Integrated alterations in inflammation, endothelial function, and extracellular matrix pathways in first-episode drug-naïve schizophrenia

Abstract

Objective peripheral biomarkers for early-stage schizophrenia are needed to improve diagnostic accuracy and treatment planning. To identify potential biomarkers, we compared multiple serum factor concentrations between 90 first-episode drug-naïve patients and healthy matched controls, and further examined associations with symptom severity and cognitive functions among patients. Serum interleukin (IL)-8, vascular cell adhesion molecule (VCAM)-1, matrix metalloproteinase (MMP)-2, and MMP-7 concentrations were quantified by Luminex multiplex immunoassays, and log10-transformed values compared with adjustment for age, sex, years of education, body mass index, and smoking status. Associations with symptoms as assessed by the Positive and Negative Syndrome Scale (PANSS) and cognitive functions as assessed by the Repeatable Battery for the Assessment of Neuropsychological Status (RBANS) were examined by Pearson’s correlation analysis. Multivariable logistic regression models were developed for case-control discrimination by receiver operating characteristic analysis with 10-fold cross-validation and calibration. Serum log10[IL-8] and log10[MMP-7] were higher while serum log10[MMP-2] and log10[VCAM-1] were lower in patients. Serum log10[MMP-2], log10[MMP-7], and log10[VCAM-1] were positively correlated among patients, whereas log10[IL-8] was not associated with this MMP/VCAM-1 axis. There were no stable linear associations with PANSS or RBANS scores. Serum log10[IL-8] demonstrated the highest single-marker discrimination (AUC = 0.742, 95%CI: 0.667–0.817), and the four-biomarker model further improved discrimination (AUC = 0.839, 95%CI: 0.781–0.897; 10-fold cross-validated AUC = 0.804). A Youden-derived threshold of 0.518 yielded 78.9% sensitivity and specificity for distinguishing cases. A serum inflammation-endothelium-extracellular matrix biomarker panel showed good discriminative performance in distinguishing first-episode drug-naïve schizophrenia from controls in a case-control sample; external validation and potential recalibration in real-world cohorts are warranted.

Introduction

Schizophrenia (SCZ) is a severe mental disorder with complex etiology, and early diagnosis and intervention are crucial for improving prognosis¹. However, current clinical diagnosis relies primarily on symptomatic assessment², which is susceptible to the influences of phenotypic heterogeneity and disease state fluctuations, leading to misdiagnosis or delayed diagnosis³. Despite progressive advances in treatment, the lack of readily accessible (non-invasive) and objective biomarkers reflecting disease-related biological processes still limits the precise diagnosis and treatment of SCZ^4,5.

Recent studies suggest that immune-inflammatory imbalance and associated vascular and blood-brain barrier (BBB) dysfunction may be involved in the pathophysiology of SCZ^6,7. Associated peripheral responses can manifest as alterations in the serum concentrations of inflammatory mediators; for instance, increased serum concentrations of the chemokine interleukin-8 (IL-8) may reflect the recruitment of inflammatory cells and immune activation status. Concurrently, peripheral inflammatory status may induce vascular endothelial cell activation and phenotypic transitions⁸. Vascular cell adhesion molecule-1 (VCAM-1), which participates in leukocyte adhesion and transendothelial migration, is widely measured as an indicator of endothelial activation, and VCAM-1 signaling is regarded as an important bridge for the transmission of peripheral inflammation to the central nervous system (CNS)^9,10. In addition to inflammatory and endothelial factors, the matrix metalloproteinases (MMPs), particularly MMP-2 and MMP-7, are key regulators of extracellular matrix remodeling, synaptic pruning, and BBB permeability^11,12. Importantly, these processes are conceptualized as functionally interconnected rather than independent. Therefore, jointly characterizing peripheral changes in serum IL-8, VCAM-1, MMP-2, and MMP-7 concentrations as a composite marker reflecting “inflammation, endothelial function, extracellular matrix (ECM) remodeling” may provide a more comprehensive understanding of the systemic abnormalities in early-stage SCZ.

Although previous studies have reported serum IL-8, VCAM-1, and MMP abnormalities in SCZ, specific findings have been inconsistent^13–15, potentially due to the inclusion of patients with chronic illness or those receiving antipsychotic treatment, as these medications possess significant anti-inflammatory or immunomodulatory effects¹⁶. Furthermore, a single biomarker is often insufficient to capture the complex biological heterogeneity of SCZ, limiting diagnostic discriminative ability^17,18. Therefore, assessing the discriminative efficacies of multi-marker combined models, while striving to minimize the confounding effects of medication and long-term illness course, may aid in the development of improved diagnostic and treatment regimens.

Based on these considerations, the present study exclusively enrolled only first-episode drug-naïve (FEDN) SCZ patients and healthy controls (HCs) well- matched for multiple demographic factors, and employed the Luminex platform to simultaneously detect the serum concentrations of IL-8, VCAM-1, MMP-2, and MMP-7. The primary objective of this study was to identify differences in serum factor concentrations between the two groups and construct a multi-marker combined diagnostic model. The secondary objective was to conduct exploratory analyses of potential associations between serum biomarker concentrations and both psychiatric symptoms and cognitive functions within the patient group. The model established in this study demonstrated reasonably high sensitivity and specificity for case-control differentiation, suggesting utility as a research-oriented auxiliary diagnostic tool. Its feasibility for translational applications warrants further validation in independent samples closer to real-world clinical differential diagnosis scenarios.

Material and methods

Study design and participants

This case-control study strictly adhered to the ethical principles of the Declaration of Helsinki, and the protocol was approved by the Ethics Committee of the Fourth People’s Hospital of Nantong. All participants (or their legal guardians) provided written informed consent prior to participation.

Ninety FEDN SCZ patients admitted to the Fourth People’s Hospital of Nantong from March 2022 to November 2024 were recruited as the case group. All were established independently by two senior psychiatrists according to Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5) criteria. Inclusion criteria were (1) Han Chinese ethnicity, (2) 18 to 65 years of age, (3) first-episode phase (duration ≤ 12 months from the onset of first psychotic symptoms), with no lifetime exposure to any antipsychotic medication, and (4) sufficient educational level to understand and comply with the cognitive and clinical assessments.

During the same period, we recruited 90 healthy volunteers from the local community as the control group. Control participants were selected using a matching strategy for age, sex, years of education, and body mass index (BMI). All control subjects were assessed by uniformly trained psychiatrists using the Structured Clinical Interview for DSM-5 (SCID-5) to rule out any current or past mental disorders. In addition, interviews confirmed no family history of psychiatric disorders.

The following exclusion criteria applicable to all participants were designed to minimize interference with inflammatory and endothelial marker measurements: (1) severe somatic comorbidities (e.g., acute infection, autoimmune diseases, uncontrolled cardiovascular or cerebrovascular diseases), (2) history of substance abuse or dependence, (3) pregnancy or lactation, and (4) use of immunomodulators, anti-inflammatory drugs, or corticosteroids within three months prior to enrollment.

Clinical and cognitive assessments

The severity of psychopathological symptoms was assessed by two trained psychiatrists, both receiving appropriate consistency training, using the Positive and Negative Syndrome Scale (PANSS)¹⁹. The PANSS comprises 30 items encompassing three dimensions: positive symptoms, negative symptoms, and general psychopathology. Each item is rated on a 7-point scale, yielding total scores ranging from 30 to 210, with higher scores indicative of greater symptom severity. During the assessment process, the raters were blinded to the participants’ biological test results. In this study, inter-rater reliability was above 0.80.

Cognitive functions were evaluated in both groups using the Repeatable Battery for the Assessment of Neuropsychological Status (RBANS, Chinese version), chosen for its ease of administration, time-efficiency, and the need for only a single tester. The RBANS comprises five index scores, Immediate Memory, Visuospatial/Constructional, Language, Attention, and Delayed Memory, derived from 12 subtests (Word List Learning, Immediate Story Retelling, Figure Copying, Line Orientation, Picture Naming, Verbal Fluency, Digit Span, Coding, Delayed Word List Recall, Word List Recognition, Delayed Story Retelling, and Delayed Figure Retrieval). Participants’ raw scores were converted to population-based scale index scores for analysis, with higher scores indicating better cognitive function. The RBANS has demonstrated good reliability and validity in patients with SCZ and in the general Chinese population²⁰.

Blood sampling and multiplex immunoassay

Venous blood samples were collected from all participants between 07:00 and 09:00 in the morning following overnight fasting. For patients, samples were collected on the day of admission. A 5 mL volume of cubital venous blood was drawn into coagulation-promoting tubes, allowed to stand at room temperature for 30 min to allow natural clotting, and then centrifuged at 3000 rpm for 15 min to separate the serum. The serum was immediately aliquoted and transferred to a −80 °C ultra-low temperature freezer for storage until unified testing. All samples were protected from repeated freeze-thaw cycles prior to assay.

Serum MMP-2, VCAM-1, IL-8, and MMP-7 concentrations were measured using the R&D Systems Human Premixed Multi-Analyte Kit (LXSAHM-04/-12/-18) on the Luminex 200 system (Luminex, Austin, TX, USA), and were expressed in pg/mL. The instrument was calibrated and validated using the Luminex 200 Calibration Kit and Validation Kit prior to testing. Before assays, serum samples were centrifuged at 10,000 rpm for 10 min to obtain the supernatant, and diluted according to panel requirements as follows: MMP-2 (LXSAHM-04) was diluted 50-fold with Calibrator Diluent RD6-52 while VCAM-1 (LXSAHM-12), IL-8, and MMP-7 (LXSAHM-18) were diluted 2-fold with RD6-52. Each well of a multi-well plate was loaded with 50 μL of sample. Samples from patients and controls were analyzed in the same experimental batches under identical conditions and tested in duplicate. Microbead incubation, detection antibody and streptavidin-phycoerythrin (SA-PE) incubation, and plate washing steps before instrument reading were performed strictly according to the manufacturer’s instructions. Standards were assayed in duplicate, and the standard curve was fitted using the software multi-parameter model. The %CV for standards was <20% and the (observed/expected) × 100 ratio was controlled within 70–130%. Assay performance was further evaluated using standard quality control procedures to ensure measurement reliability. Values below the lower limit of detection were handled based on standard curve-derived estimations. Laboratory technicians were blinded to the group allocation of the samples.

Statistical analysis

All data analyses were performed using SPSS 26.0, and figures were generated using GraphPad Prism 10.0 software. Continuous variables were first tested for normality using the Shapiro-Wilk test. Serum factor concentrations formed skewed distributions and so were log10-transformed prior to parametric statistical testing. Outliers were evaluated after log10-transformation using graphical inspection (e.g., boxplots) and standardized residuals; no extreme outliers were identified, and all observations were retained. Demographic and baseline characteristics were compared between case and control groups using the independent samples t tests, Mann–Whitney U test, or chi-square (χ²) test as indicated.

For comparison of serum factor concentrations between groups, multivariate analysis of covariance (MANCOVA) was first employed to assess the overall group effect. Subsequently, univariate analyses of covariance (ANCOVA) were used to compare group differences for each individual factor while controlling for age, sex, years of education, BMI, and smoking status. Assumptions for MANCOVA and ANCOVA, including normality of residuals, homogeneity of variances, and linearity, were assessed using standard diagnostic procedures, and no substantial violations were detected. Independent samples t tests without covariate adjustment and false discovery rate (FDR) correction using the Benjamini-Hochberg (BH) method were conducted as supplementary analyses. Effect sizes were quantified using Cohen’s d.

Associations among serum factor concentrations were first examined within the case group using Pearson’s correlation analysis; subsequently, partial correlation analyses were performed with age, sex, years of education, illness duration, smoking status, and BMI as covariates. The relationships between candidate biomarkers and both clinical symptoms (PANSS subscale scores) and cognitive functions (RBANS total score and domain scores) were examined as exploratory analysis using Pearson’s correlation analysis without correction for multiple testing.

A combined multivariable logistic regression model was constructed with group status as the dependent variable, and log10[IL-8], log10[MMP-2], log10[MMP-7], and log10[VCAM-1] entered as continuous variables (Enter method) to calculate individual predictive probabilities. This model was specified as a biomarker-based predictive model; therefore, demographic and lifestyle covariates were not included, whereas such variables were controlled for in other inferential analyses as described above. Model discrimination was evaluated by receiver operating characteristic (ROC) curve analysis and results expressed as area under the curve (AUC) with 95% confidence intervals (CIs). Internal validation was performed using stratified 10-fold cross-validation. For each fold, the model was refitted using 9/10 of the data as the training set, and out-of-fold predicted probabilities were generated for the remaining 1/10 test set. The cross-validated AUC was calculated by aggregating all out-of-fold predicted probabilities. The optimal cut-off value was determined based on the Youden index from the ROC curve of the full sample, and used to calculate the corresponding sensitivity and specificity for descriptive purposes. Calibration assessment employed apparent calibration. Briefly, a logistic calibration regression model was fitted with group status as the dependent variable and logit of the predicted probabilities as the independent variable, yielding a calibration intercept and calibration slope.

A multiple linear regression model (Enter method) was also constructed with group status (case vs. control), age, sex, BMI, smoking history, and years of education entered simultaneously as covariates for assessment of collinearity and results expressed as the variance inflation factor (VIF). All covariates included in the models were selected a priori based on their potential influence on peripheral biomarker levels. There were no missing data in the dataset; therefore, no imputation or data replacement procedures were applied. All statistical tests were two-sided, and a p < 0.05 was considered statistically significant.

Results

Demographic and clinical characteristics

There were no significant differences in general demographic variables between groups (all p > 0.05), indicating good comparability for subsequent analyses (Table 1).

Table 1.

Demographic and clinical characteristics of the patient and control groups.

	Patients (n = 90)	Controls (n = 90)	χ2/Z/t	p
Age (years)	32 (24.75, 42.25)	33 (22.75, 42.00)	−0.376a	0.707
Sex (male/female)	28/62	29/61	0.026b	0.873
Smoking (yes/no)	5/85	6/84	0.097b	0.756
BMI	21.94 ± 3.39	22.34 ± 2.50	−0.895c	0.372
Education (years)	16 (12, 16)	15 (14, 16)	−0.036a	0.971
Duration of illness (months)	3.5 (0.48, 12)	-	-	-
PANSS-T	85.73 ± 5.86	-	-	-
PANSS-P	21.63 ± 4.49	-	-	-
PANSS-N	19.86 ± 4.90	-	-	-
PANSS-G	44.24 ± 3.58	-	-	-

BMI body mass index, PANSS Positive and Negative Syndrome Scale (T total score, P positive symptom subscale, N negative symptom subscale, G general psychopathology subscale). Two-sided p < 0.05 was considered statistically significant.

^aMann–Whitney U test.

^bχ² test.

^cindependent-samples t tests.

Comparison of RBANS scores between patient and control groups

Multivariate analysis of covariance (MANCOVA) controlling for age, sex, years of education, BMI, and smoking status revealed a significant group difference across RBANS scores (Wilks’ Λ = 0.204, F (6,168) = 109.380, p < 0.001, partial η² = 0.796). Follow-up univariate analyses of covariance (ANCOVA) demonstrated that patients exhibited significantly lower RBANS scores across all domains (all p < 0.05; Table 2), with moderate to very large effect sizes (Cohen’s d ranging from −0.38 to −3.04). Detailed results from independent samples t tests without covariate adjustment, including BH-FDR correction, are provided in Supplementary Table S1.

Table 2.

RBANS total and domain scores in the patient and control groups (ANCOVA).

	Patients (n = 90)	Controls (n = 90)	Fa	p	Cohen’s d
RBANS-IM	82.09 ± 12.52	93.91 ± 13.96	39.024	<0.001	−0.89
RBANS-VS	74.63 ± 12.09	102.76 ± 12.60	267.030	<0.001	−2.28
RBANS-LAN	97.84 ± 6.83	101.14 ± 10.07	6.179	0.014	−0.38
RBANS-ATT	76.08 ± 11.74	114.86 ± 13.71	459.583	<0.001	−3.04
RBANS-DM	71.10 ± 16.69	98.92 ± 10.49	199.463	<0.001	−2.00
RBANS-T	74.80 ± 12.70	102.77 ± 10.90	314.731	<0.001	−2.36

RBANS Repeatable Battery for the Assessment of Neuropsychological Status, IM immediate memory, VS visuospatial/constructional, LAN language, ATT attention, DM delayed memory, T total score.

^aGroup differences were assessed by ANCOVA controlling for age, sex, education, BMI, and smoking status. Cohen’s d values are provided to estimate effect sizes for group differences, complementing the ANCOVA results.

Comparisons of peripheral blood IL-8, VCAM-1, MMP-2, and MMP-7 concentrations between groups

Multivariate analysis of covariance (MANCOVA) controlling for age, sex, years of education, BMI, and smoking status revealed a significant overall difference in log10-transformed serum factor concentrations between the two groups (Pillai’s Trace = 0.333, F (4,170) = 21.221, p < 0.001, partial η² = 0.333). Follow-up univariate analyses of covariance (ANCOVA) demonstrated that patients exhibited significantly higher levels of log10[IL-8] and log10[MMP-7], and lower levels of log10[MMP-2] and log10[VCAM-1] compared with controls (all p ≤ 0.001; Table 3). The corresponding effect sizes were moderate to large, with Cohen’s d values of 0.76 for IL-8, 0.44 for MMP-7, −0.85 for MMP-2, and −0.49 for VCAM-1. Detailed results from independent samples t tests without covariate adjustment, including BH-FDR correction, are provided in Supplementary Table S2.

Table 3.

Comparisons of candidate peripheral blood biomarker concentrations (log10-transformed) between patient and control groups.

	Patients (n = 90)	Controls (n = 90)	Fa	p	Cohen’s d
log10[MMP-2]	5.37 ± 0.09	5.44 ± 0.08	31.595	<0.001	−0.85
log10[MMP-7]	3.45 ± 0.22	3.36 ± 0.17	10.547	0.001	0.44
log10[VCAM-1]	5.80 ± 0.16	5.88 ± 0.18	11.093	0.001	−0.49
log10[IL-8]	0.98 ± 0.31	0.80 ± 0.16	26.495	<0.001	0.76

^aGroup differences were assessed using ANCOVA controlling for age, sex, education, BMI, and smoking status. Cohen’s d values are provided to estimate effect sizes for group differences, complementing the ANCOVA results. Positive values indicate higher levels in patients, whereas negative values indicate lower levels in patients.

Associations among biomarkers

Pairwise Pearson’s correlation analyses among the four candidate biomarkers within the patient group revealed significant and positive associations between log10[MMP-2] and log10[VCAM-1] (r = 0.508, p < 0.001), log10[MMP-2] and log10[MMP-7] (r = 0.347, p = 0.001), and between log10[MMP-7] and log10[VCAM-1] (r = 0.356, p = 0.001). Moreover, these associations remained significant in partial correlation analyses correcting for age, sex, years of education, illness duration, smoking status, and BMI as covariates (log10[MMP-2] and log10[VCAM-1]: r = 0.540, p < 0.001; log10[MMP-2] and log10[MMP-7]: r = 0.355, p = 0.001; log10[MMP-7] and log10[VCAM-1] (r = 0.289, p = 0.008), whereas no significant correlations were observed between log10[IL-8] and the other three candidate biomarkers (all p > 0.05) (Fig. 1).

Fig. 1. Partial correlation analysis between serum candidate biomarkers in patients with SCZ.

Scatter plots showing significant positive correlations between A log10[MMP-2] and log10[VCAM-1], B log10[MMP-2] and log10[MMP-7], and C log10[MMP-7] and log10[VCAM-1] after adjusting for age, sex, education, duration of illness, smoking status, and BMI. Symbols are individual patient values and solid lines represent the linear fit.

Testing for correlations of candidate biomarker concentrations with clinical symptom and cognitive function scores

Within the patient group, further Pearson’s correlation analyses were conducted to examine relationships of the four factor concentrations (log10[MMP-2], log10[MMP-7], log10[VCAM-1], log10[IL-8]) with PANSS scores (total score, positive symptoms subscore, negative symptoms subscore, general psychopathology subscore) and RBANS scores (total score and domain index scores). The results of 40 separate correlation tests revealed no significant associations (all p > 0.05; Table 4). Moreover, the correlation coefficients were generally small (| r | range approximately 0.02–0.19). This exploratory analysis was not corrected for multiple comparisons.

Table 4.

Correlations of candidate serum biomarker concentrations (log10-transformed) with clinical symptom and cognitive function scores among patients with SCZ.

	log10[MMP-2]		log10[MMP-7]		log10[VCAM-1]		log10[IL-8]
	r	p	r	p	r	p	r	p
PANSS
PANSS-T	−0.083	0.435	−0.140	0.189	0.110	0.300	−0.188	0.077
PANSS-P	−0.021	0.845	−0.076	0.477	−0.018	0.869	−0.041	0.702
PANSS-N	−0.170	0.109	−0.055	0.608	0.043	0.689	−0.114	0.284
PANSS-G	0.123	0.249	−0.058	0.586	0.144	0.176	−0.099	0.352
RBANS
RBANS-T	−0.082	0.440	−0.134	0.207	−0.120	0.258	−0.054	0.616
RBANS-IM	−0.047	0.658	−0.128	0.228	−0.102	0.337	−0.076	0.478
RBANS-VS	0.020	0.854	−0.155	0.145	−0.070	0.510	−0.042	0.696
RBANS-LAN	−0.088	0.409	−0.153	0.151	−0.152	0.153	0.052	0.624
RBANS-ATT	−0.059	0.579	−0.086	0.423	−0.058	0.584	−0.055	0.604
RBANS-DM	−0.152	0.154	−0.161	0.130	−0.127	0.231	−0.018	0.869

Associations were evaluated using Pearson’s correlation analysis. The table presents correlation coefficients (r) and p-values.

PANSS Positive and Negative Syndrome Scale (T total score, P positive symptom subscale, N negative symptom subscale, G general psychopathology subscale), RBANS Repeatable Battery for the Assessment of Neuropsychological Status (IM immediate memory, VS visuospatial/constructional, LAN language, ATT attention, DM delayed memory, T total score). These analyses were exploratory and not corrected for multiple comparisons.

Discriminative capacities of these serum biomarkers

The capacities of serum MMP-2, MMP-7, VCAM-1, and IL-8 concentrations, both individually and in combination, to discriminate FEDN patients with SCZ from matched controls were evaluated by receiver operating characteristic (ROC) curve analyses. In single-marker ROC analyses, IL-8 demonstrated the highest discriminative performance (Table 5). A combined model constructed using multivariable logistic regression (Enter method) incorporating all four candidate biomarker concentrations (log10[MMP-2], log10[MMP-7], log10[VCAM-1], log10[IL-8]) demonstrated even better discriminative ability (AUC = 0.839, 95% CI: 0.781–0.897, p < 0.001) (Fig. 2, Table 5). In stratified 10-fold cross-validation, the combined model maintained good discriminative performance as indicated by an average AUC of 0.804. The descriptive threshold value determined by the Youden index for the full-sample ROC curve was 0.518, corresponding to a sensitivity and specificity of 78.9%. In addition, apparent calibration of the combined model was assessed using a logistic calibration regression with the logit of predicted probabilities as the independent variable. The calibration intercept was near 0 and the calibration slope was near 1, suggesting no significant systematic bias in the training sample. Compared to log10[IL-8], which had the highest AUC among single markers (AUC = 0.742), the combined model yielded an AUC of 0.839 (ΔAUC = 0.097). A model including only log10[MMP-2], log10[MMP-7], and log10[VCAM-1] showed lower discriminative performance (AUC = 0.787). The regression coefficients, odds ratios (ORs), and 95% CIs for each predictor in the combined model are presented in Supplementary Table S3.

Table 5.

Efficacies of the four candidate serum biomarkers and a combined model for discriminating patients with FEDN schizophrenia from healthy matched controls.

Variables	AUC (95% CI)	p	Cut-off Value	Sensitivity (%)	Specificity (%)	Youden Index
log10[IL-8]	0.742(0.667–0.817)	<0.001	0.939	65.6	83.3	0.489
log10[MMP-2]	0.714(0.640–0.788)	<0.001	5.342	41.1	88.9	0.300
log10[MMP-7]	0.660(0.580–0.740)	<0.001	3.391	72.2	58.9	0.311
log10[VCAM-1]	0.626(0.545–0.707)	0.003	5.865	62.2	55.6	0.178
3-marker combined model	0.787(0.721–0.852)	<0.001	0.635	55.6	86.7	0.423
4-marker combined model	0.839(0.781–0.897)	<0.001	0.518	78.9	78.9	0.578

AUC, area under the curve; CI, confidence interval; 3-marker combined model includes log10[MMP-2], log10[MMP-7], and log10[VCAM-1]; 4-marker combined model includes log10[IL-8], log10[MMP-2], log10[MMP-7], and log10[VCAM-1]. Cut-off values for single biomarkers correspond to log10-transformed serum concentrations, whereas those for combined models are based on logistic regression-predicted probabilities and are therefore unitless. Cut-off values were determined using the maximum Youden index. Performance estimates are based on the full-sample ROC curve; stratified 10-fold cross-validation was used to assess model robustness.

Fig. 2. Receiver operating characteristic (ROC) curves of log10[IL-8], log10[MMP-2], log10[MMP-7], log10[VCAM-1], and the 4-marker combined model for discriminating between FEDN patients with SCZ and HCs.

The 4-marker combined model demonstrated the largest area under the curve (AUC = 0.839), indicating superior discriminative performance. The diagonal represents the reference line (AUC = 0.5).

Multiple linear regression analysis of factors influencing candidate serum biomarker concentrations

A multiple linear regression (Enter method) model was constructed with the four log10-transformed concentrations as dependent variables and group status, age, sex, BMI, smoking history, and years of education as independent variables. Illness duration was not included in the regression model as it only applies to the patient group and is not comparable in the combined modeling. The results showed that, after controlling for the aforementioned covariates, group status remained independently associated with all four candidate biomarkers (Table 6). Collinearity diagnostics revealed no significant issues (Variance Inflation Factor, VIF < 2 for all independent variables).

Table 6.

Multiple linear regression results (Enter method) for candidate serum biomarkers.

Variables	B (95% CI)	β	p	R2	Adjusted R2
log10[IL-8]	0.189 (0.116, 0.261)	0.358	<0.001	0.172	0.143
log10[MMP-2]	−0.070 (−0.094, −0.045)	−0.386	<0.001	0.194	0.166
log10[MMP-7]	0.091 (0.036, 0.147)	0.229	0.001	0.152	0.122
log10[VCAM-1]	−0.085 (−0.135, −0.034)	−0.246	0.001	0.071	0.039

Multiple linear regression (Enter method) was used to analyze the independent associations between group status and each serum factor concentration (log10[IL-8], log10[MMP-2], log10[MMP-7], log10[VCAM-1]). Each model was adjusted for age, sex, BMI, smoking history, and years of education. Collinearity diagnostics revealed no significant issues (Variance Inflation Factor, VIF < 2 for all independent variables). Group coding: Patient = 1, Control = 0; other categorical variables were set according to SPSS dummy variable coding (reference group = 0).

Discussion

In this study, we identified a consistent pattern of alterations in peripheral serum factors among FEDN SCZ patients, reflecting coordinated changes in inflammatory chemotaxis and extracellular matrix-endothelial pathways. This combined profile remained significantly different from well-matched controls even after controlling for confounding factors. Importantly, as all biomarkers were measured in peripheral blood, these findings do not directly reflect CNS processes and should be interpreted with appropriate caution.

Our finding of significantly elevated serum IL-8 concentrations in FEDN patients with SCZ is consistent with multiple previous studies on first-episode psychosis and acute-phase SCZ patients^21,22. As a key pro-inflammatory chemokine, IL-8 elevation suggests peripheral immune activation in the early stages of the disease and may reflect an inflammatory phenotype associated with innate immunity or neutrophil recruitment. Concurrently, this study also found elevated MMP-7, in contrast to the reduced serum concentrations reported in biological samples from long-term SCZ patients, likely due to the suppressive effect of antipsychotic treatment²³. The current findings suggest that peripheral MMP-7 activity may vary with disease stage and treatment status. In contrast to IL-8 and MMP-7, serum MMP-2 and VCAM-1 concentrations were significantly reduced in the patient group. MMP-2 has been reported to be elevated in the cerebrospinal fluid of patients with SCZ²⁴, which may indicate compartment-specific differences when considered alongside peripheral findings. This pattern may reflect differential regulation between central and peripheral compartments, potentially related to dynamic processes at the BBB. Similarly, systematic reviews and stratified analyses have reported decreases, increases, and stability (no significant changes) in circulating adhesion molecules (including sVCAM-1) among patients with SCZ and related psychotic disorders depending on disease stage and treatment history^25–27. In addition, some studies suggest that levels of adhesion molecules like sVCAM-1 may differ between the periphery and cerebrospinal fluid, potentially reflecting dynamic regulation at the BBB²⁸. Elevated adhesion molecules may promote leukocyte migration and other processes related to BBB integrity²⁹; however, such evidence remains largely indirect, and peripheral measurements cannot be assumed to directly represent BBB function or central neurobiological processes. Therefore, peripheral levels may more likely reflect a specific dimension of the “endothelial barrier/adhesion axis”, which may not necessarily change synchronously with chemotactic inflammatory responses represented by IL-8.

Further partial correlation analyses controlling for multiple potential covariates yielded a consistent pattern of associations among these markers in FEDN patients. Although MMP-2 was reduced and MMP-7 elevated among patients compared to matched controls, they were positively correlated at the individual level, and the concentrations of both were also positively associated with VCAM-1 concentration. This finding is consistent with the overall trend reported in previous studies on early-stage first-episode psychosis samples indicating an “abnormality in peripheral inflammation–vascular/endothelial-related humoral markers”¹⁵. From a mechanistic perspective, this synergistic variation in “matrix remodeling (MMP axis)–endothelial adhesion molecules (VCAM-1)” may reflect a vascular or barrier–inflammation-related tissue remodeling process²⁹, but this interpretation remains speculative given the lack of direct CNS measurements in the present study, although it is broadly consistent with previously reported alterations in MMP-related pathways in SCZ³⁰. However, these correlations do not imply definitive causal or directional relationships, and residual confounding factors cannot be fully excluded despite adjustment for major covariates. In contrast to these correlations among MMP-2, MMP-7, and VCAM-1, no significant correlations were observed between the MMP/VCAM-1 axis and IL-8, suggesting that the inflammatory chemotaxis axis it reflects may act independently in SCZ pathology. This “partially coupled yet partially independent” pattern indicates that IL-8 may provide complementary but non-redundant information to the MMP/VCAM axis, which may contribute to the improved discriminative performance observed in the combined model³¹.

From a systems-level perspective, the improved discriminative performance of the combined model may reflect the integration of partially independent yet biologically complementary pathways. Specifically, IL-8 may primarily capture inflammatory chemotaxis and innate immune activation^21,32,33, whereas MMP-2, MMP-7, and VCAM-1 may reflect coordinated processes involving extracellular matrix remodeling, endothelial activation, and vascular barrier-related regulation^11,34. The combination of these markers may therefore provide a more comprehensive representation of peripheral pathophysiological alterations in early-stage SCZ, which could explain why the multivariate model outperforms individual biomarkers.

Although this study confirmed significant differences in the serum concentrations of four factors between FEDN patients with SCZ and matched controls, correlation analyses revealed no significant linear associations with either PANSS psychopathological symptoms or RBANS cognitive function scores (all p > 0.05). This lack of association may reflect a dissociation between peripheral biological processes and clinical phenotypes³⁵. Specifically, these biomarkers may capture a particular pathophysiological dimension, whereas symptom severity and cognitive impairment are shaped by multiple interacting factors beyond peripheral inflammatory activity^36,37. Accordingly, these peripheral indicators may reflect general etiological processes rather than current disease state and may be more suitable as diagnostic aids than markers of symptom severity. This interpretation is consistent with previous systematic reviews and meta-analyses showing that peripheral inflammatory alterations in SCZ exhibit stage-specificity but substantial cross-sectional heterogeneity, and do not necessarily demonstrate linear relationships with clinical phenotypes^38,39. In addition, the cross-sectional design and the inclusion of acute-phase first-episode patients may introduce range restriction, potentially limiting the ability to detect associations⁴⁰. Therefore, although no significant correlations were observed in this study, longitudinal or pre/post-treatment assessments may be necessary to further evaluate the potential clinical relevance of these biomarkers.

A combined model based on multiple indicators discriminated FEDN patients with SCZ from matched HCs with relatively high specificity and sensitivity (~80%). However, it must be emphasized that the case–control study design may introduce spectrum bias, potentially overestimating model performance in real-world clinical scenarios. While the apparent calibration regression yielded a calibration intercept near zero and slope near one, indicating no significant systematic bias within the sample, this estimate was derived from the same dataset and is potentially subject to optimism bias. Although internal validation using cross-validation was performed, this does not fully eliminate the risk of overfitting. Accordingly, the possibility of model overfitting cannot be fully excluded. External cohort validation is thus required. Therefore, the predicted probabilities from the model should not be directly interpreted as individual disease risk in real clinical settings. Prior to clinical translation, model performance requires validation in independent external samples, and recalibration of the intercept and slope may be necessary to enhance generalizability and predictive stability^41,42. In light of these findings, this model is currently more suitable as a research tool or an auxiliary indicator providing objective biological information to support the comprehensive assessment of suspected cases, rather than as a replacement for the clinical diagnostic process centered on psychopathological evaluation. Future research should target control populations that more closely resemble real-world clinical differential diagnosis scenarios (e.g., affective disorders, substance-related disorders) and conduct external validation in multicenter, independent samples. Furthermore, studies should assess the incremental value of the model relative to routine clinical assessment, identify the appropriate application stages, and determine potential clinical decision thresholds, thereby clarifying the optimal positioning of this biomarker panel within the clinical diagnosis and treatment pathway.

Several limitations of this study should be acknowledged. First, the cross-sectional design precludes any causal inference regarding the relationships between peripheral biomarkers and SCZ. Second, although we included FEDN patients to minimize the potential effects of medication, residual confounding from unmeasured factors cannot be fully excluded. Third, serum samples were used in this study; although widely applied in clinical biomarker research, the coagulation process during sample preparation may influence the levels of certain inflammatory and adhesion-related markers, and caution is warranted when comparing findings with studies using plasma samples. Finally, although the combined biomarker model demonstrated promising discriminative performance, the lack of external validation limits the generalizability of these findings.

Conclusion

This study suggests that peripheral inflammatory chemotaxis (IL-8) and matrix-endothelial pathways (MMP/VCAM axis) jointly constitute a multidimensional biological signature of FEDN SCZ patients, supporting the potential of integrated serum biomarkers for improved case-control discrimination. However, these findings require validation in clinically representative settings. Future studies should assess the robustness of this biomarker panel in multicenter, independent cohorts and determine its incremental predictive value beyond routine clinical measures. Longitudinal designs and external validation will be essential to clarify its utility for early identification and biologically informed stratification of schizophrenia.

Author contributions

Xiaobin Zhang, Chuanwei Li, Qun Yang and Xiaoyu Sun were responsible for study design, statistical analysis and writing of the manuscript. Fei Jiang, Jing Cao and Peijuan Wang were responsible for laboratorial analysis. Sichang Jin, Jiancheng Qiu and Juan Bao were responsible for recruiting the patients, performing the clinical rating and collecting the samples. All authors have contributed to and have approved the final manuscript.

Data availability

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

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41537-026-00766-7.