This study was a prospective observational cohort study conducted in the neurological intensive care unit (neuroICU) of an urban academic tertiary referral hospital in Guangzhou for 1 year (staged start between February 2017 and February 2018). Patients were recruited based on the following inclusion criteria: (1) diagnosed with encephalitis by specialized physicians according to definitions from a research study published in Lancet Infect Dis (Supplementary Table S1) (2); (2) admitted to the neuroICU with a Glasgow Coma Scale (GCS) < 11; and (3) had an expected length of intensive care unit (ICU) stay (IOS) of >48 h. Disease severity scores were recorded, including the GCS, Acute Physiology and Chronic Health Evaluation-II (APACHE-II), and Sequential Organ Failure Assessment (SOFA) scores at admission. The GCS is a neurological scale which aims to give a reliable and objective way of recording the state of a person’s consciousness. Patients with low GCS scores have worse brain injury. The SOFA score is used to track a person’s status during the stay in an ICU to determine the extent of a person’s organ function or rate of failure. The APACHE-II score is a severity-of-disease classification system, one of several ICU scoring systems. Patients with high levels of SOFA and APACHE-II scores might have worse prognosis. Self-reported healthy volunteers were recruited from the Bureau of Reclamation in Guangzhou between November 2016 and January 2017. The exclusion criteria for all the subjects were as follows: (1) aged less than 18 years old or more than 80 years old; (2) had used antibiotics, prebiotics or probiotics in the last year prior to blood and feces collection; (3) had gastrointestinal disease, (4) had malignant cancer, or (5) were pregnant. Fecal samples and fasting blood samples were obtained from the patients within 72 h after admission and were collected once from individuals in the control group. Written informed consent was obtained from all healthy subjects and patients or their legal representatives. Ethical approval for both the patients and healthy subjects was received from the Medical Ethics Committee of Nanfang Hospital (No. NFEC-2018-034), and all studies were conducted in accordance with the Declaration of Helsinki.

Routine blood samples for biochemical tests were obtained within 24 h of hospital admission. All examinations were strictly performed at the laboratory in the hospital. Data were recorded from the hospital information system, including white blood cell count (WBC, ×10⁹/L), neutrophil count (NEU, ×10⁹/L), red blood cell count (RBC, ×10¹²/L), hemoglobin level (HGB, g/L), platelet count (PLT, ×10⁹/L), alanine aminotransferase level (ALT, U/L), total bilirubin level (Tbil, μmol/L), total protein level (TP, g/L), albumin level (ALB, g/L), serum chlorine level (Cl, mmol/L), serum potassium level (K,mmol/L), blood urea nitrogen level (BUN, mmol/L), serum creatinine level (SCr, μmol/L), C-reactive protein level (CRP, mg/L), procalcitonin level (PCT, ng/mL), D-dimer level (DD, mg/L), brain natriuretic peptide level (BNP, pg/mL), neuron-specific enolase level (NSE, ng/mL), and S100 calcium-binding protein B level (S100B, μg/L). Lumbar punctures were performed in patients for clinical reasons within 72 h, and cerebrospinal fluid was immediately sent to the hospital laboratory for examination. Cerebrospinal fluid albumin (CSFA, mg/L) was subsequently recorded. The cerebrospinal fluid albumin-to-serum albumin ratio (CSAR) was used to evaluate blood-brain barrier permeability, as described previously (16).

Fresh stool samples were stored at −80°C within 3 h after voiding, and 0.2 g of each was aliquoted for DNA extraction. Bacterial DNA was extracted with a magnetic bead-based stool DNA extraction kit (Shenzhen Bioeasy Biotechnology Co., Ltd., China) according to the manufacturer’s instructions (17). Using a LightCycler 480 II real-time fluorescence quantitative PCR system (Roche Diagnostics Ltd., Switzerland), the V4 region of the bacterial 16S rRNA gene was amplified by quantitative real-time polymerase chain reaction (q RT-PCR) with the bar-coded primers V4F (5′-GTGTGYCAGCMGCCGCGGTAA-3′) and V4R (5′-CCGGACTACNVGGGTWTCTAAT-3′). Samples that produced a visible product 290–310 bp in length were used for further experiments. The PCR products were mixed in equimolar ratios and purified by an EZNA Gel Extraction Kit (Omega, United States). Finally, 16S rRNA sequencing was conducted on an Illumina HiSeq 2500 platform, and 250-bp paired-end reads were generated.

Sequences longer than 200 bp were trimmed to 200 bp, and those shorter than 200 bp were removed. Depending on the overlap, we then used SeqPrep to merge the paired-end sequences and assessed the quality of the results using open-source software Quantitative Insights into Microbial Ecology (QIIME, version 1.9.1) (18). The quality of the sequences were checked in QIIME. The sequences with Phred score ≥ Q20 were considered as qualified sequences. Then, we split FASTA files based on the paired-end barcode information, which matched 100% between the barcode and the primer remained more than 200 bp after removal of the barcode and primer. After that, we removed chimeras, performed reference-based operational taxonomic unit (OTU) clustering, and finally generated a BIOM file. All samples were normalized to 7000 sequences to avoid possible errors due to the use of different sequencing depths. The α-diversity (the complexity within a community) was estimated by four indexes and calculated by QIIME (18): (a) Chao1; (b) observed species; (c) Shannon; and d) phylogenetic diversity (PD)-whole tree. The β-diversity (difference between microbial communities) was analyzed using the Bray-Curtis distance and unweighted UniFrac distance (19, 20). To determine the significantly different taxa between two groups, linear discriminant analysis (LDA) coupled with effect size measurement (LEfSe) was performed using an online utility¹ (21). Significantly different bacteria with LDA scores ≥ 3.5 were diagrammed on cladogram. Phylogenetic Investigation of Communities by Reconstruction of Unobserved States (PICRUSt) algorithm was performed in QIIME to predict the functional profiles of the bacterial metagenomes (Kyoto Encyclopedia of Genes and Genomes, KEGG) in the two groups based on the relative abundance of individual OTUs.

Fecal samples for SCFA analysis were frozen at −80°C within 3 h of voiding. Six analytes were targeted for SCFA analysis, namely, acetic acid (Dr. Ehrenstorfer, Germany), propionic acid (Dr. Ehrenstorfer, Germany), butyric acid (Dr. Ehrenstorfer, Germany), isobutyric acid (Supelco, United States), valeric acid (Nu-Chek, United States), and isovaleric acid (Sigma-Aldrich, United States). Feces were homogenized in 1.0 mL of ultrapure water containing an internal standard, 2,2-dimethylbutyric acid (Dr. Ehrenstorfer, Germany). After centrifugation, the supernatant was transferred into a new tube. Then, 10 μL of 50% sulfuric acid and 0.5 g of sodium sulfate (Macklin, China) were added to the tube along with analytically pure diethyl ether (2 mL). The solution was vortexed for 1 min and then centrifuged for 10 min at room temperature. The ether layer was collected for gas chromatography with mass selective detection (5977B GC/MSD, Agilent Technologies, Santa Clara, CA, United States) measurement (Supplementary Table S2). The GC/MS data were acquired and analyzed using MassHunter Workstation software (Agilent Technologies) running on Windows 7 (Microsoft, Redmond, WA, United States). The concentrations of fecal SCFAs were calculated with the use of external standards and are expressed as micromoles per gram of wet feces.

Intestinal permeability was determined as the serum levels of D-lactate (D-LA), intestinal fatty acid-binding protein (iFABP), lipopolysaccharide (LPS), and lipopolysaccharide-binding protein (LBP), as reported before (22, 23).

After centrifuging the blood samples, plasma-EDTA was stored at −80°C until measurement. Plasma samples used for D-LA, iFABP, LPS and LBP quantification were analyzed in duplicate using ELISA kits (Bio-swamp Life Science, Wuhan, Hubei, China) following the manufacturer’s protocols.

The continuous non-parametric data are presented as medians (interquartile ranges, IQRs) and were analyzed using Mann-Whitney U or Wilcoxon tests. The continuous parametric data are presented as the means (standard deviations, SDs) and were analyzed with Student’s t tests. The categorical data are presented as numbers (percentages, %) and were analyzed using chi-squared tests. For microbial analysis, QIIME analysis was additionally performed using the Adonis test as previously described (12). Correlations between variables were determined with Spearman’s rank correlation test. SPSS version 20 (Statistical Package for Social Sciences, Chicago, IL, United States) was used for statistical analysis. Two-tailed p values of<0.05 were considered statistically significant. The figures were generated using GraphPad Prism 7 or R version 3.4.3².

Fecal samples were collected from 28 encephalitis (ENC) patients (median age 46 years; 82.1% male; median GCS 6.5; median SOFA 6.5; median APACHE-II 14.5; median IOS 12.5). The clinical information of all encephalitis patients is shown in Table 1. Until 180 days follow-up, there were 11 patients were alive, whereas 17 patients were deceased. Twenty-eight healthy subjects served as the healthy controls (CON) and had fecal samples collected once. A principal coordinate analysis (PCoA) plot showed a significant difference in β-diversity [Bray-Curtis distance (Figure 1A) and unweighted UniFrac distance (Figure 1B)] between the ENC and CON groups (Adonis test, p > 0.05). The α-diversity, including Shannon, Chao1, PD-whole tree, observed species, and Simpson indexes, did not show significant differences between these two groups (Mann-Whitney U test, p > 0.05) (Figures 1C,D and Supplementary Table S3). As indicated by taxonomic summary (Figures 1E,F) and cladogram based on LEfSe analysis (Figure 1G), the relative abundances of the phyla Proteobacteria, Deferribacteres and Verrucomicrobia were higher in the neuroICU group than in the HC group. At the family level, Enterobacteriaceae, Porphyromonadaceae, Enterococcaceae, Verrucomicrobiaceae, Rikenellaceae and Lactobacillaceae were enriched in the neuroICU group.

To evaluate differences in microbial composition in the feces obtained from patients and controls, we compared the relative abundances in both groups, represented by read percentages (Table 2). The significantly enriched taxa in the patient group were the phylum Proteobacteria, class Bacilli, class Gammaproteobacteria, order Lactobacillales, order Erysipelotrichales, order Enterobacteriales, family Porphyromonadaceae, family Enterobacteriaceae, genus Parabacteroides, and genus Oscillospira. The significantly depleted taxa in the patient group were the class Betaproteobacteria, order Burkholderiales, family Lachnospiraceae, genus Prevotella, genus Faecalibacterium, genus Ruminococcus, and genus Sutterella.

The PICRUSt algorithm was performed to identify which pathway or mechanism is affected, based on microbial change (Supplementary Figure S1). As shown in the results, the pathways upregulated in ENC group including Transport and Catabolism, Immune System Diseases, Folding, Sorting and Degradation, Energy Metabolism, Cancers, Lipid Metabolism, Amino Acid Metabolism, Metabolism of Terpenoids and Polyketides, Enzyme Families, Genetic Information Processing, Signaling Molecules and Interaction, Metabolic Diseases, Excretory System, Transcription, Metabolism, Cellular Processes and Signaling, Metabolism of Other Amino Acids, Carbohydrate Metabolism, Poorly Characterized, Membrane Transport, Neurodegenerative Diseases, Xenobiotics Biodegradation and Metabolism, Signal Transduction, Infectious Diseases.

To identify correlations between fecal microbiota composition and health status, we first examined the correlations among microbial α-diversity indexes (Shannon, PD-whole tree, Chao1, observe species, Simpson) and clinical data (Supplementary Figure S2). The blood-brain barrier permeability is presented as the CSAR, which can reflect the degree of cerebral inflammation. PD-whole tree was positively correlated with the serum concentrations of potassium (r = 0.391, p = 0.040) and S100β (r = 0.394, p = 0.038) but negatively correlated with levels of total bilirubin (r = −0.386, p = 0.042). Observed species was significantly correlated with S100β levels (r = 0.433, p = 0.021). The Shannon, PD-whole tree and observed species indexes had positive correlations with the CSAR (r = 0.468, p = 0.018; r = 0.449, p = 0.024; and r = 0.395, p < 0.05, respectively).

Correlation analysis was subsequently performed among clinical parameters and relative abundances of bacterial groups detected in the feces of encephalitis patients (Figure 2). A positive correlation of the family Ruminococcaceae reads with GCS score was observed (r = 0.384, p = 0.044). The phylum Firmicutes and order Clostridiales were positively associated with IOS (r = −0.387, p = 0.042 and r = −0.383, p = 0.044, respectively). The genus Akkermansia showed negative correlations with IOS (r = 0.404, p = 0.033).

Survival analysis was further performed to explore the association between mortality and microbial diversity (Figure 3). When the cohort was divided into two groups with low bacterial diversity (α-diversity < Median, n = 14) and high diversity (α-diversity > Median, n = 14), there was no intergroup difference in the short-term mortality. However, when the patients were divided into two groups based on the median of observed species, the survival analysis had a trend toward significance [Log Rank p = 0.056, HR = 0.4035, 95%CI = (0.1543, 1.055)].

To evaluate the SCFAs in fecal samples from encephalitis patients and healthy subjects, we quantified the fecal concentrations of acetate, propionate, butyrate, isobutyrate, valerate and isovalerate by GC-MS (Figure 4). The concentrations of acetate, propionate and butyrate were significantly increased in the fecal samples from encephalitis patients (acetate: 41.11 ± 25.71 μmol/g; propionate: 14.44 ± 12.28 μmol/g; butyrate: 4.144 ± 5.509 μmol/g) compared with those in samples from healthy subjects (acetate: 82.64 ± 43.01 μmol/g; propionate: 26.48 ± 18.34 μmol/g; butyrate: 15.84 ± 13.41 μmol/g). Isobutyrate, valerate and isovalerate were nearly undetectable in the vast majority of patient and control samples.

Spearman’s tests were performed to identify correlations between fecal SCFAs and clinical parameters. Results were shown in Supplementary Figure S3. We found that acetate was negatively correlated with age (r = −0.433, p = 0.027), BUN (r = −0.498, p = 0.010) and CRP (r = −0.432, p = 0.028); propionate showed negative correlation with age (r = −0.532, p = 0.005); butyrate was negatively correlated with CRP (r = −0.433, p = 0.027), age (r = −0.534, p = 0.005), CSAR (r = −0.539, p = 0.008) and D-LA (r = −0.390, p = 0.049) while positively correlated with ALB (r = 0.488, p = 0.011).

To evaluate intestinal permeability in encephalitis patients and healthy controls, we quantified the plasma concentrations of D-LA, iFABP, LPS and LBP, which were previously reported as intestinal integrity biomarkers (22–24). The concentrations of D-LA, iFABP, LPS, and LBP were significantly higher in plasma samples from encephalitis patients (D-LA: 6430.2 ± 1056.2 ng/mL; iFABP: 7.779 ± 1.714 ng/mL; LPS: 1218.3 ± 229.9 pg/mL; LBP: 157.9 ± 23.3 ng/mL) than in samples from healthy subjects (D-LA: 3006.6 ± 2123.4 ng/mL; iFABP: 3.813 ± 1.952 ng/mL; LPS: 585.7 ± 297.4 pg/mL; LBP: 73.5 ± 35.8 ng/mL), indicating that intestinal mucosal integrity was significantly reduced during cerebral inflammation (Figure 5).

Spearman’s correlation analyses were further performed to identify correlations among intestinal integrity biomarkers and clinical parameters. Results were shown in Supplementary Figure S4. The iFABP showed positive correlations with LPS (r = 0.586, p = 0.001) and S100B (r = 0.439, p = 0.019); LPS correlated positively with CSAR (r = 0.435, p = 0.030); LBP has negative correlation with TP (r = −0.411, p = 0.030).