The effectiveness of artemisinin-based combination therapies (ACTs) to treat Plasmodium falciparum malaria is threatened by resistance. The complex interplay between sources of selective pressure—treatment properties, biological factors, transmission intensity, and access to treatment—obscures understanding how, when, and why resistance establishes and spreads across different locations. We developed a disease modelling approach with emulator-based global sensitivity analysis to systematically quantify which of these factors drive establishment and spread of drug resistance. Drug resistance was more likely to evolve in low transmission settings due to the lower levels of (i) immunity and (ii) within-host competition between genotypes. Spread of parasites resistant to artemisinin partner drugs depended on the period of low drug concentration (known as the selection window). Spread of partial artemisinin resistance was slowed with prolonged parasite exposure to artemisinin derivatives and accelerated when the parasite was also resistant to the partner drug. Thus, to slow the spread of partial artemisinin resistance, molecular surveillance should be supported to detect resistance to partner drugs and to change ACTs accordingly. Furthermore, implementing more sustainable artemisinin-based therapies will require extending parasite exposure to artemisinin derivatives, and mitigating the selection windows of partner drugs, which could be achieved by including an additional long-acting drug.

Background Mathematical models provide an understanding of the dynamics of a Plasmodium falciparum blood-stage infection (within-host models), and can predict the impact of control strategies that affect the blood-stage of malaria. However, the dynamics of P. falciparum blood-stage infections are highly variable between individuals. Within-host models use different techniques to capture this inter-individual variation. This struggle may be unnecessary because patients can be clustered according to similar key within-host dynamics. This study aimed to identify clusters of patients with similar parasitaemia profiles so that future mathematical models can include an improved understanding of within-host variation. Methods Patients’ parasitaemia data were analyzed to identify (i) clusters of patients (from 35 patients) that have a similar overall parasitaemia profile and (ii) clusters of patients (from 100 patients) that have a similar first wave of parasitaemia. For each cluster analysis, patients were clustered based on key features which previous models used to summarize parasitaemia dynamics. The clustering analyses were performed using a finite mixture model. The centroid values of the clusters were used to parameterize two established within-host models to generate parasitaemia profiles. These profiles (that used the novel centroid parameterization) were compared with profiles that used individual-specific parameterization (as in the original models), as well as profiles that ignored individual variation (using overall means for parameterization). Results To capture the variation of within-host dynamics, when studying the overall parasitaemia profile, two clusters efficiently grouped patients based on their infection length and the height of the first parasitaemia peak. When studying the first wave of parasitaemia, five clusters efficiently grouped patients based on the height of the peak and the speed of the clearance following the peak of parasitaemia. The clusters were based on features that summarize the strength of patient innate and adaptive immune responses. Parameterizing previous within host-models based on cluster centroid values accurately predict individual patient parasitaemia profiles. Conclusion This study confirms that patients have personalized immune responses, which explains the variation of parasitaemia dynamics. Clustering can guide the optimal inclusion of within-host variation in future studies, and inform the design and parameterization of population-based models.

No abstract