Network meta-analysis of transcriptome expression changes in different manifestations of dengue virus infection

Background Several studies have been performed to study transcriptome profiles after dengue virus infections with partly different results. Due to slightly different settings of the individual studies, different genes and enriched gene sets are reported in these studies. The main aim of this network meta-analysis was to aggregate a selection of these studies to identify genes and gene sets that are more generally associated with dengue virus infection, i.e. with less dependence on the individual study settings. Methods We performed network meta-analysis by different approaches using publicly available gene expression data of five selected studies from the Gene Expression Omnibus database. The study network includes dengue fever (DF), hemorrhagic fever (DHF), shock syndrome (DSS) patients as well as convalescent and healthy control individuals. After data merging and missing value imputation, study-specific batch effects were removed. Pairwise differential expression analysis and subsequent gene-set enrichment analysis were performed between the five study groups. Furthermore, mutual information networks were derived from the top genes of each group comparison, and the separability between the three patient groups was studied by machine learning models. Results From the 10 possible pairwise group comparisons in the study network, six genes (IFI27, TPX2, CDT1, DTL, KCTD14 and CDCA3) occur with a noticeable frequency among the top listed genes of each comparison. Thus, there is an increased evidence that these genes play a general role in dengue virus infections. IFI27 and TPX2 have also been highlighted in the context of dengue virus infection by other studies. A few of the identified gene sets from the network meta-analysis overlap with findings from the original studies. Mutual information networks yield additional genes for which the observed pairwise correlation is different between the patient groups. Machine learning analysis shows a moderate separability of samples from the DF, DHF and DSS groups (accuracy about 80%). Conclusions Due to an increased sample size, the network meta-analysis could reveal additional genes which are called differentially expressed between the studied groups and that may help to better understand the molecular basis of this disease. Supplementary Information The online version contains supplementary material available at 10.1186/s12864-022-08390-2.


Background
Dengue is a mosquito-borne viral infection mainly distributed throughout the tropical and sub-tropical areas of the globe ( [59]; https:// apps. who. int/ iris/ handle/ 10665/ 44188). Being responsible for around 400 million infections every year it is the most prevalent arboviral disease Open Access *Correspondence: Klaus.jung@tiho-hannover.de 1 Institute for Animal Breeding and Genetics, University of Veterinary Medicine Hannover, Bünteweg 17p, D-30559 Hannover, Germany Full list of author information is available at the end of the article of humans [2], and in recent years, its global incidence has grown dramatically [12]. Dengue is caused by four different dengue virus serotypes (DENV1-4) which are transmitted by infected female Aedes aegypti and Ae. albopictus mosquitoes [12]. According to the current World Health Organisation (WHO) criteria, dengue can be classified into (a) dengue with or without warning signs and (b) severe dengue. The clinical spectrum of dengue infections ranges from an unapparent subclinical picture to two well defined syndromes historically known as dengue fever (DF) and dengue hemorrhagic fever/ dengue shock syndrome (DHF/DSS) [13]. DF is an acute self-limited systemic disease characterized by fever, headache, nausea, vomiting, myalgia, arthralgia, rash, and leukopenia. Individuals that may develop warning signs might present gingival bleeding, lethargy, hepatomegaly, thrombocytopenia and ascites and/or pleural effusion [59]. Ultimately, during the critical phase of infection the clinical picture may progress to a more severe form of disease characterized by capillary permeability and organ impairment [13]. Although virological and epidemiological risk factors contributing to the development of severe dengue have been identified, the molecular mechanisms underlying pathogenesis have yet to be characterized [13]. Nevertheless, it is noteworthy that for such a complex disease in its clinical manifestations the simple therapeutic management is highly effective in saving lives [60], as early detection of progression to the severe form of disease drops fatality rates of severe dengue to below 1% [35,60]. The host transcriptional profile may be correlated with disease severity [15], hence, the characterization of the specific transcriptional patterns associated with each clinical form of disease could not only help anticipating the clinical progression and adequate management but also extend our knowledge on the molecular mechanisms underlying pathogenesis.
Transcriptome expression analyses by means of DNA microarrays [46] or next-generation sequencing (NGS) technology [32] can help to better understand the pathogenesis of virus infections as well as the role of genes and pathways in host responses [11,17].
However, transcriptome expression data has a highdimensional character, i.e. thousands of genes are studied in samples of usually small size, especially in cell line experiments [27]. Different techniques such as p-value correction have been proposed to reduce false positive findings [4], but transcriptome expression studies are often underpowered, increasing the chance of false negative results [22]. Therefore, several approaches for metaanalysis of transcriptome expression studies have been considered in order to achieve increased power and thus increased scientific evidence of findings [31,41,53]. In the context of infectious diseases, such methods have for example been applied for meta-analysis of transcriptome expression data from West Nile virus infected [20] or SARS-CoV2 infected samples [28]. Furthermore, metaanalysis can increase the reproducibility of gene expression analysis [49], and can increase the generalizability of results when data is taken from a variety of sources [10].
While meta-analysis as performed on clinical trials is usually done by two-stage approaches, i.e. results of individual analyses are merged by means of p-value-or effect-size-combination, meta-analysis of transcriptome expression data is often also possible in a single stage approach after merging the raw data of the involved studies. The single stage approach in the latter case is possible due to the free availability of high-dimensional gene expression data in public repositories such as Gene Expression Omnibus (GEO) [9] or ArrayEpress (AE) [6]. When merging gene expression data from independent studies additional steps for batch effect removal are usually necessary [14,25].
In this meta-analysis on gene expression profiles from studies on dengue fever, we use the data merging approach which also allows to make group comparisons that were not performed in the original studies. This is obtained by building a network of study groups from the original data. We have previously shown that so called network meta-analysis on merged data (single-stage analysis) is highly correlated with network meta-analysis by means of two-stage analysis [56]. Network meta-analysis has first been presented in the context of clinical trials [45], and has also been used to mine protein expression databases [57]. Our analysis comprises differential expression analysis as well as gene ontology (GO) term enrichment analysis. We performed differential expression analysis once based on the merged data and once by a rank-based approach [19] on the results from the individual studies. The advantage of the merged data approach is an increased power but loses large numbers of genes in the merging step. The rank-based approach does not profit that much from an increased power but does not lose genes.
Using the top genes identified by both approaches, we derive mutual information networks to study changes in the pairwise correlation between genes, and fit machine learning models to the merged data to study the separability of different patient groups. Finally, we compare our new results with the results reported in the original studies.

Database search and building of the study network
In March 2021, searches in the AE and GEO databases by the keyword 'Dengue' yielded 69 database entries, 31 of which referred to samples from Homo sapiens. Further selection left 29 datasets generated by either RNA-seq of coding RNA or by array-based transcription profiling. Because the samples of the 29 datasets were taken from different types of tissues, we manually screened the publications linked with the database entries. As result, 5 datasets remained where dengue fever was studied in whole blood samples (Table 1). All 5 datasets were generated by array-based transcription profiling. A flow diagram according to recommendations of the PRISMA guidelines for systematic reviews and meta-analyses [38] that depicts the selection process is provided as Supplementary Fig. S1. After outlier removal (see next subsection) 329 individuals remained from all groups.
The resulting study network is illustrated in Fig. 1A, and is comprised of five different group: samples of 1) healthy control individuals, 2) of patients with dengue hemorrhagic fever (DHF), 3) of patients with dengue shock syndrome (DSS), 4) of patients with dengue fever (DF), and 5) of convalescent individuals. Only in the study by Loke et al. [29] data from all five groups was available in the databases. The other studies contributed with subsets of these five groups.
The studies by Long  The female/male ratio was in general well balanced, individual groups in some studies had slightly more males or females.

Data merging, missing values imputation and batch effect removal
In total, a union of 35.695 genes were involved in all five selected studies, with an intersect of only 12.046 that were studied in each of the studies (Fig. 1B). Before data merging, 9.180 genes which occurred in only one of the studies were removed from the data sets. Thus, only genes which occurred in at least two studies were considered for the merged data set which had a final size of 26.515 genes. The merged data matrix contained approximately 22% of empty entries which were filled by missing value imputation. Finally, batch effects, i.e. systematic differences between groups, were removed. The merging, imputation and batch effect removal steps are illustrated in Fig. 2. Note again, that the loss of genes concerns the differential expression analysis based on the merged data, but not the analyses based on rank aggregation.
Using group specific bag plots on the data projected to the space of the first two principal components, 15 outliers were detected in total and subsequently removed (Fig. 3) leading to a slightly reduced sample size. Four outliers were removed from the DHF group, 9 from the DF group and 2 from the DSS group (Table 1). Although, it is a statistically not recommended procedure to remove outliers from a data set just because these observations are extreme, we decided for our network meta-analysis to remove some outliers for several reasons. When fusing data from multiple independent studies there is a higher chance that some individuals may not belong to the particular disease group, and in consequence they do not fit the model well [43,44]. Furthermore, it has been observed that in data mining single abnormal measurements might dominate the calculation of averages [5], which in the case of transcriptomics would lead to wrong estimates of the fold changes.

Overview of results from differential gene expression
In this subsection, we report the results based on the merged data set. Additional results for differential expression analysis based on rank aggregation are reported in the subsequent subsection. As expected, the largest numbers of differentially expressed genes were found in the comparisons of the disease groups versus either the group of normal or versus the group of convalescent patients ( Table 2). Taking only the FDR-adjusted p-values as selection criterion, the number varied between several hundred up to few thousand genes. The union of genes selected in the different group comparisons by an FDRthreshold of 5% were 7387 genes. For these genes, the p-values and log2 fold changes from all 10 pairwise group comparisons are provided as Supplementary Tables T1. When applying additionally a moderate threshold of +/− 1 for the log2 fold change, the numbers reduced clearly. In total, a union of 72 genes were found in all comparisons when using these strict selection criteria. According to these strict criteria (FDR-adjusted p < 0.05 & log2 fold change > +/− 1), no genes were selected when comparing DHF, DSS and DF groups among each other, or normal versus convalescent samples. A clustered heatmap based on these 72 genes and all 329 individuals is presented in Supplementary Fig. S2.
For further analyses (such as the mutual information networks and classifier training) we used the ranking provided by the FDR-adjusted p-values without the log2 fold change as selection criterion. While a p-value comprises information about mean, variance and sample size, the fold change is only constructed using mean expression levels. In fact, a gene with a small fold change can still be significant if the within-group variances are small, while a gene with a large fold change can be nonsignificant if within-group variances are large. Therefore, significant genes can falsely be excluded by using a too strict fold change threshold.

Top noticeable genes from differential expression analysis of the network meta-analysis
Among the different lists of top 10 genes from the two approaches of network meta-analysis (merged data analysis and analysis by rank aggregation), six genes occur more than three times (Table 3): IFI27 (7x), TPX2 (6x), CDT1 (5x), DTL (5x), KCTD14 (5x) and CDCA3 (4x). The high frequency of their occurrence among the top 10 lists provides evidence that these genes could play an important role in the context of DENV infections. IFI27, TPX2, CDT1 and CDCA3 are particularly known for their role in cancer [8,36,58,61], and TPX2 also for its role in the context of infectious diseases, in particular dengue fever [51].
In order to limit the findings not only to the top 10 lists, we also analysed the top 20 and top 50 lists. From all group comparison, 266 individual genes occur among the top 20, where 19 of them occur more than three times. If looking at the top 50, 618 individual genes are detected in total, where 48 of them occur more than 3 times. The additional 48 genes are listed in Supplementary Table T3. Still, the six above mentioned genes are those with the highest frequency among the top 20 and top 50 lists.

Genes with strong correlation changes in mutual information networks
Here, we selectively describe mutual information networks with the largest changes between two groups. For each pair of study groups, we took the genes listed in Table 3, i.e. the top 10 from the merged data approach, and the top 10 from the rank aggregation approach.
The union of these genes then had a size of ≤20 genes.
The changes in the mutual information networks for all group comparisons are provided as Supplementary Figs. S3. While the presented Figure depict the changes in a mutual information network between two group, the group-specific mutual information adjacency matrices are presented as Supplementary Tables T4.
In the mutual information network for DF versus DHF (Fig. 4) the correlation between LOC288086 and ENPP5 is strongly lowered in samples from DHF compared to the correlation in DF. Vice versa, the correlation between LOC153682 and RNF 157-AS1 is increased in DHF compared to DF.
In the network constructed from the comparison of DF versus DSS, the correlation of PSMA6 with NELFE and with DNAJC15 is stronger in DHF than in DF.
An increased correlation between NELFE and DNAJC15 in DSS samples is observed in the network constructed from the comparison of DHF versus DSS. In this network PSMA6 and NME2 lose their correlation in DSS samples compared to their expression in DHF samples. Table 2 also presents the number of significant GO terms from each comparison. Detailed results of the union of 111 selected GO-term analyses are also provided as Supplementary Tables T2. The most GO-terms were identified as significantly enriched in the comparisons with the normal group or with the group of convalescent patients. Again, no significant GO-terms were selected in the comparisons between DHF, DSS and DF group.  Table 3 Top 10 of differentially expressed genes from the pairwise comparisons in the study network. The first three columns result from the analysis of the merged data set, the last two columns results from the rank-based aggregation of differential expression analysis of the individual studies. The latter analysis does not yield p-values or fold changes but a score for each gene

DSS versus DHF Convalescent versus Normal
Among the significant GO terms are the general infection related terms such as "defense response to virus" (GO:0051607), "viral process" (GO:0016032) and "innate immune response" (GO:0045087). The latter two GO terms were also mentioned by the contributing study of Sun et al. [48]. Other overlapping GO terms with the results of Sun et al. are "mitotic spindle organization" (GO:0007052), "cell division" (GO:0051301), "mitotic spindle organization" (GO:0007052), "regulation of translation" (GO:0006417).
Among GO terms identified in the study by Loke et al. [29] was "nucleic acid binding" (GO:0003676). Our GO enrichment analysis identified the more specific terms "DNA binding" (GO:0003677) and "RNA binding" (GO:0005524) which are subcategories of "nucleic acid binding".

Separability of different dengue manifestations by transcriptome signatures
Genes were ranked in each pairwise comparison between DF, DHF and DSS by their p-values. Classifier models were trained using unions of genes from each of the three ordered gene lists. Thus, if for example the top 10 genes of each of the three ranking lists were taken their union was ≤ 30 genes (note, that due to overlaps between three sets, the size of their union can be smaller than the sum of the individual sizes). At maximum, the 500 top ranked genes from each pairwise comparison were involved as predictors in each model. Due to overlaps, the maximum size of a classifier was given by 1203 genes (instead of 3*500 = 1500 genes). The final number of genes in each model is displayed on the x-axis at the bottom of Fig. 5. As can be seen, there, all classifier models yielded an accuracy bigger than 33.3% which would be achieved by a random classifier for three subtypes. The best performance was achieved by the support vector machine with which an accuracy of nearly 80% was obtained when taking approximately 250 genes as signature. Linear discriminant analysis yielded an accuracy of about 75%, however with 750 genes. Finally, nearest shrunken centroids yielded a nearly constant accuracy of about 67%, independently of how many genes were incorporated.
Looking at the individual groups, the DF group has the highest sensitivity. I.e., 117 out of 122 (96%) of DF patients were correctly classified by the support vector machine (with 250 genes) within cross vaidation as DF patients, while 1 DF patients was wrongly classified as DHF patients and 4 wrongly as DSS patient. The sensitivity of DHF was 60%, i.e. 49 out of 81 DHF patients were correctly classified while 31 DHF patients were classified as DF patients and 1 as DSS patient. Finally, 29 out of 45 (64%) of DSS patients were correctly classified the remaining 16 wrongly as DF patients. Thus, DHF and DSS patients tend to be wrongly classified as DF patients, and in consequence DF has a low positive predictive value. In cross validation, only 117 out of 164 (71%) patients classified as DF, where truly DF patients. In contrast, positive predictive values of DHF and DSS were 98% (49/50) and 85% (29/34), respectively. The individual curves of sensitivity, specificity, positive and negative predictive values versus the size of a classifier are provided as Supplementary Figs. S4.

Methodical issues and reproducibility of findings from contributing studies
Performing meta-analyses of multiple independent studies has several benefits, foremost the increased power for statistical analysis. This is of special importance for high-throughput data where thousands of statistical tests bear the risk of false positive and false negative conclusions. In addition, the special character of a network meta-analysis facilitates indirect group comparisons that have not been performed in the individual studies. In the case of this work, each group comparison in the study network was directly performed by at least two of the individual studies. However, by using the data of all studies a particular group comparison can be supported by data from other studies. E.g., the comparison between DSS and convalescent individuals was only carried out in two studies originally. In the network meta-analysis by means of the merged data, all five studies contribute to this particular comparison.
Here, we have used two approaches for network metaanalysis, one based on the merged data and one rankbased approach. The advantages and disadvantages of either approach were given in the introduction. Besides, network meta-analysis can be performed using p-values, estimates of fold changes and their variances from the original data. Then, a new set of p-values and fold changes can be calculated by a graph theoretical framework [45]. While this framework avoids the problem of batch effect removal, inconsistencies within the network can occur. The chance of inconsistencies is of course increased in high-dimensional settings, where a network has to be built for each single gene.
Having different study settings or patient characteristics in the individual studies can be regarded as either an additional source for false conclusions, but also as an argument for producing results that are more generalizable. Here, different age groups contributed in the five original studies. Thus, one can argue that the genes we identified are more generally related to DENV infections, and do not depend on particular age groups.
When performing meta-analyses of high-dimensional data with the purpose of gene selection, there will be gains and losses in the new gene lists compared to those produced in the original studies. We don't want to claim that the findings of the individual studies were false positive, just because there was little overlap with our new results. Besides the reason of different patient characteristics and sampling time points, the process of data merging and processing and the increased power can lead to different results. Also the perfect matching of gene lists by different aliases of gene names is still a problem for meta-analyses, not on the data level but on the text level of published manuscripts. Difficulties for meta-analysis also arise by incomplete or heterogeneously reporting of results in the original studies. Therefore, we tend to interpret our new findings rather as false negatives in the original studies, and suggest to incorporate the newly identified genes in the research on dengue infections.

Reproducibility of findings from contributing studies
To further evaluate how the findings of the individual studies are in line with the findings of the network meta-analysis, we looked at the most important genes reported by these studies. In particular, we first compared our findings with individual genes highlighted in the manuscripts of the different publications. Loke et al. [29] mentioned 1525 genes in total that were differentially expressed between samples from DF, DHF and DSS individuals. Among these, the authors highlighted LTBR, PRAM1, CD14. None of these was found among the differentially expressed genes between the three patient classes in the network meta-analysis.
Above, we already mentioned the signature reported by Sun et al. [48] included CACNA2D3 which was also found among the top genes in the network metaanalysis. Besides, their signature included reported LOC286087, SLC4A4, PSPH, MYOM2, CD244, SMAD5, where LOC286087 was also identified among the top genes in the network meta-analysis.
Besides genes highlighted in their manuscripts, Long et al. [30] present a supplementary list of top 100 selected genes (ranked by fold changes), Loke et al. [29] also present three supplementary lists of top 100 selected genes from the comparisons between DF, DHF and DSS groups, Popper et al. [39] present a supplementary list of about 300 genes from the different group comparisons, and Sun et al. [48] present two signatures (in total 142 genes) they identified from classifier analyses. Kwissa et al. [23] did not present additional tables. The overlap between these additional lists from the original studies and the union of the top 50 lists from the network meta-analysis are given in Supplementary Fig. S5. The overlap of reported genes among all of these reported lists was zero, even the overlap among all individual studies, showing a large heterogeneity of findings and little robustness. Nevertheless, some genes occurred as overlap between pairs, triplets and quadruplets of lists.

Biological implications
As one of the mentioned six genes, and as a general noticeable player in infectious diseases, the function of IFI27 as part of the innate immune response in connection with viral diseases has also been reported in diverse organisms, including mice [24] and chicken [26]. Also in humans, the protein level increased after virus infection. Therefore, IFI27 expression was discussed as biomarker for viral infections [50]. This elevated expression is caused by various members of the flavivirus family. It has been associated with Zika, West Nile and dengue viruses infection [1,40,55]. During DENV infection, IFI27 is not only differentially expressed but further negatively correlated with the severity of the disease [37].
Except for DTL, the other five genes have been reported to distinguish between DSS and convalescent individuals by Long et al. [30] whose data also contribute to this meta-analysis. DTL itself has been shown to play role in oncogenic virus pathogenesis.
One gene that occurs only three times among the top 10 lists, CACNA2D3 has been reported by the contributing study of Sun et al. [48] as member of a molecular signature to distinguish between DHF and DF individuals. Besides, these mentioned genes have rarely been reported by the studies of the other contributing data sets, and could therefore be new candidates for further research on DENV infections.
The longitudinal characterization of the host transcriptome over time enables the assessment of the dynamic nature of the transcriptional profiles during the onset of the clinical disease upon DENV infection [48]. In line with this longitudinal study, others reported specific association of the transcript abundance pattern with the time course of disease [23,39]. Generally, genes associated with innate immune sensing and pro-inflammatory responses to viral infection, in particular those encoding type I IFN-related proteins, cytokine/chemokine-mediated signaling and complement activity coincide with high viremia during the initial clinical disease phase [23,48]. On the other hand, in a later phase of illness, genes associated with pathways involved in mitotic cell metabolic processes, translational control of protein biosynthesis and B cell differentiation are more prominent [29,48]. In fact, the host transcriptional profile during DENV infections has been associated with the clinical manifestation/disease severity [29,30,39], time of sampling/day of illness [23,48], viral load [23] and serological evidence of previous DENV exposure [39]. However, different studies favor certain features over others. For instance, in the study carried out by Kwissa and colleagues the transcriptional profile was not able to discriminate DF from DHF and was rather associated with the viral load than the clinical manifestation [23]. On the contrary, the differential transcriptome signatures observed between individuals with DF and DHF sampled on the same day after onset of disease suggests an association with the clinical manifestation rather than time of sampling [30]. Differentially expressed genes associated with a specific clinical manifestation were also described elsewhere, namely expression of mitochondrial ribosomal proteins was associated with DSS and genes encoding neutrophil-derived anti-microbial peptides were associated with DHF [29]. Remarkably, it is consistent that genes related with innate and adaptive immunity, such as IFN-mediated responses and antigen presentation and T cell priming respectively are down-regulated in individuals that develop severe dengue compared to those with uncomplicated disease [23,39,48]. Of particular importance in the context of severe dengue pathogenesis is the presumption that transcriptional events associated with capillary permeability and consequent hypovolemic shock may take place before cardiovascular collapse [30] which would have important prognostic potential.

Conclusions
The meta-analysis of host genome-wide transcript expression profiles can be particularly valuable in the identification of candidate genes that may be used as biomarkers, supporting prognostic capacity and adequate clinical management in the context of DENV infection. Compared to the findings of the original studies, the findings of this meta-analysis may be more robust due to increased sample sizes in the individual groups. Nevertheless, we don't argue that the new findings should replace the old once but rather complement. I.e., we assume that newly identified genes from the meta-analysis should be treated as false negatives in the original studies. As network meta-analyses have been very rarely performed on transcriptomics data, our study has also revealed methodical challenges and can provide ideas for methodical improvement.

Database search and study selection
Both repositories, AE and GEO, were queried using the search terms "Dengue". For selection of studies, we followed the reporting criteria of the PRISMA statement [34] where possible. Inclusion criteria for the final selection were that samples of whole blood where taken from human individuals. We defined no selection exclusion criteria regarding age, gender or social or pathological variables.

Data fusion and preparation for network meta-analysis
After determining a union of 35.695 genes that were in total regarded in the five studies a merged data set was built. After removing 9.180 genes that were solely present in only one of the studies, 26.515 genes remained for analysis. The k-nearest neighbour averaging method [54], implemented in the R-package 'impute' , was used to impute missing values in the merged data set, i.e. for genes that were only present in two, three or four studies. Next, study specific batch effects were removed using the ComBat methods from the R-package 'sva' which implements the model of Johnson et al. [16]. This model accounts for study specific additive and multiplicative effects. In order to obtain more homogeneous study groups, outliers were removed using the bagplot approach on the principal components [21].

Differential gene expression, mutual information networks gene ontology analysis
Pairwise group comparisons using the merged data set were performed with the R-package 'limma' to identify differentially expressed genes [47]. Resulting p-values were adjusted using the method of Benjamini and Hochberg to control for a false discovery rate of 5% [4]. In addition, confidence intervals for the log fold change were calculated [18]. In addition to the analysis of the merged data set, rank-based analysis based on the differential expression analysis of the individual data sets was performed using the R-packages 'RobustRankAggreg' [19].
The union of top10 genes from the merged-dataapproach and the rank-based approach were subjected to analysis by mutual information networks with the R-package 'minet' [33], which themselves, were built on the merged data. Thus, each mutual information network was comprised by ≤20 genes. First mutual information networks of the selected genes were determined for each of the five study groups individually. For each network, the 'minet' function generates an adjacency matrix which stores the mutual information between each pair of genes in this network, normalized to values of the interval [0, 1]. In contrast to linear correlation, mutual information reflects more generally the difference between the joint distribution of two genes to the product of their marginal distributions. We visualized the results by overlapping the networks for each pair of study groups, so that changes in inter-gene-correlation between two study groups can be seen. In particular, the difference between the two adjacency matrices from two study groups was taken as basis of each network plot. Thus, gains and losses in the size of mutual information between two genes can be illustrated in the plots. The R-package 'Rgraphviz' was used to plot the networks.
Genes were annotated with gene ontology (GO) terms, and enrichment of GO terms among the differentially expressed genes was assessed by Fisher's exact test [3]. GO annotations were taken from the UniProt database (www. unipr ot. org).

Training and evaluation of classifiers to separate different manifestations of dengue fever
Linear discriminant analysis [42], support vector machines [7] and nearest shrunken centroids [52] were used to train classifier models to study the separability of DF, DHF and DFF samples based on their expression profiles. Default settings of the R-packages 'MASS' , 'e1071' and 'pamr' were used to fit the classifiers to the training data. As only hyperparameter, the number of genes involved in each transcriptome signature was additionally studied. Genes were selected from the lists of differentially expressed genes from all three pairwise comparisons between the three subtypes, and ranked by their raw p-values. Leaveone-out-cross validation was used to assess the accuracy of correct classification, as well as sensitivity, specificity and positive and negative predictive values.