Transcriptional level of inflammation markers associates with short-term brain structural changes in first-episode schizophrenia

In this study, 38 first-episode patients with schizophrenia were recruited during their hospitalization in the Department of Psychiatry, Xijing Hospital, between May 2015 and October 2017 [18] (Table 1). The diagnosis was based on the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition [19]. Patients were included if this was the first time they were in the hospital or sought outpatient clinical help for their mental disorders. Additional inclusion criteria for patients require no more than 2 weeks of cumulative exposure to antipsychotics [20]. The exclusion criteria were listed in the Additional file 1: Supplementary Methods [20, 21]. All patients received second-generation antipsychotic medications determined by the clinicians during the study period, and 36 out of 38 patients received combined rTMS treatment during hospitalization (Additional file 1: Supplementary Methods) [22]. After a period of treatment (mean treatment duration: 148.29 ± 56.20 days), patients received a follow-up MRI scanning and assessment of clinical symptoms and cognitive functions.

Fifty-one healthy controls were recruited from the local community through advertisement and social media, with gender- and age-matched to the patient group (Table 1). The healthy control group received MRI scanning and cognitive assessment at recruitment and follow-up periods (131.51 ± 4.46 days). Written informed consent was signed by all subjects or parents of subjects under 18 years of age, after understanding this study in detail.

Symptom severity of the enrolled patients was evaluated using the Positive and Negative Syndrome Scale (PANSS) at baseline and follow-up time points [23]. PANSS subscale scores were calculated including positive symptoms, negative symptoms, and general psychopathology. We used ΔPANSS to indicate the degree of symptom relief in patients after treatment (T₁) and before treatment (T₀): ∆PANSS = (PANSS_T₁ − PANSS_T₀)/(PANSS_T₀ − 30). The Wechsler Adult Intelligence Scale (WAIS), revised in China [24], was applied to patients and controls to evaluate their cognitive functioning. We performed the WAIS digital span (forward and backward) test and digital symbol coding test on all participants. In addition to the MRI scans, we also collected routine laboratory records of blood samples from the clinical database. Specifically, 2 ml of intravenous blood was collected from the upper forearm in an EDTA vacuum tube under fasting condition after admission, and the samples were measured using a Sysmex XS Automatic Hematology Analyzer. The data on leukocyte count and percentage were extracted from baseline laboratory tests to analyze the inflammatory indicators.

High-resolution T1-weighted MRI was acquired using a GE Discovery MR750 3.0 T scanner located in the Department of Radiology, and all subjects underwent T2WI scans to rule out organic diseases. Participants showing gross artifacts and/or excessive head motion were excluded, following protocols established previously [21]. The MRI acquisition was accomplished at baseline and follow-up, at the same time as the clinical assessment. The scanning parameters were repetition time 8.2 ms, echo time 3.2 ms, flip angle 12°, field of view 256 × 256 mm, matrix 256 × 256, slice thickness 1 mm, and sagittal slices 196. The image preprocessing followed the analysis process of FreeSurfer (version 6.0.0, https://​surfer.​nmr.​mgh.​harvard.​edu) and was performed on the Linux operating system. The brief preprocessing steps included correcting the uniformity of the bias field; removing the skull; segmenting the brain tissue into gray matter, white matter, and cerebrospinal fluid; cortical surface reconstruction; registration of individual space to the standard space by nonlinear transformation; the registered image was segmented into 68 cortices based on Desikan-Killiany atlas and 14 subcortical regions. For each subject, 150 regional structural measurements were obtained, including 68 regional cortical thicknesses, 68 surface areas, and 14 subcortical volumes.

The statistical analysis was performed on MATLAB (R2020b) platform. A linear mixed-effect model was applied to simultaneously test the group differences (group), time point (time), and their interaction effect on global and regional morphological measurements (Y). Age and sex were included as covariates. When comparing regional measurements, we additionally included global measurements to adjust for global differences between individuals. Each subject was given a subject ID (SubID) which was included as the random factor. Multiple testing effect was corrected using false discovery rate (FDR) correction with q < 0.05 for regional measurement comparisons.

Brain transcriptional data were obtained from the Allen Human Brain Atlas (AHBA) [25] (http://​human.​brain-map.​org/​), including 20,734 genes across the brain from six healthy human donors (five males, age range of 24 ~ 57). For each tissue sample, gene expression data represented by 58,692 probes were quantified, normalized, and averaged across probes following a previous study [26]. We used gene expression data from the left hemisphere as they are available from all donors. Tissue samples were mapped to the Desikan-Killiany atlas [27] based on Euclidean distance to the nearest voxel in the MNI 152 template [26]. For each gene, the expression level was averaged across the six donors and normalized to z score across all cortical regions, resulting in a normalized gene expression matrix of size 34 × 20,734 (region × genes) for further analysis. Detailed processing steps are described in Additional file 1: Supplementary Methods [25‐28].

To identify inflammation-related genes, we selected genes from a previous systematic review on leukocytes and subtypes (monocyte, lymphocyte, neutrophil, eosinophil, and basophil) [29]. To reduce potential bias in gene set selection, we further verified our results using inflammation-related genes reported in another review [30]. Genes related to antipsychotic treatment response were selected from a genome-wide association study (GWAS) on antipsychotic treatment in schizophrenia [31]. Included genes are listed in Additional file 1: Table S1.

We next examined the associations between brain structural changes and baseline peripheral inflammation markers in patients. Inflammation markers were quantified including leukocyte count and percentage of leukocyte subtypes (monocyte, lymphocyte, neutrophil, eosinophil, and basophil) (Table 2). As some inflammation markers do not follow a normal distribution (Additional file 1: Table S2), Spearman’s rank correlation was performed between inflammation markers and structural changes in the brain regions that showed a significant group main effect or group by time interaction effect. We further examined the associations between inflammation markers and subcortical volume changes, as some previous studies reported an association between subcortical volume and peripheral inflammatory markers [32], but not consistently [33].

We next investigated whether brain structural changes were associated with gene expression patterns related to inflammation and antipsychotic response (Additional file 1: Table S1). For each set of genes of interest, normalized gene expression was averaged across samples. The normalized gene expression maps were cross-correlated with group-averaged cortical thickness and surface area changes in patients. The same correlation analysis was also performed in controls to examine whether the effects were unique to patients.

As recent studies pointed out that spatial autocorrelation of brain maps results in false-positive effects [34], we verified our correlation results between brain and transcriptional maps using generative null models proposed by Burt et al. [35]. Specifically, 5000 surrogate maps preserving spatial autocorrelation of brain structural changes in patients were generated using the BrainSMASH toolkit [35]. The null distribution of correlations between gene expression and surrogate brain maps was calculated and compared with our empirical results. The two-sided p value was determined as the proportion of random permutations that exceeded the empirical correlation coefficient between gene expression and brain maps.

Spearman’s rank correlation was performed between inflammation markers (leukocyte and leukocyte subtypes) and patients’ behavioral scores (PANSS and WAIS subscales) at baseline and follow-up.

Demographic and clinical information of all participants were shown in Table 1. Controls were matched with patients on age (p = 0.09) and sex (p = 0.37), and controls had a higher education level (p = 0.01) and shorter follow-up interval (p = 0.04) than patients. The patient’s symptom severity at baseline and follow-up period was illustrated in Table 3. The majority of patients (33 out of 38) had an improvement in their symptoms, with PANSS total score reduced by more than 30% (Additional file 1: Fig. S1). On cognitive assessments, patients showed significantly lower scores on WAIS digital symbol coding and digital span (forward and backward) tests compared to controls (all p < 0.001). Patients did not show significant improvements in WAIS scores at follow-up. Information on patients who were medicated before admission is listed in Additional file 1: Table S3.

We first tested whether the brain regions that showed a significant group main effect or group by time interaction effect would correlate with patients’ peripheral inflammation levels. Next, we performed an exploratory analysis in the subcortical regions. We did not observe a significant association between structural changes and inflammation markers.

Next, we investigated whether brain structural changes correlated with the gene expression profile of inflammation and antipsychotic treatment (Additional file 1: Fig. S2). Our analysis revealed a positive correlation between cortical thickness changes and gene expression of monocyte [29] in patients (r = 0.54, p = 8.8 × 10⁻⁴, FDR corrected) but not in healthy controls (r =  − 0.05, p = 0.76) (Fig. 3 and Table 6). This result was replicated using inflammation genes from another study, which showed a positive correlation between cortical thickness changes and gene expression of monocyte in patients (r = 0.51, p = 0.002) [30]. We further tested whether this micro–macro association was driven by spatial autocorrelation of neuroimaging measurements. Our empirical results exceeded the null distribution of correlations between 5000 surrogate maps and monocyte gene expression (two-tail p = 0.001 for [29] and p = 0.006 for [30]). The correlation between cortical thickness changes and antipsychotic genes in patients is r = 0.37, p = 0.03, but did not survive FDR correction [31]. We also conducted a sub-group analysis of patients who were medicated at baseline, which showed a significant correlation with monocyte gene expression level (Additional file 1: Table S5).

For WAIS tests, we found that the left superior parietal lobule thickness change correlated with WAIS digital span (backward) change in patients (r = 0.57, p = 3.4 × 10⁻⁴) but not in controls (r = 0.09, p = 0.54) (Fig. 4). For PANSS score changes and medication dosage, we did not find any significant correlations with brain structural changes.

We found an uncorrected association between monocyte percentage and PANSS positive score (Spearman’s r = 0.44, p = 0.006), as well as between monocyte percentage and digital span (backward) score (Spearman’s r = 0.38, p = 0.02) at baseline (Table 7). Yet, these correlations did not survive multiple corrections.