Metabolic syndrome-associated gut microbiota and plasma metabolite profiles in schizophrenia

Abstract

Metabolic syndrome (MetS) is a common and clinically important comorbidity in schizophrenia (SCZ), but the biological basis remains unclear. We aimed to characterize the gut microbial and plasma metabolites associated with comorbid MetS in SCZ and to evaluate their potential as integrated biomarkers. A total of 210 participants were enrolled, including 79 patients with SCZ+MetS, 71 patients with SCZ alone, and 60 matched controls. Gut microbiota was analyzed by 16S rRNA gene sequencing, and plasma metabolites were profiled using GC-MS and UPLC-MS. Differential analysis, microbiota-metabolite-clinical phenotype association analysis, and classifier modelling based on gut microbiota and plasma metabolites were mainly performed. We identified thirteen gut microbial genera, including Raoultella, Muribaculum, Pantoea, Paramuribaculum, and Acinetobacter, that showed higher levels in SCZ+MetS patients than in patients with SCZ alone and controls. Amino acids and related derivatives including branched-chain amino acids, L-Glutamic acid, and 3-Methyl-2-Oxovaleric acid showed higher levels in SCZ+MetS patients than in both patients with SCZ alone and control, whereas the levels of phosphatidylcholine were lower relative to the other two groups. Muribaculum, Pantoea, and Raoultella were positively correlated with plasma amino acids and related derivatives including branched-chain amino acids, L-Glutamic acid, and 3-Methyl-2-Oxovaleric acid. A random forest classifier integrating gut microbiota and plasma metabolites achieved good predictive ability for SCZ+MetS, with an AUC of 0.975. In summary, SCZ+MetS is characterized by distinct gut dysbiosis and some disruptions in amino acid and phospholipid metabolism, which may serve as clinically relevant biomarkers of this comorbidity.

Introduction

Patients with schizophrenia (SCZ) are at increased risk of developing metabolic syndrome (MetS), with a prevalence of 35.3%, regardless of the specific definition of MetS [1]. Comorbid MetS in SCZ (SCZ+MetS) is closely correlated with cognitive impairment, disease relapse, and increased cardiovascular disease-related death [2–4]. Elucidating the mechanisms underlying the SCZ+MetS is essential for improving clinical outcomes. Studies have disclosed several potential contributors to this comorbidity, including unhealthy lifestyle (e.g.: physical inactivity and imbalanced diet), the use of antipsychotic drugs such as clozapine and olanzapine, shared genetic vulnerability, gut microbiota dysbiosis, and inflammation and endocrine dysregulations [5–9].

Specifically, accumulating studies have posited the important role of gut microbiota across SCZ and metabolic disorders (Supplementary Table 1 and 2). In SCZ patients, most studies have reported lower relative abundances of Roseburia, Faecalibacterium, and Coprococcus and higher relative abundance of Collinsella than in healthy controls (Supplementary Table 1). Similarly, metabolic disorders such as type 2 diabetes, hyperlipidemia, obesity, and hypertension exhibited distinct gut bacterial composition as well, including lower relative abundance of Faecalibacterium and Oscillospira and higher relative abundance of Alistipes (Supplementary Table 2). The concordant direction of some microbiota differences (e.g.: lower level of Faecalibacterium), across SCZ and metabolic disorders suggests a potential involvement of the gut microbiota in their co-occurrence. One study compared the gut microbiota composition in SCZ patients with central obesity and those with normal weight, and reported that female SCZ patients with central obesity possessed higher relative abundance of Prevotella and Roseburia and lower relative abundance of Akkermansia [10]. However, studies particularly characterizing the gut microbiota in SCZ patients with comorbid metabolic disorders remain limited (Supplementary Table 3).

Disease-associated circulating metabolic changes could provide insights into pathophysiology and clinical intervention [11]. Previous studies have investigated characteristics of circulating metabolites in individuals with SCZ and metabolic disorders [12–19]. Most have consistently reported higher levels of valine and glutamate and lower levels of tryptophan, n3 polyunsaturated fatty acids, and n6 polyunsaturated fatty acids in SCZ patients compared with healthy individuals [18]. In individuals with MetS and related metabolic diseases, branched-chain amino acids (BCAAs), including valine, isoleucine, and leucine, as well as phenylalanine and tyrosine, have been reported to be at higher levels [13, 15]. Nevertheless, evidence on metabolic profiles in patients with SCZ+MetS is scarce. Two studies reported that C5 acylcarnitine level was higher in patients with SCZ+MetS compared with SCZ alone [20, 21]. Given the critical role of the gut microbiota in host systemic metabolism [22], elucidating associations between the gut microbiome and circulating metabolites in individuals with SCZ+MetS is important, yet remains largely unexplored. Conducting multidimensional studies to further investigate mechanisms underlying SCZ+MetS is needed.

In this study, we aimed to integrate 16S rRNA gene sequencing of the gut microbiota with untargeted plasma metabolomics to identify gut microbiota and circulating metabolites associated with SCZ+MetS. Additionally, we sought to investigate correlations between gut microbiota and plasma metabolites to uncover potential candidate markers for SCZ+MetS.

Methods

Subject recruitment and data collection

From September 2021 to August 2022, a total of 79 patients with SCZ+MetS, 71 patients with SCZ alone, and 60 controls were enrolled. Based on α level of 0.05 and the statistical power of 0.80, a priori power analysis using GPower (v3.1) estimated a minimum required sample size of 155 (55 participants per group). All patients were admitted to Shanghai Mental Health Center, and the controls were recruited from the same area.

All SCZ patients were diagnosed according to the Diagnostic and Statistical Manual of Mental Disorders (5th Edition), were aged 18-70 years, of Han Chinese ethnicity, and had been receiving stable antipsychotic treatment for at least six months. According to the diagnostic criteria for MetS outlined in the Chinese Guidelines for the Prevention and Treatment of Type 2 Diabetes (2017 edition), SCZ patients were further categorized into two groups: SCZ+MetS and SCZ alone. Exclusion criteria were as follows: suffering from severe physical diseases or infectious diseases; experiencing significant changes in living environment or dietary conditions, using probiotics or antibiotics, or experiencing diarrhea or constipation in the last month. Additionally, healthy controls with a history of mental disorders were also excluded from study.

Clinical data including general demographic information, disease duration, medications, family history, individual medical history, anthropometric indicators, and blood biochemical indicators were collected.

For biological sample collection, approximately 3 g of fresh feces were collected after overnight fasting by collection tube containing fecal protection liquid (Guangdong Longsee Biomedical LTD, China), following the standard operating procedures. Samples were immediately frozen at −80 °C and stored for 16S rRNA gene sequencing. 5 mL of peripheral venous blood was collected after overnight fasting by lithium heparin anticoagulant tubes and EDTA anticoagulant tubes, respectively. Samples were centrifuged at 4 °C within 4 h (3000 rpm, 15 min). 300 μL of the supernatant was transferred into cryotube and frozen at −80 °C for non-targeted metabolomics and inflammatory factor detections. The method of meso-scale discovery electrochemiluminescence (K15396S-1, Rockville, MD, USA) was used to detect inflammatory factors (Supplementary method).

16SrRNA gene sequencing

The genomic DNA of fecal samples was extracted by MagPure Soil DNA KF Kit (Magen, China) following the manufacturer’s instructions. V3-V4 variable regions of 16S rRNA genes were amplified with universal primers 343 F (5’-TACGGRAGGCAGCAG-3’) and 798 R (5’-AGGGTATCTAATCCT-3’). Then PCR products were purified with AMPure XP beads (Agencourt, USA) and quantified with Qubit dsDNA HS assay kit (Yeasen, China). DNA concentrations were adjusted to the same level for subsequent sequencing, which was performed on the Illumina NovaSeq6000 platform with two paired-end read cycles of 250 bases each by OE Biotech LTD (Shanghai, China).

Untargeted metabolomics profiling

Gas chromatography-mass spectrometry (GC-MS) and Ultra-performance liquid chromatography-mass (UPLC-MS) measurements were performed by OE Biotech LTD (Shanghai, China) with detailed sample preparation and running procedures in supplementary method. Briefly, GC-MS was conducted by Agilent 7890B gas chromatography system and Agilent 5977 A MSD. The DB-5MS fused-silica capillary column (30 m × 0.25 mm × 0.25 μm, Agilent J & W Scientific) was utilized. UPLC-MS was performed by ACQUITY UPLC I-Class plus system (Waters) and Q-Exactive quadrupole-Orbitrap mass spectrometer with heated electrospray ionization (ESI) source (Thermo Fisher Scientific). ACQUITY UPLC HSS T3 column (100 mm × 2.1 mm, 1.8 μm) was used. Both ESI positive and ESI negative ion modes were run.

16S rRNA gene amplicon data processing

Raw 16S rRNA gene sequencing data in FASTQ format were processed by Cutadapt (v4.0, https://cutadapt.readthedocs.io/en/stable/) to remove adapter sequence. Paired-end reads were merged and quality-filtered by VSEARCH (v2.15.2, https://peerj.com/articles/2584/) [23], requiring a minimum overlap of 10 bases between forward and reverse reads, a maximum mean error rate of 1% and sequences containing ambiguous bases been removed. Sequence quality was further assessed using FastQC (v0.11.9, https://www.bioinformatics.babraham.ac.uk/projects/fastqc/). Dereplication, denoising, chimera detection, and amplicon sequence variant (ASV) feature table construction were performed by VSEARCH and USEARCH (v10.0.240, https://www.drive5.com/usearch/), with denoising implemented by UNOISE3 algorithm, which provides balanced sensitivity and specificity [24]. Taxonomic assignment of ASVs was conducted by aligning sequences against the RDP database (v: rdp_16s_v18) [25] using VSEARCH, with a minimum confidence threshold of 0.6.

Untargeted metabolomics data processing

Raw data obtained by GC-MS were converted to the.abf format by AnalysisBaseFileConverter software (v4.0.0, https://www.reifycs.com/abfconverter/) and subsequently imported to MS-DIAL software (v4.9.221218, https://systemsomicslab.github.io/compms/msdial/main.html) [26] for peak detection, deconvolution, peak identification, peak alignment, wave filtering, and missing value interpolation. Annotation of the metabolite was conducted by referring to the in-house database LUG (Oebiotech, Shanghai, China) with the value of identification score cut-off set at 0.7. Metabolite intensities were segmented and normalized according to the internal standard with a relative standard deviation (RSD) lower than 0.1.

Raw data obtained by UPLC-MS were processed by Progenesis QI (v2.3, Nonlinear Dynamics, Newcastle, UK) for baseline filtering, peak identification, integral, retention time correction, peak alignment, and normalization. The Human Metabolome Database (HMDB, https://hmdb.ca/) [27], LIPID MAPS (https://lipidmaps.org/) [28], METLIN (http://metlin.scripps.edu) [29], and in-house EMDB2.0 database (Oebiotech, Shanghai, China) were used as references for metabolite annotation based on precise mass-to-charge ratio, secondary fragments, retention time, and isotopic distribution.

Statistical analysis

Clinical characteristics analysis

Statistical analysis was performed by R 3.4.2. Continuous variables were first examined for the normal distribution. For those with normal distribution, ANOVA with post-hoc analysis was used, otherwise, Kruskal-Wallis test was used. Continuous variables were reported as means and standard deviations. Categorical variables were reported as absolute values and percentages. Chi-square analysis was utilized for categorical data comparisons. All tests were two-sided and Bonferroni method was utilized to account for multiple comparisons with threshold set as < 0.05.

Between-group gut microbiota composition analysis

Rarefaction analysis was performed using USEARCH based on the normalized ASV abundance table. Alpha diversity was estimated by Shannon-Weaver and Simpson indices using ASV features by the R package vegan (v2.6-2). Beta diversity between groups was calculated using Bray-Curtis distance and tested by permutational multivariate analysis of variance (PERMANOVA, 99999 permutations). Multivariate homogeneity of group dispersions (PERMDISP) was conducted to evaluate the homogeneity of variances between groups. Taxa were filtered by removing low prevalence genera (present in < 20% samples) for downstream analysis. Between-group differences in taxonomic relative abundance were evaluated using Dunn’s Kruskal-Wallis test, followed by Holm correction for multiple comparisons.

Random forest classification based on gut microbiota data

All genera with prevalence > 20% were used to construct the random forest (RF) classification model. Firstly, samples were randomly split into a training set (60%) and a test set (40%). Next, within the training set, we performed five dependent repeats of five-fold cross-validation for feature selection (function rfcv from R package randomForest, v4.7-1.1). Specifically, the mean prediction errors of different numbers of selected genera were estimated via the cross-validation. The subset of genera with the lower mean prediction error was then used to build the final prediction model based on the training set. Finally, the classification performance of the model was evaluated using the hold-out test set and visualized using receiver operating characteristic (ROC) curves generated with the R package multiROC (v1.1.1). Area under the ROC (AUC) was calculated for each group and overall three groups (Micro and Macro algorithm from multiROC).

Between-group metabolite difference analysis

Data from GC-MS and UPLC-MS were log10 transformed and standardized for downstream analysis, respectively. Firstly, principal component analysis (PCA) was performed to evaluate the general stability of measurements of GC-MS and UPLC-MS. Subsequently, partial least squares-discriminant analysis (PLS-DA) was conducted by R package rolls (v1.28.2), and the variable projection importance (VIP) scores were calculated. PLS-DA model performance was evaluated using seven-fold cross-validation to test both explanatory power and predictive ability, and 1000 permutations were performed to assess potential overfitting. Metabolites with VIP > 1 were selected as potential variables for group discrimination. Furthermore, One-way ANOVA was performed to compare the metabolite intensity differences between groups and Fisher’s least significant difference (LSD) was used for post-hoc analysis. Subsequently, fold change (FC) was calculated between SCZ+MetS and SCZ alone, SCZ+MetS and controls. The False discovery rate (FDR) was set as 0.05 for multiple comparison corrections.

Weighted gene co-expression network analysis of metabolites

Clusters of co-occurrence plasma metabolites from GC-MS analysis were constructed by weighted gene co-expression network analysis (WGCNA) using R package WGCNA (v1.71). A signed weighted network was chosen. The soft threshold was set as 18 according to the scale-free network criteria. Modules were identified using a combination of dynamic tree cutting and partitioning around medoids (PAM). Signed eigengene-based connectivity (KME) was calculated to determine hub metabolites in each module, with KME > 0.8 as the cutoff. Clusters of co-occurrence metabolites from UPLC-MS were constructed in the same way with the soft threshold set as 8.

Correlation analysis among gut microbiota, plasma metabolites, metabolic modules, and clinical phenotypes

Regarding the relationships between gut microbiota, plasma metabolites, and clinical phenotypes, Spearman’s partial correlation analysis (R package ppcor v1.1) was conducted to estimate the correlations with adjustment of potential confounders, including gender, age, medications, and disease duration. Genera in intersection between differential genera and genera selected by RF and with consistent directional changes across comparisons (SCZ+MetS vs. SCZ alone and SCZ+MetS vs. controls), and metabolites in the intersection between differential metabolites and metabolites with VIP > 1, also with consistent directional changes across the two comparisons were used for the analysis. Metabolite co-occurrence modules generated by the WGCNA were utilized for clinical phenotype correlation analyses as well. Finaly, the associations among the screened genera and metabolites including the metabolite modules were performed with the same confounding factors controlled.

Functional prediction of 16S marker sequences

Functional inference from the 16S rRNA amplicon data was performed using PICRUSt2 (v2.5.1, https://github.com/picrust/picrust2) [30], with gene functions annotated against the Kyoto Encyclopedia of Genes and Genomes (KEGG) database (https://www.genome.jp/kegg/). To identify microbial functions that significantly differed between groups, we used the Linear Discriminant Analysis Effect Size (LEfSe) method with an LDA score threshold of ≥2.

Functional enrichment analysis of differential metabolites and co-occurrence metabolites

Differential metabolites screened by the aforementioned methods and co-occurrence metabolites obtained through WGCNA were collapsed to pathway enrichment analysis based on KEGG and the Small Molecule Pathway Database (SMPDB, https://smpdb.ca/) by the online website MetaboAnalyst5.0 (https://www.metaboanalyst.ca/). Over-representation analysis was utilized to test the significance of statistics (P < 0.05).

Random forest classifier using gut microbiota and plasma metabolites

To assess the group discrimination ability of combined gut microbiota and plasma metabolites, we constructed the RF classifier. Samples were randomly split into a training set (60%) and a test set (40%). Candidate features for model construction using training set included: 1) differential gut microbial genera that were associated with clinical phenotypes and 2) differential plasma metabolites (amino acids, fatty lipids, and phospholipids) that were also associated with clinical phenotype. Here differential genera and metabolites were screened by the same consistent-direction strategy and functional analysis highlighted variables. Evaluation of model performance was conducted in the hold-out test set using the same approach described above.

Results

Clinical characteristics of subjects

A total of 210 subjects were enrolled, including 79 patients with SCZ+MetS (30 female, mean [SD] age, 56.47 [10.51] years), 71 patients with SZC alone (32 female, mean [SD] age, 54.11 [13.38] years), and 60 controls (27 female, mean [SD] age, 55.32 [8.96] years). Age and gender distributions were comparable between these three groups. Clinical metabolic indices were higher in SCZ+MetS than SCZ alone and controls. Different types of antipsychotic drugs and sedative drugs used did not differ between the two SCZ groups. Additionally, patients with SCZ+MetS had higher levels of inflammatory factors, including IL-17A, TNF-α, IL-6, and IL-1β, than controls. No significant differences of these inflammatory factors were observed between SCZ patients with and without comorbid MetS. Details are provided in Table 1.

Table 1.

Clinical characteristics of subjects.

	SCZ+MetS	SCZ alone	Controls	P value
Age (year)	56.47 ± 10.51	54.11 ± 13.38	55.32 ± 8.96	0.437
Gender (female, n, %)	30 (38)	32 (45)	27 (45)	0.604
Body weight (kg)	77.19 ± 12.97a	58.33 ± 8.17c	62.91 ± 7.76b	<0.001
BMI (kg/m2)	27.17 ± 3.48a	21.5 ± 2.49c	23.26 ± 1.62b	<0.001
WC (cm)	97.44 ± 8.8a	79.73 ± 8.78b	81.25 ± 5.07b	<0.001
Systolic BP (mmHg)	123.68 ± 14.83a	117.23 ± 12.68b	120.21 ± 12.34ab	0.017
Diastolic BP (mmHg)	78.96 ± 8.56a	75.25 ± 7.93b	75.28 ± 8.05b	0.009
Glucose (mmol/L)	6.10 ± 1.72a	4.72 ± 0.48b	5.14 ± 0.55b	<0.001
TG (mmol/L)	2.33 ± 1.67a	1.02 ± 0.35b	1.27 ± 0.72b	<0.001
HDL-C (mmol/L)	1.06 ± 0.29b	1.34 ± 0.33a	1.35 ± 0.31a	<0.001
Disease course (year)	25.37 ± 10.64	26.48 ± 13.7	/	0.578
Type of Antipsychotic Drugs				0.943
Containing olanzapine and clozapine	56 (71)	49 (69)	/
Others	23 (29)	22 (31)	/
Use of sedative and hypnotic drugs				0.529
Yes	14 (18)	9 (13)	/
No	65 (82)	62 (87)	/
Use of drugs treating extrapyramidal side effects				0.251
Yes	21 (27)	26 (37)	/
No	58 (73)	45 (63)	/
Use of drugs treating metabolic disturbance				<0.001
Yes	52 (66)	6 (8)	/
No	27 (34)	65 (92)	/
Family history				<0.001
Without family history of psychiatric or metabolic diseases	55 (82)b	47 (72)b	60 (100)a
With family history of psychiatric diseases alone	10 (15)	18 (28)	0 (0)
With family history of metabolic diseases alone	1 (1)	0 (0)	0 (0)
With family history of psychiatric and metabolic diseases	1 (1)	0 (0)	0 (0)
IL-12p70 (fg/ml)	258.70 (216.35, 318.25)	322.77 (209.16, 471.11)	296.29 (223.37, 329.38)	0.26
IL-17A (fg/ml)	788.059 (613.04, 1294.19)a	659.33 (466.67, 1121.70)a	393.14 (302.81, 569.04)b	0.001
IL-2 (fg/ml)	189.9 (144.57, 322.56)	162.08 (135.05, 303.97)	151.63 (119.59, 212.13)	0.365
IFN-γ (fg/ml)	863.68 (734.02, 1263.89)	848.38 (548.73, 1512.88)	667.46 (507.54, 916.64)	0.253
IL-10 (fg/ml)	793.19 (698.58, 1225.55)a	988.44 (683.09, 1305.36)a	496.67 (269.01, 520.37)b	<0.001
TNF-α(fg/ml)	1427.35 (1147.30, 1853.15)a	1079.31 (982.24, 1407.22)ab	1011.17 (830.94, 1170.63)b	0.001
IL-6 (fg/ml)	3033.86 (2078.09, 4126.58)a	2088.7 (1489.24, 4303.54)a	1655.51 (1173.41, 1972.40)b	<0.001
IL-1β (fg/ml)	331.49 (224.36, 498.42)a	292.85 (156.03, 615.84)ab	141.27 (86.57, 242.84)b	0.011

WC waist circumference, BMI body mass index, TG triglyceride, HDL-C High-density lipoprotein cholesterol, BP blood pressure. SCZ+MetS: schizophrenia patients with comorbid MetS, SCZ alone: schizophrenia patients without MetS.

a/b/c indicates the results of multiple comparisons. For the same variable, the corner marked with the same letter indicates no statistical difference between the two groups; otherwise, there exists a difference.

Differential compositions of gut microbiota between groups

200 out of the 210 subjects had available gut microbiota data. Each sample produced more than 70,000 reads (Supplementary Table 4). 7726 ASV features were obtained and annotated to 132 high prevalence genera (present in > 20% samples). General compositions of the gut microbiota between groups were described at Supplementary Fig. S1A-D.

Alpha diversities assessed using Shannon and Simpson indices did not differ significantly among these three groups (Supplementary Fig. S2A-B). Beta diversity using the Bray-Curtis distance demonstrated a significant dissimilarity between all three groups with adjustment of age and gender (P < 0.001, Fig. 1A). The multivariate homogeneity of dispersion showed significant differences between SCZ+MetS and control (P < 0.001), as well as between SCZ alone and control (P = 0.013). The dispersion between SCZ+MetS and SCZ alone was comparable (P = 0.34, Supplementary Fig. S2C). Besides, the influences of medications on the overall composition of gut microbiota in SCZ patients were also explored and no obvious influences were observed (Supplementary Fig. S2D-E).

Fig. 1. Differential gut microbiota compositions between groups.

A Principal coordinates analysis (PCoA) of beta diversity based on Bray-Curtis distances. Differences in beta diversity among SCZ+MetS group (red), SCZ-only group (purple), and control group (green) were assessed using PERMANOVA, with the P value shown in the figure. B Differential gut microbial genera between groups. The y-axis shows the names of bacterial families and genera. Color indicates differences in relative abundance (e.g., for the comparison of SCZ+MetS vs. con, the value was calculated as SCZ+MetS - control and multiplied by 100), and asterisks denote the significance level after multiple-comparison correction (*** < 0.001, ** <0.01 and * <0.05). C Thirty-three gut microbial genera were selected by the random forest classifier through cross-validation test and ranked according to the standardized mean decrease in accuracy. The color of each genus indicates its corresponding phylum. D Performance of the random forest classifier was assessed by ROC analysis on the hold-out test set, demonstrating moderate to good discriminatory ability. AUC were calculated for each group (SCZ+MetS, SCZ alone, and control) and for the overall three-group classification using micro- and macro-averaging.

Out of the 132 gut microbial genera, 60 (31 between SCZ+MetS and SCZ alone; 40 between SCZ+MetS and control; 41 between SCZ alone and control) showed significant differences in relative abundances between these three groups (Fig. 1B). Relative abundances of gut microbial genera, such as Paramuribaculum, Acinetobacter, Muribaculum, and Raoultella, exhibited consistent higher levels in the two comparisons of SCZ+MetS versus SCZ alone and SCZ+MetS versus controls. However, some genera including Roseburia, Faecalibacterium, and Lactobacillus only differed between SCZ patients and controls but did not show significant differences between SCZ patients with and without MetS.

Random forest modeling identified 33 genera of the 132 genera to construct the classification model (Fig. 1C). These genera showed moderate to good discriminatory ability across the three groups (Fig. 1D). The AUC values were 0.927 for SCZ+MetS (CI [0.851, 0.975]), 0.795 for SCZ alone (CI [0.654, 0.892]), and 0.941 for controls (CI [0.845, 0.983]). Among the 33 selected genera, 28 also overlapped with differential genera identified across these three groups including Paramuribaculum, Acinetobacter, Muribaculum, and Raoultella.

Differential compositions of plasma metabolites between groups

206 of 210 subjects had available plasma metabolite data. A total of 312 metabolites were obtained by GC-MS and 773 metabolites by UPLC-MS. Lipid and lipid-like molecules, as well as organic acids and derivatives comprised the majority of metabolites, although their proportions differed between GC-MS and UPLC-MS (Supplementary Fig. S3A-D). Downstream data analysis of metabolites detected from the two platforms were performed separately.

Differential metabolite profiles among groups were evaluated. At the overall level, PLS-DA showed significant discriminations between the three groups (Fig. 2A-D), with 94 GC-MS and 291 UPLC-MS metabolites identified respectively contributing to the group separation (VIP > 1). Regarding the individual differential metabolites between groups, 64 were identified between SCZ+MetS and SCZ alone, 130 between SCZ+MetS and controls, and 121 between SCZ alone and controls (adjusted P < 0.05, GC-MS, Supplementary Table 5). For data from the UPLC-MS, 131 differential metabolites were identified between SCZ+MetS and SCZ alone, 411 between SCZ+MetS and controls, and 384 between SCZ alone and controls (adjusted P < 0.05, Supplementary Table 6). To assess whether the direction of differential metabolite changes was consistent across comparisons, FC were performed, followed by Pearson correlation analysis between the log2FC values of differential metabolites in comparisons 1) SCZ+MetS versus SCZ alone and 2) SCZ+MetS versus controls (Fig. 2E, G). The correlation coefficients were 0.25 (P = 0.129, GC-MS) and 0.683 (P < 0.001, UPLC-MS). The top 20 metabolites among these consistent-direction intersections, ranked by VIP, were presented in Fig. 2F and H. Notably, levels of amino acids and derivatives, including BCAA (e.g.: L-Valine and L-Isoleucine/L-Leucine), L-Glutamic acid, and N-Lactoylphenylalanine, were higher in SCZ+MetS patients than in both SCZ alone patients and controls. Conversely, phospholipids including phosphatidylcholines (PC) and lysophosphatidylcholine (LysoPC) like PC (18:2(2E,4E)/0:0), PC (18:0(11cp)/0:0) and LysoPC (20:1(11Z)/0:0) were lower in SCZ+MetS patients compared with SCZ alone patients and controls.

Fig. 2. Differential plasma metabolite compositions between groups.

A PLS-DA score plot showing differences in the overall plasma metabolite profiles (GC-MS) among the SCZ+MetS group (red), SCZ alone group (purple), and control group (green). B Permutation test used to evaluate PLS-DA model (GC-MS). C PLS-DA score plot showing differences in the overall plasma metabolite profiles (UPLC-MS) among the SCZ+MetS group (red), SCZ alone group (purple), and control group (green). D Permutation test used to evaluate PLS-DA model (UPLC-MS). E Pearson correlation analysis of the log2FC values of differential metabolites identified by GC-MS in the comparisons of SCZ+MetS vs. SCZ alone and SCZ+MetS vs. controls. F Top 20 differential metabolites from the two comparisons ranked by VIP values (GC-MS). G Pearson correlation analysis of the log2FC values of differential metabolites identified by UPLC-MS in the comparisons of SCZ+MetS vs. SCZ alone and SCZ+MetS vs. controls. H Top 20 differential metabolites from the two comparisons ranked by VIP values (UPLC-MS).

Associations among gut microbiota, plasma metabolites, and clinical phenotypes

To identify gut microbial genera and plasma metabolites associated with the co-occurrence of SCZ and MetS, we applied the consistent-direction filtering strategy (details in method and Fig. 3A). As a result, 13 of 14 screened gut microbial genera, 29 of 30 screened metabolites (GC-MS) and 64 of 71 screened metabolites (UPLC-MS) showed significant correlations with clinical metabolic indicators (Fig. 3B-D and Supplementary Fig. S4A & C). Overall, the correlation directions of the identified genera and metabolites with metabolic indices (e.g.: BMI and triglycerides) were generally consistent with the categorical group outcomes.

Fig. 3. Associations among gut microbiota, plasma metabolites, and clinical phenotypes.

A Consistent-direction filtering strategy used to identify gut microbial genera and plasma metabolites associated with the co-occurrence of SCZ and MetS. B-F Spearman partial correlation analyses of the filtered gut microbial genera, filtered plasma metabolite, and clinical phenotypes. Metabolites shown in (C and E) were detected by GC-MS, whereas metabolites shown in (D and F) were detected by UPLC-MS. Red indicates a positive correlation, whereas blue indicates a negative correlation. Darker color intensity represents a stronger correlation, and asterisks denote statistically significant correlations after multiple comparison corrections.

Regarding genus-metabolite relationships, genera like Muribaculum, Pantoea, and Raoultella were positively correlated with plasma amino acids and derivatives like BCAA, L-Glutamic acid, and 3-Methyl-2-Oxovaleric acid. Pantoea negatively correlated with plasma fatty lipid 7-tetradecynoic acid (Fig. 3E, F, Supplementary Fig. S4B & D).

Furthermore, we assessed whether medications and demographic factors influenced gut microbiota and plasma metabolite profiles (Supplementary Fig. S5A-B and Figure S6A-D). Correlations between these factors and gut microbiota were not observed. However, gender appeared to influence the metabolism of lysine, with metabolites including Alpha-Aminoadipic acid and 4-trimethylammoniobutanoic acid being higher in males.

Associations of plasma metabolite clusters with gut microbiota and clinical phenotypes

We further performed WGCNA to identify modules of co-occurrence metabolites (Supplementary Fig. S7A-D and Figure S8A-D, Supplementary Table 7 and 8). The Modules represented by core metabolites including amino acids, fatty acids, phospholipids, which had KME > 0.8, showed significant correlations with disease states and clinical metabolic indicators (Fig. 4A, B). Similarly, positive correlations between genera Muribaculum, Pantoea, and Raoultella and metabolite modules represented by amino acids and related derivatives were observed, while negative correlations between Muribaculum and Pantoea and modules represented by fatty acids were observed (Fig. 4C, D).

Fig. 4. Correlation analysis of plasma metabolite clusters with clinical phenotypes and gut microbiota.

A–D Spearman correlation analysis of metabolic clusters, represented by their core metabolites, with clinical indicators and gut microbiota. The metabolite clusters shown in (A and B) were detected by GC-MS, whereas those shown in (C and D) were detected by UPLC-MS. Red indicates a positive correlation, whereas blue indicates a negative correlation. Darker colors represent stronger correlations. The correlation coefficient (upper) and significant level (lower) were shown in each cell.

Gut microbiota functional differences and metabolite pathway enrichment analysis

PICRUST2 was used to predict gut microbiota functions. Tyrosine metabolism, PPAR signaling pathway, and some other pathways of amino acid and lipid metabolism were enriched in SCZ+MetS, whereas GABAergic synapse and pathways related to glucose metabolism were enriched in SCZ alone (Supplementary Fig. S9). For plasma metabolites, the differential metabolites and co-existing metabolites were enriched in pathways including Valin, Leucine and Isoleucine Degradation, Phenylalanine and Tyrosine Metabolism, and Gluconeogenesis (Supplementary Fig. S10A-B). These function analyses together with differential composition analyses indicated the potential disturbances in amino acid and lipid metabolism of SCZ+MetS.

Potential of gut microbiota and plasma metabolites for group classification

We then constructed the RF prediction model using the 13 gut microbial genera and 33 plasma metabolites (amino acids, fatty acids, phospholipids, and related derivates were retained from GC-MS and LC-MS, Supplementary Table 9) screened by the consistent-direction filtering strategy (Fig. 5A). The model showed good prediction ability across the three groups. The AUC values were 0.975 for SCZ+MetS (CI [0.926, 0.995]), 0.853 for SCZ alone (CI [0.751, 0.924]), and 0.976 for controls (CI [0.907, 0.995]).

Fig. 5. Potential pathways that gut microbiota and plasma metabolites involved in the co-occurrence of schizophrenia and metabolic syndrome.

A Random forest prediction model based on 13 gut microbial genera and 33 plasma metabolites, including amino acids, fatty acids, phospholipids, and related derivatives. Metabolites detected by both GC-MS and LC-MS were both included. B Potential pathways of the differential metabolites and gut microbiota for the comorbidity of SCZ and MetS. FA fatty acid, TG triglyceride, PC phosphatidylcholine, LPC lysophosphatidylcholine, LPE lysophosphatidylethanolamin.

Discussion

MetS in SCZ patients is generally associated with poorer clinical outcomes. The increasingly recognized role of the gut microbiota in both psychiatric and metabolic disorders suggests that microbial dysbiosis may contribute to the co-occurrence of SCZ and MetS. In this cross-sectional study, we compared gut microbiota composition and plasma metabolite profiles among patients with SCZ+MetS, patients with SCZ alone, and healthy controls. We further examined the associations of gut microbiota and plasma metabolites with the co-existence of SCZ and MetS, and evaluated their combined potential to discriminate among the groups.

At the community level, both gut microbial composition and plasma metabolite profiles differed across the three groups. However, the dispersion of microbial composition was higher in the two SCZ groups than in controls, which indicates the higher heterogeneity of SCZ groups than controls and may have influenced the results of two SCZ groups comparisons with controls. In contrast, no significant difference in dispersion was observed between the SCZ+MetS and SCZ alone groups, suggesting that the observed microbial compositional differences between these two groups are more likely to reflect microbial variation related to MetS rather than within-group heterogeneity.

At the individual gut microbiota level, differential genera including Barnesiella, Duncaniella, Paramuribaculum, Acinetobacter, Alloprevotella, Pseudomonas, Helicobacter, Raoultella, Pantoea, and Muribaculum were identified in both comparisons of SCZ+MetS versus SCZ along and SCZ+MetS versus controls. Among these, Barnesiella, Duncaniella, Acinetobacter, Alloprevotella, and Pseudomonas have previously been linked to metabolic disorders such as type 2 diabetes mellitus [31–36]. The relationship between Barnesiella and schizophrenia, however, has been inconsistent across previous studies [37, 38]. Pantoea and Muribaculum, which have been considered potentially beneficial in metabolic disorders [39, 40], showed contrasting trends in our research, which may highlight the context-dependent changes of microbiota in SCZ+MetS or some compensatory effects. In addition, evidence regarding Alloprevotella and Helicobacter in SCZ patients remains limited, although animal studies have suggested increased higher relative abundance in SCZ-like mice [41, 42]. Paramuribaculum has also been associated with disease progression in mouse models of type 2 diabetes [43].

We also identified several genera that appeared to be more specifically related to SCZ rather than metabolic comorbidity. Abundance of Faecalibacterium, Roseburia, Coprococcus, Dorea, Fusicatenibacter, and Faecalibacillus were lower in the two SCZ groups than in controls, with no difference between SCZ patients with or without comorbid MetS. Previous studies of schizophrenia and metabolic disorders have reported lower levels in Faecalibacterium, Coprococcus, and Fusicatenibacter, whereas findings for Dorea and Roseburia have been less consistent [44–50], which may highlight the complexity of SCZ+MetS rather than the single combination of SCZ and MetS.

Regarding the individual differential metabolites, SCZ+MetS was characterized by higher levels of amino acids and derivatives like BCAA, L-Glutamic acid, 3-methyl-2-oxovaleric acid, N-Lactoylleucine, and N-Lactoylphenylalanine compared with SCZ alone and controls. Co-occurrence metabolite clusters containing these compounds were also significantly associated with group status and metabolic clinical indicators. On the other hand, the levels of multiple phospholipids and fatty acids, including PC (18:2(2E,4E)/0:0), LysoPC (18:2(9Z,12Z)/0:0), and 12Z,16-heptadecadienoic acid, were lower in SCZ+MetS patients compared with SCZ alone and controls, and the corresponding metabolite clusters also showed some negative correlations with clinical metabolic indices. Previous studies more pinpointed the BCAA had higher levels, and PC and LysoPC had lower levels in individuals with SCZ or metabolic disturbances as well. For example, one study identified two amino acid and organic acid clusters (BCAA, phenylalanine and tyrosine in cluster 1; proline, glutamic, lactic and pyruvic acids in cluster 2) showing higher levels in SCZ patients than in matched controls [51]. Orešič et al. [52] reported SCZ patients with elevated triglycerides had lower levels of LysoPC. Choi et al. [53] summarized that BCAA levels were higher in individuals with obesity and insulin-resistance. Zhong et al. [54] conducted a prospective cohort study and observed that LysoPC was independently associated with decreased type 2 diabetes risk. Bellot et al. [14] summarized that most PC and LysoPC species were negatively correlated with BMI, although some unsaturated species showed positive correlations. Taken together, these findings and our results suggest that abnormalities in amino acid and phospholipid metabolism may be shared features of SCZ and metabolic comorbidities.

Regarding the microbiota-metabolite correlations, we found several differential genera including Barnesiella, Muribaculum, and Raoultella that were significantly correlated with differential amino acids or their derivatives including BCAAs, L-glutamic acid, L-alanine, and N-lactoylleucine, as well as metabolite clusters containing these compounds, while few correlations were observed with lipid metabolites and clusters. Functional prediction of the gut microbiota and pathway enrichment analysis of differential plasma metabolites both highlighted amino acid- and lipid metabolism-related pathways. The combination of differential gut microbiota and plasma metabolites showed good potential for discriminating SCZ+MetS from the other two groups. Previous studies reported gut microbiota may contribute to the changes in plasma metabolites. Liang et al. [55] proposed that SCZ-associated gut dysbiosis could promote protein fermentation and increase amino acid production. Saad et al. [56] summarized that gut microbiota played a role in the development of obesity, which was associated with higher levels of circulating lipopolysaccharide and BCAA. As our study was cross-sectional design, prospective studies are needed to elucidate the directions between gut microbiota and the disturbed plasma metabolites.

Based on the differential and functional analyses, we summarized potential pathways that may be related to the co-occurrence of SCZ and MetS. In particular, the higher BCAA levels observed in SCZ+MetS suggest potentially disturbed BCAA catabolism. BCAAs are converted by branched-chain aminotransferase into branched-chain α-ketoacid (BCKA) and glutamic acid. BCKAs are subsequently metabolized by branched chain α-ketoacid dehydrogenase and downstream enzymes to generate intermediates such as acetyl-CoA and propionyl-CoA that could participate in the tricarboxylic acid (TCA) cycle. Aromatic amino acids such as phenylalanine and tyrosine can contribute to TCA-related intermediates like fumarate, while alanine is closely involved in gluconeogenesis. Fatty acids can also be metabolized to form acetyl-CoA, which fuels TCA cycle activity and influences overall energy metabolism. Additionally, fatty acids, the essential substrates for phospholipid synthesis, can directly affect phospholipid metabolism. Because amino acid, lipid, and carbohydrate metabolism are tightly interconnected, disruption of one pathway may affect the others. Our results suggest that this metabolic balance may be disturbed in SCZ+MetS patients, leading to abnormalities in amino acid metabolism, fatty acid metabolism, and phospholipid homeostasis. Figure 5B visualized these pathways with metabolites highlighted by different colors (Red: higher levels in SCZ+MetS vs. SCZ and controls, blue: in the reverse direction). Regarding the involvement of gut microbiota, our cross-sectional study suggests that gut microbiota was associated with host metabolic disturbances in SCZ+MetS, which need validation of prospective studies.

To our knowledge, this is the first study including SCZ+MetS, SCZ alone, and controls to investigate the associations of gut microbiota and plasma metabolites with comorbid MetS in SCZ. This design allowed us to distinguish gut microbial and plasma metabolic features potentially related to SCZ+MetS from those more closely associated with SCZ itself. The integrated analysis of gut microbiota, plasma metabolomics, and clinical metabolic indices provided a relatively comprehensive view of host-microbiota-metabolism associations. The classifier integrating microbial and metabolic features highlighted their potential value as clinically relevant biomarkers. In addition, the combined use of both GC-MS and UPLC-MS broadened metabolite coverage, although differences in analytical scope and sensitivity between the two measurements may raise inconsistencies such as the correlation analyses between Log2FC values of differential metabolites in the comparisons of SCZ+MetS versus SCZ along and SCZ+MetS versus control from the two measurements.

Our study also has some limitations. First, because of the cross-sectional design, we could not determine the causal relationships among gut microbiota, plasma metabolites, and the co-occurrence of SCZ and MetS. Further prospective follow-up studies are warranted. Second, the use of antipsychotics and medications for metabolic disturbances may have introduced bias, although these factors were considered as potential confounders in the analyses. Third, dietary preferences may also have affected the results. Although the patients were all hospitalized and likely shared relatively similar dietary patterns, and the controls were recruited from the same geographic area, residual dietary confounding cannot be excluded. Finally, the microbiota- and metabolite-based model also requires external validation in larger cohort.

In summary, our findings indicate that patients with SCZ+MetS exhibit distinct gut microbial and plasma metabolic profiles, especially the amino acids, fatty lipids, and phospholipids, compared with patients with SCZ alone and controls. Differential microbial taxa (e.g.: Barnesiella, Muribaculum, and Raoultella) were positively associated with plasma metabolites, particularly the amino acids such as BCAA, L-glutamic acid, and L-alanine. The combination of gut microbiota and plasma metabolites showed good predictive performance for SCZ+MetS, highlighting their potential as clinically relevant biomarkers and providing insights into possible underlying mechanisms and intervention strategies.

Author contributions

LNN and LC were responsible for patient recruitment, data analysis and manuscript draft. DYL and LY were involved in patient recruitment and data collection. JPJ and CJH were in charge of the proposal design and manuscript revision. CJH coordinated the whole project.

Funding

This work was supported by the National Natural Science Foundation of China (82071500), the Program of Shanghai Academic/Technology Research Leader (21XD1423300), Shanghai Science and Technology Innovation Program (21Y11921100), Shanghai Clinical Research Center for Mental Health (19MC1911100), Shanghai Key Laboratory of Psychotic Disorders (13dz2260500), the Programe of Chen Frontier Lab for AI and Mental Health (TCCI) - Shanghai Mental Health Center (SMHC) (2023-TX-018), Open Project Funded by the Shanghai Institute of Traditional Chinese Medicine for Mental Health (SZB2023204), the Natural Science Foundation of Shanghai, China (23ZR1454600), and the Fundamental Research Funds for the Central Universities (YG2025QNB32).

Data availability

The data that support the findings of this study have been deposited into CNGB Sequence Archive (CNSA) of China National GeneBank DataBase (CNGBdb) with the project number of CNP0007407.

Code availability

We used the open source program language R (version 3.4.2) with detailed packages and parameters described in methods and open source website MetaboAnalyst (https://www.metaboanalyst.ca/) to analyze the data and create the plots.

Competing interests

The authors declare no competing interests.

Ethics approval and consent to participate

This study was approved by the Institutional Review Board of Shanghai Mental Health Center and was conducted adhering to the guidelines of the Declaration of Helsinki (2020-82C1). The written informed consent forms were obtained from all participants. Our study was registered in ChiCTR with No. ChiCTR2100052446.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41398-026-04252-7.