This study received approval from the Ethics Committee of Beijing Children’s Hospital, Capital Medical University (Ethical approval number: [2023]-E-156-Y). All CSF samples were obtained from the biobank of Beijing Children’s Hospital. Written informed consent for participation in this study was provided by the participants’ legal guardian.

A total of 104 CSF archived samples from hospitalized patients were analyzed. Patients enrolled at Beijing Children’s Hospital (National Children’s Medical Center) between December 2016 and March 2021, with a clinical suspicion of meningitis who underwent a lumbar puncture (LP) and had CSF samples collected, were included. Patients were categorized into the following groups: TBM, PM, VM, CCM, BD, and Ctrl. Patients with febrile convulsions, traumatic brain injury, and without etiological evidence were excluded. The inclusion criteria for the four types of meningitis required the isolation of a pathogen from CSF or blood culture, or the detection of the virus using PCR. Finally, 7 TBM, 28 PM, 20 VM, 9 CCM, 10 BD, and 30 Ctrls were enrolled in this research. All the patients were negative for HIV infections.

Each CSF sample was collected using a syringe and placed into a polypropylene sample collection tube. The total volume of CSF (approximately 1–2 ml) was retained, then centrifuged at 4°C for 10 minutes at 2,000 × g. Then, the supernatant was removed and stored at -80°C until analysis. Before proteomic analysis, the TBM samples were filtered by 0.22μm sterilizing filter and processed in a CL2 lab with CL3 standards to ensure biosafety. Clinical data were collected retrospectively from the patient’s medical records by clinicians.

Protein digestion was carried out using the filter-aided sample preparation (FASP) method with minor modifications (Wiśniewski et al., 2009). Briefly, 100 μL of each CSF sample was reduced with 20 mM dithiothreitol (DTT) at 95°C for 5 minutes, followed by alkylation with 50 mM iodoacetamide (IAA) at room temperature for 45 minutes in the dark. Next, the samples were transferred to 30-KD ultrafiltration filters and centrifuged at 14,000 × g for 15 minutes at 4°C. Then, the samples were washed three times with 25 mM ammonium bicarbonate (ABC). The treated samples were digested with trypsin (enzyme-to-protein ratio of 1:50) in 25 mM ABC and incubated at 37°C for 14 hours. Finally, the digested peptides were collected after centrifugation at 14,000 g for 20 minutes. The peptide concentration was quantified using the BCA method and loaded with an equal amount for LC-MS/MS analysis.

LC-MS/MS analysis was conducted on an UltiMate 3000 coupled to an Exploris 480 Orbitrap mass spectrometer using an electrospray ion source (all Thermo Fisher Scientific). Purified peptides were separated at 60°C on 50 cm columns with an inner diameter of 75 µm C18 analytical column (Omitech). Mobile phases A and B consisted of 99.9/0.1% water/formic acid (v/v) and 80/20/0.1% acetonitrile/water/formic acid (v/v/v). The flow rate was maintained at 1500 nl/min, with the initial concentration of 5% B being increased linearly to 30% B over 20 minutes, followed by a further increase to 90% within 2.5 minutes, and a 2.5-minute plateau at the end.

MS data were acquired using the data-independent acquisition (DIA) scan mode for single-shot patient samples. The DIA method employed a variable isolation window of 60 windows for acquisition ( Supplementary Table S1 ). A full MS scan was performed within the m/z range of 350-1,200 at a resolution of 120,000, while the DIA scan was acquired at a resolution of 30,000. The automatic gain control (AGC) target was set to a custom value. The maximum injection time was 50 ms. The high-energy collision dissociation (HCD) energy was set to 30%.

Raw data generated by MS underwent quantitative analysis using Spectronaut Pulsar (v18.6, Biognosys AG) with a direct DIA analysis strategy. The database searched was the SwissProt human database (released in March 2024, containing 20,433 sequences). Searches utilized carbamidomethylation as a fixed modification, acetylation of the protein N-terminus as a variable modification, and oxidation of methionines as another variable modification. The Trypsin/P proteolytic cleavage rule permitted a maximum of 2 missed cleavages. The FDR for PSM, peptides, and protein groups was set at 0.01. DIA analysis was performed with parameters set as follows: the precision iRT was calibrated based on local non-linear regression; the Q-value cutoff for precursors and proteins was 0.01; the protein label-free quantification (LFQ) method was MaxLFQ; cross-run normalization was carried out using a local normalization strategy to adjust for the systematic variance of LC-MS performance (Callister et al., 2006); the protein inference algorithm was IDPicker (Zhang et al., 2007); and the quantification MS level was MS1.

The pre-processing of the proteomic data was performed using the wkomics (https://www.omicsolution.com/wkomics/main/) analysis platform. Proteins with a missing ratio ≥ 40% across all three groups (Meningitis, BD, and Ctrl) were removed. The remaining proteins, which still had a missing ratio of ≥ 40% in each group, were imputed with the minimum value for that group. Proteins with a missing ratio < 40% in each group were imputed by the Sequential K-Nearest Neighbors (SeqKNN) method (Kim et al., 2004). The imputation method was validated by NAguiderR to evaluate method priority (Wang et al., 2020).

The abundance of a protein in each sample was normalized to its median abundance among all proteins in that sample to get the relative protein abundance. Then, the data matrices were automatically scaled as z-score values for subsequent statistical analysis. The pairwise DEPs between TBM, PM, VM, CCM, or BD, and Ctrl were defined by adjusting age and sex as confounding factors by performing a linear modeling limma analysis. Differential expressed proteins (DEPs) were defined as fold change ≥2 or ≤0.5, and along with a Benjamini-Hochberg adjusted p-value <0.05 (Benjamini and Hochberg, 1995). DEPs between TBM and VM were corrected using the Benjamini–Hochberg method. DEPs between TBM and PM, and TBM and CCM, were not corrected. GraphPad Prism (version 9) was used to create scatter plots, box plots, violin plots, and columns for data visualization.

To generate CSF protein biomarker panels that distinguish TBM patients from other CNS infections (discriminative model, including PM, VM, and CCM) as well as from non-CNS infection patients (diagnostic model), we employed a multilevel screening methodology described in a previous study (Sun et al., 2022) with minor modifications. Firstly, Spearman’s rank correlation was used to assess the intercorrelation among proteins, excluding proteins with a moderately high correlation with more than four or five other proteins (rho ≥ 0.6). Secondly, we utilized proteins with low expression similarity to determine their importance in distinguishing between the two classes in the diagnostic and discriminative models, employing the random Forest R package. To minimize variability, we computed 100 random forests, each consisting of 150,000 trees, to generate averaged mean decrease accuracy values for each protein. The mean decrease accuracy values were averaged across the 100 random forest replicates for each protein. Thirdly, ROC analysis was performed to evaluate the diagnostic performance of the candidate biomarkers. Then, the permutation test was conducted with 1,000 iterations as the performance measure for ROC analysis to select the robust DEPs. Finally, the completeness of each protein in either of the two classification groups was evaluated; only proteins identified in every TBM patient without any imputation and with a completeness of more than 90% completeness in the other groups were retained ( Supplementary Figure S1 ).

The biomarker panel evaluation methodology includes internal cross-validation and PCA analysis to evaluate the classification capability of panel proteins. Additionally, post-hoc power analyses for Statistical power were calculated for every biomarker in classification models to assess the differentiation capability at the current sample sizes. The receiver operating characteristic (ROC) analysis was performed using the “Biomarker Discovery” module of the MetaboAnalyst 6.0 platform.

The Ingenuine Pathway Analysis (IPA) (Krämer et al., 2014) software was used to enrich DEPs to signaling pathways. Log2(FC) of DEPs was used as the observation value for IPA analysis. The p-value of IPA analysis was calculated with the right-tailed Fisher’s exact test and was considered significant if less than 0.05. Protein-protein interaction (PPI) network analysis was conducted utilizing the STRING database (https://string-db.org/) and was visualized through Cytoscape (V.3.8.2) (Franz et al., 2016). The heatmap was rendered using TBtools. Pattern recognition analysis (PCA and OPLS-DA) was administered using SIMCA 14.1 (Umetrics, Sweden) software. Visualization of the hypothetical model was executed utilizing the BioRender platform (https://www.biorender.com).

We classified the 310 cytokines into six types based on the IMMPORT database (Updated: July 2020) (https://www.immport.org/shared/home). The statistical significance of cytokines was determined as they were DEPs in CSF proteomic data. According to a previously published research article (Mo et al., 2024), we matched the association between the 310 cytokines and immune cells. Fifty-three cytokines from our data were involved in the function of multiple immune cells and are highlighted. The shinyCircos (Yu et al., 2018) was used to visualize the proteomic data.

to systematically identify the characteristic proteomic patterns associated with TBM, we carried out a retrospective DIA-MS-based proteomics workflow ( Figure 1A ). We recruited a total of 104 individuals, who were classified into six groups for DIA-MS quantitative proteomic analysis: TBM (n=7), purulent meningitis (PM, n=28), viral meningitis (VM, n=20), cryptococcus neoformans meningitis (CCM, n=9), non-infectious brain diseases (BD, n=10), and non-CNS infection controls (Ctrl, n=30). The detailed clinical information for these six groups is shown in Supplementary Table S2 . There were no statistically significant differences in gender (P = 0.44) between these groups. The age distributions were balanced among the TBM, PM, VM, and CM groups, but the Ctrl group samples were younger. The clinical indicators, such as the protein concentration, were higher in TBM than in VM, while the concentrations of glucose and chloride were lower than those in VM, which may reflect the severity differences between TBM and VM ( Table 1 ; Supplementary Figure S2 ).

Proteins differentially expressed in TBM, PM, VM, and CCM were identified by comparing them to Ctrl and excluding proteins that changed compared to BD. We then used the IPA database for functional annotation to identify key functional terms altered in TBM. Cytokine analysis revealed changes in cytokines and receptors in TBM. Ultimately, we developed diagnostic biomarker panels and evaluated them using receiver operating characteristic (ROC) curves to distinguish TBM from PM, VM, CCM, and Ctrl ( Figure 1A ).

We conducted proteome profiling of CSF samples using a high-throughput DIA-MS-based proteomics strategy. By applying this workflow, a total of 2,781 proteins were identified ( Supplementary Table S3 ). The average number of protein groups detected for Ctrl, BD, TBM, PM, VM, and CCM groups was 1284, 1405, 1291, 1548, 1503, and 1293, respectively ( Figure 1B ). In all the samples, the quantitative protein intensities across the six groups spanned over six orders of magnitude, and the top ten most abundant proteins accounted for 45.3% of all CSF protein abundance in our datasets ( Figure 1C ). Among the data acquired by DIA, there were 325 proteins with 100% completeness, 970 proteins with 75% completeness, and 1,343 proteins with 50% completeness ( Figure 1D ).

To estimate system stability throughout the entire analysis process, the pooled peptides from all samples were used as a QC to observe the stability of the instrument signal. During the analysis, the QC was analyzed before and after all samples and between every 8–10 samples, resulting in 10 QC samples in our research. To avoid system errors, samples were analyzed in random order, and different groups of samples were interleaved in analysis. The median coefficients of variation (CVs) and correlation coefficients of the QC samples were 0.18 and 0.99 ( Supplementary Figure S3A, B ), demonstrating the consistent stability of the mass spectrometry platform. The PCA analysis of the 10 batches indicated that there was no batch effect in this study ( Supplementary Figure S3C ).

A total of 1429 CSF proteins were retained for further analysis after imputation. Unsupervised principal component analysis (PCA) was conducted among the three groups to visualize the differences in CSF protein profiling among the Ctrl, BD, and meningitis patients. The results indicated a clear separation tendency between the Ctrl and meningitis groups through PCA analysis ( Figure 1E ; Supplementary Figure S4A ). Despite the unclear separation between the four types of meningitis ( Figure 1F ), the orthogonal partial least squares discriminant analysis (OPLS-DA) mode revealed apparent differences between the four groups ( Supplementary Figure S3C ). One hundred permutation tests confirmed no overfitting of the models ( Supplementary Figures S4B , S4D ). Additionally, we assessed the heterogeneity of patients across six groups. The coefficient of variation (CV) values for PM and VM were higher than those of the other groups ( Supplementary Figure S4E ), aligning with the correlation of identified proteins among the six groups and indicating high variability within patients, especially in the PM and VM groups ( Supplementary Figure S4F ). Notably, the TBM and CCM heterogeneity was lower than that of the other two meningitides. The above results demonstrate the potential for discrimination against TBM across different groups.

We used linear modeling with limma analysis, adjusting for age and sex as confounding factors, to identify DEPs between TBM vs. Ctrl, PM vs. Ctrl, VM vs. Ctrl, CCM vs. Ctrl, and BD vs. Ctrl. DEPs were defined as fold change ≥2 or ≤0.5, and along with a Benjamini-Hochberg adjusted p-value <0.05. Consequently, a total of 712, 757, 189, 538, and 206 proteins were identified ( Figures 2A–E ), respectively. The changed CSF proteins between BD vs. Ctrl were further excluded to avoid bias. Finally, we selected 512, 558, 84, and 333 DEPs in TBM, PM, VM, and CCM, respectively, for the subsequent cross-comparison analysis ( Supplementary Table S4 ).

To evaluate differences in the host response between the sample groups, we plotted the differential proteins in scatter plots, with the corresponding log2 fold change for TBM/CTRL on the x-axis and the log2 fold change for PM/CTRL, VM/CTRL, and CCM/CTRL on the y-axis. The linear regression analysis yielded correlation coefficients of r = 0.76, r = 0.60, and r = 0.68 for the comparisons of TBM and PM, TBM and VM, and TBM and CCM, respectively ( Figures 2F–H ). This cross-comparison reveals that TBM and PM or CCM evoke a more similar response compared to VM.

We then annotated the canonical pathway of the 512, 558, 84, and 333 proteins in TBM, PM, VM, and CCM patients using the IPA software to compare the altered functional terms in these four CNS infections and identify the characteristic functional terms associated with TBM. As a result, we identified that the complement cascade, coagulation system, complement system, leukocyte extravasation signaling, and acute phase response signaling were prominent due to their strong association and activation levels in TBM patients. In addition, we observed that pathways such as natural killer cell signaling were enriched exclusively in TBM patients ( Supplementary Figure S4A ).

We annotated 16 representative pathways enriched in TBM as they stood out for their association or activation levels, ranking the top 23 regulated proteins by their frequency of enrollment ( Figure 3A ). Notably, collagen proteins COL1A2, COL1A1, and COL6A1 emerged as the most significantly dysregulated proteins. The collagens compose the organized scaffold of the extracellular matrix (ECM) and play crucial roles in the proliferation and differentiation of neuronal progenitors, as well as in dendritic and axonal growth and guidance (Jovanov Milošević et al., 2014; Long and Huttner, 2019). We then conducted a PPI analysis among the regulated proteins involved in the 16 pathways to identify proteins that interact with collagen. We noticed that the COL6A1 interacted with the 72 kDa type IV collagenase (MMP2) and COL1A1. The COL1A1 can interact with vascular cell adhesion protein 1 (VCAM-1) and fibrinogen beta chain (FGB) ( Figure 3B ; Supplementary Figure S4B ).

The M. tb may bypass barriers via infected macrophages and neutrophils (Davis et al., 2019). Protein VCAM-1 and intercellular adhesion molecule 1 (ICAM-1) are reported to be associated with the inflammation leukocyte extravasation process (Mo et al., 2024), and are upregulated in our proteomic data in TBM patients ( Figure 3C ). In addition, it has been reported that inflammation may initiate coagulation (Esmon, 2005), while in our proteomic data, we identified the upregulated FGB and F2 were enriched in the formation of fibrin clots (clotting cascade). The coagulation factors were reported to interact with the ECM and function as ECM components (Fu et al., 2024). Our proteomic data indicated that the upregulation of coagulation factors, along with the significant downregulation of collagens, may reflect an imbalanced ECM organization. Because the NCAM signaling for neurite outgrowth showed an inhibitory trend (z score = -0.63), and the ECM was reported to play an essential role in neurite outgrowth (Sun et al., 2022), we hypothesized that COL6A1, which is enriched in the NCAM signaling for neurite outgrowth, warrants further functional validation.

We further clustered 22 differential proteins involved in the complement cascade and complement system, as they emerged as the most dominant pathways when compared to PM and CCM ( Supplementary Figure S4C ). The top 10 key hub interactive proteins were analyzed. Notably, CFI, CFH, C8B, and C8A stood out due to their higher fold changes compared to PM and CCM ( Figure 3D ).

Overall, the proteomic data indicate widespread association of the complement cascade, the formation of fibrin clot, the leukocyte extravasation signaling, and the collagen degradation process. The altered proteins in these processes may contribute to the TBM pathogenesis.

Based on the selection criteria in the Methods, we conducted a multilevel analysis to define CSF protein signatures to distinguish TBM from PM, VM, CCM, and the non-CNS-infection cohort (Ctrl). First, we selected 15, 17, 11, and 74 DEPs as candidates for feature selection to differentiate TBM vs. PM, TBM vs. VM, TBM vs. CCM, and with the Ctrl group ( Supplementary Figure S1 ). To obtain complementary biomarker combinations, we evaluated these proteins using Spearman’s rank correlation. For the differentiation with PM, we excluded 1 protein that had a moderately high correlation with more than four other proteins (rho ≥ 0.6), and the 14 remaining proteins with less interdependency (median correlation coefficient of 0.09) were chosen for subsequent analyses. Similarly, we retained 13 proteins for differentiation with VM, 11 proteins for differentiation with CCM, and 30 proteins for differentiation with Ctrl ( Supplementary Figure S6 ).

Next, we evaluated these proteins as input variables and identified the most important features in the discriminative and diagnostic models using the random forest algorithm ( Supplementary Figure S7 ). Meanwhile, we measured the classification performance of each protein in the discriminative and diagnostic stratification using ROC analysis ( Supplementary Figure S8A ). The permutation test was conducted with 1,000 iterations as the performance measure for ROC analysis to select the robust DEPs. The ratio that was larger than the original AUC values was calculated as a P-value. Only Proteins with a P-value < 0.05 were retained as candidates for panel construction ( Supplementary Figure S8B ). Furthermore, the completeness of each protein in its corresponding group was also evaluated ( Supplementary Figure S9 ). Finally, a CSF protein signature for discrimination with PM consisted of F2 and TRMP; for VM, it included ENPP2 and WARS1; for CCM, it comprised F12, APOM and CD163; and a classifier for TBM diagnosis consisted of HLA-B and MGAT1 was generated, respectively.

The panel for classifying TBM and PM achieved an AUC of 0.874 (95% CI, 0.748-0.999) with a specificity of 0.882(95% CI, 0.729-1.0) and a sensitivity of 0.929(95% CI, 0.929-1), as determined by 10-fold cross-validation ( Figure 4A ). Similarly, the panel for classifying TBM and VM achieved an AUC of 0.928 (95% CI, 0.827-1) with a specificity of 0.882(95% CI, 0.729-1.0) and a sensitivity of 0.950(95% CI, 0.950-1), as determined by 10-fold cross-validation ( Figure 4D ). We used the linear SVM algorithm to evaluate the panel classifying TBM and CCM, which achieved an AUC of 0.993 (95% CI, 0.869-1) with a specificity of 1(95% CI, 0.59-1.0) and a sensitivity of 1(95% CI, 0.66-1) ( Figure 4G ). Finally, the TBM diagnostic panel achieved an AUC of 0.934 (95% CI, 0.825-1) with a specificity of 0.941(95% CI, 0.829-1.0) and a sensitivity of 0.933(95% CI, 0.933-1) ( Figure 4J ).

The results of PCA of these biomarker combinations showed clustering of different groups ( Figures 4B, E, H, K ). The normalized expression of each biomarker in different groups is shown in Figures 4C, F, I, L . Biomarkers distinguishing TBM from PM showed statistical powers ranging from 0.74 (CBR1) to 0.97 (F2), with values of 0.99 (WARS1) and 0.9 (ENPP2) in differentiating TBM from VM, values of 0.97 (APOM), 0.92 (F12), and 0.76 (CD163) in distinguishing TBM from CCM, and values of 0.99 (HLA-B) and 0.97 (MGAT1) in differentiating TBM from Ctrl. This result indicates that most proteins effectively differentiate patient groups at current sample sizes ( Supplementary Table S5 ).

The activated microglia that secrete cytokines may have caused excessive inflammation, contributing to poor outcomes of TBM (Barnacle et al., 2023). In this study, we identified 310 cytokines and their receptors in the CSF. They were categorized into six types: chemokines, interferons, interleukins, transforming growth factor-β (TGF-β) family, tumor necrosis factor (TNF) family, and other cytokines ( Figure 5A , track 1).

We identified 120, 115, 12, and 68 significantly dysregulated cytokines and receptors from TBM, PM, VM, and CCM, respectively, when compared with the Ctrl group ( Figure 5A , track 2, 4, 6, 8), totaling 155 significantly dysregulated cytokines and receptors. These modulated cytokines and receptors were enriched for the Complement and coagulation cascades and MAPK signaling pathway ( Supplementary Figure S10 ). Most cytokines and receptors in CSF were upregulated (i.e., 91 of 120, 75.8% in TBM; 81 of 115, 70.4% in PM; 7 of 12, 58.3% in VM; 44 of 68, 64.7% in CCM) in meningitis patients compared to Ctrls ( Figure 5A , track 3, 5, 7, 9).

Cytokines produced by immune cells mediate diverse immune processes. In our data, 53 significantly dysregulated cytokines were involved in the functions of multiple immune cell types ( Figure 1A , track 10), including 41 in TBM patients, as described in the Materials and Methods. We then focused on the distinct dysregulated cytokines and receptors in TBM. Notably, we identified that A2M and ADA2 exhibited higher levels in TBM than in PM, VM, and CCM when compared to Ctrl, whereas PDGFB displayed the lowest levels in TBM; all were involved in macrophage function ( Figures 5B–D ).