Background
The first malaria vaccine against
Plasmodium falciparum to reach Phase III clinical trials, RTS,S/AS01, has demonstrated moderate levels of efficacy against both clinical and severe malaria in young children in the 18 month follow-up of Phase III trials across 11 African sites and in several Phase II trials in Africa [
1‐
5]. Site- and time-specific data from the Phase III trial recently published [
1] indicated a vaccine efficacy against clinical cases over 18 months post third dose of 46 % (95 % CI 42–50) in children 5–17 months at first vaccination and 27 % (95 % CI 20–32) in infants (6 weeks at first immunisation, 12 weeks at third dose) [
1], with much higher observed efficacy at 6 months post third dose (5–17 months: 68.3 % (95 % CI 64.3–71.8), 6–12 weeks: 47.2 % (95 % CI 39.4–54.1)) indicating an initial quick decay [
1]. Since the malaria burden in many countries is still high, even a vaccine whose efficacy decays quickly may be of public health benefit. A WHO policy recommendation on the implementation of RTS,S vaccination in a number of malaria endemic countries in Africa is possible earliest at the end of 2015 [
6]. Quantitative predictions of the expected public health impact and cost-effectiveness for different immunisation schedules may partly inform this recommendation.
Plasmodium falciparum malaria is transmitted to humans through bites from infected mosquitoes and has a complex life cycle in the human host. An infected mosquito injects sporozoites into subcutaneous tissue of the host; the sporozoites then travel to the liver. Successful invasion of hepatocytes depends on the circumsporozoite protein (CSP) of the sporozoite [
7]. Following replication in the liver the parasite enters the blood stream, infecting erythrocytes and multiplying. It is the erythrocytic cycle of
Plasmodium falciparum that causes clinical disease.
The RTS,S vaccine induces antibodies in the host against CSP and thus, with high enough antibody titre, prevents liver infection and subsequent clinical malaria that would have resulted from a blood stage infection. RTS,S has been shown to be efficacious and safe [
1], but as antibody titres to CSP wane so does protection against successful infection of the liver [
8], and observed efficacy against clinical disease decays relatively rapidly in the trial [
1]. Repeated malaria infections induce natural, but not complete, immunity in the host to many stages of the parasite life cycle, predominantly to the blood stage causing clinical disease. There is a tendency for efficacy against clinical malaria to wane more rapidly in sites where exposure is higher [
1], which is to be expected, because natural immunity to blood stage parasites is more rapidly acquired by non-vaccinated individuals. Any partially protective malaria infection blocking intervention, such as RTS,S or seasonal malaria chemoprophylaxis, aimed at infants and young children will give rise to age-shifts of burden and susceptibility to infection for this reason.
A moderately efficacious, leaky vaccine such as RTS,S, that reduces probability of infection but faces a high force of infection has complicated dynamics, including effects that cannot be detected in field trials [
9] and, prior to Phase IV follow-up studies, mathematical models are essential to predict long-term outcomes of vaccination programs when delivered to populations outside trial settings. Such models indicate how population-level outcomes relate to vaccine properties (efficacy and duration of protection) or to the schedule of delivery, age at vaccination, exposure and other contextual factors. Models can address the question of whether different clinical efficacies observed in different transmission settings [
4] are a result of differences in the challenge or due to differences in the vaccine effect. By identifying key long-term drivers of differences in public health impact and cost-effectiveness between possible immunisation schedules, or between different health system contexts, models can also help optimise vaccination schedules.
A number of micro-simulation models of malaria in humans have been specifically designed for predicting the public health impact of interventions, including malaria vaccines [
10‐
13]. These models take into account levels of herd immunity and long-term effects of vaccination or other infection blocking interventions, such as deferral of events to older ages, and suggest that vaccination with pre-erythrocytic vaccines such as RTS,S via the Expanded Program of Immunisation (EPI), could substantially reduce paediatric morbidity and mortality during the first decade of vaccine use. The benefits of RTS,S are likely to be highest with levels of transmission entomological inoculation rate (EIR) between 2 to 50, which corresponds to intermediate levels of transmission in our model [
9,
11,
14‐
17]. The EIR at the start of a vaccination program is critical, irrespective of how it has come about, or of whether transmission is increasing or decreasing [
18], while herd immunity is likely to be negligible [
9,
11]. A probabilistic sensitivity analysis [
17] indicated that the EIR distribution, decay rate of vaccine effects and model for severe disease are important drivers of uncertainty in public health impact.
Development of these models focused heavily on fitting models to field data because of the need for quantitative predictions, but only limited data were available on the actual profile of the RTS,S vaccine. Previously published results from Phase III clinical trials of RTS,S [
3,
4] have been of limited value for parameterising mathematical models of public health impact. Site- and time-specific data recently published [
1] for the first 18 months of follow-up of Phase III trials now make it possible to carry out comprehensive fitting of models of vaccine action and their validation.
This paper reports the use of models within the OpenMalaria platform [
11] to obtain precise estimates of underlying vaccine properties given 18 months follow-up, with observations every six months, across 11 trial sites. Using a Bayesian Markov chain Monte Carlo (MCMC) approach the likely profile of the RTS,S vaccine is determined, estimating the rate of decay of efficacy in the Phase III trial, thus allowing one to project longer term trial outcomes. Investigations of validity and the consequent estimates of vaccine properties and of clinical efficacy expected in each of the trial sites for follow-up longer than 18 months are also investigated.
In addition and beyond previous analyses, country-specific estimates are made of the likely public health impact of RTS,S programs in 43 sub-Saharan African countries, with vaccine properties aligned with the latest results from Phase III trials of RTS,S resulting from the fitting analysis in this paper. The predictions are made via a weighted averaging approach over a large database of simulations that take into account country-specific context of current malaria burden, intervention coverage, demographics, and health system capacity. Several possible implementation strategies of immunising infants and children are considered. The parameterisation and prediction approach using micro-simulations provides us with uncertainty estimates around both the vaccine profile and the predictions of public health impact, highlighting where additional data are needed. Although a very large number of computationally expensive simulations are needed, the method will allow the estimates to be updated once final Phase III data are available, without re-running these simulations.
The predictions for both public health impact and for clinical trial efficacy fitting make use of an ensemble of structurally different models [
11], each a variant on a single baseline model [
10], with final results derived by aggregating many simulation runs. Further examination of public health impact includes analysis of result sensitivity to vaccine properties (initial efficacy against infection, vaccine half-life of efficacy against infection, decay shape) and country-specific properties (transmission, access to care) and considers the effects of structural and stochastic uncertainty on our predictions.
Discussion and conclusions
Simulation models of the public health impact of pre-erythrocytic vaccines against malaria are not new, but there is new urgency in making specific predictions for RTS,S/AS01 linked to the malaria situation in endemic countries using available Phase III data to parameterise models. The reason is that a recommendation on the use of RTS,S is expected as early as the end of 2015. Previously the public health impact of introducing the RTS,S vaccine into routine vaccination schedules in Africa has been difficult to predict because available clinical trial data were inadequate for accurately estimating the kinetics of vaccine protection, and this uncertainty in the vaccine profile meant that geographically specific predictions of likely impact [
17] were mainly of value to indicate general principles and data gaps. Site- and time- specific data from 18 months of follow-up of the Phase III trials [
1] have now enabled us to estimate the vaccine profile accurately enough for quantitative predictions of impact at national level to have sufficient plausibility for guiding policy decision as well as for informing subsequent implementation decisions by ministries of health.
Using available clinical trial data, the estimate of the initial efficacy against infection of RTS,S/AS01 is around 63 % (95 % CI 39.5–80.3 %) for infants and 79.2 % (95 % CI 67.3–84.8 %) for children, and is slightly higher than the efficacy in challenge trials which directly estimate the same quantity. In challenge trials with RTS,S in adults, 42 % [
28] and 47 % [
29] protection against an infection challenge was observed with adjuvant AS02, and 50 % observed when using adjuvant AS01B [
30]. Consistent with our results is the almost equivalent estimate obtained with natural challenge of 65.9 % (95 % CI 42.6–79.8 %) protection against first infection in a Phase I/IIb trial immunising infants with RTS,S/AS01 [
31]. The model estimates for RTS,S/AS01 initial efficacy against infection in this work are substantially higher than those previously estimated by modelling from the initial Phase II RTS,S/AS02 of 52 % [
16], and, as expected, higher than the directly measured efficacy against clinical episodes at 18 months follow-up [
1]. However, there is considerable uncertainty around them, especially for the 6–12 week cohort.
The underlying vaccine profile of efficacy against infection and decay, which reflects the induced pre-erythrocytic immunity, is most likely the same across the trial sites, even though the measured clinical efficacy which also depends on secondary effects on blood stage immunity, appears to be lower in sites with higher exposure [
1]. This effect can be accounted for by between-site variation in transmission level, the extent of transmission heterogeneity, and in levels of access to care, all of which modify the relationship between the underlying efficacy in preventing infection and efficacy against clinical disease, justifying our use of site-independent estimates of the underlying initial efficacy and decay.
RTS,S initial protection is high and decays relatively rapidly and although clinical efficacy over time might seem low, RTS,S implemented in addition to current malaria control measures across endemic countries in Africa will have substantial impact in averting malaria cases. RTS,S would avert 100–580 malaria deaths and 45,000 to 80,000 clinical events for every 100,000 fully vaccinated child in the first 10 years of the program. This would potentially increase if boosting doses are added. The uncertainty in the vaccine profile is compounded in these predictions of public health impact by the uncertainty in the distributions of transmission levels in the different countries. This does not even take into consideration the uncertainties in demographic projections, in future trends of malaria and control, and in the assumptions about vaccination coverage; with coverage levels and population growth in higher transmission areas predicted to have a much larger impact than uncertainty in future trends of transmission. In addition, the differences in predicted impact between the vaccination schedules are small in relation to the uncertainty ranges. In particular, the predictions of public health impact of EPI vaccination and vaccination at 6–9 months are very similar, with the former averting overall slightly more episodes of illness, and the latter overall more deaths depending on coverage (a consequence of the age dependence in the case fatality rate and association with indirect mortality due to co-morbidities at younger ages [
26]).
Previous simulations of the effects of paediatric malaria vaccination programs demonstrated minimal herd immunity effects [
9], meaning that this intervention strategy will not have any substantial effect on overall levels of malaria transmission. This is a consequence of the targeting of a narrow age range (those at highest risk of life-threatening disease) to vaccinate, not of the vaccine profile per se. Indeed, the high initial efficacy of RTS,S/AS01 is similar to the profile aimed at for vaccines aiming to interrupt transmission [
32], and mass administration of a vaccine with such a high efficacy would have substantial transmission effects [
9]. However, the current strategy for licensure of RTS,S does not envisage mass vaccination, and this is outside the scope of this paper, but previous efforts have indicated the potential benefits in low transmission settings [
9]. Post-registration use of the vaccine will be important, as will further modelling investigations.
The availability of very extensive data on prevalence from MAP [
21] means that there is a better basis for estimating the vaccine-avertable burden of disease for malaria than for other major childhood infections. The high burden of
Plasmodium falciparum disease means that we predict the public health impact of RTS,S to be comparable to that of other new childhood vaccines, such as those against Haemophilus influenza type b and pneumococcus, despite the leakiness and relatively low efficacy of the vaccine. Such a large public health impact is based on much higher rates of severe disease and mortality than have been observed in the trials (where severe disease rates were low and malaria mortality almost absent, presumably because very high standards of care were achieved [
1]). These higher levels of disease are those measured in the non-trial datasets to which the OpenMalaria models were originally fitted [
11,
26]. For comparisons with other vaccines it is also relevant to consider that some deaths arising from co-infections could be averted by vaccination against either of the pathogens concerned. This particularly applies to our simulated numbers of indirect malaria deaths, which are intended to capture the effects of interactions between
Plasmodium falciparum and co-infections, especially respiratory bacteria.
A very important source of uncertainty in our predictions is in the kinetics of the vaccine effect on infection rates. The analysis suggests that the efficacy in preventing infections decays exponentially with a half-life of decay of around 1 year (Table
2), which is much faster than was previously thought but is in line with published data of IgM serum concentrations [
8]. The public health impact will depend on not just the half-life, but also the functional form of efficacy decay. Once data from longer follow-up periods of the trial are available it should become possible to estimate whether decay curves belonging to families other than the exponential are more appropriate. In line with previous analyses [
11] we infer that the efficacy measured against clinical malaria in the trial is declining over time even more rapidly than the underlying effect in preventing new infections, so the superficial interpretation that the decline in efficacy means that vaccination has only a transient effect should be resisted. Conversely, the temptation should be resisted to present efficacy as values cumulated up to specific time points, which makes waning of efficacy less evident. It is essential to compare incidence between the arms of the trial over each time interval, allowing recurrent events in the same children. However, the prediction that the time-period specific efficacy in some trial sites may fall below zero by the end of the trial, based on extrapolating the existing decay, highlights the need to manage expectations so that such a result is not misinterpreted. This is an unavoidable property of a leaky vaccine combatting recurrent challenges from a pathogen that stimulates partial immunity. Some clinical events in vaccinated children will be delayed, rather than averted, a phenomenon that must be taken into account in predicting public health impact of all partially protective malaria interventions, but which should not be interpreted as an adverse effect of vaccination.
Data are still being accrued that will be crucial for estimating the shape of the efficacy decay, and the estimation will be repeated when the results from the full follow-up of 32 months are available. This analysis will also enable us to assess whether a different efficacy for the boosting dose is expected compared to the third dose given 18 months prior to boost. This will considerably reduce the uncertainty in predictions of the effect of boosting.
All models assume no rapid evolution of the parasite sensitivity to the RTS,S antigen, and fears over resistance are actually small, but this should not impact the evaluation of a new intervention with the potential to prevent malaria morbidity and mortality.
Since the computational requirements of our analysis were enormous, with each of the simulations from OpenMalaria requiring significant computing time, repeating the analysis is not a trivial exercise. However, a distinct benefit of our data-basing and weighting approach is that estimates for different countries, trial sites or geographical areas with different transmission and health system parameters can be made without running new micro-simulations. Only the fitting and weighting steps will need to be repeated when new trial data are available, and these have comparatively low computational requirements. Bayesian MCMC estimation of weighting factors also provides a way to fit the very complex OpenMalaria models simultaneously to multiple outcomes from the trials (prevalence and clinical incidence) without the computationally expensive need to re-run the simulations iteratively. Other advantages offered by the model averaging approach over estimates based on single parameterisations include the propagation of the uncertainty in the vaccine profile through to the public health impact predictions, allowing the influence of these factors to be compared with the sensitivity to assumptions about transmission and health systems. Weighted averaging of simulations also provides a straightforward approach for analysis by repeating the calculation of public health impact, using different weight vectors. The use of a model ensemble capturing different assumptions around development of immunity and degree of transmission heterogeneity also provides lower-bound estimates of the impact of structural uncertainty [
11], and replicating simulations with random number seeds tells us how much stochasticity influences our results.
An additional key message from this analysis is that the decay in efficacy is the parameter contributing the most uncertainty to the prediction of public health impact of RTS,S and for second generation malaria pre-erythrocytic vaccines. Other promising pre-erythrocytic vaccines have already demonstrated near 100 % efficacy in challenge trials [
33] before rechallenge. The developers of these vaccines also need to consider that, while a high initial efficacy is clearly highly desirable, the temporal pattern of decay in efficacy is of equal, if not more, importance as a determinant of the likely public health impact of vaccination programs.
Competing interests
The analysis for this study was conducted following a call initiated and facilitated by the PATH-Malaria Vaccine Initiative. Employees of PATH-Malaria Vaccine Initiative reviewed and commented on early draft manuscripts, but were not involved in the final approval of the manuscript or final results.
Authors’ contributions
MAP and TAS designed the study. MAP performed the analysis. MAP, KG, TAS, MTanner and MTarantino provided input on simulations and data on model parameters and helped with interpretation of results. MAP and TAS prepared the manuscript. All authors read and approved the final manuscript.
Funding
This work was funded by Bill & Melinda Gates Foundation project number 1032350 and the PATH-Malaria Vaccine Initiative (MVI). No funding bodies had any role in the study design, data analysis, decision to publish, or preparation of the manuscript. The clinical trial data was previously published by The RTS,S Clinical Trials Partnership (2014) [
1], but it did not fund the investigators to undertake the analysis.