the cutoff of FDR = 0.05.

(F and G) Recaptured DEMs reported in the literature. p value was calculated using the ropls package.

Since disease duration and antipsychotic drug dosage may additionally influence metabolite abundance specifically within the patient group, we further evaluated metabolite intensities in patients with EOS using a covariate model incorporating sex, age, disease duration, and medication dosage. No significant associations were observed after Benjamini-Hochberg correction, suggesting that neither disease duration nor drug dosage represented major confounding factors in our dataset (Table S4). Following metabolite annotation, 117 DEMs with MS2-level annotations were retained for downstream analyses (Figure 3C).

Chemical classification using HMDB revealed that “lipids and lipid-like molecules” constituted the largest superclass of dysregulated metabolites, representing 36.75% (43/117) of high-confidence DEMs (Figure 3D). When setting the FDR filtering criteria to 0.1, 23 terms were detected using MetaboAnalyst.¹⁸ These enriched pathways revealed dysregulation mainly focused on purine metabolism and lipid homeostasis (Figure 3E and Table S5). Key dysregulated metabolites contributing to these enrichments included hypoxanthine, xanthine, xanthosine, linoleic acid, docosahexaenoic acid (DHA), and alpha-dimorphecolic acid.

Tissue enrichment analysis identified 32 significant tissues (FDR<0.05). Although the dysregulated metabolites were detected from peripheral blood, substantial overrepresentation was also observed in cerebrospinal fluid (CSF), myelin sheath, brain, and nervous tissue (Table S5). Several DEMs, including linoleic acid, were enriched in both CSF and myelin, whereas L-kynurenine, 5-hydroxyindoleacetic acid (5-HIAA), inosine, and propionic acid were enriched in both CSF and brain tissues. Stearic acid, creatine, and bilirubin were simultaneously enriched in CSF, myelin, and brain-associated tissues, while DHA was shared across all four brain-related tissue categories (Table S5).

Since we detected ASD enriched terms based on DEMs from patients with SCZ, this may be caused by incomplete knowledge about psychiatric disease-related metabolites. To contextualize our findings, we compiled 134 previously reported DEMs of schizophrenia from 44 publications. Nine were recapitulated in our EOS cohort (Figures 3F and 3G): stearic acid, L-kynurenine, 5-HIAA, creatine, DHA, eicosadienoic acid, linoleic acid, and two lysophosphatidylcholines (lysoPCs 18:0 and 18:2). Strikingly, 66.67% (6/9) were lipids or lipid-like molecules, validating the disease-relevance of the profound dysregulation of lipid metabolism. DHA, a very important omega-3 fatty acid for human brain function, was detected to be decreased both in our data (FC = 0.44, p = 6.15E−03) and public evidence.^19,20 Linoleic acid was significantly decreased (FC = 0.69, p = 1.85E−04), consistent with reports of its reduction in schizophrenia brain tissue.^21,22,23 As a member of membrane lipids, linoleic acid level was reported as a potential biomarker of conversion to psychosis in ultra-high risk of patients with psychosis.²⁴ Eicosadienoic acid, a potential DHA precursor, was also decreased, aligning with its reported reduction in patient brains.²⁵ Conversely, stearic acid was elevated, mirroring findings in brain tissue and erythrocyte membranes.^22,23 The dysregulated lysoPCs (18:0 and 18:2) corroborated previous reports of their altered levels in schizophrenia blood.^26,27 The increased plasma level of lysoPC 18:2 following olanzapine treatment has been reported to be associated with improvement of positive symptoms in schizophrenia,²⁸ which further supports our observation that the decreased level of lysoPC 18:2 in EOS is unlikely to be driven by antipsychotic drug effects. Among non-lipid biomarkers, L-kynurenine was elevated, matching its reported increase in the dorsolateral prefrontal cortex,²⁹ while 5-HIAA showed altered levels consistent with its complex association with schizophrenia pathophysiology.³⁰ Creatine levels were elevated in the serum of patients with EOS, which contrasts with the reduced cortical creatine signals previously reported by magnetic resonance spectroscopy studies in patients with schizophrenia.³¹ This discrepancy may reflect tissue-specific metabolic differences, particularly given that most previous studies have focused on creatine kinase activity rather than circulating creatine levels.

Differential metabolites, particularly lipids, correlate with schizophrenia symptom severity

To find out whether these metabolic changes were related to the disease, we calculated correlations between the 117 high-confidence DEMs and PANSS scores. A substantial proportion of metabolites associated with symptom severity (|r²| ≥0.3, p < 0.1): 22/117 (18.80%) correlated with total PANSS, 8/117 (6.84%) with positive symptoms, and 18/117 (15.38%) with negative symptoms (Figures 4 and S1 and Table S6). Consistent with our overall findings, lipids and lipid-like molecules predominated among metabolites correlated with total and positive symptom scores (Figure 4A).

Figure 4.

Metabolic dysregulation correlates with schizophrenia symptom severity

(A) The superclasses of DEMs correlated with PANSS symptom scores. Superclass information was retrieved from HMDB annotation.

(B) DEMs correlated with PANSS symptom scores.

(C) Correlation between the expression level of DEMs and PANSS negative symptom scores. The correlation line was fitted according to the dots in the figure using the linear regression method. Correlation coefficient r and p value were calculated using the cor.test function in R. The Benjamini-Hochberg correction was applied to adjust the p value.

Of the 18 DEMs correlated with PANSS negative symptom scores, 11 showed negative correlation with PANSS negative symptom scores: Phe-Thr (r = −0.60, p = 8.08E−03), gamma−hydroxybutyric acid (r = −0.55, p = 1.93E−02), trans−traumatic acid (r = −0.51, p = 3.10E−02), Gly-Val (r = −0.49, p = 2.92E−02), 3−hydroxy−3−(3−hydroxyphenyl) propanoic acid−o−sulfate (r = −0.48, p = 4.56E−02), trimethyloxazole (r = −0.47, p = 5.16E−02), Val-Leu (r = −0.44, p = 6.94E−02), hypoxanthine (r = −0.44, p = 7.08E−02), Asn−Phe (r = −0.43, p = 7.13E−02), DHA (r = −0.43, p = 7.08E−02) and blennin A (r = −0.41, p = 9.49E−02) (Figure 4C). Trimethyloxazole, dipeptide Gly-Val, Val-Leu, and DHA were also negatively correlated with PANSS total scores (Figure S1B). 7 DEMs, 27−norcholestanehexol (r = 0.44, p = 6.62E−02), styrene (r = 0.46, p = 5.71E−02), 2,6-dihydroxybenzoic acid (r = 0.49, p = 3.86E−02), citpressine II (r = 0.50, p = 3.46E−02), 2-methylpentanal (r = 0.59, p = 1.01E−02), 4,6−dinitro−o−cresol (r = 0.67, p = 2.57E−03) and m-guaiacol (r = 0.72, p = 7.73E−04), were positively correlated with PANSS negative symptom scores.

Of the 8 DEMs correlated with PANSS positive scores, 6 exhibited significant negative correlation with clinical symptoms, and two showed positive correlation (Figure S1A), including alpha-dimorphecolic acid (r = 0.51, p = 2.95E−02), which also exhibited significant positive correlation with PANSS total scores (r = 0.47, p = 5.10E−02; Figure S1B). Notably, different from a study that reported decreased levels of alpha-dimorphecolic acid in schizophrenia plasma,³² we observed a marked increase (FC = 5.34, FDR = 3.91E−15), suggesting potential differences between early-onset and adult-onset disease. Among 22 DEMs correlated with PANSS total scores, 77.27% (17/22) exhibited negative correlation (Figure S1B).

These findings demonstrate that metabolic dysregulation, particularly in lipid metabolism, associates with clinical symptom severity, consistent with our observation that genetic variants of risk genes related to lipid metabolism were also associated with symptom score.

High-risk genes with genetic variants associated with lipid dysregulation

To further search for genetic variants directly linking to specific metabolites differentially expressed in patients, we integrated our genetic and metabolic datasets by testing the association between variants and metabolites. We compared the expression level of metabolites between variant-carrying patients and non-variant-carrying patients. We also calculated metabolite-metabolite correlations to capture coordinated metabolic modules. After filtering out the non-annotated metabolites, we obtained a protein-metabolite network containing 118 protein-metabolite associations and 42 metabolite-metabolite correlations between 31 high-risk genes and 94 metabolites (81 DEMs and 13 DEM-correlated non-DEMs).

To focus on abnormal lipid metabolism in EOS, which was a predominant metabolic signature of the patients, we retrieved lipids and lipid-like molecules from this network along with their associated high-risk genes, and obtained an EOS protein-lipid subnetwork, with 27 lipids associated with 19 high-risk genes and 4 DEMs correlated with gene-associated lipids (Figure 5A). Several genes emerged as key hubs, among which DDX3X, FAXDC2, SIGMAR1, COQ10A and TTC39A exhibited association with at least 3 DEMs. A small deletion (chrX:41344421–41344425) in DDX3X (dead-box helicase 3 X-linked), present in two patients with EOS, associated with four lipids, including lysoPC 18:2. DDX3X was not previously reported to be linked to schizophrenia, but mutations in this gene was reported to cause neurodevelopmental disorders including intellectual disability and autism,³³ suggesting a potential role of this gene in brain neurodevelopment associated with EOS pathogenesis. A missense variant (chr5:154,834,677; C→A) in FAXDC2 (Fatty acid hydroxylase domain containing 2), found in two patients, was associated with multiple lipids. FAXDC2 regulates lipid metabolism in cancer and has been linked to postoperative neurocognitive decline and Parkinson’s disease,^34,35,36 positioning it as a novel candidate gene in EOS with a functional connection to cognitive function. A missense variant (chr9:34635682; G→A) in SIGMAR1 (Sigma non-opioid intracellular receptor 1) was identified in two patients. A polymorphism in SIGMAR1 has been associated with prefrontal cortex activation in schizophrenia,³⁷ and our data now link it directly to lipid dysregulation. A PTV (chr12:56,267,121, G>T) of COQ10A was identified in two patients. Although no direct association between COQ10A and schizophrenia has yet been established, the COQ gene family, which encodes enzymes involved in CoQ10 biosynthesis, has previously been implicated in schizophrenia susceptibility.³⁸ In addition, a frameshift variant (chr1:51,302,241, C > −) of TTC39A was identified in two patients. Therefore, this protein-lipid subnetwork represents an EOS-relevant lipid metabolism regulatory network. Dysregulation of this network may perturb lipid metabolism in the brain cells.

Figure 5.

Lipid dysregulation network of EOS revealed the association between mutations in high-risk genes and differential abundance of metabolites

(A) The protein-lipid subnetwork was retrieved from the lipids and their neighboring nodes from the protein-metabolite network. Cyan circles represent DEMs; tawny diamonds represent high-risk genes with variants in at least two SCZ samples; gray lines indicate the protein-metabolite association, and light blue lines indicate the metabolite-metabolite correlation.

(B–T) Further expression level changes of lipids in variant carriers. Mc SCZ, Non-mc SCZ, and HC represent patients with mutation carrier schizophrenia, non-mutation carrier schizophrenia patients, and healthy controls, respectively. A patient harboring mutations in one or more genes that have a connection with a particular metabolite was defined as a mutation carrier. Statistics were applied using a t test. ∗, p < 0.05; ∗∗, p < 0.01; ∗∗∗, p < 0.001; ∗∗∗∗, p < 0.0001; ∗∗∗∗∗, p < 0.00001; n.s., no significance.

We further validated the links between genetic variants and metabolite expression changes, focusing on lipid metabolism by comparing the lipid expression levels between variant-carrier patients, non-carrier patients, and healthy controls. Strikingly, 20 of 27 (74.07%) variant-associated DEMs in this subnetwork showed significant down-regulation in patients, and 16 showed further down-regulation in variant-carrier patients (Figures 5B–5P and Table S6): lysoPC 22:6 (p = 2.54E−08, variants of RNF126 and FZD6), 3alpha−Corosolic acid (p = 4.40E−06, variants of COQ10A and FAXDC2), punicic acid (p = 5.10E−05, variants of NEURL4 and UBAP2), 27−norcholestanehexol (p = 9.35E−06, variants of PRX and C5orf49), lysoPC 20:5 (p = 6.78E−04, variants of FZD6 and FAXDC2), 4,8,12,15,19−docosapentaenoic acid (p = 2.39E−04, variants of MICAL2, SIGMAR1 and CCDC110), 27−nor−5b−cholestane−3a,7a,12a,24,25−pentol (p = 1.43E−04, variants of DDX3X and COQ10A), dodecanedioic acid (p = 1.11E−04, variants of NEURL4, COQ10A and FAXDC2), 8−epixanthatin (p = 1.26E−04, variants of CMYA5), ar−Artemisene (p = 1.50E−03, variants of SIGMAR1), DHA (p = 5.60E−04, variants of SIGMAR1), isoursodeoxycholic acid (p = 1.96E−05, variants of CMYA5), lysoPC 18:2 (p = 5.86E−06, variants of DDX3X), octadecanedioic acid (p = 1.32E−04, variants of FAXDC2), traumatic acid (p = 2.77E−04, variants of FAXDC2), and (s)-oleuropeic acid (p = 7.29E−07, variants of FAXDC2) (Figures 5 and S3 and Table S7).

While most of the lipids showed further down-regulation in variant-carrier patients, four lipids, gamma−hydroxybutyric acid (variants of PRX and C5orf49), 3−carboxy−4−methyl− 5−pentyl−2−furanpropanoic acid (variants of IKZF2, DDX3X, and FAXDC2), 11'−carboxy−alpha−chromanol (variants of CBLB, ARID1A, and FAXDC2), and pregnanetriolone (variants of CREM, SIGMAR1, and TTC39A), were not further down-regulated in variant-carrier patients (Figures S3A–S3D). Gamma-hydroxybutyric acid is an endogenous neuromodulator of GABA and dopamine systems³⁹; while exogenous gamma-hydroxybutyric acid has been explored as a treatment for schizophrenia,^40,41 our data suggest its endogenous levels may be influenced by specific genetic variants in EOS.

Of the seven DEMs that were significantly up-regulated in patients, sebacic acid (p = 2.55E−03, variants of DDX3X), lysoPC 20:2 (p = 3.55E−05, variants of OSMR) and belnnin A (p = 1.01E−11, variants of OSMR) were further up-regulated in variant-carrier patients (Figures 5R–5T). However, ent-16b-19-kauranediol 19-acetate (variants of MICAL2 and CCDC110), (3beta,22E,24R)−3−Hydroxyergosta−5,8,22−trien−7−one (variants of TTC39A) and glycerol 1-hexadecanoate (variants of CBLB) showed an inconsistent change trend in variant carriers, and norethindrone was not further up-regulated in variant carriers (variants of CBLB and TTC39A) (Figures S3E–S3H).

The variant carriers of 13 risk genes showed further downregulation of 16 lipids, and variant carriers of 2 risk genes showed further upregulation of 3 lipids, suggesting that these variants may perturb lipid metabolism. These findings provide direct evidence that genetic mutations in high-risk genes functionally impact lipid metabolism in patients with EOS, providing a potential mechanistic link between genotype and metabolic phenotype.

Discussion

Using whole-exome sequencing and untargeted metabolomics, we found that damaging mutations in high-risk genes are linked to metabolic abnormalities in patients with EOS. First, we established that both genetic variants and the associated DEMs show more correlation with the severity of negative symptoms than with positive symptoms, suggesting that metabolic dysregulation may be more closely linked to the neurodevelopmental and cognitive dimensions of EOS pathology. Next, by integrating protein-metabolite interactions and metabolite-metabolite correlations, we connected specific high-risk genes to disruptions in lipid metabolism. We further validated these links by comparing lipid expression levels between mutation carriers and non-carriers. Previous studies have either focused on the identification of disease-related variants or dysregulated metabolites, while we attempted to build the connection between them. Although there is no literature evidence that these high-risk genes are involved in abnormal metabolism in schizophrenia, they regulate metabolic processes that may be important for brain function.

The metabolic signature for patients with EOS is dominated by lipids and lipid-like molecules (36.75% of differentially abundant metabolites). Pathway enrichment analyses further highlighted adipogenesis, transcriptional regulation of adipocyte differentiation, and linoleic acid metabolism as major altered processes. Interestingly, a small fraction of high-risk genes identified through genetic analysis are located in metabolic pathways, mainly in lipid metabolic pathways. This convergence on lipid metabolism from both metabolomic and genomic analyses suggests that lipid dysregulation is not merely a comorbidity but is intrinsically linked to the genetic architecture of EOS. Lipids are critical for synaptic membrane composition, myelination, neurotransmitter signaling, and neuroinflammatory regulation, all of which have been implicated in schizophrenia pathophysiology. Thus, disruption of lipid metabolism may provide an important mechanistic link between genetic susceptibility and abnormal neurodevelopment in EOS.

Several metabolites identified in this study displayed expression patterns inconsistent with those reported in previous schizophrenia studies. For example, alpha-dimorphecolic acid exhibited opposite directional changes compared with findings from adult schizophrenia cohorts.³² As one of the linoleic acid-derived lipid mediators and endogenous PPARγ agonist implicated in inflammatory and lipid regulatory pathways, the altered directionality of alpha-dimorphecolic acid in EOS may reflect developmental stage-specific lipid metabolic remodeling. In addition, differences in age distribution, disease duration, medication exposure, dietary factors, and clinical heterogeneity across studies may also contribute to inconsistent metabolic signatures. Although our multivariable analyses suggested that disease duration and antipsychotic dosage did not significantly influence the observed metabolomic alterations, longitudinal studies may be needed to investigate whether these metabolites, such as alpha-dimorphecolic acid, represent EOS-specific signatures.

Our network analysis moves beyond correlation to suggest functional mechanisms linking genetic variants to metabolic phenotypes. We identified 27 lipids directly associated with 19 high-risk genes, and crucially, 85.19% (23/27) of tested lipids showed significant and directionally consistent abundance changes in SCZ versus healthy control and mutation carrier SCZ versus non-mutation carrier SCZ comparison. DDX3X, encoding an RNA helicase implicated in neurodevelopmental disorders,³³ is associated with four lipids, including lysophosphatidylcholines—suggesting that neurodevelopmental risk genes may exert effects partly through disrupted lipid metabolism. FAXDC2, a fatty acid hydroxylase domain-containing protein linked to lipid metabolism in cancer,³⁴ represents a novel candidate in psychiatric disease with a clear mechanistic connection to lipid processing. Collectively, these findings support a model in which genetically driven disturbances in lipid metabolism contribute to the molecular pathology of EOS.

A key strength of our present study is the integration of rare-variant genomics with metabolomic profiling in the same individuals. Recent studies have increasingly applied multi-omics approaches to investigate schizophrenia, including cross-omics analyses of genomic, transcriptomic, metabolomic, epigenetic, and microbiome-related data.^42,43,44,45 However, these studies did not directly assess the relationship between rare deleterious variants and metabolic abnormalities within the same individuals. By integrating whole-exome sequencing, serum metabolomics, protein-metabolite interactions, and carrier-versus-non-carrier comparisons, our individual-level analysis provides preliminary evidence supporting a functional genetic-metabolic axis linking rare mutations to lipid dysregulation in EOS.

In summary, our findings address a long-standing question of whether there is a genetic factor behind the metabolic abnormalities observed in patients with schizophrenia. The convergence of genetic mutations and metabolic phenotypes provides a candidate mechanism for how genetic risk may lead to measurable metabolic disruption in EOS.

Limitations of the study

Several limitations should be considered in interpreting the findings of this study.

• •Sample Size and Genetic Association Analysis: The relatively modest sample size may limit the statistical power for identifying rare genetic variants and detecting subtle metabolomic alterations. Future studies involving larger and independent cohorts are needed to further validate the identified genetic-metabolic associations and improve the detection of additional disease-related molecular features.
• •Clinical Phenotype Characterization: PANSS scores were available for only a subset of patients, which may reduce the robustness of correlations between risk genes, metabolites, and clinical symptoms. Expanded clinical assessments and longitudinal follow-up studies will be valuable for clarifying the relationship between molecular alterations and symptom progression.
• •Tissue Specificity of Metabolic Alterations: Serum metabolomic profiles reflect systemic metabolic states and may not fully represent brain-specific metabolic changes associated with EOS. Future studies integrating central nervous system-related molecular data, multi-omics approaches, and functional experimental validation will help further elucidate the biological mechanisms linking genetic variation to metabolic dysregulation.

Resource availability

Lead contact

Any additional information required to reanalyze the data reported in this paper is available from the lead contact, Xinping Yang (xpyang1@smu.edu.cn), upon request.

Materials availability

This study did not generate new unique reagents or materials.

Data and code availability

• •Whole-exome sequencing data have been deposited in the Genome Variation Map at the National Genomics Data Center (NGDC: PRJCA048048, GVM: GVM001175).
• •Metabolomics data have been deposited at Metabolomics Workbench (Metabolomics Workbench: ST005005).
• •The code used for data analysis has been deposited at Zenodo (Zenodo: https://doi.org/10.5281/zenodo.21367499) and is publicly available.
• •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Acknowledgments

This work was supported by the Regional Joint Fund of Guangdong Province (grant no. 2019B1515120080), the National Natural Science Foundation of China (grant no. 82071503), the Key Scientific, Technological Projects of Guangdong Province (grant no. 2018B030335001), and the President Foundation of Nanfang Hospital, Southern Medical University (grant nos. 2024B052, 2024B045 and 2025B028). The funding organizations had no role in designing and conducting the study, collecting, managing, analyzing, and interpreting the data, or preparing, reviewing, or approving the manuscript.

Author contributions

X.Y.: conceptualization, data curation, formal analysis, funding acquisition, investigation, visualization, writing – original draft, and writing – review and editing. W.Y.: data curation, formal analysis, funding acquisition, methodology, and writing – review and editing. H.P.: data curation, formal analysis, visualization, and writing – review and editing. W.Y.: investigation, funding acquisition, resources, and writing – review and editing. J.X.: investigation, resources, and writing – review and editing. H.L.: investigation, resources, and writing – review and editing. C.F.: investigation, resources, and writing – review and editing. G.L.: investigation, resources, and writing – review and editing. Y.G.: investigation, resources, and writing – review and editing. Y.C.: investigation, resources, and writing – review and editing. X.Z.: resources, supervision, and writing – review and editing. B.Z.: conceptualization, resources, supervision, and writing – review and editing. X.Y.: conceptualization, funding acquisition, supervision, writing – original draft, and writing – review and editing.

Declaration of interests

The authors declare no competing interests.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Biological samples
Human serum from EOS patients and healthy individuals	Southern Medical University Nanfang Hospital	N/A
Deposited data
Whole-exome sequencing data	This paper; National Genomics Data Center (NGDC) -GVM	NGDC:PRJCA048048, GVM: GVM001175
LC-MS/MS metabolic data	This paper; Metabolomics Workbench	ST005005
Software and algorithms
R	https://cran.r-project.org/src/base/R-4/	v4.1.3
BWA-MEM	https://github.com/lh3/bwa/releases/tag/v0.7.17	v0.7.17
Picard	https://github.com/broadinstitute/picard/releases/tag/2.20.3	v2.20.3
BedTools	https://github.com/arq5x/bedtools2/releases/tag/v2.28.0	v2.28.0
Genome Analysis Toolkit (GATK)	https://github.com/broadinstitute/gatk/releases	v4.1.2
Variant Effect Predictor	https://www.ensembl.org/vep	v104
XHMM	https://github.com/RRafiee/XHMM	v1.0.0
bcftools	https://github.com/samtools/bcftools/releases?page=3#release-1.9	V1.9
Hail	https://pypi.org/project/hail/0.2.123/	V0.2.123
ProteoWizard	https://proteowizard.sourceforge.io/download.html	v3.0.7414
XCMS	https://github.com/sneumann/xcms/tree/RELEASE_3_9	V3.9.3
masscleaner	https://masscleaner.tidymass.org/	v1.0.4
CAMERA	https://bioconductor.org/packages/3.12/bioc/html/CAMERA.html	v1.46.0
metaX	https://github.com/wenbostar/metaX	v2.0.0
MetaboAnalyst	https://www.metaboanalyst.ca/	v6.0
Cytoscape	https://cytoscape.org/	v3.10.4
ggplot2	https://cran.r-project.org/src/contrib/Archive/ggplot2/	v3.4.2
Hmisc	https://cran.r-project.org/src/contrib/Archive/Hmisc/	v5.1.3
Exome analysis pipeline	https://github.com/yang91/Combination-of-Exome-Sequencing-and-Metabolite-Profiling-in-early-onset-schizophrenia https://doi.org/10.5281/zenodo.21367499	Zenodo: https://doi.org/10.5281/zenodo.21367499
LC-MS/MS analysis pipeline	https://github.com/yang91/Combination-of-Exome-Sequencing-and-Metabolite-Profiling-in-early-onset-schizophrenia https://doi.org/10.5281/zenodo.21367499	Zenodo: https://doi.org/10.5281/zenodo.21367499
Other in-house code	https://github.com/yang91/Combination-of-Exome-Sequencing-and-Metabolite-Profiling-in-early-onset-schizophrenia https://doi.org/10.5281/zenodo.21367499	Zenodo: https://doi.org/10.5281/zenodo.21367499

Experimental model and study participant details

The study protocol was approved by the Ethics Committee of Southern Medical University Nanfang Hospital (approval number: NFEC-202306-K4), and all subjects provided consent for research participation. Blood samples were collected at the Southern Medical University Nanfang Hospital and processed under anonymized numerical codes to protect participant identity.

A total of 28 patients with early-onset schizophrenia (EOS) and 20 healthy controls were included in this study. Clinical diagnosis was established by qualified psychiatrists according to DSM-5 criteria. EOS was defined as the onset of persistent psychotic symptoms at ≤20 years of age, with enrollment occurring before age 35. Healthy controls were enrolled between the ages of 15 and 35 years. Clinical symptoms were assessed using the Positive and Negative Syndrome Scale (PANSS) in 18 of 28 patients. Demographic and clinical characteristics are provided in Table S1.

Method details

Whole-exome sequencing and alignment

Whole-exome sequencing was performed using the Igtech T086 V4 capture kit and NovaSeq platform (iGeneTech) in paired-end mode. Raw reads were quality-filtered and aligned to the human reference genome (GRCh38) using BWA-MEM⁴⁶ (v0.7.17). Duplicate reads were marked and removed using Picard MarkDuplicates (http://broadinstitute.github.io/picard; v2.20.3). Mean depth of coverage and target region capture rates were calculated using BEDTools⁴⁷ (v2.28.0) .

Variant identification

Single-nucleotide variants (SNVs) and small insertions/deletions (indels) were detected using GATK⁴⁸ (v4.1.2) following the best practice protocol. Variant Quality Score Recalibration (VQSR) was performed using training sets from HapMap, 1000 Genomes Project, dbSNP, Mills indels, and Axiom Exome data, with truth sensitivity threshold set at 99.9.

Variant filtering

Variants passing VQSR were further filtered using Python hail packakge (https://github.com/hail-is/hail/releases/tag/0.2.13; v0.2.13) based on the following criteria: (i) not located in tandem repeat regions; (ii) for the variant site with multiple genotypes, read depth ∼ 10-1,000×, ≤90% reads supporting wild-type or homozogous allele, or ∼ 25%-75% reads supporting heterozogous allele; (iii) genotype quality ≥25; (iv) minor allele frequency (MAF) <0.05 in East Asian populations (gnomAD v2.1.1). Odds ratios were calculated by comparing allele frequencies between EOS patients and controls, and variants with an odds ratio >1 were selected as EOS-related variants.

Functional annotation

Variants were annotated using Variant Effect Predictor⁴⁹ (v104) with SIFT, PolyPhen, dbNSFP, CADD, and MutationTasser databases (Ensembl v104, SIFT v5.2.2, PolyPhen v2.2.2, dbSNP v154 and CADD v1.6). We prioritized protein-truncating variants (PTVs; high impact) and single amino acid variants (SAVs; moderate impact) predicted as deleterious by at least three prediction tools.

Copy number variant (CNV) detection

CNVs were detected using XHMM⁵⁰ (v1.0.0) following mean-center coverage calculation, principal component analysis, and depth normalization. CNVs detected in control samples were excluded. Remaining CNVs were compared to known schizophrenia-associated CNV regions from published studies.⁵¹ Only CNVs overlapping known schizophrenia loci with an odds ratio >1 in our cohort were retained for further analysis.

Mutation score (MS) calculation

For each gene, we calculated a mutation score based on: (i) number of PTVs (m) and SAVs (n); (ii) predicted deleteriousness (pLI score for PTVs, MPC score for SAVs); (iii) number of patients carrying each variant; and (iv) presence in EOS-specific CNVs. The mutation score was calculated as:

$$MS={\sum }_{i=1}^m{e}^{V_{ptv\_i}}\ast S_{ptv\_i}+{\sum }_{j=1}^n{e}^{V_{sav\_j}}\ast S_{sav\_j}+C\text{,}$$

where V represents variant deleteriousness scores, S represents the number of patients with each variant, and C is the CNV count.

Literature support score (CS) calculation

We compiled 3,097 schizophrenia candidate genes from six categories of genetic studies: genome-wide association study (GWAS), CNV studies, whole-exome/genome sequencing (WES/WGS), differential expression gene (DEG), differential methylation gene (DMG), and linkage analyses. For each category, a candidate score (CS) was calculated based on the number of supporting publications (n):

$$CS=\{\begin{array}{l}1+\frac{1}{1+e^{-n}},n>0\\ 1,n=0\end{array}$$

Gene risk score (GRS) calculation

The GRS was calculated by combining mutation and literature scores:

$$GRS=MS\ast \prod _{k=1}^6{CS}_k$$

Genes with GRS >4 were classified as high-risk genes for subsequent analyses.

Metabolic pathway gene collection

Metabolic pathway genes were obtained from the Reactome database¹⁷ (v87) following the approach of Sinkala et al.⁵² We extracted proteins from 15 first-tier subcategories under the ‘Metabolism’ category, including carbohydrate metabolism, lipid metabolism, amino acid metabolism, nucleotide metabolism, and others.

Serum metabolomic profiling

Serum samples were analyzed by liquid chromatography-tandem mass spectrometry (LC-MS/MS) in both positive and negative ion modes (Wininnovate Bio). Quality control (QC) samples were prepared by pooling equal aliquots of all serum samples. Before formal data acquisition, repeated QC injections were performed to equilibrate the LC-MS/MS system. During acquisition, QC samples were intermittently injected at regular intervals throughout the analytical sequence to monitor instrument stability, retention time consistency, and signal reproducibility. Raw data were converted to mzXML format using ProteoWizard⁵³ (v3.0.7414).

Peak detection and normalization

Raw LC-MS/MS data were processed using the XCMS⁵⁴ (v3.9.3) package for peak detection, retention time alignment, and feature extraction. Peak detection was performed using the centWave algorithm with optimized parameters for mass accuracy, peak width, and signal-to-noise ratio. Retention time correction was performed using the ObiWarp method. The resulting feature table was further processed using the masscleaner⁵⁵ (v1.0.4) package for quality control, missing value handling, and normalization. Low-quality peaks were defined as peaks with >20% missing values in quality control (QC) samples or >50% missing values in either patient or control groups. Missing values were imputed using the knn method, and intensities were normalized using the pqn method. Peaks with a coefficient of variation (CV) ≥30% in QC samples after normalization were excluded.

Metabolite identification

Compound spectra extraction and adduct ion annotation were performed using the CAMERA⁵⁶ (v1.46.0) package. The following adduct ions were considered during annotation: positive ion mode, [M+H]⁺, [M+Na]⁺, [M+K]⁺, and [M+NH4]⁺; negative ion mode, [M−H]⁻, [M+NH4−2H]⁻, [M+2Cl]²⁻, and [2M−3H]³⁻.

MS1 and MS2 metabolite annotations were performed using metaX⁵⁷ (v2.0.0) against in-house spectral libraries (Wininnovate Bio) and public databases, including MassBank, HMDB, KEGG and LipidBlast. The annotation parameters were set as follows: MS1 mass tolerance, 10 ppm and 0.01 Da; MS2 mass tolerance, 0.05 Da; and metabolite identification score cutoff, 75%.

High-confidence metabolite annotations required: (i) MS2 spectral data; and (ii) consistency between MS1 and MS2 annotations. Specifically, for a metabolite to be considered accurately annotated, MS2 identifications (HMDB ID, KEGG ID, or metabolite name) must match corresponding MS1 entries. Metabolite superclasses were assigned based on HMDB classifications.

Identification of differentially expressed metabolites (DEMs)

Differential metabolite analysis was performed in R using the lm function based on multivariable linear regression models. Metabolite intensities were compared between EOS patients and healthy controls while adjusting for age and sex as covariates. Multiple testing correction was performed using the Benjamini-Hochberg false discovery rate (FDR) method. DEMs were identified using a significance threshold of false discovery rate (FDR) < 0.05, resulting in the identification of 3,256 and 1,815 DEMs under positive and negative ion model respectively.

To further evaluate the potential effects of clinical variables on metabolite alterations, disease duration and antipsychotic drug dose were additionally included as covariates in the regression models. No significant associations were observed after multiple testing correction, indicating that disease duration and drug dose did not substantially influence the identification of DEMs in this study.

Annotation refinement

For pathway enrichment analysis, we prioritized metabolites with high-confidence MS2-level annotation. Where multiple peaks corresponded to the same metabolite annotation, mean intensity across peaks was calculated to avoid redundancy. This resulted in 117 unique, accurately annotated DEMs used for biological interpretation.

Pathway and enrichment analysis

Pathway enrichment analysis of the 117 high-confidence DEMs was performed using MetaboAnalyst¹⁸ (v6.0). We tested for enrichment in: (i) metabolic pathways (SMPDB, KEGG, RaMP databases); (ii) disease signatures in blood and cerebrospinal fluid; and (iii) tissue/cellular locations and environmental exposures. Statistical significance was assessed using hypergeometric tests with FDR correction (FDR <0.05).

Schizophrenia associated metabolic biomarker compilation

Known schizophrenia metabolic biomarkers were compiled from 44 publications identified through PubMed searches using the keywords “metabolic biomarker in schizophrenia” and “differentially expressed metabolites in schizophrenia.” Metabolite annotations were manually curated across PubChem, HMDB, and KEGG databases, yielding 134 unique biomarkers for comparison with our DEMs.

Correlation analysis between metabolites and clinical symptoms

Among the 28 EOS patients, 18 had complete PANSS assessments. We calculated Pearson correlation coefficients between DEM intensities and PANSS total scores, positive symptom scores, and negative symptom scores using the cor. test function in R. Metabolites were considered significantly correlated if they met the criteria: |r²| ≥0.3 and p <0.1.

Protein-metabolite associations analyses

To link genetic variants to metabolite dysregulation, we selected high-risk genes with PTVs or SAVs present in ≥2 patients. For each gene, we stratified EOS patients into mutation carriers (Mc SCZ) and non-mutation carriers (No-mc SCZ) groups. We then tested whether each of the 3,604 DEMs and 15,428 non-DEMs showed differential abundance between: (i) Mc SCZ versus No-mc SCZ, and (ii) Mc SCZ versus controls. The intensity of metabolites could differ by several orders of magnitude, and analyses based solely on raw intensity values would be influenced by extreme values and non-normal distributions. To improve the robustness and reliability of the association analyses, we additionally performed rank transformation paralleling raw intensity comparison, which could reduce sensitivity to outliers and large-scale intensity differences while preserving the relative ordering of metabolite abundance across samples.

A metabolite was classified as associated with a gene's variants if it showed significant differences (FDR-adjusted p <0.05) in both comparisons using both intensity and rank-based tests. We further filtered non-DEMs to retain only those correlated with at least one DEM (see below). After applying MS2 annotation filtering, we identified 118 high-confidence protein-metabolite associations involving 31 high-risk genes and 73 metabolites (62 DEMs, 11 non-DEMs).

Metabolite-metabolite correlations analyses

Spearman pairwise correlations between all metabolites were calculated using rcorr function in Hmisc package (v5.1.3) with Benjamini-Hochberg FDR correction. Metabolite pairs with |r²| ≥0.9 and FDR-adjusted p <0.05 were classified as correlated. For non-DEMs, we require correlation with ≥1 DEM. After filtering accurate MS2 annotations, 42 correlation pairs involving 30 DEMs and 4 non-DEMs were retained.

Genetic–metabolomic network construction and lipid subnetwork extraction

We integrated protein-metabolite associations and metabolite-metabolite correlations to construct a comprehensive regulatory network. Network visualization and layout were performed using Cytoscape⁵⁸ (v3.10.4).

A lipid dysregulation subnetwork was subsequently extracted by selecting metabolites classified as “Lipids and lipid-like molecules” (based on HMDB superclass) together with their interacting proteins. Three non-DEMs (lauroy-L-carnitine, lysoPE 18:2 and capric acid) included in this subnetwork but not correlated with lipids were removed. LysoPC 18:0 and stearic acid remained metabolite-metabolite correlations without association with any high-risk genes were also removed. The final lipid dysregulation subnetwork contains 31 lipids, 19 associated high-risk genes and 63 regulation edges.

Validation of mutation-metabolite associations

For validation of the network-level associations, a mutation-collapsing strategy was further applied for each metabolite. Specifically, when a metabolite was associated with multiple genes in the mutation–metabolite network, variants in all metabolite-associated genes were combined. EOS samples carrying at least one variant in any metabolite-associated gene were classified as mutation-carrier SCZ samples (McSCZ), whereas EOS samples lacking variants in all genes associated with given metabolite were classified as non-mutation carrier SCZ samples (No-mcSCZ). Therefore, the composition of the McSCZ and No-mcSCZ groups varied depending on the set of genes associated with each metabolite. Metabolite intensities were compared between groups using Student's t-test. This analysis tested whether the cumulative burden of mutations in lipid-associated genes corresponded to altered lipid levels.

Quantification and statistical analysis

Unless otherwise stated, statistical comparisons were performed using Student's t-test or Wilcoxon rank-sum test as appropriate. Fisher’s exact test was performed when comparing sex difference between EOSCZ and normal samples. Multiple testing correction was applied using the Benjamini-Hochberg false discovery rate (FDR) method. Statistical significance was defined as adjusted p <0.05 for primary analyses, with p <0.1 used for exploratory correlation analyses with PANSS scores due to limited sample size. All analyses were performed in R (v4.1.3). Data visualization was performed using ggplot2 (v3.4.2) and custom scripts.

Supplemental information

Table S2. High-risk genes in patients with EOS, related to Figure 1

PTV_number and SAV_number indicate the numbers of PTVs and SAVs detected in EOS samples, respectively. EOS_with_PTV, EOS_with_SAV, and EOS_with_variants indicate the numbers of EOS samples carrying PTVs, SAVs, or either variant type. pLI_VS and MPC_VS represent variant constraint scores for PTVs and SAVs, respectively. EOS_with_CNV indicates the number of EOS samples carrying schizophrenia-associated copy number variations (CNVs). GWAS, CNV, WES, DEG, DMG, imputation, and linkage indicate the number of supporting studies using corresponding approaches. GRS represents the gene risk score calculated by integrating variant burden, variant scores, CNV information, and supporting evidence. See STAR Methods for details.

Table S4. DEMs with accurate annotation, related to Figure 3

DEMs marked in bold were identified in both positive and negative ionization modes and were merged as a single metabolite for subsequent analyses. Differences in metabolite abundance were estimated using linear regression models. The regression models included different covariate combinations: disease status, gender, and age were included as covariates in the first model; disease duration was included as an independent variable in the second model; and medication dose was included as an independent variable in the third model. β represents the estimated regression coefficient. Adjusted p values were calculated using Benjamini-Hochberg correction. Drug dose was normalized to chlorpromazine equivalents (CPZ equivalents). DEM, differentially expressed metabolite; m/z, mass-to-charge ratio.

Table S5. Enriched terms of DEMs from MetaboAnalyst, related to Figure 3

DEMs involved pathways and tissues were listed. DEM, differentially expressed metabolite.

Table S6. DEMs significantly correlated with PANSS scores, related to Figure 4

DEMs with an absolute Pearson’s correlation coefficient value (|r|) ≥ 0.3 and p < 0.1 were considered significantly correlated with PANSS scores. r.total, r.positive, and r.negative represent the Pearson’s correlation coefficients between metabolite abundance and PANSS total, positive symptom, and negative symptom scores, respectively. p value.total, p value.positive, and p value.negative represent the corresponding p value from Pearson’s correlation analyses calculated using the cor.test function in R. PANSS, Positive and Negative Syndrome Scale; DEM, differentially expressed metabolite.