Enteroviral meningitis-specific transcriptomic profile
We compared the transcriptomic profile of acute enteroviral meningitis samples from 35 patients to their own convalescent sample to identify pathways altered due to the enteroviral infection (EVM1 group). 2380 DEGs (2108 upregulated and 272 downregulated genes) were found between acute and convalescence samples (Additional file 2). For further interpretation, the resulting differentially expressed genes (DEGs) obtained after a differential gene expression analysis were translated to corresponding Gene Ontology (GO) categories with a minimum log2 fold enrichment of 1, thus at least twofold log change (Additional file 3). Interestingly, the GO analysis showed a predominant upregulation of the type I interferon (IFN) signature (Additional file 3) with three type I IFN related GO categories within the top 4 GO terms with a fold enrichment value of 5.09 to 5.60: (1) type I interferon signaling pathway (5.60), (2) cellular response to type I interferon (5.60), and (3) response to type I interferon (5.09). The 10 strongest IFN-related single DEGs are (in decreasing order): IFIT1, IFI44L, RSAD2, OAS3, OASL, MX1, IFIT3, EIF2AK2, IFITM3 and IFI44 (Additional file 2). Furthermore, the GO analysis showed that enteroviral meningitis might affect the negative regulation of viral genome replication, protein targeting to ER, the detection and response to foreign virus particles and many more biological and metabolic processes (Additional file 3).
Bacterial meningitis-specific transcriptomic profile
Six patients in our cohort were diagnosed with bacterial meningitis (BM1), not taking into account the case of neuroborreliosis (BM2)(Table 1). For two of these BM1 patients, we were able to obtain a second convalescent sample. Given that only one significant differential expressed gene, namely the pseudogene FTH1P11, was found between the bacterial convalescent and enteroviral convalescent samples (Additional file 4: Figure S1), we proceeded using two methods: (A) a paired comparison between the two samples for which both a bacterial meningitis and convalescent sample were available and (B) an unpaired comparison between the six bacterial meningitis samples and all convalescent samples (originating from patients with either enteroviral or bacterial meningitis). We only retained the DEGs and GO categories for results that were found in both methods.
Method A: BM1 longitudinal DEG analysis
We found GO categories related to innate immunity including macrophage activation (Top 1 category; fold enrichment of 9.16), and mainly granulocytes activation and degranulation, in especially neutrophil activation (regulation IL-8 = neutrophil-activating factor (NAF) production) with neutrophilic marker CD177 as strongest single DEG. As expected, we noticed the detection of DEGs involved in bacterial lipopeptides and –proteins (and cellular response against it) as well. Furthermore, we found signs of T cell immunity with T cell activation in immune response (2nd GO, 7.04), T cell differentiation (20th GO, 5.25), T cell activation (32nd GO, 4.55) and three more T cell regulation GO terms with a lower-fold enrichment. In addition, we found GO categories with a lower-fold enrichment that are related to general cytokine production and regulation, inflammatory responses and more general (myeloid) leukocyte/lymphocyte related regulation, differentiation and activation. The lowest fold enrichments are reserved for more non-immune general pathways as signaling, metabolic reactions and enzyme/transcription factor activities (Additional file 5 and 6).
Method B: acute BM1 vs. pooled convalescence samples
Top responses in this analysis were similar to those found in the BM1 longitudinal analysis: innate immunity related GO categories such as granulocyte activation and degranulation and the more general (myeloid) leukocyte/lymphocyte categories and the T cell related GOs (Fig. 1). Also, single DEGs were quite similar as for Method A (Additional files 7 and 8).
Summarized trends in both Method A and B analyses
A clear trend could be seen within both comparisons between bacterial and convalescent samples towards an activation of the innate immune system, in particular macrophages and neutrophils, and a sign of T cell activation was noted for patients with bacterial meningitis.
Gene expression differences between enteroviral and bacterial meningitis
680 DEGs (349 downregulated and 331 upregulated genes) could be identified in 47 acute EVM1/EVM2 samples compared to the six BM1 samples (Additional file 9). Based on all measured DEGs, the most significant GO term is the regulation of tumor necrosis factor (TNF) secretion with a fold enrichment of 11.68. This is directly followed by an upregulation of three type I IFN related terms in EVM patients compared to BM patients [response to type I IFN (10.90; 22 DEGs), type I IFN signaling pathway (10.88; 20 DEGs) and cellular response to type I IFN (10.88; 20 DEGs), all DEGs upregulated with positive log2fold changes], which was also found in the longitudinal EVM analysis (Additional files 2 and 10). Furthermore, we noted other viral related GO terms concerning viral life cycle, replication and regulation with a high fold enrichment, as expected (Additional file 10), followed by more general protein localization and targeting GO terms. In the lower GO terms with a fold enrichment below 3.5 neutrophilic and leukocyte responses appear, together with cytokines regulation and production. These lower GO terms were also found in the longitudinal BM analysis (Additional files 5 and 7). However here the associated single DEGs have a negative log2fold change, meaning that they are downregulated in the EVM samples compared to the upregulation in BM samples (Additional file 10). To study the DEGs in more detail, we performed a separate GO analysis on the 331 upregulated and the 349 downregulated DEGs (Additional file 10). As expected, the upregulated DEGs leaded to four immune-related GOs: defense response to virus, regulation of multi-organism process, immune effector process and immune system process. The downregulated DEGs were traced to 12 different GO terms, where only the last two terms were immune-related: inflammatory response and defense response (Additional file 11).
A similar analysis was performed using all enteroviral and bacterial samples (including BM2), which is discussed in Additional file 12: Results.
Enteroviral versus bacterial meningitis classifier
In the last step we used the normalized gene expression values from the EVM1/2 versus BM1 samples to build a random forest classifier that would be able to distinguish enteroviral meningitis cases from acute bacterial cases. After cross-validation, we obtained an AUC value of 0.982 (Fig. 2a), indicating that the classifier is able to discriminate enteroviral from bacterial samples. To determine which genes were indicative of the meningitis type of the sample, the genes that passed the feature selection step (Bonferroni α = 0.05) were extracted from a random forest model that was trained on all the EVM and BM samples. In total, a set of 56 predictive genes were identified in this way (Additional file 13). To assess whether a smaller set of genes could be equally or comparably performant, the feature selection step was made stricter (Bonferroni α = 0.001). This stricter method identified a predictive gene set of 37 genes with an AUC of 0.982 (Fig. 2b and Additional file 13, Additional file 14: Figure S2). Most of the 56 and 37 predictive classifier genes are present as DEG, found in the EVM1/EVM2 versus BM1 analysis (Additional file 13).
In addition, we gathered a set of 41 genes that had previously been implicated with viral versus bacterial infections from a recently published “general” classifier from Herberg et al., 2016 [13]. Using the same leave-one-out cross validation strategy, we tested how well this set of 41 genes was able to predict whether an unknown sample was diagnosed as enteroviral or bacterial meningitis. We obtained an AUC value of 0.979 (Fig. 3), which shows that the two sets of signature genes are equally performant.
Finally, for potential clinical applications, we investigated whether it is possible to specifically identify enteroviral meningitis samples from any infectious sample. To attain this goal, we trained a classifier on all our viral meningitis samples versus all other samples (including background samples from paediatric patients with non-infectious inflammatory conditions (Additional file 1) and patients with other viral meningitis causes) following the same method as described for the enteroviral versus bacterial classifier. We obtained an AUC value of 0.928 (Fig. 4), indicating excellent performance, and identified a set of 61 genes that were predictive of the sample being an enteroviral meningitis sample (Additional file 15). Only five of those classifier genes are not present as DEG in the EVM1 versus control analysis (Additional file 15).