A Computer Modelling Approach To Evaluate the Accuracy of Microsatellite Markers for Classification of Recurrent Infections during Routine Monitoring of Antimalarial Drug Efficacy

Abstract

Antimalarial drugs have long half-lives, so clinical trials to monitor their efficacy require long periods of follow-up to capture drug failure that may become patent only weeks after treatment. Reinfections often occur during follow-up, so robust methods of distinguishing drug failures (recrudescence) from emerging new infections are needed to produce accurate failure rate estimates. Molecular correction aims to achieve this by comparing the genotype of a patient's pretreatment (initial) blood sample with that of any infection that occurs during follow-up, with matching genotypes indicating drug failure. We use an in silico approach to show that the widely used match-counting method of molecular correction with microsatellite markers is likely to be highly unreliable and may lead to gross under- or overestimates of the true failure rates, depending on the choice of matching criterion. A Bayesian algorithm for molecular correction was previously developed and utilized for analysis of in vivo efficacy trials. We validated this algorithm using in silico data and showed it had high specificity and generated accurate failure rate estimates. This conclusion was robust for multiple drugs, different levels of drug failure rates, different levels of transmission intensity in the study sites, and microsatellite genetic diversity. The Bayesian algorithm was inherently unable to accurately identify low-density recrudescence that occurred in a small number of patients, but this did not appear to compromise its utility as a highly effective molecular correction method for analyzing microsatellite genotypes. Strong consideration should be given to using Bayesian methodology to obtain accurate failure rate estimates during routine monitoring trials of antimalarial efficacy that use microsatellite markers.

Keywords: Bayesian; efficacy; malaria; microsatellite; molecular correction; recrudescence; reinfection; resistance; therapeutic efficacy study.

FIG 1

Failure rate estimates obtained using the match-counting algorithm and the Bayesian analysis algorithm for failing AR-LF therapy under low-, medium-, and high-transmission scenarios. The true failure rate is denoted in each plot by a horizontal gray line. For the match-counting algorithm, the threshold for the number of matching loci at which a recurrence was classified as a recrudescence varied between 2 and 7. For the Bayesian analysis, the cutoff for posterior probability at which a recurrence was classified as a recrudescence varied between ≥0.1 and ≥0.9.

FIG 2

Failure rate estimates obtained using the match-counting algorithm and the Bayesian analysis algorithm for nonfailing AR-LF therapy under low-, medium-, and high-transmission scenarios. The true failure rate is denoted in each plot by a horizontal gray line. For the match-counting algorithm, the threshold for the number of matches at which a recurrence was classified as a recrudescence varied between 2 and 7. For the Bayesian analysis, the cutoff for posterior probability at which a recurrence was classified as a recrudescence varied between ≥0.1 and ≥0.9.

FIG 3

ROC curves showing the diagnostic ability of the Bayesian analysis method for 3 scenarios of transmission intensity for nonfailing and failing AR-LF therapy. ROC curves and AUC are shown for all recrudescence and for high-density recrudescence. A high-density recrudescence was defined as explained in the text.

FIG 4

Distributions of the posterior probabilities of recrudescence estimated by the Bayesian algorithm for 3 scenarios of transmission intensity for nonfailing and failing AR-LF therapy. A high-density recrudescence was defined as explained in the text.

FIG 5

Contour plot of the posterior probability of recrudescence estimated by the Bayesian algorithm as a function of the density of recrudescent clones (i.e., the proportion of recrudescent clones in the total recurrent infection biomass) in the initial sample and the recurrent sample. The plot shows the combined data from all 6 scenarios modeled for AR-LF. Each contour line indicates the posterior probability of recrudescence, and the areas between the lines represent the numbers of recurrent infections in the population with those posterior probabilities.