Host Transcriptional Meta-signatures Reveal Diagnostic Biomarkers for Plasmodium falciparum Malaria

Abstract Background Transcriptomics has been used to evaluate immune responses during malaria in diverse cohorts worldwide. However, the high heterogeneity of cohorts and poor generalization of transcriptional signatures reported in each study limit their potential clinical applications. Methods We compiled 28 public data sets containing 1556 whole-blood or peripheral blood mononuclear cell transcriptome samples. We estimated effect sizes with Hedge's g value and the DerSimonian-Laird random-effects model for meta-analyses of uncomplicated malaria. Random forest models identified gene signatures that discriminate malaria from bacterial infections or malaria severity. Parasitological, hematological, immunological, and metabolomics data were used for validation. Results We identified 3 gene signatures: the uncomplicated Malaria Meta-Signature, which discriminates Plasmodium falciparum malaria from uninfected controls; the Malaria or Bacteria Signature, which distinguishes malaria from sepsis and enteric fever; and the cerebral Malaria Meta-Signature, which characterizes individuals with cerebral malaria. These signatures correlate with clinical hallmark features of malaria. Blood transcription modules indicate immune regulation by glucocorticoids, whereas cell development and adhesion are associated with cerebral malaria. Conclusions Transcriptional meta-signatures reflecting immune cell responses provide potential biomarkers for translational innovation and suggest critical roles for metabolic regulators of inflammation during malaria.

Malaria causes thousands of deaths worldwide and accurate diagnosis combined to effective treatment are critical to control Plasmodium transmission.Malaria diagnosis is a straightforward process when trained microscopists and clinicians are available.Nevertheless, biomarkers are desired to predict outcomes of infection, severity, and therapeutical success, and to discriminate malaria from other febrile diseases.Because inflammation causes clinical manifestations of malaria, biomarkers underlying the immune response can provide indicators of disease, point to molecular mechanisms of pathogenesis, and reveal novel pharmacological targets [1].
Integration of transcriptomics data obtained from multiple independent cohorts is a powerful approach to identify biomarkers with diagnostic and prognostic applications [16,17].To overcome the limitations of analyses using single cohorts, we compiled 28 public transcriptomic data sets that span >1500 samples to perform multicohort analyses.We explored this compendium with meta-analyses and random forest models, which revealed gene signatures that differentiate uncomplicated P. falciparum malaria from controls or culture-confirmed bacterial infections; or characterize patients with cerebral malaria (CM).Importantly, the results suggest significant activity of immunometabolic regulators during malaria.

Data Compilation and Processing
Public microarray and RNA sequencing data from humans infected with Plasmodium were downloaded from the Gene Expression Omnibus (GEO), ArrayExpress and European Nucleotide Archive repositories.Each gene expression data set contained ≥5 whole-blood or peripheral blood mononuclear cell (PBMC) samples for case patients and controls.The cohorts included patients with uncomplicated malaria (UM) and participants from controlled human malaria infections (CHMIs).We also downloaded data sets including samples from patients with severe malarial anemia (SMA) or CM.Some data sets included samples from individuals with asymptomatic infection, presymptomatic infection, enteric fever (EF), community-acquired pneumonia (CAP), and chronic obstructive pulmonary disease (COPD).Additional data sets included samples from individuals with EF or sepsis.Uninfected controls included samples obtained from healthy individuals, obtained before CHMI (baseline), or obtained after treatment for malaria (Tables 1 and 2).
Robust multi-array average-normalized data were downloaded for Affymetrix arrays, and quantile normalization was used for other array platforms.Probes were mapped to genes using biomaRt R package v. 2.56.1 [30].We used the "MaxMean" method of the collapseRows function from Weighted Correlation Network Analysis R package v. 1.72.1 to summarize probes.Data were log 2 transformed.Normalized RNA sequencing data were downloaded from GEO.For some data sets, raw data were downloaded from the NCBI Sequence Read Archive using the SRA toolkit v. 3.0.1.Quality control was performed with FastQC v. 0.11.9, and adapter sequence and trimming were performed with fastp v. 0.23.0.Trimmed reads were mapped to the human genome (GRCh38.p13)and quantified using the Rsubread package v. 2.16.0 [31].Data were normalized to counts per million reads mapped and log 2 -transformed using edgeR v. 3.42.4[32], and gene annotations were performed with biomaRt v. 2.56.1 [30].

Transcriptional Meta-analysis of Uncomplicated P. falciparum Malaria
We split the data sets into 7 discovery and 6 validation cohorts.Discovery cohorts included 193 individuals with UM and 176 healthy controls.Validation cohorts included 59 individuals with UM and 57 controls (Table 1).Extended validations included samples from individuals with asymptomatic infection, presymptomatic infection, mixed asymptomatic and symptomatic infection, Plasmodium vivax malaria, EF, CAP, and COPD (Table 1).We used the MetaIntegrator R package v. 2.1.3[33] to analyze discovery cohorts by calculating the Hedge's adjusted g value as the effect size for each gene in each data set, which were latter combined with DerSimonian-Laird random-effects model.The Benjamini-Hochberg false discovery rate (FDR) was used to correct for multiple hypothesis testing.To account for unbalanced sample sizes and different sample types (PBMCs or whole blood), we used a leave-one-out approach that removed one data set at a time from the meta-analysis.An FDR-adjusted P value threshold of <.001, an effect size of 1.6-fold, and presence in ≥6 data sets identified 16 genes that were significant at all iterations.

Identification of Diagnostic Signatures With Random Forest models
To identify a signature that discriminates malaria from bacterial infections, we used the COmbat CO-Normalization Using conTrols (COCONUT) method from the metaIntegrator R package to combine microarray data sets acquired from wholeblood samples from patients with P. falciparum UM, SMA, CM, sepsis, or EF and respective uninfected controls (Table 2).This resulted in a "superset" of 714 samples and 6209 genes.Discovery cohorts included arrays from Illumina platforms, while validation cohorts included other platforms.To identify a signature that discriminates CM from other clinical phenotypes, we also combined arrays from patients with UM, SMA, or CM and uninfected controls, using the Combat method (Table 2).This resulted in a "superset" of 190 samples and 11 070 genes.Discovery cohorts included arrays from diverse platforms and the one independent validation cohort (E-MTAB-6413) was acquired with RNA sequencing and normalized separately.Discovery cohorts were used for feature selection using random forest analysis and built-in feature selection according to implementations in MetaboAnalyst 5.0 [34].Selected models were refined by forward search method in the MetaIntegrator R package.

Blood Transcription Module Meta-analyses
Gene expression in each data set was reduced to the activity of 346 blood transcription modules (BTMs) [35] using the tmod R package [36], which performs principal component analysis (PCA) for each BTM in each data set, retaining the PC1 score as the module activity.We used the MetaIntegrator package to perform BTM-driven meta-analyses or random forest analyses ,as described in the previous sections.For meta-analysis, the significance threshold was set at P < .05(FDR adjusted).Random forest analyses were also applied to BTMs to identify differences between malaria and bacterial infections or to determine malaria severity.The forward search method in the MetaIntegrator R package was used to refine selected signatures.

Meta-signature scores
The uncomplicated Malaria Meta-Signature (uMMS), Malaria or Bacteria Signature (MoBS), cerebral Malaria Meta-Signature (cMMS), and CM BTM scores were calculated as the z score of the difference between the geometric mean of upregulated genes/BTMs and the geometric mean of downregulated genes/BTMs.and tumor necrosis factor.These data were used in correlation analyses with identified transcriptional signatures.We also downloaded metabolomics data available for the data set GSE156791, which are deposited at the Metabolomics Workbench repository (under accession no.ST001400).Metabolomics data were normalized by the mean, and all data were log 2 transformed.These data were used in correlation analyses to obtain functional insights.

Statistical Analyses and Visualization
We used t tests, paired t tests, analysis of variance with the Bonferroni multicomparison test, and Spearman correlations to evaluate statistical significance where applicable.Meta-analyses of correlations between gene signatures and parasitemia were performed using Fisher's method.Receiver operating characteristic (ROC) curve analysis was used to test gene signatures, and the area under the curve (AUC) was used as a performance metric.Brier scores were used to evaluate accuracy of probabilistic predictions (Supplementary Table 1).Heat maps, forest plots, and ROC analysis were performed with the MetaIntegrator R package v. 2.1.3.Bubble plots were generated with the ggplot2 R package v. 3.4.4.

Identifying a Conserved Gene Signature for Uncomplicated P. falciparum Malaria With Meta-analysis
To identify a unified transcriptional signature characterizing the human immune response during P. falciparum UM, we downloaded 13 gene expression data sets in public repositories (Table 1).The cohorts include individuals with P. falciparum UM due to natural infection or CHMI.We used 7 discovery cohorts in a meta-analysis that resulted in 11 up-regulated and 5 down-regulated genes (P < 1.5 × 10 −8 ; FDR < 6.6 × 10 −7 ; effect size, 1.6-fold), whose differential expression was highly reproducible in the validation cohorts (P < .004;FDR < 0.05) (Figure 1A).This gene set was termed the uMMS.Effect sizes of ABLIM1 and SMPDL3A are shown for all UM cohorts in Figure 1B; see Supplementary Figure 1 for the remaining genes.ROC analysis based on the uMMS resulted in a summary AUC of 0.98 (95% confidence interval [CI], .91-.99) for discovery (Figure 1C) and 0.96 (.89-.99) for validation cohorts (Figure 1D).Infection with Plasmodium can be asymptomatic or cause severe complications and death.We tested whether the uMMS could distinguish between different clinical phenotypes.A score based on the uMMS distinguished UM from asymptomatic infection (Supplementary Figure 2), but it was not associated with severity (Supplementary Figure 3).The uMMS score increased in samples from individuals with presymptomatic malaria in only 2 of 5 cohorts (Supplementary Figure 4) but clearly distinguished mixed asymptomatic/symptomatic malaria from control samples (Supplementary Figure 4B and  4C).The uMMS displayed lower performance for discriminating P. vivax malaria from control samples (Supplementary Figure 5).Malaria can be confounded with other inflammatory diseases, especially febrile infections [16,18].The uMMS showed significant ability to discriminate between UM and EF, CAP, and COPD (Supplementary Figure 6).

Defining a Diagnostic Signature That Differs Between Malaria and Bacterial Infections
Meta-analysis relies on the heterogeneity of diverse cohorts containing samples from relevant phenotypes (eg, malaria and controls; malaria and EF).Therefore, we were unable to use this strategy to find a signature discriminating malaria from other bacterial diseases such as sepsis.To overcome this issue, we merged multiple cohorts from patients with malaria, EF, sepsis, and controls into a "superset" and split them into discovery and validation cohorts (Table 2).We applied random forest analysis in a 2-class approach, using malaria as cases and all the other conditions as controls, as described elsewhere [16].The analysis resulted in diverse models containing different numbers of classifier genes.We chose a signature containing 25 genes, which was submitted to a forward search approach within the MetaIntegrator package.We identified an 8-gene signature that discriminates malaria from EF, sepsis, and controls with highperformance in discovery and validation cohorts (Figure 2A).
The MoBS score showed significant differences between malaria and the other groups in discovery and validation cohorts (Figure 2B).It also distinguished UM from controls in 8 independent data sets, with a summary AUC of 0.9 (95% CI, .76-.96) (Figure 2C).Effect sizes for genes composing MoBS revealed higher variability compared with uMMS, but pooled effect sizes are in accordance with random forest results (Figure 2D).The MoBS score differed in 1 of 5 cohorts of presymptomatic malaria (Supplementary Figure 7), but it did not differ between controls and mixed asymptomatic/symptomatic samples (Supplementary Figure 7B and 7C) or P. vivax UM (Supplementary Figure 8).MoBS score differed between SMA and CM in one data set included in the discovery cohort but not in an independent data set (Supplementary Figure 9).

Transcriptional Meta-signature Characterizing Cerebral Malaria
Because neither uMMS nor MoBS could reliably distinguish malaria severity, we also used random forest analysis to derive a cMMS.For that, we merged microarray data sets containing samples from individuals with UM, SMA, and CM and from controls into a "superset" of 190 samples that was used as discovery cohort (Table 2).Random forest analysis using a 2-class classifier differentiating CM from the rest resulted in a model with 25 genes that was manually curated by removing genes whose contribution was not relevant for final AUC.We identified a 19-gene signature that distinguished CM from other phenotypes, with AUCs of 0.95 in the discovery cohort and 0.82 in an independent validation data set (E-MTAB-6413) (Figure 3A).The cMMS score differed between CM and all other groups, but it also differed between controls and UM or SMA (Figure 3B).Selection frequency and rank importance of genes included in the cMMS are shown in Figure 3C.cMMS distinguished UM from controls for most cohorts but displayed reduced discriminatory power for 3 of them (Supplementary Figure 10A).It also demonstrated reduced ability to simultaneously discriminate between malaria and EF, sepsis, and controls (Supplementary Figure 10B).

Correlation Between Meta-signatures and Hallmarks of Clinical Malaria
Plasmodium infection causes RBC disruption as a direct function of parasite replication, which triggers the inflammatory response.Meta-analyses of Spearman rank correlation between the signature scores and parasitemia revealed significant positive correlations for all of them, with uMMS displaying the highest meta-coefficient (Supplementary Figure 11A).We also evaluated the associations between the identified signature scores and other parasitological, hematological, and cytokine and chemokine data available for the data set GSE117613 [8].There were positive correlations with PfHRP2 and hemoglobin levels (Supplementary Figure 11A and 11B) and negative associations with RBC, MCV, and platelet counts for all of them, but only the cMMS score was significantly correlated with EPO levels (Supplementary Figure 11B).The 3 signature scores were correlated positively with plasmas abundance of interleukin 1RA, 6, 8, and 10, tumor necrosis factor, CCL2, CCL4, and CXCL10 (Supplementary Figure 11C and 11D).

BTM Activity Reflecting Conserved Human Immune Responses to Malaria
The unbiased BTM framework captures gene expression activity that translates into immunological and metabolic processes in the human blood [35].We performed a BTM-driven meta-analysis to identify conserved activity of gene modules during UM.Meta-analysis with 7 discovery cohorts (Table 1), using a leave-one-out and forward search approach, resulted in a set of 6 down-regulated and 3 up-regulated BTMs related to nuclear pore complex, SMAD2/3 signaling, transcriptional targets of glucocorticoid receptor (GR), T-cell activation, B cells, proinflammatory myeloid cells, cell cycles, and immune activation (Figure 4A).ROC analysis using this BTM set revealed summary AUCs of 0.97 (95% CI, .85-1)for discovery cohorts and 0.95 (.8-.99) for validation cohorts (Supplementary Figure 12A).This BTM set displayed poor discriminatory performance in both natural infection and CHMI in P. vivax malaria cohorts (Supplementary Figure 12B).To gather further insights into the transcriptional profiles that underlie the immune responses to infection, we also used BTM-driven random forest models to identify conserved modules that differ between malaria and other conditions, such as EF, sepsis, and healthy controls.We used the MoBS discovery superset (Table 2) in random forest analysis to select 14 BTMs that discriminate malaria from other groups with relevant performance in discovery and validation cohorts (Supplementary Figure 12C).These BTMs are related to T and B lymphocytes, transcriptional targets of GR, SMAD2/ 3 signaling, antigen presentation, platelets, extracellular matrix, and complement (Figure 4B).We also used the cMMS discovery "superset" in BTM-driven random forest analysis and a forward search method to unveil 6 BTMs, mostly related to immune activation, cell development, and adhesion (Figure 4C).This BTM set discriminates CM from other groups in the discovery cohort and validation data set (Supplementary Figure 12D).Calculation of a CM BTM score revealed significant differences between CM and UM or SMA in the discovery cohort and validation data set (Figure 4D).

Enzymatic Activity and Metabolite Signaling During Malaria Suggested by Conserved Host Gene Expression
The genes composing the signatures and the BTMs associated with malaria are presumably translated into proteins and enzymes involved in metabolism and cell signaling in blood leukocytes.To gain insights into the functional activity of genes encoding enzymes and BTMs associated with metabolism, we integrated pairwise transcriptomics and metabolomics data available for the data set GSE156791.SMPDL3A encodes sphingomyelin phosphodiesterase acidlike 3A, an enzyme that hydrolyses nucleotide diphosphate and triphosphate (eg, adenosine diphosphate, adenosine triphosphate [ATP]) into nucleotide monophosphate (eg, adenosine monophosphate [AMP]) (Supplementary Figure 13A) [37].SMPDL3A expression was consistently up-regulated in the blood of individuals with malaria when compared with uninfected controls in  1B).Of note, the relative levels of AMP were increased in the plasma (Figure 5A) and were correlated positively with SMPDL3A expression during malaria but not before infection (Figure 5B).
The BTM transcriptional targets of GR (M74) showed reduced activity during UM (Figure 4A and Supplementary Figure 14A).At the same time, the relative abundance of glucocorticoids, such as cortisol and cortisone, increased in plasma with malaria (Figure 5C).Glucocorticoids diffuse through cell membranes and interact with GR in the cytoplasm, which is transported to the nucleus and regulates the expression of transcriptional targets of GR (Supplementary Figure 13B) [38].Strikingly, M74 activity exhibited strong negative correlations with the abundance of cortisol and cortisone only after clinical malaria (Figure 5D).Those associations are driven by genes such as CD3E, ZC3H12D, HLA-DOA, and MEF2C (Supplementary Figure 14B).

DISCUSSION
Malaria diagnosis can be accomplished without difficulty by trained microscopists and clinicians in structured settings.However, there have been emerging concerns about resistance to rapid diagnostic tests in highly endemic areas, which motivates searches for alternatives [39].Differentiating symptomatic malaria from febrile illness with coincidental parasitemia is also critical, as severe bacterial infection of patients asymptomatically infected with P. falciparum can confound diagnosis and misguide appropriate treatment.Here, we identified transcriptional biomarkers of disease caused by P. falciparum that differentiate UM or CM or that discriminate malaria from bacterial infection with significant performance.Importantly, we identified associations between different data types that suggest relevant immunometabolic activity during malaria.
We used both meta-analyses and data set merging followed by random forest analysis to identify gene meta-signatures, all of which present advantages and limitations.The uMMS displays the best performance for distinguishing Plasmodium infection from uninfected controls, including P. vivax malaria.However, it had a moderate performance for distinguishing malaria from bacterial infection, and it could not discriminate malaria severity.Therefore, the uMMS represents a broad transcriptional signature that has significant overlap with gene expression induced by other inflammatory conditions.The MoBS scores exhibits good performance for discriminating malaria from EF, sepsis, and controls simultaneously, but the magnitude of scores and not whether the gene is expressed in a determined condition is what drives those differences.MoBS also seems to be specific for clinical P. falciparum malaria, shown by reduced performance for differentiating presymptomatic and mixed samples from controls.However, as it was generated with whole-blood samples, it displayed lower generalizability for data acquired from PBMCs.The cMMS identified samples from CM with good performance both in discovery and validation cohorts, but it also showed lower generalizability for discriminating malaria from other conditions.
Anemia is a typical clinical manifestation of malaria that is caused in part by parasite replication.Positive associations with parasitemia and levels of PfHRP2 concomitant with negative associations with RBC counts and MCV support the significant role of genes composing the different signatures.Furthermore, cMMS had the strongest association with MCV and was significantly associated with EPO, which promotes RBC production and function and plays an important role in the pathogenesis of CM [40].The data suggest that the expression of selected genes fluctuates according to the interactions between host and Plasmodium, whereby parasite growth causes RBC destruction, which, in turn, activates the inflammatory response.Associations with cytokines and chemokines in the plasma corroborate this hypothesis.
We identified positive association between SMPDL3A expression in leukocytes and relative levels of AMP in the plasma during malaria.Sphingomyelin phosphodiesterase acidlike 3A (encoded by SMPDL3A) is an enzyme secreted by macrophages, which displays nucleotide phosphodiesterase activity against nucleotide triphosphates, including ATP [37].These data suggest that in response to Plasmodium infection, leukocytes secrete SMPDL3A to convert ATP into AMP to regulate the inflammatory signaling of purinergic receptors [41].Moreover, SMPDL3A is a cyclic guanosine monophosphate-AMP degrading enzyme that restricts cGAS-STING signaling [42] and may contribute to regulate type I interferon responses during malaria [43].Single-cell transcriptomics of PBMCs from P. falciparum UM demonstrate that type I interferon responses are activated in all cell types, suggesting it is linked to induction of immunoregulatory networks during malaria [44].
We found significant modulation and associations between transcriptional targets of GR (M74) and the abundance of cortisol and cortisone, suggesting that, in addition to pregnenolone [19], glucocorticoids also regulate immune cell responses during malaria, but this hypothesis requires further validation using other methods.BTM analyses also revealed significant associations of modules underlying immune activation, cell development, and adhesion during CM.Mouse models suggest that the dynamics of leukocyte development, activation, and adhesion to brain vasculature can be linked to the severity of CM [45].However, adherent properties of leukocytes during CM may also affect other organs to promote pathology [46].
Both gene-level and BTM-level analyses suggest that the transcriptional signatures identified in this study are mostly associated with P. falciparum malaria, because they fail to characterize individuals infected with P. vivax with the same performance.Those differences might be driven by higher parasite densities during P. falciparum malaria, which is supported by significant associations between gene signatures and parasitemia.However, specific technical limitations may also influence the results, including sequencing depth and coverage for both data sets with P. vivax malaria.
In conclusion, we identified host transcriptional responses with potential for translational innovation that also contribute to a better understanding of the human immune response to P. falciparum malaria.

e474•
JID 2024:230 (15 August) • Silva et al The Journal of Infectious Diseases M A J O R A R T I C L E Abbreviations: Asymp, asymptomatic infection; CAP, community-acquired pneumonia; CHMI, controlled human malaria infection; COPD, chronic obstructive pulmonary disease; EF, enteric fever; ID, identification number; PBMCs, peripheral blood mononuclear cells; Presymp, presymptomatic infection; RNA-seq, RNA sequencing; UM, uncomplicated malaria; uMMS, uncomplicated Malaria Meta-Signature; WB, whole blood.aAccession IDs are from Gene Expression Omnibus (GEO) database, except for PRJEB45911, which is deposited at the European Nucleotide Archive.bPlatforms according to GEO database.

Figure 1 .
Figure 1.Uncomplicated malaria induces a conserved transcriptional signature in blood leukocytes.A, Heat map of the effect sizes of genes composing the uncomplicated Malaria Meta-Signature in discovery (n = 369) and validation (n = 116) cohorts.B, Representative forest plots of one down-regulated gene, ABLIM1, and one up-regulated gene, SMPDL3A.The standardized mean difference in the x-axis is computed as the log 2 -transformed Hedge's adjusted g value.The standard error of the mean for each study is inversely proportional to the size of blue rectangles, and 95% confidence intervals (CIs) are represented by whiskers.Red diamonds reflect the combined mean difference for each gene, and their width represents the 95% CI.C, D, Receiver operating characteristic (ROC) curves comparing individuals with uncomplicated Plasmodium falciparum malaria and controls in discovery (D) and validation (E) cohorts.Abbreviation: AUC, area under the curve.

Figure 2 .
Figure 2. Random forest analysis identifies a signature discriminating malaria or bacterial infections.A, Receiver operating characteristic (ROC) curves comparing individuals with Plasmodium falciparum malaria and individuals with enteric fever, sepsis, and controls in discovery and validation cohorts using the Malaria or Bacteria Signature (MoBS).Abbreviations: AUC, area under the curve; CI, confidence interval.B, MoBS score comparing individuals with P. falciparum malaria and individuals with enteric fever or sepsis and controls in discovery and validation cohorts.***P < .001.C, ROC curves comparing individuals with uncomplicated malaria and controls using MoBS.D, Heat map of the effect sizes of genes composing MoBS in individuals with malaria compared with controls (n = 531).MoBS scores were compared using analysis of variance followed by the Bonferroni multicomparison test.Error bars represent means with standard errors of the mean.

Figure 3 .
Figure 3. Cerebral malaria (CM) is characterized by a systemic transcriptional signature.A, Receiver operating characteristic (ROC) curves comparing individuals with CM and individuals with uncomplicated malaria (UM) or severe malarial anemia (SMA) and controls in discovery and validation cohorts using the cerebral Malaria Meta-Signature (cMMS).Abbreviations: AUC, area under the curve; CI, confidence interval.B, cMMS score comparing individuals with CM and individuals with UM or SMA and controls in discovery and validation cohorts.C, Genes composing the cMMS ordered by selection frequency and rank importance in the random forest model.cMMS scores were compared using analysis of variance followed by the Bonferroni multicomparison test.Error bars represent means with standard errors of the mean.**P < .01;***P < .001.

Figure 4 .
Figure 4. Blood transcription module (BTM) underlying host responses to Plasmodium falciparum malaria.A, Bubble plots of the effect sizes of significant BTMs in discovery and validation cohorts, using meta-analysis of uncomplicated P. falciparum malaria and controls shown in Table 1.The bubble size and color scale are proportional to the standardized mean difference (effect size) between individuals with malaria and controls.B, Heat map of BTM activity in individuals with P. falciparum malaria and individuals with enteric fever or sepsis and controls in discovery and validation cohorts shown in Table 2. C, BTMs differing between individuals with cerebral malaria (CM) and individuals with uncomplicated malaria (UM) or severe malarial anemia (SMA) and controls, ordered by selection frequency and rank importance in the random forest model.D, CM BTM signature score differentiating CM from other clinical phenotypes in discovery and validation cohorts.*P < .05;**P < .01;***P < .001.Abbreviations: MHC, major histocompatibility complex; TBA, to be assigned; Th1, T-helper 1; SMAD, Mother against Dpp (MAD) homology (MH) domain transcription factor; TLR, Toll-like receptor; uMMS, uncomplicated Malaria Meta-Signature.

Figure 5 .
Figure 5. Functional validation via integration with metabolomics data.A, Comparison of the relative abundance of 5´-adenosine monophosphate (AMP) in the plasma before infection (healthy individuals) and during Plasmodium falciparum malaria.B, Associations between SMPDL3A expression and AMP before (healthy) and during P. falciparum malaria.C, Comparison of the relative abundance of cortisol and cortisone in the plasma before (healthy) and during P. falciparum malaria.D, Associations between M74 activity and cortisol, and cortisone before (healthy) and during P. falciparum malaria.Data were analyzed using paired t tests and Spearman rank correlation.*P < .05;**P < .01;***P < .001.

Table 2 . Data Sets Used for Random Forest Analyses
Abbreviations: CHI, controlled human infection; CM, cerebral malaria; cMMS, cerebral Malaria Meta-Signature; EF, enteric fever; ID, identification number; MoBS, Malaria or Bacteria Signature; PBMCs, peripheral blood mononuclear cells; RNA-seq, RNA sequencing; SMA, severe malarial anemia; UM, uncomplicated malaria; WB, whole blood.a Accession IDs are from the Gene Expression Omnibus (GEO) database, except for E-MTAB-6413, which is deposited at the ArrayExpress database.b Platforms according to the GEO database.