Malaria remains a leading cause of global morbidity and mortality, and disproportionately affects children <5 years old in the World Health Organization (WHO) African Region (World Health Organization, 2020). The causative agents of malaria, Plasmodium parasites (especially P. falciparum and P. vivax), are transmitted to humans through the bite of infected female Anopheles mosquitoes (World Health Organization, 2020). Although some effective prevention tools (e.g., long-lasting insecticidal nets) have been developed, a durably efficacious preventive vaccine will be required to meet the challenge of eradicating malaria (Wilson et al., 2019).

The RTS,S/AS01 vaccine (GSK) is the most clinically advanced vaccine to date (Laurens, 2020) and targets the pre-erythrocytic stage of P. falciparum by eliciting antibody and T-cell responses to circumsporozoite protein (CSP), the major surface protein of sporozoites (Hoffman et al., 2015). In a phase 3 trial in African infants and children, estimated vaccine efficacy (VE) of RTS,S/AS01E against clinical malaria over a follow-up time of 12 months post-final dose was 56% in young children (5–17 months old) and 31% in infants (6–12 weeks old) (RTS, S Clinical Trials Partnership, 2011; RTS, S Clinical Trials Partnership, 2012; RTS, S Clinical Trials Partnership, 2014; RTS, S Clinical Trials Partnership, 2015). RTS,S/AS01 obtained a positive scientific opinion from the European Medicines Agency in 2015 (Hawkes, 2015) and was recommended by the WHO for a pilot implementation program in Ghana, Malawi, and Kenya that initiated in 2019 (World Health Organization, 2019).

Controlled human malaria infection (CHMI) studies are a valuable tool for identifying correlates of protection, including in the context of vaccine development (Stanisic et al., 2018). MAL068 [NCT01366534 (Ockenhouse et al., 2015)] and MAL071 [NCT01857869 (Regules et al., 2016)] were randomized phase 2a studies (MAL068 observer blind, MAL071 open-label) designed to evaluate the safety, reactogenicity, immunogenicity, and efficacy of RTS,S/AS01 against sporozoite challenge of healthy, malaria-naïve adults. A schema of each study and a summary of the results are shown in Figure 1.

Each trial included a standard arm with three RTS,S/AS01 doses administered at one-month intervals (RRR arm), as well as an alternative arm with a modified version of this regimen (Figure 1). In MAL068, the alternative arm consisted of one dose of replication-incompetent, recombinant adenovirus 35 expressing circumsporozoite protein (Ad35.CS.01), followed by two doses of RTS,S/AS01 (ARR arm). In MAL071, the alternative arm was a delayed fractional-third-dose arm consisting of two doses of RTS,S/AS01 one month apart, and a one-fifth fractional dose 6 months later (RRr arm). At three weeks post-last vaccination, all participants were challenged with infected P. falciparum mosquitoes and monitored for parasitemia through Day 28 post-challenge. In the RRR arms, VE against malaria infection was 52% (95% CI 25–70%) in MAL068 and 63% (29–80%) in MAL071. In the alternative arms, VE against malaria infection was 44% (21–60%) in the ARR arm of MAL068 and 87% (67–95%) in the RRr arm of MAL071.

A large body of work (Ockenhouse et al., 2015; Chaudhury et al., 2016; Kazmin et al., 2017; Dennison et al., 2020; Du et al., 2020; Pallikkuth et al., 2020; Suscovich et al., 2020; Seaton et al. under revision) ¹ has been conducted using immune response data from MAL068 and/or MAL071 to identify correlates of RTS,S/AS01-mediated protection in malaria-naïve adults, yielding many insights and generating hypotheses about the mechanism of protection in this population. The immunodominant amino acid repeat region (NANP) of CSP is the main target of the B-cell response to CSP (White et al., 2015), and day-of-challenge anti-NANP repeat region IgG antibody titers have been shown to associate with protection (Ockenhouse et al., 2015). However, there does not appear to be a quantitative cut-off associated with protection (Ockenhouse et al., 2015; Suscovich et al., 2020), indicating that other immune responses also play a role as discussed further below.

Du et al. reported that the MX2/GPR183 expression ratio in peripheral blood mononuclear cells (PBMCs) at 1 day post-third dose discriminated protected from non-protected RTS,S/AS01 recipients in five independent CHMI trials, including MAL068 and MAL071 (Du et al., 2020). Specifically, participants with higher fold-change from baseline to 1 day after the third immunization were more likely to be protected. GPR183 (also known as EBI2), along with its ligand 7α,25-dihydroxycholesterol, plays an important role in T-cell dependent antibody responses by positioning different immune cell types at the appropriate compartments in secondary lymphoid organs such as the spleen (Barington et al., 2018). Upregulation of GPR183 helps localize activated CD4⁺ T cells, B cells, and dendritic cells to the T zone—B-cell follicle interface (Gatto and Brink, 2013; Li et al., 2016; Lu et al., 2017). In CD4⁺ T cells, GPR183 upregulation also augments T peripheral follicular helper cell (Tfh) cell development at early time points and is crucial for the generation of extrafollicular plasma cell responses (Gatto and Brink, 2013; Li et al., 2016). Relatedly, Pallikuth et al. showed that early (day 6 post-first dose) CSP-specific pTfh responses, along with durable CSP-specific memory B cell responses emerging post-dose two, were associated with protection in MAL071 (Pallikkuth et al., 2020). These findings are consistent with a model wherein localization of GPR183-expressing CD4⁺ T cells, B cells, and/or dendritic cells to the T zone—B-cell follicle interface (leading to a higher MX2/GPR183 ratio, as GPR183-expressing cells leave the peripheral compartment) is crucial for development of the protective CSP-specific pTfh and memory B cell responses seen by Pallikkuth et al. (2020).

Kazmin et al. reported an inverse correlation with protection of natural killer (NK) cell-related blood transcriptional module expression in PBMCs, in both arms of MAL068 (Kazmin et al., 2017), generating the hypothesis that vaccination stimulates migration of circulating NK cells to lymphoid tissues/organs such as the liver (i.e., leading to a reduction of NK cells in the peripheral blood) where the NK cells might contribute to the killing of infected cells through mechanisms such as antibody-dependent cellular cytotoxicity (ADCC).

Phagocytosis via opsonophagocytic pathways such as antibody-dependent cellular phagocytosis (ADCP) of parasites can have widely divergent outcomes [reviewed in Leitner et al. (2020)]. On the one hand, phagocytic cell uptake of opsonized parasites can result in parasite clearance by antibody-dependent cellular phagocytosis as well as immune-cell activation and secretion of inflammatory molecules. This outcome was mechanistically demonstrated in a recent “systems serology” analysis of >100 antibody features (Fc receptor binding, effector functions, etc.) of anti-CSP (full length), anti-NANP6, and anti-CSP C-terminal region antibodies in MAL068 vaccinees (Suscovich et al., 2020), which implicated both NANP6-specific ADCP and FcGRIIIA [expressed by natural killer (NK) cells (Vivier et al., 2008)] binding of NANP6-specific antibodies in protection (Suscovich et al., 2020). Similarly, opsonization by CSP-specific antibodies was proposed to contribute to RTS,S/AS01-induced protection (Schwenk et al., 2003).

On the other hand, parasites can also both evade and actively manipulate the host immune response via “phagosomal escape”, where phagocytosed parasites reside in the phagocytic immune cell and may also influence its function to undermine the host defense (Leitner et al., 2020). Chaudhury et al. reported that the “opsonization index” (corresponding to phagocytic activity normalized to anti-CSP antibody titer) was an inverse correlate of protection in MAL068 and that opsonophagocytic activity was likely mediated by C-terminal-specific antibodies (Chaudhury et al., 2016). A subsequent study by Chaudhury et al. presented evidence that, in MAL071, anti-CSP IgG4 antibodies inhibited opsonophagocytic activity of other anti-CSP IgG antibody subclasses. Subsequent analyses have shown that most opsonophagocytic antibodies are C-terminal-specific (rather than NANP6-specific) and that lower opsonophagocytic activity was associated with protection across all RTS,S/AS01 trials analyzed. The apparently conflicting findings regarding opsonophagocytic activity in RTS,S/AS01 vaccinees could be due to the different approaches taken: dividing CSP-targeted phagocytosis by CSP antigen titer (Chaudhury et al., 2016) vs. directly assessing CSP antigen-targeted phagocytosis (Suscovich et al., 2020).

Most recently, Seaton et al. examined a large number of CSP-specific antibody features (magnitude, subclass, and avidity) as humoral correlates of protection in the MAL068 and MAL071 trials, reporting that magnitude of NANP6-specific IgG1 antibody was the most predictive variable of protection in univariate models (Seaton et al., under revision) ¹ .

This substantial body of MAL068/MAL071 correlates work has included antibody, cellular, and transcriptomic assays performed by many groups across different research organizations. Papers have focused on limited subsets of data and have not integrated all available immune response data across the two studies, nor have they used the same approach or mathematical tools for assessing a given immune response measurement as a correlate of protection (CoP).

Given that CHMI trials are resource-intensive to conduct, it is worthwhile to mine the few valuable datasets that are currently available, using a variety of approaches and techniques. For the first time, here we considered the entire repertoire of available MAL068 and MAL071 assay data in a correlates analysis, including all immune response variables in a unified data analysis procedure. We focused on day-of-challenge data because of the broad availability of data at this time-point, in order to highlight the integration aspect across assays. We evaluated the performance of different predictive computational models, including univariate and bivariate, in predicting individual-level outcomes in the MAL068 and the MAL071 trials. In this analysis, we identify NANP6-targeted antibody dependent cellular phagocytosis (ADCP) and anti-NANP6 IgG1 antibody titer as consistently predictive immune response measurements, with higher levels of each biomarker on day of challenge associated with individual-level protection after controlled malaria challenge. Our findings suggest that univariate models perform as well or better (in terms of predictive probability) than bi- or trivariate models, at least in this dataset, and provide the first head-to-head comparison (where the same pre-processing and statistical methodology has been used) between the previously published CoPs corresponding to day-of-challenge immune response measurements.

Building upon the wealth of previous RTS,S/AS01 immune correlates analyses, here we took a predictive modeling approach that included immune response data across many different assays and assessed the contribution of each immune biomarker to prediction of individual-level outcomes of RTS,S/AS01 vaccinees in two different controlled human malaria infection studies. We identified univariate and bivariate models with good predictive accuracy across studies [validation area under the receiver operating characteristic curve (AUC) > 0.8], and our first major conclusion is that there was little to no improvement in the bivariate models compared to the univariate models, when combining all the day-of-challenge biomarker data into a comprehensive dataset. We hypothesize that the relatively small sample sizes, which in turn limited the size of the training/testing/validation sets, hindered the identification of well-performing complex models. Applying the same approach to future studies with larger sample size may enable such models to be identified. Another potential explanation for the lack of significant improvement in bivariate vs univariate models (when combining all day-of-challenge biomarker data into a comprehensive dataset) is the high correlation between variables, as highly correlated variables will generally not improve in logistic regression over single variable models.

We identified central NANP repeat region (NANP6)-targeted antibody dependent cellular phagocytosis (ADCP), binding of NANP6-targeting antibodies to the FcGRIIIa receptor, and anti-NANP6 IgG1 antibody titer as consistently predictive immune response measurements, with higher levels of each biomarker on day of challenge associated with individual-level protection after controlled malaria challenge. Our second major conclusion is thus that our data support a “quality as well as quantity” hypothesis for how RTS,S/AS01-induced antibodies may protect, in line with the previous findings of Suscovich et al. (2020) (discussed further below). Regarding full-length CSP-specific antibodies, we found contrasting associations of avidity with protection, depending on antibody subclass (IgG1 with non-protection; IgG2 with protection).

We next relate our results to the published correlates analyses of MAL068 and MAL071 that were introduced earlier. In the systems serology analysis by Suscovich et al. (2020), LASSO was used to downselect from a broad array of qualitative and quantitative antibody variables measured on the day of challenge in MAL068 vaccinees, build multivariate models, and assess how well each model could discriminate protected from non-protected vaccinees. Similar to our findings, Suscovich et al. reported that parsimonious models were able to accurately discriminate protected from non-protected vaccinees across studies, with a bivariate model (trained on MAL068 RRR) containing NANP6-specific ADCP and FCGRIIIA binding of NANP6-specific antibodies well-predicting protection in both the RRR arm of MAL071 and in the RRR arm of MAL027, another controlled human malaria infection study (Kester et al., 2009). While none of the top-performing bivariate models in our analogous cross-study validation (built on MAL068 RRR and validated on MAL071 RRR) contained the exact same combination of variables as the model in Suscovich et al. (ADCP NANP6 and FCGRIIIA NANP6), FcGRIIIAV NANP6, FcGRIIIAF NANP6, and ADCP NANP6 all ranked highly in our top-performing univariate and bivariate models. Therefore, our results are broadly consistent with the conclusion of Suscovich et al. that NANP6-specific antibody engagement of multiple effector cell types may be important for protection. However, our findings differed from Suscovich et al. in that we identified NANP6-specific antibody engagement of FcGRIIA, FcGRIIB, and FcGRIIIB as associated with protection in our top-performing models in the cross-study validation, whereas the FCGRII-related variables (NANP6 FcGRIIA, NANP6 FcGRIIB) were not among the five-feature LASSO-selected model, nor the two-feature model, reported by Suscovich et al. to discriminate protected from nonprotected participants, nor were NANP6 FcGRIIA, NANP6 FcGRIIB, or NANP6 FcGRIIIB found to be significantly different in univariate profiling of protected vs nonprotected participants by Suscovich et al. FcGRIIA is expressed by neutrophils and mediates antibody-dependent cellular cytotoxicity (ADCC) (Derer et al., 2014; Treffers et al., 2018) and is also expressed by monocytes, dendritic cells and macrophages and mediates ADCP (Anania et al., 2019); FcGRIIB is an inhibitory FcγR expressed by B cells (Amigorena et al., 1989), dendritic cells, macrophages, activated neutrophils, mast calls and basophils (Nimmerjahn and Ravetch, 2008) and FcGRIIIB is expressed by neutrophils and mediates endocytosis of immune complexes (Chen et al., 2012). The shared expression of FcGRIIA and FcGRIIIB by neutrophils suggests a potential role in NANP6-targeted ADNP in protection, which is consistent with the network linkage of NANP6 ADCP with ADNP in Suscovich et al.

Moreover, consistent with the findings of Seaton et al. (under revision) ¹ , NANP6-specific IgG1 scored highly in the top-performing models. IgG1 antibodies are known to bind to FcGRs, including FcGRII and FcGRIII and subsequently mediate downstream effector functions such as ADCC and ADCP. Taken together, these data indicate a role for NANP6-specific antibody activation of multiple FcγR-expressing immune cell effector functions. While we did not identify ADCC NANP6 in any of our top-performing models, the identification of NANP6-specific FCGRIII-related Fc array variables in our top-performing models suggests that effector functions beyond phagocytosis may be important.

Dennison et al. recently analyzed antibody avidity data from MAL071 and reported that, while no association with protection was seen with the avidity of antibody responses to CSP, NANP6, C-term, or NPNA3 in the RRR arm, higher-avidity antibodies to CSP and C-term tended to associate with protection in the RRr arm, suggesting a potential difference in the mechanisms of protection in the two arms. Our findings are consistent with those of Dennison et al. in that BLI variables related to avidity of anti-CSP or anti-C-term antibodies did not appear in top-performing (as assessed by cross-validated AUC) models trained on MAL068 RRR.

We next discuss our findings in the context of the previous transcriptional correlates studies (Kazmin et al., 2017; Du et al., 2020). A major difference between our study and the previous studies is that ours is the first to have included day-of-challenge PBMC transcriptional data in a correlates analysis. Kazmin et al. identified relatively early (within one week) transcriptional signatures of protection in the ARR and RRR arms of MAL068 and reported day-of-challenge functional enrichment (compared to baseline) for many BTMs in both study arms, but did not perform the correlates analysis on day-of-challenge levels. Likewise, Du et al. focused on the 1 day post-third dose time point, which falls within the innate response window (Du et al., 2020), and complements our focus on the day of challenge, at which point the innate adjuvant-driven response has waned (Kazmin et al., 2017). While we identified only a few BTMs in our top-performing models, two of the top three two-variable models (in terms of validation AUC on MAL071 RRR) each included a lysosome-related BTM (“lysosome M209” in the top-performing model and “lysosomal endosomal proteins M139” in the third-best performing two-variable model). Interestingly, day-of-challenge PBMC expression of each of these BTMs was associated with non-protection. The M209 BTM was recently proposed to represent the peroxisome (Courtland et al., 2021) based on high-resolution spatial proteomic data (Geladaki et al., 2019). There is now abundant evidence that peroxisomes regulate and influence both immune function and inflammation (Di Cara et al., 2019), and while peroxisomes have been shown to play an important role in antiviral signaling (Dixit et al., 2010), to our knowledge, an association between peroxisomes and malaria susceptibility has not previously been described. Further studies would be needed to confirm such an association.

Du et al. found that the discriminatory ability of the MX2/GPR183 signature varied over time, with little to no difference between log₂ fold-change (vs baseline) MX2/GPR183 ratio in protected vs. nonprotected vaccinees on day of challenge. It is thus not surprising that we did not identify any BTMs related to MX2 (a prototypical interferon-induced gene) or GPR183 (an oxysterol receptor) among our top-predicting variables in our day-of-challenge analyses. However, similar to Du et al., we did find that our top-predicting variables were not correlated with the known correlate anti-CSP titer (Dobaño et al., 2019), suggesting that our study also identified a separate aspect of the RTS,S/AS01-induced immune response implicated in protection.

Pallikkuth et al. used machine learning to identify combinations of immune response biomarkers, focusing on B-cell and peripheral T follicular helper (pTfh) cell responses, that were predictive of individual-level protection in the MAL071 trial (Pallikkuth et al., 2020). While we only analyzed day-of-challenge biomarkers, Pallikkuth et al. also analyzed immune responses throughout the 3-dose vaccination schedule up to the day of challenge: many of their most predictive variables were from relatively early time points, e.g., Day 6 post-dose one and Day 28 post-dose 2 (these time points were designed to obtain samples at critical periods of follicular helper T cell development). In their best-performing model, which contained 18 parameters, 85% of all individual-level outcomes were predicted correctly. Their analysis thus complements ours and reinforces the conclusion that molecular signatures of protection can vary greatly depending on the specific time point(s) post vaccination that are examined.

Our focus on day-of-challenge immune response measurements meant that our models did not incorporate all previously identified correlates of protection, e.g., the many transcriptional modules whose early post-vaccination (pre-challenge) expression was found to associate with protection (or non-protection) reported by Kazmin et al. (2017), or the early CSP-specific pTfh responses reported by Pallikkuth et al. to associate with protection (Pallikkuth et al., 2020). We did not investigate here the extent to which these earlier responses linked to protection (or non-protection) correlate with our top-performing variables, and thus it is possible that some individuals may have mounted an earlier protective response whose transcriptional “stamp” did not persist until day-of-challenge. We conjectured that post-challenge outcomes of such individuals would be harder to predict, since they would be determined by immune responses not included in the analysis set used here. In a leave-one-out prediction analysis, we did find evidence that cross-study post-challenge outcomes are easier to predict (using a day-of-challenge dataset) for some participants than others (Supplementary Figure S7). In particular, in the validation set, individuals who became infected after challenge tended to be correctly predicted by higher proportions of models compared to individuals who did not become infected after challenge. The presence of a cluster of four individuals whose MAL071 RRR challenge outcomes were particularly poorly-predicted (< 25% of all models) by models built on MAL068 RRR is consistent with the hypothesis that immune response kinetics and/or qualities may have varied across participants, and that this subset may have mounted quicker protective responses that were not reflected in the day-of-challenge dataset used here.

Additional limitations of our work include that our study is in the context of a highly controlled challenge environment, and there are important differences between the CHMI studies [malaria-naïve adults, CHMI done using a three-dose schedule with a one-month interval (with the exception of the RRr arm)] and the current target population for which RTS,S/AS01 is licensed (young children in malaria-endemic areas, RTS,S/AS01 is recommended to be used as a four-dose regimen, including a booster dose). Thus, it remains to be seen whether the biomarkers highlighted by our study are similarly relevant biomarkers of protection in malaria-endemic field settings. However, in the light of our finding that anti-NANP6 IgG1 titer was consistently associated with protection, Dobaño et al. (2019) reported that the magnitude of anti-NANP6 IgG antibodies was significantly associated with vaccine efficacy against clinical malaria, and Ubillos et al. (2018) reported that the RTS,S/AS01-induced increase (baseline to Month 3) in anti-NANP6 IgG1 antibodies was associated with protection from clinical malaria, in subsets of participants drawn from the phase 3 efficacy study of RTS,S/AS01 in infants and young children in malaria-endemic regions in Africa (RTS, S Clinical Trials Partnership, 2011; RTS, S Clinical Trials Partnership, 2012; RTS, S Clinical Trials Partnership, 2014; RTS, S Clinical Trials Partnership, 2015). These findings provide some evidence that our findings may apply in diverse populations that differ in important respects, e.g., baseline malaria immune status, although correlates were not assessed in post-booster samples in either Dobaño et al. or Ubillos et al.

Second, the high efficacy of the RRr regimen in MAL071 hindered the development of models that could discriminate protected from non-protected individuals; thus, we had to restrict our cross-study analysis to the RRR arm of MAL071. Thirdly, by focusing on day of challenge, we likely missed the opportunity to pick up earlier immune response biomarkers–especially transcriptional modules–that may have waned by day of challenge. Additional limitations include the necessity of restricting our analysis to variables that were common across the two trials, the fact that we did not assess predictive ability of individual genes in addition to the BTM analysis, and the lack of inclusion of alternative gene sets [such as Gene Ontology (GO) gene sets related to e.g., immunological signatures (Godec et al., 2016) or to canonical pathways/processes (Liberzon et al., 2015)] in the gene set enrichment analysis. Despite these limitations, our study further highlights Fc effector-related attributes of NANP6- and subclass-specific antibodies, over all other immune features investigated, that merit further confirmation in field trials. Our study is thus valuable in potentially helping prioritize resource-intensive laboratory measurements for assessment in future studies.