Research ArticleMalaria

Antigen-stimulated PBMC transcriptional protective signatures for malaria immunization

See allHide authors and affiliations

Science Translational Medicine  13 May 2020:
Vol. 12, Issue 543, eaay8924
DOI: 10.1126/scitranslmed.aay8924

Predicting protection

A highly effective malaria vaccine would save many lives but correlates of protection remain ill-defined. Moncunill et al. studied peripheral blood cells isolated from individuals that had received sporozoites under chemoprophylaxis or the RTS,S vaccine. Transcriptomic analysis of gene expression after in vitro stimulation of cells revealed preimmunization and postimmunization signatures, which were validated with separate cohorts. The preimmunization signatures hint at mechanisms of differential vaccine responses between individuals; once validated in additional studies, the postimmunization signatures could be used as a surrogate for protection in clinical trials, possibly accelerating vaccine development.


Identifying immune correlates of protection and mechanisms of immunity accelerates and streamlines the development of vaccines. RTS,S/AS01E, the most clinically advanced malaria vaccine, has moderate efficacy in African children. In contrast, immunization with sporozoites under antimalarial chemoprophylaxis (CPS immunization) can provide 100% sterile protection in naïve adults. We used systems biology approaches to identifying correlates of vaccine-induced immunity based on transcriptomes of peripheral blood mononuclear cells from individuals immunized with RTS,S/AS01E or chemoattenuated sporozoites stimulated with parasite antigens in vitro. Specifically, we used samples of individuals from two age cohorts and three African countries participating in an RTS,S/AS01E pediatric phase 3 trial and malaria-naïve individuals participating in a CPS trial. We identified both preimmunization and postimmunization transcriptomic signatures correlating with protection. Signatures were validated in independent children and infants from the RTS,S/AS01E phase 3 trial and individuals from an independent CPS trial with high accuracies (>70%). Transcription modules revealed interferon, NF-κB, Toll-like receptor (TLR), and monocyte-related signatures associated with protection. Preimmunization signatures suggest that priming the immune system before vaccination could potentially improve vaccine immunogenicity and efficacy. Last, signatures of protection could be useful to determine efficacy in clinical trials, accelerating vaccine candidate testing. Nevertheless, signatures should be tested more extensively across multiple cohorts and trials to demonstrate their universal predictive capacity.


Malaria remains a major public health problem, causing an estimated 218 million cases and 405,000 deaths in 2018 (1). Although a vaccine is considered a crucial tool in combating this infectious disease (2), decades of research have thus far only resulted in a single subunit vaccine candidate, RTS,S/AS01E, which has been recommended by the World Health Organization for pilot implementation studies starting in 2019 in three African countries. The efficacy of RTS,S/AS01E against clinical malaria in a phase 3 trial for licensure over 12 months of follow-up was moderate, about 31% in infants of 6 to 12 weeks and about 56% in children of 5 to 17 months (3). Vaccines based on attenuated Plasmodium falciparum sporozoites through irradiation (PfSPZ vaccine) and chemoprophylaxis (PfSPZ-CVac) are promising candidates (4) that have shown up to 100% efficacy in controlled human malaria infection (CHMI) trials in malaria-naïve adults. However, efficacy under natural exposure or in a pediatric population has not been evaluated extensively and may be lower compared to CHMI (5).

Major hurdles in the development of an effective malaria vaccine are the absence of immune correlates of protection and understanding of the mechanisms of protective immunity among other difficulties, including the complexity of the Plasmodium parasite’s cycle and its high polymorphism (4, 6). Efficacy testing of each vaccine candidate in humans requires conducting complex, long, and expensive clinical trials, in which protection is assessed by subjecting vaccinated individuals to experimental (CHMI) or natural P. falciparum challenge. Identification of in vitro surrogate markers of protection would simplify clinical trial design and accelerate the development of new or improved vaccines. The best correlate thus far of RTS,S/AS01E is the titer of immunoglobulin G (IgG) against the vaccine antigen (7), a region of the P. falciparum circumsporozoite protein (CSP). IgG against CSP targets the sporozoites and is thought to prevent liver-stage infection and subsequent infection of red blood cells (RBCs). However, no IgG threshold for protection has been identified. IgG to CSP is also associated with protection induced by irradiated PfSPZ in malaria-naïve adults (4) but not in preexposed populations (5).

A powerful model to dissect immune correlates of sterile, preerythrocytic immunity to malaria is the chemoprophylaxis and sporozoite (CPS) immunization regimen (4, 8), which inspired the PfSPZ-CVac. In PfSPZ-CVac, purified sporozoites previously cryopreserved are given as direct venous inoculation to volunteers on a prophylactic regimen of chloroquine. In CPS immunization, fresh sporozoites are delivered by repeated bites of infected mosquitoes to volunteers also on a prophylactic regimen of chloroquine (8). Optimal dosing confers 100% protection in malaria-naïve adults (4, 8), similar to PfSPZ-CVac. Suboptimal dosing of CPS immunization results in different degrees of protection, ranging from sterile protection to delayed patency or no protection compared to malaria-naïve infection controls (9, 10). This heterogeneity and the controlled conditions, in comparison with trials in endemic areas where vaccinees are subject to natural exposure at unknown time points, make the CPS regimen a useful tool to study malaria immunity.

Currently, an emerging approach based on omics technologies and systems biology analyses is being used to decipher immune signatures that predict vaccine immunogenicity (11) and has the potential to also identify signatures of protection (12) and provide insights into mechanisms relevant for vaccine efficacy. Successful studies have identified transcriptional signatures induced early after immunization that correlate with and predict the later adaptive immune responses in humans for vaccines targeting yellow fever virus (13), influenza (14, 15), meningococcus (16), HIV (17), and even for RTS,S (12). However, only a few studies have investigated signatures of protection, and these were performed precisely in RTS,S and CPS CHMI studies in malaria-naïve adults (12, 1820). In this study, we used peripheral blood mononuclear cell (PBMC) samples from volunteers (malaria-naïve adults) immunized by CPS (9, 10) and from naturally exposed African children and infants who participated in the RTS,S/AS01E phase 3 trial (3), combined with a systems biology approach to identifying gene signatures of protective immunity against malaria.


CPS and RTS,S cohorts for the identification of candidate signatures

In the CPS trial (9), 24 malaria-naïve adults were immunized with suboptimal doses of P. falciparum–infected mosquitoes under a chloroquine prophylaxis. After 5 months, they were challenged with parasites through the bites of infected mosquitoes (Fig. 1A and fig. S1). Sterile protection was observed in 17 out of 24 volunteers. In the African phase 3 trial, we selected a subset of 255 infants and children who received three doses of RTS,S/AS01E or a comparator vaccine (rabies vaccine for children and meningococcal C conjugate vaccine for infants) in three sites (Bagamoyo in Tanzania, Lambaréné in Gabon, and Manhiça in Mozambique). On the basis of a case-control design, we included all children who had clinical malaria episodes during a follow-up period of 12 months after the third primary vaccination (n = 50; Fig. 1B and fig. S1B) and selected up to four nonmalaria controls for each malaria case. After sample processing, we ended up analyzing all 24 CPS-immunized volunteers (fig. S2) and 178 phase 3 trial vaccinees: 127 RTS,S individuals (24 nonprotected and 70 protected) and 51 comparator individuals (7 nonprotected and 32 protected) (fig. S3).

Fig. 1 Study design for the identification of signatures of protection.

(A) Samples from a chemoprophylaxis and sporozoite (CPS) immunization clinical trial (9) and (B) the RTS,S/AS01E vaccine phase 3 clinical trial (3) were used. The CPS trial involved immunization of malaria-naïve adult volunteers by bites from P. falciparum (Pf)–infected mosquitoes during chloroquine (CQ) chemoprophylaxis. Three groups of volunteers received three different doses of bites from infected mosquitoes. After infectious challenge, 17 of 24 volunteers were protected from infection. Blood samples were collected at baseline (before immunization) and 5 months after the last dose (after immunization). Children and infants from three different African countries in the RTS,S/AS01 phase 3 trial received three doses of the RTS,S vaccine or a comparator vaccine 1 month apart and were followed up for detection of clinical malaria episodes. Samples from 50 volunteers who had malaria (nonprotected) and 205 volunteers who did not have any malaria episode (protected) were selected. Blood samples were collected at baseline for children and 1 month after third vaccine dose for children and infants. (C) Previously cryopreserved PBMCs were stimulated in vitro with circumsporozoite protein (CSP) peptide pool or P. falciparum–infected red blood cells (PfRBCs) and their respective background controls, dimethyl sulfoxide (DMSO), and uninfected red blood cells (uRBCs). Gene expression was measured by microarrays. (D) Differential gene expression analysis using linear regression models, and subsequently, GSEA was performed. Microarray data were also used for protein network–based models. Proteins behaving differently in the models were down-selected using data science methods, leading to identification of pre- and postimmunization signatures of protection, consisting of three to five proteins, with accuracies of >70%. TPMS, Therapeutic Performance Mapping System.

We analyzed the transcriptomic profile of in vitro P. falciparum recall responses for CPS volunteers both before immunization and just before challenge with infected mosquitoes and for African volunteers before immunization (only children) and 1 month after RTS,S immunization. This approach (Fig. 1C) was selected instead of ex vivo transcriptomic responses because it maximized the chances to detect acquired immune responses induced by immunization and mechanisms of protection. We stimulated PBMC in vitro using a CSP peptide pool to assess preerythrocytic responses and whole asexual blood-stage parasites [P. falciparum–infected RBCs (PfRBCs)]. This strategy was chosen to increase the chances to detect antigen-specific responses to CPS (due to the large overlap between late liver-stage and blood-stage antigens) and to evaluate naturally acquired immunity in African participants. CSP- and PfRBC-stimulated gene expression was background-corrected with the gene expression profile of PBMCs from the same participant and time point treated with dimethyl sulfoxide (DMSO; present in the peptide pools) and uninfected RBCs (uRBCs), respectively, to control for participant-specific differences. Thus, the differential gene expression measured (Fig. 1D) was specifically induced by the antigen stimulations.

Differential gene expression associated with CPS immunogenicity

We found 60 genes significantly up-regulated [q < 0.1 and |fold change (FC)| > 1.5] and 18 down-regulated after CPS immunization compared to preimmunization upon PfRBC recall but no significant differences upon CSP stimulation (data file S1). Many of the changes corresponded to genes involved in immune responses. Among the 10 top up-regulated genes were IFNG (interferon gamma) (FC = 4.68 and q < 0.0001) and GZMB (granzyme B) (FC = 2.39 and q < 0.0001), which have been previously associated with CPS immunogenicity (9), IL-22 (FC = 6.64 and q < 0.0001), CISH (FC = 2.23 and q < 0.0001), CCL4 (FC = 2.14 and q < 0.0001) and LTA (lymphotoxin alpha; FC = 1.98 and q < 1.71 × 10−6). No differential gene expression was identified comparing protected with unprotected volunteers either before or after immunization (data file S1).

Consistently, principal components analysis (PCA) using FC expression of CSP/DMSO and PfRBC/uRBC stimulations did not reveal any differences by immunization, dose, sex, or protection in the CPS cohort. Samples showed some clustering only by stimulation condition (PfRBC/uRBC versus CSP/DMSO) (fig. S4). Similarly, unsupervised hierarchical clustering and heatmaps did not reveal any clustering by study conditions besides stimulation (fig. S5).

CPS immunization often leads to detectable parasitemia after the immunizing mosquito bites (9), especially after the first dose, and an antibody and cellular recall response to PfRBCs (21, 22) may be related to this initial blood-stage exposure. We assessed the correlation of PfRBC/uRBC FC gene expression in CPS-immunized volunteers with cumulative parasitemia during the immunization period to identify genes associated with blood-stage parasite exposure and immunogenicity. Many genes had |ρ values| > 0.4 and P < 0.05, but none of them were significant after adjusting for multiple testing (data file S2). In addition, some nonsterilely protected individuals showed partial protection (longer prepatency period) (9). We found only 10 and 4 significant genes after adjusting for multiple testing (q < 0.1), correlating with time to positive quantitative real-time polymerase chain reaction (qRT-PCR) and thick blood smear prepatency, respectively (data file S2). Annotation of genes for both cumulative parasitemia and prepatency analyses revealed immune responses that are likely to be antigen-specific (data file S2).

Lack of differential gene expression associated with RTS,S/AS01E

Gene expression of RTS,S-vaccinated individuals was analyzed in relation to comparator vaccinees and not contrasting to the same individuals at preimmunization because these transcriptional data were available only for a subset of children and the comparator group allowed controlling for malaria exposure and age. There were no significantly up- or down-regulated genes upon CSP or PfRBC stimulation in RTS,S vaccinees in relation to comparators (data file S3). Similarly, no differential gene expression was detected in protected versus nonprotected individuals who received the RTS,S or the comparator vaccine (data file S3). PCA and unsupervised hierarchical clustering and heatmaps using FC expression of CSP/DMSO and PfRBC/uRBC stimulations did not reveal any differences between vaccine groups, time points, protection, site, age, and sex either (figs. S6 and S7), and samples were only clustered by stimulation condition (PfRBC/uRBC versus CSP/DMSO).

Gene sets associated with CPS immunization and sterile protection

We next performed gene set enrichment analysis (GSEA), which allows us to investigate sets of genes when the individual gene associations are not strong. As gene sets, we used blood transcriptional modules (BTMs) (16) that have been successfully applied in previous systems vaccinology studies (12), to enable functional interpretation of the transcriptional responses. We considered only the BTMs that were significantly enriched also using two alternative enrichment analyses, correlation-adjusted mean rank (CAMERA) and Tmod, to increase the confidence in the findings. For all these analyses, we used the lists of all genes regardless of statistical significance (data file S1) ranked according to their average expression FC in volunteers after CPS immunization relative to preimmunization for each recall response (considering its stimulation background control). A summary of all gene set results can be found in the data file S4.

CPS immunization induced some common BTM activity upon CSP and PfRBC stimulation including dendritic cell (DC) activation, T cells, natural killer (NK) cells and IFN/antiviral sensing (Fig. 2A). Two of the genes found in “signaling in T cells (M35.0 and M35.1)” BTMs were IFNG and GZMB, which were among the top up-regulated genes upon immunization and have been previously associated with CPS (9). In addition, upon CSP stimulation, BTMs related to B cells, inflammatory/Toll-like receptor (TLR)/chemokines and signal transduction were up-regulated. These BTMs indicate both innate and acquired responses, probably related to antigen-specific responses, subsequent signaling amplification, and bystander activation. Curiously, for the PfRBC recall response after CPS relative to preimmunization, there was a repression of BTMs of antigen presentation, DC surface receptors, monocytes and some BTMs of inflammatory/TLR/chemokines, and signal transduction. The simultaneous positive and negative enrichment of related BTMs and the positively enriched module “immune regulation–monocytes, T and B cells (M57)” may indicate immunomodulatory and inhibitory processes in response to PfRBC stimulation in CPS-immunized volunteers.

Fig. 2 Transcriptional responses associated with CPS immunogenicity and protection.

Each square represents a blood transcription module (BTM). The color shading indicates normalized enrichment scores obtained by GSEA analysis for BTMs. Assignment of a BTM to a high-level annotation group is illustrated by a colored sidebar. GSEA, CAMERA, and Tmod were run with genes ranked by (A) the expression of postimmunization (Post) relative to preimmunization (Pre) and (B) protected relative to nonprotected individuals for CSP and PfRBC recall stimulations after and before immunization. Modules that did not reach the significance cutoff of an FDR q value of 0.1 in all three enrichment methods or a minimum of 10 matched genes were eliminated. Modules without annotation are not shown. Modules that represent common associations of both immunogenicity and protection at postimmunization are highlighted with a circle for CSP recall responses and triangle for PfRBC recall responses, and filled symbols indicate that enrichment had the same direction, whereas empty symbols indicate that enrichment had the opposite direction. AP-1, activator protein-1; MHC, major histocompatibility complex; ECM, extracellular matrix.

Unexpectedly, gene set analysis comparing protected with nonprotected CPS-immunized individuals (Fig. 2B) upon CSP recall showed several negatively enriched BTMs related to innate responses (antigen presentation, DC activation, inflammatory/TLR/chemokines, and IFN/antiviral sensing among others) and also some BTMs related to acquired immune responses (B and T cells). There was a large overlap between BTMs associated with protection as assessed postimmunization on the one hand, and immunogenicity on the other hand, although in the opposite direction (28 of the 34 BTMs negatively associated with protection were found positively enriched at postimmunization compared to preimmunization; Fig. 2B). Analysis before immunization revealed not only negative enrichment of some of the same BTMs related to DC activation, chemokine clusters, and viral sensing in protected relative to unprotected volunteers but also positive enrichment of some other BTMs related to T cells and NK cells.

In contrast, in CPS-immunized individuals, a more robust response was induced upon PfRBC recall compared to CSP stimulation. Only positively enriched BTMs in the protected individuals were found relative to unprotected volunteers (Fig. 2B). These BTMs were related to antigen presentation, DC activation, B cells, T cells, cell cycle, extracellular matrix and migration, inflammatory/TLR/chemokines, signal transduction, and monocytes. Several postimmunization BTMs that correlated with protection were also correlated with immunogenicity (Fig. 2B), although many of these were negatively enriched by CPS immunization, which seems contradictory. Some of the BTMs associated with protection in postimmunization samples were already positively associated with protection in preimmunization samples (Fig. 2B) such as the ones related to DC activation, monocytes or inflammation, and TLR signaling. When looking at the FC in the expression of genes upon PfRBC stimulation relative to uRBC background, we found that although individuals had many genes more down-regulated after immunization than before immunization, protected immunized volunteers displayed consistently less down-regulation than unprotected immunized volunteers (Fig. 3).

Fig. 3 Fold-change gene expression in CPS volunteers for genes from BTMs negatively enriched upon immunization and positively enriched in protected compared to unprotected immunized volunteers.

Charts show the log2 fold change expression upon PfRBC stimulation relative to uRBC background for genes that are found in the leading edge of enrichment in the GSEA analysis in each represented module: (A) BTM “myeloid cell–enriched receptors and transporters (M4.3),” (B) BTM “regulation of antigen presentation and immune response (M5),” (C) BTM “enriched in monocytes (M11.0),” and (D) BTM “TLR and inflammatory signaling (M16).” Different colored lines represent volunteers before immunization (green) and CPS immunized protected (pink) and unprotected (blue) volunteers.

We also performed the gene set analyses using Kyoto Encyclopedia of Genes and Genomes (KEGG) (23), BioCarta, Reactome (24), and another set of modules defined by Chaussabel et al. (25). Results (data file S4) are also consistent with the identified BTM responses (Fig. 2). For instance, for CSP stimulations, gene sets positively enriched with immunization included IFN responses (Reactome IFN-γ signaling, IFN-αβ signaling, IFN signaling, or regulation of IFN-γ signaling; Chaussabel’s modules M1.2, M3.4, and M5.12) or responses related to T cells [BioCarta cytotoxic T cells, IL-22BP (interleukin-22BP), TH1TH2, or IL-2RB pathways]. These gene sets were negatively enriched in protected compared to unprotected CPS-immunized volunteers. For PfRBC stimulation, there were also many enriched gene sets similar to the BTM identified, e.g., those related to inflammatory responses that were positively associated with protection or Chaussabel’s monocyte module M4.14, which was negatively enriched upon CPS immunization but positively enriched in protected versus unprotected immunized individuals.

We also performed a weighted gene correlation network analysis (WGCNA) (26) to identify groups of coexpressed genes and applied GSEA to ranked genes according to the differential correlation. The results were consistent with the GSEA analysis applied to differentially expressed genes, particularly for CPS immunogenicity upon CSP stimulation (data file S5).

Gene sets associated with RTS,S/AS01E immunization and clinical protection

As expected, gene set analysis of RTS,S immunogenicity revealed significantly enriched BTMs in response to CSP (q ≤ 0.1; Fig. 4A) but almost no BTMs in response to PfRBC. CSP stimulation in RTS,S vaccinees relative to comparators induced innate and acquired responses, including DC activation, NK cells, inflammatory/TLR/chemokines, IFN/antiviral sensing, T cells, and cell cycle BTMs. However, it also induced the repression of BTMs mainly related to inflammatory/TLR/chemokines, signal transduction, and monocytes. Similar BTMs were also repressed in PfRBC recall responses in CPS immunization compared to preimmunization.

Fig. 4 Transcriptional responses associated with RTS,S/AS01 immunogenicity and protection in RTS,S/AS01 and comparator vaccinees.

Each square represents a BTM. The color shading indicates normalized enrichment scores obtained by GSEA analysis for BTMs. Assignment of a BTM to a high-level annotation group is illustrated by a colored sidebar. GSEA, CAMERA, and Tmod were run with genes ranked by the expression of (A) RTS,S/AS01E vaccinees (after immunization) relative to comparator vaccinees and (B) protected relative to nonprotected individuals in RTS,S/AS01E and comparator vaccinees for CSP recall stimulations and (C) for PfRBC recall stimulations. Modules that did not reach the significance cutoff of FDR q value of 0.1 in all three enrichment methods or a minimum of 10 matched genes were eliminated. Modules without annotation are not shown. Modules that represent common associations of both immunogenicity and protection are highlighted with a circle for CSP recall responses and a triangle for PfRBC recall responses, filled symbols when enrichment had the same direction, and empty symbols when enrichment had the opposite direction.

For CSP stimulation in RTS,S vaccinees, BTMs related to DC activation, cell cycle, complement, inflammatory/TLR/chemokines, IFN/antiviral sensing, monocytes, and neutrophils were positively associated with protection from clinical malaria. Of these BTMs, the ones related to DC activation, cell cycle, and IFN/antiviral sensing were associated with RTS,S immunogenicity and were not found in comparator vaccinees (Fig. 4B), suggesting vaccine (antigen) specificity in these responses. Curiously, BTM related to inflammatory/TLR/chemokine responses, monocytes, and neutrophils that were associated with protection in RTS,S vaccinees were found to be inversely associated with RTS,S immunogenicity.

In comparator vaccinees, additional BTMs were associated with protection, suggesting innate and naturally acquired responses in these children. Positively enriched BTMs were related to DC activation, cell cycle, signal transduction, and monocytes. Three different T cell BTMs were negatively associated with protection.

Despite the scarce significant associations of BTMs with RTS,S immunogenicity for PfRBC recall responses, some BTMs were negatively associated with protection in RTS,S vaccinees (Fig. 4C). These BTMs were mainly related to cell cycle and inflammatory/TLR/chemokine response and included “proinflammatory cytokines and chemokines (M29),” which was also inversely correlated with vaccination. These responses could be related to the decrease of PfRBC exposure due to the protection induced by the vaccine in RTS,S vaccinees. Instead, comparator vaccinees had some BTMs positively correlated with protection (Fig. 4C) that could be related to naturally acquired responses.

As expected, most of the genes in the leading edge of the BTMs associated with protection and RTS,S immunogenicity were more up-regulated in protected compared to unprotected RTS,S vaccinees (Fig. 5A). When looking at the FC of genes in BTMs negatively enriched in RTS,S vaccinees upon CSP stimulations but positively associated with protection, such as “enriched in monocytes (M11.0),” we found that despite vaccination-induced down-regulation of genes, protected RTS,S vaccinees had less down-regulation than unprotected RTS,S vaccinees (Fig. 5B). Thus, this vaccine-induced down-regulation may be negative for protection. These BTMs were associated with protection regardless whether children or infants had been vaccinated with RTS,S or the comparator vaccine, and we had found similar associations with protection for PfRBC recall responses in CPS volunteers (both before and after immunization). Therefore, these responses seem vaccine nonspecific.

Fig. 5 Higher fold change gene expression in protected than unprotected RTS,S vaccinees for genes from BTM related to IFN signatures and monocytes.

Charts show the log2 fold change of genes upon CSP recall stimulation related to DMSO background for genes that are found in the leading edge of enrichment in the GSEA analysis in the BTMs related to (A) IFN signatures (M67, M75, M111.1, M127, and M150) and (B) the genes contributing to enrichment of the BTM “enriched in monocytes (II) (M11.0).” Different colored lines represent comparator vaccinees (green), protected (pink), and unprotected (blue) RTS,S/AS01E vaccinees.

Gene set analyses using alternative data files to BTMs resulted in the identification of similar responses (data file S4). For instance, for CSP stimulations, RTS,S vaccinees had enriched gene sets related to IFN responses (e.g., Reactome “IFN-γ signaling” or “IL-2 signaling” and Chaussabel’s modules M1.2, M5.12, or M3.4), T cells (BioCarta “cytotoxic T cells” or “TH1TH2 responses”), or cell cycle (BioCarta, Reactome, and KEGG “cell cycle”) that reflect antigen-specific responses. Some of Chaussabel’s modules were negatively enriched, mainly related to inflammatory responses (M3.2 and M4.13) and monocytes (M4.14), which were also consistent with the results obtained with BTMs. Curiously, only six gene sets were positively enriched in protected versus unprotected RTS,S vaccinees for CSP stimulations, but these included the Chaussabel’s IFN modules, as well as the Reactome IFN-γ signaling and IFN-αβ signaling sets, reinforcing the importance of these responses on protection.

Similarly, when we performed GSEA of WGCNA, we found four BTMs significantly (q ≤ 0.1) enriched in RTS,S versus comparator vaccinees for CSP stimulations, all related to IFN/antiviral responses (data file S5). Regarding RTS,S protection, for CSP stimulation, two gene sets were identified and were related to recognition of antigen by B cells (BioCarta “BCR signaling pathway” and Reactome “antigen activates B cell receptor, leading to generation of second messengers”), which is interesting given the role of IgG antibodies to CSP in RTS,S-induced protection (7). This could indicate that there are CSP-specific memory B cells in samples from RTS,S vaccinees.

Changes in frequencies of leukocyte populations

We performed leukocyte phenotyping (fig. S8A) to identify changes in the frequencies of cell subsets due to CPS or RTS,S immunization and between protected and nonprotected volunteers, because these changes may drive, in part, the observed differences in module responses. Consistent with previous observations (9), we detected an expansion of γδ T cells after CPS immunization (adjusted P < 0.001) and higher frequencies of B cells in protected than in nonprotected CPS-immunized volunteers (adjusted P = 0.031) (fig. S8B). In the RTS,S study, there were no statistically significant changes between RTS,S and comparators or between protected and nonprotected volunteers (fig. S8C).

Predictive transcriptional signatures of immunization-induced protection

We aimed to identify transcriptional signatures that predict protection in each immunization strategy separately using a previously developed systems biology approach (27, 28). This approach consisted of a network-based analysis, which overlays the gene differential expression values on its corresponding protein in a human protein functional network. This analysis increases the signal-to-noise ratio and takes into consideration the interactions among multiple components to obtain a more robust and biologically relevant output. In addition, the network collects previously published biological data on the proteins in relation to the immune response to malaria, reducing the solution space and focusing the analysis on biologically meaningful regions. The network-based analysis produced a list of proteins that behaved differently in protected and unprotected individuals, and signatures of protection, consisting in combinations of three to five of these proteins, were identified through data science analysis.

We identified several signatures of protection not only after immunization but also before immunization, when using CSP or PfRBC stimulations as recall response and for both immunization strategies. Table 1 shows the best signature for each experimental condition for the CPS and RTS,S immunization studies based on generalization capabilities and predictive accuracies calculated through leave-one-out (LOO) cross-validation. Accuracy of the signatures corresponds to the proportion of samples correctly classified, whereas generalization capability corresponds to the probability to correctly predict a sample that was not used in the training sets of LOO cross-validation. For the CPS top signatures shown, generalization capabilities ranged from 71 to 90% and accuracies from 71 to 100%. For the RTS,S top signatures shown, generalization capabilities ranged from 79 to 83%, and accuracies ranged from 84 to 100%. Many of the genes from these signatures encode proteins involved in immune response with different functions. CPS protection signatures included genes encoding for cytokines [CXCL10, CXCL16, colony-stimulating factor 2 (CSF2) or granulocyte-macrophage colony-stimulating factor (GM-CSF), transforming growth factor–β3 (TGF-β3), and IL-7], receptors [F2RL2, IL-3RA, and TNFRSF11B (tumor necrosis factor receptor superfamily member 11B)], cell signaling/signal transduction [F2, KL (Klotho), G protein subunit alpha 11 (GNA11), protein tyrosine kinase 2 (PTK2), calcium voltage-gated channel auxiliary subunit alpha2delta 2 (CACNA2D2), CACNA1F, rho GDP dissociation inhibitor beta (ARHGDIB), potassium voltage-gated channel subfamily B member 1 (KCNB1), and adrenoceptor alpha 2B (ADRA2B)], an inhibitor of nuclear factor κB (NF-κB) transcription factor [NFκBIE (NF-κB inhibitor epsilon)], chaperones [heat shock protein family A member 8 (HSPA8) and calreticulin (CALR)], and cell adhesion molecules (ITGA2 and ITGB7). Some of these genes were found in BTMs associated with CPS-induced protection: CXCL10 in “chemokine cluster (I) and (II) (M27.0 and M27.1),” “antiviral IFN signature (M75),” and “innate activation by cytosolic DNA sensing (M13)”; CXCL16 in “DC surface signature (S5)” and “antiviral sensing and immunity/IRF2 targets network (I) (M111.0)”; IL3RA in “activated (LPS) DC surface signature (S11)”; and PTK2 in “immune activation-generic cluster (M37.0),” although PTK2 was not found at the leading edge of the GSEA enrichment. The other genes did not appear in the identified BTMs. In the top RTS,S signatures, there were genes encoding for cytokines (IL-18, IL-3, and IL-12B), cell receptors [PRLR (prolactin receptor), TLR4, FCGR2A, and ciliary neurotrophic factor receptor (CNTFR)], apoptosis [caspase 6 (CASP6) and tumor necrosis factor receptor superfamily member 10A (TR10A)], cell cycle [cyclin E2 (CCNE2)], cell signaling/signal transduction {Erb-B2 receptor tyrosine kinase 2 (ERBB2), FAK2 (focal adhesion kinase 2), GNAT3, serum/glucocorticoid regulated kinase 1 (SGK1), mitogen-activated protein kinase kinase kinase 7 (M3K7), Cbl proto-oncogene C (CBLC), JAK2 (janus kinase 2), GNA11, P85B (PIK3R2, Phosphoinositide-3-Kinase Regulatory Subunit 2), SH2B2 (SH2B Adaptor Protein 2), WASL (WASP Like Actin Nucleation Promoting Factor), MAP2K7 [mitogen-activated protein kinas (MAPK) kinase 7], and IKKB [inhibitor of NF-κB (IκB) kinase B]}, and cell adhesion molecules [ITB1 (integrin beta 1), NCAM1 (neural cell adhesion molecule–1), and VCAM1 (vascular cell adhesion molecule–1)]. Some of these genes were found in BTMs associated with RTS,S-induced protection: SGK1 in “immune activation–generic cluster (M37.0)” and “enriched in monocytes (II) (M11.0)”; TLR4 in “TLR and inflammatory signaling (M16)” and “enriched in myeloid cells and monocytes (M81)”; and IL-18 in “enriched in activated DC (M165),” although only the latter was found in the leading edge of the enrichment.

Table 1 Signatures of protection induced by CPS and RTS,S/AS01E identified from mathematical network models based on the transcriptional data.

View this table:

Prospective validation of signatures of protection

Each of the experimental conditions and immunization strategies provided a list of signatures with good accuracies and generalization capabilities, not only the ones shown in Table 1. We applied a prioritization algorithm to select a reduced number of proteins, with the aim of enhancing the probability of identifying signatures that predict protection in the validation and covering most of the conditions analyzed. We chose 70 genes from the CPS study and 74 genes from the RTS,S study (data file S6), 13 of which were in both lists (MACROD2, SGK1, MEN1, SMAD6, TYROBP, F2, CXCL4, CXCL10, JAM2, YOD1, CD276, RMDN2, and KL). The list of selected genes was used for validation through qRT-PCR (Fig. 6A) on an additional set of 15 volunteers (120 samples) from a second human CPS immunization trial and an independent set of 23 children and 10 infants from the RTS,S phase 3 trial (41 samples, only for CSP recall responses and with some different experimental conditions, which we cannot exclude to have an impact on gene expression). We validated 32 and 37 protein combinations (signatures) with accuracies of >70% for CPS and RTS,S (data file S7), respectively, using the focused set of validation genes.

Fig. 6 Validation of signatures of protection.

(A) Samples for validation were obtained from a CPS trial in malaria-naïve adults that used chloroquine (CQ) and mefloquine (MQ) and a separate group of children and infants from two different African countries from the RTS,S phase 3 trial. PBMCs were similarly stimulated in vitro with CSP peptide pool and DMSO, or in only CPS samples, PfRBC, and uRBC. Gene expression was measured by qRT-PCR from about 70 genes involved in the signatures of protection of each immunization strategy and that were selected on the basis of Greedy algorithms. Normalized qRT-PCR data and data science methods were used to validate 32 and 37 previously identified signatures for CPS and RTS,S, respectively. Predictive mathematical models were developed using artificial neural network (ANN) and LOO cross-validation (LOOCV) for three selected signatures. (B) Potential mechanisms by which the genes of the three selected validated signatures may confer protection against malaria. Order of signatures: CPS preimmunization signature, CPS postimmunization signature, and RTS,S postimmunization signature. Purple and orange nodes indicate genes that belong to the signature (purple and orange represent up-regulated and down-regulated genes, respectively). White nodes indicate genes participating in the mode of action of the signature. Broken-lined nodes contain more than one gene, all of them acting in the mode of action in the same way. Purple arrows show activation, whereas orange lines show inhibition. NFAC1, nuclear factor of activated T-cells, cytoplasmic 1; VWF, von willebrand factor. VEGFA, vascular endothelial growth factor A. (C) ROC curves predicting malaria protection (based on the qRT-PCR validation data) for the selected CPS pre- and postimmunization and RTS,S postvaccination signatures of protection.

For CSP stimulation, two of the originally top performing signatures for CPS immunization (Table 1) predicted the volunteers from the validation set who were going to be protected before (CSF2, NFkBIE, TNFRSF11B, and ITGA2) and after CPS immunization (F2, CXCL10, and KL) with 100% accuracies. Some genes were found in several pre- and postimmunization CPS signatures such as CXCL10 and F2RL2 among others. The two CPS signatures of protection above indicate a prominent role of the NF-κB canonical pathway and an activation of the noncanonical pathway (Fig. 6B). ITGA2 (CD49b) is an integrin that elicits several cellular signals, including the p38 MAPK cascade (29) that controls the transcriptional activity of RelA (NF-κB p65) (30). Further contributing to the canonical NF-κB pathway, down-regulation of NFκBIE (or IkBε) may sustain the activation of NF-κB (31). GM-CSF (CSF2) signaling may also activate the canonical NF-κB pathway (32) through IKK that leads to the degradation of the NF-κB inhibitor IκBα (32). NF-κB is a nuclear factor that, when activated through the canonical pathway, induces an inflammatory response and is involved in innate and acquired immune responses. Down-regulation of osteoprotegerin (TNFRSF11B or TR11B), which inhibits the activation of the noncanonical NF-κB pathway through receptor activator of NF-κB-receptor activator of NF-κB ligand (RANK-RANKL) binding (33), would result in activation of the noncanonical NF-κB pathway. In addition, GM-CSF (CSF2, a target gene of NF-κB) (34) is important for development and maturation of DCs (35) that are key for antigen presentation upon vaccination and infections and is critical for T cell functions. GM-CSF also increases IL-2R (36), potentiating IL-2 signaling, which, in turn, promotes the expression of IL-2Rα subunit through signal transducers and activators of transcription 5B (STAT5B) (37) in a positive feedback loop that can increase expansion of lymphocytes and IFN-γ production (38). Therefore, GM-CSF would lead to a better vaccine take or enhanced protective responses upon vaccination because GM-CSF is related to IFN-γ and other T helper 1 cell (TH1) responses (35). Last, p38 MAPK also activates forkhead box O (FOXO) transcription factors (39, 40), among other proteins. FOXO is involved in regulating immune responses to inflammation, in response to growth factors and T cell–specific functions (41). FOXO3 has been associated with susceptibility to severe malaria (42). The mode of action of the postimmunization signature indicates that F2 (protothrombin) results in thrombin by enzymatic cleavage. Thrombin activates the canonical NF-κB pathway (43), has a proinflammatory effect, induces leukocyte recruitment, and stimulates chemotaxis and cytokine production by DCs (44, 45). Thrombin also promotes the activation of platelets through binding of von Willebrand factor (46). In addition, activation of platelets is marked by an increase in vascular endothelial growth factor, which can induce the secretion of GM-CSF that activates the canonical NF-κB pathway. CXCL10 is a target gene of NF-κB induced by IFN-γ that stimulates monocytes and NK cells, modulates T cell migration through modulation of adhesion molecules, and activates CXCR3 (47) that is implicated in T cell homing to the liver (48). KL is a transmembrane protein involved in endoplasmic reticulum stress (49), calcium homeostasis, modulation of inflammation, and attenuation of the activation of the canonical NF-κB pathway (50). Therefore, KL down-regulation in protected individuals could result in increased activation of NF-κB.

For RTS,S, one postimmunization signature for children and one for infants that were among the best signatures obtained initially (ERBB2, PRLR, FAK2 and GNAT3, SGK1, and TLR4; Table 1) were validated with accuracies of 90 and 96%, although GNAT3 could not be detected in any validation sample, probably due to the different experimental conditions (e.g., shorter duration of the in vitro stimulations; see Materials and Methods). Because of the lack of detection of GNAT3 for the infants’ signature and because the target population of RTS,S/AS01E is children, we selected the postimmunization signature of this age group. The mode of action of the signature (ERBB2, PRLR, and FAK2) also indicates a role of the canonical NF-κB pathway (Fig. 6B). Signaling through the kinase ERBB2 and the PRLR and inhibition of FAK2 could prevent the inhibition of p50 (NFKB1) by catenin beta 1 (CTNB1) (51), increasing NF-κB activity. However, ERBB2 may provide protection by increasing the rate of hypoxia-inducible factor 1A synthesis (52) that activates transcription of hypoxia-inducible genes, including heme oxygenase 1 (HMOX1) (53), which, in malaria and other infections, is induced to limit iron availability and to provide other anti-inflammatory functions (54), possibly contributing to the acquisition of malaria protection. Furthermore, ERBB2 activation might lead to the transcription factor STAT3 activation through JAK3 (55). STAT3 participates in the induction of several molecules involved in immune response, such as IL-6 (56, 57). In addition, STAT3 might up-regulate FOXO1/FOXO3 expression (58) and further contribute to protective responses as mentioned before. The activation of JAK3 also results in phosphorylation of STAT5B (59, 60). Last, ERBB2 can also dimerize with epidermal growth factor receptor (EGFR) (61) that might also promote the activation of STAT3 (62, 63) and STAT5B (64). On the other hand, PRLR activation induces JAK2 (65), which can phosphorylate EGFR (66). Last, reduced activity of FAK2 might decrease the RAC1, RAC2, and RAC3 activation (23), which can dampen MAPK signaling (MK01 and MK03) (23), allowing higher FOXO1 and FOXO3 expression (23, 39, 40). MAPK are also involved in SMAD2 and SMAD3 inhibition (67); thus, reduced MAPK signaling pathways might increase SMAD2 and SMAD3 activity, which are involved in the regulatory T cell differentiation process to avoid inflammation exacerbation in parasitic infection (68, 69). Reduced activity of FAK2 can also diminish the FAK2-induced activation of SRC (proto-oncogene tyrosine protein kinase Src) (70). Since SRC is involved in the activation of MAPK signaling (23) and CTNB1 (71), its reduced activation might contribute further to the inhibition of CTNB1 and MK01 and MK03.

Some genes appeared in several RTS,S signatures. The most frequent ones were TLR4, present in 24.3% of the signatures, SGK1 in 18.9%, MAPK13 in 8%, and EREG in 6%, the latter only in preimmunization signatures, whereas the others were found in both pre- and postimmunization signatures (data file S7).

With the selected validated signatures, we constructed predictive models by artificial neural networks (table S1), which are flexible and can easily provide mathematical formulas to be used as tools in clinical trials. Tables S2 and S3 show the validation qRT-PCR values and prediction results using the models. Area under the curve (AUC) in receiver operating characteristics (ROC) curves of both CPS signature models was excellent, with an AUC of 1 (Fig. 6C), and the ROC curve of the RTS,S signature model also showed a remarkable predictive capacity, with an AUC of 0.8304 (Fig. 6C).


We identified and validated genes and signatures that predict vaccine-induced protection in naïve adults and young African children from different countries, before and after immunization, with high generalization capabilities and accuracies, using samples from a series of malaria vaccine trials of two different malaria immunization strategies targeting the preerythrocytic phase of P. falciparum. Although the identification of predictive signatures of protection in samples collected before immunization for both immunization strategies was unexpected, previous studies have also found baseline predictors of vaccination responses (72, 73) and RTS,S efficacy (74). In addition, African infants and children could have malaria-specific responses at baseline as they may be previously exposed to Plasmodium in utero and during the first months of life, although malaria transmission intensity at that time for the three sites was low/moderate (3). In contrast, CPS volunteers were true malaria-naïve adults. The genes of the preimmunization signatures in both immunization regimens and the BTMs associated with CPS protection at preimmunization reveal innate, inflammatory responses and cell activation rather than antigen-specific memory responses. Thus, CSP and PfRBC recall responses at baseline could reflect the intrinsic capacity of the volunteers to respond effectively to immunization (vaccine take). This is of utmost importance because the absence of such signatures of protection at baseline could help identify individuals who require, for instance, a higher immunization dose to become protected [we know that all CPS volunteers likely would have been protected if immunized by optimal doses (8, 9)] or an additional immunomodulatory stimulus or different adjuvants.

The genes and signatures of both CPS immunization and RTS,S vaccine provided insights into the molecular mechanism necessary to induce malaria protection through immunization. The two selected CPS signatures of protection (CSF2, NFkBIE, TNFRSF11B, and ITGA2 and F2, CXCL10, and KL) indicate a prominent role of the NF-κB canonical pathway in priming the immune system for a protective vaccination response and an activation of the noncanonical pathway via the RANK-RANKL system. The RANK-RANKL system is expressed in a wide variety of tissues, including T cells, DCs, and memory B cells (75, 76). NF-κB has a central role in inflammation, inducing the expression of proinflammatory molecules, including cytokines and chemokines, and regulates the activation and differentiation of innate and acquired immune cells. The canonical pathway leads to acute responses, whereas the noncanonical NF-κB activation leads to sustained responses. We speculate that higher NF-κB activity leads to a more efficient immunization response, and NF-κB–induced cytokines, such as IFN-γ, TNF and GzmB (9, 77), may directly contribute to malaria-protective responses. The correlation network analysis of CPS-induced protection also revealed some gene sets related with the mode of action of the signatures and NF-κB, e.g., Reactome “TAK1 activates NFKB by phosphorylation and activation of IKKS complex,” BioCarta “RELA pathway,” or the BioCarta “RANKL pathway.” In addition, we found the BioCarta pathways “protothrombin (TPO) pathway” and “platelet-derived growth factor (PDGF) pathway,” which are related to F2 found in the postimmunization CPS signature. Recently, platelets have been shown to control Plasmodium parasitemia through killing of intraerythrocytic parasites (78).

The RTS,S postvaccination signature of protection (ERBB2, PRLR, and FAK2) and other genes found in the other signatures (e.g., TLR4) also indicates an important role of the canonical NF-κB pathway and activation of STAT3 and STAT5B transcription factors. These transcription factors are also involved in the expression of inflammatory cytokines, such as IL-6, and receptors (37, 56) and differentiation and generation of TH17, T follicular helper (TFH) cells, and B cells (79, 80). Studies in murine models have shown that IL-6 mediates protective immunity against the preerythrocytic stages of malaria by inducing IL-1β and TNF, and during the erythrocytic stage of disease by controlling parasitemia through boosting of specific IgG antibodies (81). TLR4 is an innate receptor that recognizes pathogen-associated molecular patterns such as lipopolysaccharide (LPS) and induces proinflammatory responses through the canonical NF-κB pathway and type I IFN responses through interferon regulatory factor 3 (IRF3) (both responses found in the BTMs associated with protection) (82). It has also been shown that LPS induces the noncanonical NF-κB pathway (83). The AS01E adjuvant contains monophosphoryl lipid A (MPL), which is a derivative of LPS and is therefore recognized by TLR4 (84). Thus, higher expression or activation of TLR4 in protected individuals could be involved in a better response to RTS,S/AS01E vaccination. In another systems biology study, TLR5 expression was associated with the magnitude of antibody responses to an inactivated influenza vaccine, and it was shown that TLR5-gut microbiota sensing was necessary for antibody responses to this and the inactivated polio vaccine (85). Thus, the presence of TLR4 in several preimmunization RTS,S signatures points to this receptor as a possible target to increase vaccine take, at least for TLR4-targeting vaccine adjuvants. Previous or simultaneous treatments that increase TLR4 expression and signaling to all or selected infants and children could increase vaccine take and efficacy.

CPS-induced protection is mediated by preerythrocytic immunity (8), although the underlying immunological mechanisms are not known. In previous studies, CD107a+ CD4+ T cells and CD8+ T cells producing GzmB upon PfRBC stimulation were associated with protection upon PfRBC stimulation (9), whereas IFN-γ+ CD4+ T cells and γδ T cells, as well as CSP and MSP1-specific antibody responses, were associated with exposure during immunization but not associated with protection (9, 21). Thus far, CPS-specific T cell responses to the hallmark preerythrocytic antigen CSP have not consistently been detected by either flow cytometry or enzyme-linked immune absorbent spot. Our differential gene expression analysis reporting no CSP responses and stronger PfRBC responses is consistent with these data. Therefore, the negative association of BTMs with protection upon CSP recall responses, which seems contradictory, could be due to noisy responses because of low signal to background ratios. In contrast, BTMs induced upon PfRBC stimulations related to signaling in T cells, DC activation, and signal transduction were found to be positively associated with protection in immunized volunteers and not before immunization and are thus in agreement with an adaptive malaria-specific CPS-induced response. Therefore, only modules relevant to protection against late-liver (and blood-stage) infection seem to be induced by CPS immunization, whereas no modules relevant to protection against the sporozoite infection appear to be elicited.

Regarding RTS,S, previous studies mainly showed antibody responses (7) and some T cell responses (86, 87), with only weak associations of cell responses with protection. We found many BTMs correlating with protection in RTS,S vaccinees and not in comparators that were associated with immunogenicity, suggesting vaccine-induced antigen-specific responses. Most of these BTMs were related to DC activation, cell cycle, and IFN/antiviral sensing. The genes of these BTMs were comprised in the whole blood IFN signatures that were associated with RTS,S vaccination and protection in a trial with naïve adults (18). Another analysis of the same dataset also found IFN-γ signaling and NF-κB pathways associated with RTS,S-induced protection (19). RTS,S/AS01 studies have shown no antigen-specific production of IFN-γ from CD4+ T cells (86), but IFN-γ could be produced by NK cells via bystander activation through CD4+ T cells producing IL-2 (88). In another RTS,S trial with malaria-naïve adults, Kazmin et al. (12) found BTMs of B and plasma cells positively associated with immunogenicity and protection, and NK cells negatively associated with both protection and antibody responses. We also detected RTS,S-induced responses in some NK BTMs, although they were not associated with protection. Our experimental strategy is different from the previous studies in which they performed transcriptomics in whole blood or PBMCs without stimulation, directly ex vivo, and early after vaccination. In contrast, we used in vitro recall responses at 1 month after the third vaccination, and we corrected for background responses for each individual using the same number of cells.

Some BTMs associated with protection in RTS,S vaccinees were also found in comparator vaccinees, particularly BTMs related to inflammatory/TLR/chemokine responses and monocytes. This probably reflects non–vaccine-specific responses. Most of these particular BTMs were the ones negatively enriched in RTS,S compared to comparator vaccinees. Similar to RTS,S, we found repression of BTMs of inflammatory/TLR responses and monocytes, and additionally, antigen presentation and DC activation for CPS immunization. However, as in RTS,S, many of these BTMs were positively associated with protection. Thus, despite down-regulation by immunization, protected immunized individuals had less down-regulation than unprotected immunized volunteers. Many of these BTMs were also found before immunization (as in comparators for the RTS,S study), suggesting protection mechanisms independent of immunization. For instance, protected individuals may have monocytes of higher capacity to mediate Fc receptor/antibody-dependent responses, leading to better control of infection. This finding highlights a potentially crucial role of monocytes in malaria protective responses.

Limitations of our study include the lack of additional cohorts to prove the universal predictive capacity of the identified signatures. Additionally, the transcriptomic analysis was conducted on in vitro stimulations instead of ex vivo whole blood. This experimental design may introduce more variability and difficult the interpretation of the identified signatures, particularly the ones at baseline. However, it allows to attribute differential gene expression to the antigen-specific recall response, and we are more certain that differences in gene expression were less likely due to basal differences, in cell subset counts for instance, than in an ex vivo approach without stimulation. Nevertheless, we did observe some changes in cell subsets upon CPS immunization and between protected and nonprotected CPS-immunized volunteers, which could affect in part some of the module responses. A related limitation is the modest acquired cellular responses induced by the immunizations, particularly for RTS,S. Weak CSP-specific cell responses (87) and low frequency of CSP-specific CD4+ T cells were detected in RTS,S vaccinees in the phase 3 trial using similar stimulation conditions (86), which is probably the reason behind the absence of statistical significance in the differential gene expression analysis. Despite the overall weak differential responses elicited upon in vitro recall, we have obtained meaningful responses through both gene set analyses and the systems biology approach. The fact that we did not detect a differential expression in the genes encoding the proteins from the signatures of protection obtained with the network-based analysis is not unexpected, because the network-based analysis can highlight other proteins that are highly related to differentially expressed genes, despite not being found significant by conventional differential gene expression analysis.

Our results confirm the potential and feasibility of systems biology approaches to identifying molecular signatures of protection against complex diseases in field studies, where individuals are under heterogeneous conditions, and even in vulnerable populations such as children. The finding that BTMs induced by CSP recall responses and associated with RTS,S protection in infants and children from endemic areas were also found in PfRBC recall responses associated with protection in the CPS trial in malaria-naïve adults suggests potential for global protective responses to malaria. Preimmunization signatures and BTMs associated with protection could indicate that the physiological status of the host at baseline could determine vaccine efficacy and infection outcomes. Therefore, our results could be valuable for identifying upfront low or nonvaccine responders and finding strategies to make nonresponders more reactive and hence improve vaccine take and protection. In addition, the postvaccination signatures of protection identified could be useful for vaccine clinical trials as surrogate markers to predict protection.


Detailed Supplementary Materials and Methods are available in the Supplementary Materials.

Study design

We used samples from malaria-naïve adults from a CPS clinical trial (9) and children and infants (two different age cohorts) from three African countries participating in the RTS,S/AS01E phase 3 clinical trial (MAL055, trial registered with, number NCT00866619) (3). Samples from all individuals participating in the CPS clinical trial were included (fig. S1). Samples from three sites and two age cohorts from the RTS,S/AS01E phase 3 trial were selected. Specifically, all malaria cases detected 1 to 12 months after vaccination from whom at least two aliquots of PBMC samples were available 1 month after or before vaccination were included (fig. S2). Malaria cases were defined as individuals who sought care at a health facility and had any P. falciparum asexual parasitemia by blood smear. Up to four controls were selected for each case based on site and age group. At the time of sample selection, all study investigators were blinded to vaccination, and selection was performed randomly. As expected, per MAL055 trial design, about two-third of children had received RTS,S/AS01E vaccination, and one-third had received the comparator vaccine. Sample sizes and study design are detailed in Fig. 1, figs. S1 to S3, and the Supplementary Materials. Previously cryopreserved PBMCs were stimulated in vitro with CSP peptide pool or PfRBC and their respective background controls, DMSO, and uRBC. RNA was extracted for gene microarray analysis. Differential gene expression analysis using linear regression models and subsequently GSEA was performed (see Supplementary Materials and Methods). Of 688 hybridized samples from the RTS,S study, 2 had no CEL data, and 31 were discarded because they were outliers (did not pass the quality check criteria). All microarray data were also used for protein network–based models. Proteins behaving differently in the models were down-selected using data science methods, leading to identification of pre- and postimmunization signatures of protection, consisting of three to five proteins, with accuracies of >70%. For validation of signatures of protection, samples from another CPS trial (10) and samples from independent children and infants participating in the RTS,S/AS01E phase 3 trial were used. The gene expression of 70 selected genes from the signatures of protection was measured by qRT-PCR. Among the validated signatures, three were selected for development of predictive models using artificial neural networks.


We used cryopreserved PBMCs obtained by density gradient centrifugation from venous whole blood. PBMCs were stimulated with P. falciparum NF54-infected RBC for 16 hours and with CSP peptides for 24 hours. uRBCs and DMSO were used as negative control stimuli for each individual and time point. RNA was extracted for microarray analysis. For RTS,S/AS01E transcriptional validation, RNA from 12-hour stimulations with CSP and DMSO performed freshly on site of blood collection were used. Remaining PBMCs were stained and acquired on a Cyan ADP nine-color flow cytometer (Beckman Coulter) at Radboud University Medical Center (RUMC) and an LSRII cytometer (BD Biosciences) at ISGlobal. Flow cytometry data were analyzed using FlowJo software. A representative gating strategy for the panel is shown in fig. S10A.

Gene expression microarrays and qRT-PCR

Human Gene 2.1 ST microarrays in 96-array plates (Affymetrix) and qRT-PCR (Fluidigm System Dynamics Array) for validation were performed using provider’s recommendations.

Statistical analysis

Wilcoxon signed rank and Wilcoxon rank-sum tests were used to compare cell frequencies between groups. Differential gene expression (DGE) analysis was performed using linear models using the limma package in R/Bioconductor (89), and empirical Bayes methods were used. Spearman correlations were used to assess the association of FC gene expression and total cumulative parasitemia during CPS immunization and the time of patency. For GSEA, CAMERA, and Tmod, analyses were performed using preranked analyses of DGE (all genes regardless of statistical significance) and were performed in R. Fisher’s exact tests were used to assess whether enrichments were due to genes relatively relevant in terms of differential gene expression. Network-based mathematical models were built for nonprotected and protected individuals for each of the cohorts using the Therapeutic Performance Mapping System (Anaxomics Biotech) (27, 28, 9096). To center the network analysis, the known key proteins with a relevant role in malaria immune responses were identified by a review of the scientific literature (data file S8). Proteins that behaved differently in the protected and nonprotected models were selected. Combinations of three to five proteins were selected after applying a data science analysis (different feature selection approaches and base classifier methods using LOO cross-validation). To select the best gene candidates among the identified signatures for prospective validation, an approach based on Greedy algorithms (97) was applied. The qRT-PCR data from the validation individuals were used to confirm the signatures of protection. The predictive mathematical functions developed for selected validated signatures were obtained through artificial neural networks. A LOO cross-validation strategy was used to assess the accuracy. P values were adjusted by the false discovery rate (FDR) as described by Benjamini and Hochberg and Bonferroni methods.


Materials and Methods

Fig. S1. Experimental design and sample sizes.

Fig. S2. CPS study profile and sample flow chart.

Fig. S3. RTS,S study profile and sample flow chart.

Fig. S4. PCA of CPS gene expression data.

Fig. S5. Heatmaps and unsupervised hierarchical clustering in gene expression after cell stimulation for all CPS immunized volunteers.

Fig. S6. PCA of RTS,S/AS01E gene expression data.

Fig. S7. Heatmaps and unsupervised hierarchical clustering in gene expression after cell stimulation for all RTS,S and comparator vaccinees.

Fig. S8. Frequencies of leukocyte subsets after resting of PBMCs.

Table S1. Predictive models obtained by artificial neural networks.

Table S2. Validation qRT-PCR values and prediction results for the selected CPS preimmunization signature of protection.

Table S3. Validation qRT-PCR values and prediction results for the selected CPS postimmunization signature of protection.

Table S4. Validation qRT-PCR values and prediction results for the selected RTS,S/AS01E postimmunization signature of protection.

Data file S1. Gene lists differential gene expression CPS immunogenicity and protection.

Data file S2. Gene list correlations PfRBC FC and CPS outcomes and frequencies of gene ontology biological processes terms and BTMs of genes correlated with CPS outcomes.

Data file S3. Gene lists differential gene expression RTS,S/AS01E immunogenicity and protection.

Data file S4. Summary of GSEA and differential gene expression results.

Data file S5. Summary of GSEA and WGCNA results.

Data file S6. Selected genes for validation for CPS and RTS,S/AS01E.

Data file S7. Postvalidation signatures of protection with accuracy data for CPS and RTS,S/AS01.

Data file S8. CSP 15-mer peptides, corresponding to CSP region of RTS,S and predicted CD4+ and CD8+ T cell epitopes.

Data file S9. Comparisons and limma contrasts.

Data file S10. Gene list for validation.

References (98133)


Acknowledgments: We thank all the trial volunteers and the staff from the Radboud University Medical Center, Ifakara Health Institute and Bagamoyo Research and Training Centre (IHI-BRTC in Tanzania), Centre de Recherches Médicales de Lambaréné (CERMEL, Gabon) and Manhiça Health Research Center, and Fundação Manhiça (FM-CISM, Mozambique), all of whom made this study possible. Specifically, we thank the principal investigators of the RTS,S/AS01 phase 3 trial: S. Abdulla, P. Alonso, J. Sacarlal, P. Aide, and E. Bijker, who lead the CPS trials. We also thank L. Puyol, D. Barrios, C. Amroune, and C. Hernández from ISGlobal for laboratory logistics and management support and K. Stuart and J. Carnes from the Center for Infectious Disease Research (CIDR, Seattle) who performed some RNA extraction for validation. We thank the GlaxoSmithKline Biologicals S.A. for support in the conduct of the RTS,S study. Funding: This work was supported by the EU FP7-founded SysMalVac project (grant number 305869-2); the Instituto de Salud Carlos III (ISCIII) (PI11 00423); the Netherlands Organisation for Health Research and Development (ZonMw) (project 95110086); and the Agència de Gestió d’Ajuts Universitaris i de Recerca AGAUR (2014SGR991). RTS,S sample collection was supported by the PATH Malaria Vaccine Initiative (MVI). C. Dobaño was recipient of a Ramon y Cajal Contract from the Ministerio de Economía y Competitividad (RYC-2008-02631). G.M. was recipient of a Sara Borrell–ISCIII fellowship (CD010/00156) and work was performed with the support of Department of Health, Catalan Government grant (SLT006/17/00109). ISGlobal is a member of the CERCA Programme, Generalitat de Catalunya. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Author contributions: G.M., A.S., J.J.C., L.O., L.P., J.M., E.M.P., C.H.M.K., C. Daubenberger, B.M., S.T.A., R.S., and C. Dobaño designed the studies and/or experiments. A.N., C.J., J.J.C., M.M., C. Daubenberger, B.M., S.T.A., R.S., G.M., and C. Dobaño performed the clinical studies and collected samples and clinical data. G.M., A.S., M.M., and A.B.H. conducted laboratory experiments. G.M., A.S., L.O., J.M., L.P., D.A., R.V., J.M.M., and M.D.-F. analyzed the data. H.S. and J.J.A. performed data management and sample selection. G.M., N.A.W., and C. Dobaño coordinated the study. G.M., C. Dobaño, R.V., and J.M.M. interpreted the data. G.M., A.S., and C. Dobaño wrote the manuscript. All authors read and approved the final manuscript. Competing interests: The authors declare that they have no competing interests. Data and materials availability: All data associated with this study are present in the paper or the Supplementary Materials. Microarrays data are available in the manuscript and at the Gene Expression Omnibus (GEO) database, (accession no. GSE144826). Code for the systems biology analysis is based on the Therapeutic Performance Mapping System technology proprietary of Anaxomics Biotech. Code used for the rest of analyses is available at GitHub (, DOI: 10.5281/zenodo.3630760).

Stay Connected to Science Translational Medicine

Navigate This Article