Comparison of techniques and markers to distinguish Plasmodium falciparum recrudescence from new infection in Rwanda Authors: Sara L. Cantoreggi, Monica Golumbeanu, Michaela Zwyer, Aline Uwimana, Jean Damascene Niyonzima, Aimable Mbituyumuremyi, Naomi W. Lucchi, Mateusz M. Plucinski, Christian Nsanzabana Published: 2026 ABSTRACT Accurate estimation of antimalarial drug efficacy against P. falciparum requires PCR correction to distinguish recrudescence from new infection in recurrent infections. Here, we compared two different genotyping techniques and different decision algorithms used for PCR correction using samples collected in a Therapeutic Efficacy Study (TES) conducted in Rwanda in 2018. We first optimized an assay to genotype four microsatellites by capillary electrophoresis and assessed its sensitivity in detecting minority clones with laboratory parasite strain mixtures. We then analyzed patient samples by capillary electrophoresis using msp1, msp2, glurp and microsatellites (length-polymorphic markers) and amplicon deep sequencing using cpmp, ama1, cpp and csp (SNP-rich markers), and assessed their diversity and allelic frequency. We classified the recurrent infections as recrudescence or new infection based on different marker combinations and decision algorithms (WHO, 2/3, Bayesian algorithm). Among microsatellites, TA40 and PfPK2 had the highest sensitivity in detecting minority clones. Msp2/3D7, glurp and SNP-rich markers had the highest genetic diversity and lowest allelic frequencies. The WHO algorithm identified the fewest recrudescences across all marker combinations, while the 2/3 identified the most. The Bayesian algorithm yielded intermediate results. Replacing glurp with any of the microsatellites or using SNP-rich markers did not significantly alter recrudescence estimates. Amplicon sequencing of SNP-rich markers provided the most consistent results regardless of the decision algorithm used and therefore holds great potential for reliable PCR correction. Additionally, probabilistic approaches may provide more robust results compared to match counting algorithms; however, they require rigorous validation with datasets from different transmission settings.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-026-57063-7. Introduction Gains in malaria control and elimination are threatened by the development of resistance to antimalarial drugs against Plasmodium falciparum. Indeed, artemisinin (ART) partial resistance and subsequent selection of resistance to the partner drug in Artemisinin-based Combination Therapies (ACT) are a cause of concern1,2. After emerging in Southeast Asia in the late 2000s3, partial ART resistance quickly developed into full ACT resistance, leading to treatment failures in multiple countries2,4. In recent years, ART resistance has been increasingly reported in East Africa and the Horn of Africa5–8, although high levels of ACT treatment failure have not been reported yet. It is therefore crucial to strengthen drug efficacy monitoring through therapeutic efficacy studies (TESs). The assessment of antimalarial drug efficacy requires PCR correction to distinguish recrudescences from new infections in recurrent infections9,10. Highly polymorphic markers, such as merozoite surface protein 1 (msp1), merozoite surface protein 2 (msp2), glutamate-rich protein (glurp), and multiple microsatellites11–14 have been used for this purpose. To overcome the limited sensitivity of glurp in detecting minority clones in polyclonal infections15, the World Health Organization (WHO) now recommends replacing this marker with one microsatellite16. For the transition period, the WHO recommends using both the msp1/msp2/glurp and msp1/msp2/microsatellite methods for PCR correction to allow for comparison and evaluation16. Microsatellites are highly polymorphic17,18, but their sensitivity in detecting minority clones has been only partially assessed19. On the other hand, amplicon deep sequencing (AmpSeq) offers a promising alternative, which has been shown to be highly sensitive in detecting minority clones, reproducible and robust in distinguishing recrudescence from new infection19,20. Markers proposed for AmpSeq, such as cpmp, ama1-D2, ama1-D3, cpp and csp, have shown very high diversity in previous studies19–21. However, more data is required to establish a standardized protocol to be applied on a larger scale for TESs. A decision algorithm allows for classification of sample pairs as recrudescence or new infection11,13,16. Three different decision algorithms have been commonly used in this context: 1) the WHO-recommended (3/3) algorithm, which requires all markers to give concordant results to classify a sample pair as a recrudescence; 2) the 2/3 algorithm, where concordance of two markers is sufficient to determine a recrudescence; and 3) an algorithm based on Bayesian statistics, which calculates the posterior probability (range 0.0 to 1.0) of recrudescence while considering the allelic diversity of each marker in the dataset22. PCR correction outcomes may vary greatly based on markers and decision algorithms used, especially in high transmission settings where high complexity of infection is common. Therefore, it is essential to use PCR correction methods that accurately distinguish recrudescences from new infections to ensure a precise estimation of drug efficacy. Here, we genotyped samples from a TES conducted in Rwanda in 2018 to evaluate the efficacy of Artemether-Lumefantrine (AL). Seven microsatellites had been previously analyzed with the Bayesian algorithm, and PCR-corrected efficacy of AL was found to be > 90%5. We repeated genotyping with msp1/msp2/glurp, msp1/msp2/microsatellite (Poly-α, PfPK2, TA40, and TA81) and five single nucleotide polymorphism (SNP)-rich markers (cpmp, ama1-D2, ama1-D3, cpp and csp). We assessed the genetic diversity of each marker and analyzed the different marker combinations with three decision algorithms, with the aim of comparing markers and algorithms to identify the most robust ones for PCR correction. Methods Sample collection Samples were collected in a TES conducted in Rwanda in 2018 by the Rwanda Biomedical Center (RBC) supported by the United States President’s Malaria Initiative (PMI) and Centers for Disease Control and Prevention (US-CDC). Dried blood spots (DBS) on Whatman 903 filter paper were collected in three sentinel sites: Masaka (Kicukiro district, City of Kigali), Rukara (Kayonza district, Eastern Province) and Bugarama (Rusizi district, Western Province). Detailed information about site characteristics and sample collection has been published elsewhere5. DNA extraction Three punches of 3 mm diameter were obtained from each DBS. DNA extraction was performed with the QIAamp 96 DNA Blood Kit (Qiagen) according to the manufacturer’s instructions23. Genotyping of length polymorphic markers: msp1, msp2 and glurp Msp1, msp2, and glurp, as well as their respective allelic families (MAD20, K1 and Ro33 for msp1, 3D7 and FC27 for msp2), were amplified by primary and nested PCR as previously described15,19. Amplicon sizes were assessed by high-resolution capillary electrophoresis (H-CE) on ABI 3730xl® (Thermo Fisher Scientific). For Ro33, considered to be monomorphic24,25, sample positivity was assessed with the DNA Fast Analysis Kit on QIAxcel (Qiagen). H-CE data was analyzed using GeneMapper v. 6 (Thermo Fisher Scientific). Any peak with fluorescence intensity < 500 RFU was considered background noise. Additionally, a cut-off of 10% and 20% on the highest peak’s fluorescence intensity was applied for msp1/msp2 allelic families and glurp, respectively, to remove stutter peaks, as indicated in15. Genotyping of length polymorphic markers: microsatellites A nested PCR assay for Poly-α, PfPK2, TA40 and TA81 was optimized based on previously published protocols14,26,27. Details are available in Supplementary Tables 1–4. Amplicon sizes were assessed by H-CE on ABI 3730xl® (Thermo Fisher Scientific). H-CE data was analyzed using GeneMapper v. 6 (Thermo Fisher Scientific). Any peak with fluorescence intensity < 500 RFU was considered background noise. Peaks were identified by visual inspection of the electropherograms and comparison with the peak pattern in the positive control samples, complemented by a peak-calling algorithm adapted from28 to remove stutter peaks. The algorithm applies five distinct thresholds, each of which defines the relative height a peak must have with respect to other peaks in the same sample in order to be considered a real peak28. The thresholds were determined by analyzing single strain samples containing the 3D7, K1, HB3, and FCB1 P. falciparum laboratory strains with known microsatellite lengths, and calculating the thresholds necessary to discard the stutter peaks. More details can be found in Supplementary Tables 9 and 10. Limit of detection (LOD) for microsatellites We have previously assessed the limit of detection (LOD) for minority clones for msp1, msp2, glurp and SNP-rich markers19. Here, we determined the LOD of the Poly-α, PfPK2, TA40 and TA81 assay with the same methodology. Briefly, we analyzed mixtures of four different laboratory strains of P. falciparum (3D7, K1, HB3, and FCB1) in different ratios as described in19. The assay was run in triplicate. We only considered strain mixture samples in which the strain with the longest amplicon was in minority (Supplementary Table 5). A strain mixture sample was classified as positive if the minority clones were detected (i.e., if all strains present in the sample were detected) in two out of three replicates. The LOD was defined as the lowest ratio of laboratory strains at which a strain mixture sample was classified as positive. Intra-assay reproducibility was assessed by calculating the percentage of samples returning the same result for all three replicates. Amplicon deep sequencing of SNP-rich markers: cpmp, ama1-D2, ama1-D3, cpp and csp Cpmp (430 bp), ama1-D2 (479 bp), ama1-D3 (516 bp), cpp (395 bp), and csp (329 bp) were sequenced as described previously20,21. Library preparation consisted of primary, nested and adapter PCR, followed by magnetic purification, normalization and sample pooling. Amplicons were then deep sequenced on MiSeq (Illumina) in paired-end mode with the MiSeq reagent kit v3 (600 cycles). Bioinformatics analysis of amplicon sequencing data The R package HaplotypR v0.3.129 was employed to analyze the amplicon sequencing data and identify haplotypes. Sequencing reads were first demultiplexed by sample and marker. Primer sequences were then removed, and the forward and reverse reads were merged. These merged reads were aligned to the 3D7 P. falciparum reference genome, and SNPs were identified within each sample. To ensure reliability, only dominant SNPs (occurring at 50% or higher mismatch rate in at least two samples) were considered. Based on these identified SNPs, haplotypes were constructed. Further analysis included only the haplotypes occurring in at least two of the three replicates, having a frequency of 1% or higher, and supported by at least 20 reads. We further determined the minimum number of reads necessary for a haplotype to be considered true to discard spurious haplotypes. The thresholds were determined based on the haplotypes detected in the positive controls. These contained four laboratory strains with known haplotype sequences mixed in different ratios. For most markers, we detected between one and three spurious haplotypes in some controls, and we hypothesized that these were due to contamination and/or to sequencing errors (Supplementary Table 12). To remove them, we applied marker-specific cut-offs. For ama1-D2, a haplotype had to be supported by at least 15% of the number of reads of the haplotype with the highest number of reads within a sample. For ama1-D3, cpmp and cpp a cut-off of 25%, 4%, and 8% was defined, respectively. Haplotypes detected in the same sample with a number of reads below the threshold were discarded. No threshold was applied for csp because no spurious haplotypes were detected in the positive controls. The same thresholds were applied to patient samples. Genetic diversity and expected heterozygosity of length polymorphic and SNP-rich markers Genetic diversity was assessed by counting the number of genotypes (for length polymorphic markers) and haplotypes (for SNP-rich markers), and the frequency at which they were present in the samples from the dataset. Amplicons that were within 3 bp (for msp1/msp2/glurp) and 2 bp (for microsatellites) of difference were considered identical. Additionally, the expected heterozygosity (He) was calculated for each marker with the following formula: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{H}_{e}=\:\left[n/\left(n-1\right)\right]\left[1-\sum\:{{p}_{i}}^{2}\right]$$\end{document}, where n is the number of genotyped samples and pi is the frequency of each allele30. Multiplicity of infection analysis Multiplicity of infection (MOI) for each sample, defined as the number of clones which co-infect a single host, was calculated as the count of all clones detected in the respective sample for each marker. Only samples collected before treatment (day 0) were included in the analysis. The MOI of the different markers was then stratified by site (Masaka, Rukara, and Bugarama) and used as a proxy for marker diversity, sensitivity and transmission intensity. The Kruskal-Wallis test31 was used to assess whether the differences between MOI values among all sites or markers were statistically significant. When this test yielded significant results, the post-hoc Dunn’s test with Bonferroni adjustment32 was used for pairwise comparisons between individual sites or markers. Three markers with the highest MOI values (one from each marker type: msp1/msp2/glurp, microsatellites, and SNP-rich markers) were selected for further analysis of MOI in individual day 0 and day X (where "X" corresponds to the day of recurrence) samples. Concordance between two markers was defined as the percentage of samples in which both markers detected the same MOI. Classification of recurrences Recrudescence vs. new infection analysis was performed by comparing genotypes from the sample collected before treatment (day 0) and from the sample collected in the recurrent infection (day X). For length polymorphic markers, the size of the amplicons detected in the day 0 and day X samples was compared. Amplicons that were within 3 bp (for msp1/msp2/glurp) and 2 bp (for microsatellites) of difference were considered as having the same size. For SNP-rich markers, the haplotypes detected in the day 0 and day X samples were compared. When at least one common haplotype was found on both day 0 and day X, the sample pair was considered a recrudescence. Different marker combinations were used to classify recurrent infections as either recrudescent or new infections, namely msp1/msp2/glurp, msp1/msp2/Poly-α, msp1/msp2/TA81, msp1/msp2/TA40, msp1/msp2/PfPK2, cpmp/ama1-D3/csp, cpmp/ama1-D2/csp, cpmp/ama1-D3/cpp, cpmp/ama1-D2/cpp, csp/ama1-D3/cpp, csp/ama1-D2/cpp, ama1-D2/ama1-D3/csp, ama1-D2/ama1-D3/cpmp, ama1-D2/ama1-D3/cpp, and cpmp/cpp/csp. The different decision algorithms, namely the WHO, 2/3 and Bayesian algorithms, were applied to each marker combination. For the Bayesian algorithm we used 100’000 iterations and 2 chains to assess and ensure convergence. The probability of recrudescence for each sample pair was calculated by taking the mean of the probabilities estimated in the 2 chains. Sample pairs with probability of recrudescence > 0.5 were considered recrudescent. To calculate the total number of recrudescences within the dataset, all probabilities of recrudescence were summed up. Identity-by-descent (IBD) analysis Identity-by-descent (IBD) analysis was performed for SNP-rich markers with Dcifer, a method allowing to infer the degree of shared ancestry between polyclonal infections33. The corresponding R package dcifer, version 1.2.1 (https://cran.r-project.org/web/packages/dcifer/index.html) with default settings was used. Given the low number of markers in analysis, which deviates from the typical application of the method, the “lrank” parameter was set to 1. Therefore, this analysis should be considered exploratory. Results Limit of detection of nested PCR assay for microsatellites Fig. 1Limit of detection and intra-assay variability of microsatellites TA81, Poly-α, TA40 and PfPK2. The assay was tested with four laboratory strains (3D7, K1, HB3, FCB1) mixed in different ratios. (a) Limit of detection (LOD) of the assay for each microsatellite (x-axis). Y-axis: ratios of the strain mixture samples. The height of the bars corresponds to the LOD value. (b) Intra-assay reproducibility for each microsatellite (x-axis). The height of the bars corresponds to the reproducibility value (0.0 to 1.0, y-axis). Limit of detection and intra-assay variability of microsatellites TA81, Poly-α, TA40 and PfPK2. The assay was tested with four laboratory strains (3D7, K1, HB3, FCB1) mixed in different ratios. (a) Limit of detection (LOD) of the assay for each microsatellite (x-axis). Y-axis: ratios of the strain mixture samples. The height of the bars corresponds to the LOD value. (b) Intra-assay reproducibility for each microsatellite (x-axis). The height of the bars corresponds to the reproducibility value (0.0 to 1.0, y-axis). The limit of detection (LOD) for minority clones varied among markers and replicates. In general, the LOD was very low for TA81 and Poly-α, as all four laboratory strains were detectable only up to the 5:1:1:1 ratio. For TA40, the LOD was three times higher, with all strains detected up to a 1:1:15:1 ratio. PfPK2 had the highest LOD, detecting all strains at a 1:50:1:50 ratio, about ten times higher than TA81 and Poly-α (Fig. 1a). Intra-assay reproducibility was moderately high for TA40, PfPK2 and Poly-α (concordance of 63.6%), and lower for TA81 (concordance of 18.2%) (Fig. 1b). Diversity and allelic frequency of markers used A total of 256 patient samples, including 188 baseline and 34 paired samples, were successfully analyzed for msp1, msp2, glurp, TA40, TA81, Poly-α, and PfPK2. Additionally, the 34 paired samples were successfully sequenced for SNP-rich markers. The full sample list and a flow chart illustrating the number and origin of samples analyzed are available in Supplementary Tables 6 and Supplementary Fig. 1. Genetic diversity and expected heterozygosity (He) varied across the markers (Fig. 2; Table 1). When considering msp1, msp2 and glurp, msp2/3D7 and glurp had the highest diversity, with 65 (He = 0.976) and 39 (He = 0.954) genotypes detected, respectively. Msp1/MAD20, msp1/K1 and msp2/FC27 were less diverse, with 25 (He = 0.927), 24 (He = 0.908), and 33 (He = 0.883) genotypes, respectively. For msp2/FC27, 3 genotypes had high allelic frequencies of 23%, 18%, and 15%, and were present in more than half of the isolates. Microsatellites generally exhibited lower diversity than the other length-polymorphic markers. Poly-α was the most diverse, with 23 distinct genotypes (He = 0.923), followed closely by TA40 (23 genotypes, He = 0.881), and TA81 (12 genotypes, He = 0.853). PfPK2 was the least diverse microsatellite, with one dominant genotype exhibiting an allelic frequency of 34% (16 genotypes, He = 0.834). All SNP-rich markers had high diversity. Cpmp had the highest diversity, with 52 haplotypes (He = 0.981), followed by cpp with 39 haplotypes (He = 0.961). Ama1-D2 and ama1-D3, two fragments of the ama1 gene, yielded very similar diversity values, as expected, with each 38 haplotypes and expected heterozygosity, He, of 0.955 and 0.959, respectively. Csp also showed similar diversity with 35 haplotypes (He = 0.958). The He for SNP-rich markers was higher compared to length polymorphic markers, except msp2/3D7. Fig. 2Diversity and allelic frequency of different markers. Each pie chart represents the allelic diversity of one marker. The detected alleles are indicated with different shades of each color, while their frequency is represented by the size of the slice and indicated in percentage (%). Only frequencies > 10% are indicated. (a) msp1 (MAD20, K1), msp2 (3D7, FC27), and glurp. (b) Microsatellites (Poly-α, TA40, TA81, and PfPK2). (c) SNP-rich markers (cpmp, ama1-D2, ama1-D3, cpp, and csp). Each row is ordered by decreasing expected heterozygosity (i.e., decreasing diversity). n = number of genotypes/haplotypes. Diversity and allelic frequency of different markers. Each pie chart represents the allelic diversity of one marker. The detected alleles are indicated with different shades of each color, while their frequency is represented by the size of the slice and indicated in percentage (%). Only frequencies > 10% are indicated. (a) msp1 (MAD20, K1), msp2 (3D7, FC27), and glurp. (b) Microsatellites (Poly-α, TA40, TA81, and PfPK2). (c) SNP-rich markers (cpmp, ama1-D2, ama1-D3, cpp, and csp). Each row is ordered by decreasing expected heterozygosity (i.e., decreasing diversity). n = number of genotypes/haplotypes. Table 1Expected heterozygosity of markers. Values vary between 0 (no diversity) and 1 (high diversity). Markers are listed by decreasing expected heterozygosity.MarkerExpected heterozygosity (He) msp1/msp2/ glurp 3D7 0.976 glurp 0.954 MAD20 0.927 K1 0.908 FC27 0.883Microsatellites Poly-α 0.923 TA40 0.881 TA81 0.853 PfPK2 0.834SNP-rich markers cpmp 0.981 cpp 0.961 ama1-D3 0.959 csp 0.958 ama1-D2 0.955 Table 1Expected heterozygosity of markers. Values vary between 0 (no diversity) and 1 (high diversity). Markers are listed by decreasing expected heterozygosity.MarkerExpected heterozygosity (He) msp1/msp2/ glurp 3D7 0.976 glurp 0.954 MAD20 0.927 K1 0.908 FC27 0.883Microsatellites Poly-α 0.923 TA40 0.881 TA81 0.853 PfPK2 0.834SNP-rich markers cpmp 0.981 cpp 0.961 ama1-D3 0.959 csp 0.958 ama1-D2 0.955 Expected heterozygosity of markers. Values vary between 0 (no diversity) and 1 (high diversity). Markers are listed by decreasing expected heterozygosity. msp1/msp2/ glurp Multiplicity of infection MOI varied among sites and markers (Fig. 3). In general, the MOI was slightly lower in Masaka (mean across all markers: 1.47, range 1–4), followed by Bugarama (mean: 1.64, range 1–6), and Rukara (mean: 1.77, range 1–7). Differences in MOI were found to be statistically significant between Bugarama and Masaka (adj. p-value = 1.47·10− 3) and between Rukara and Masaka (adj. p-value = 1.38·10− 5). Among the markers, msp1 revealed the highest MOI (mean: 2.48, range 1–7), followed by msp2 (mean: 2.47, range 1–7), csp (mean: 2.03, range 1–6), cpmp (mean: 2.00, range 1–6), and cpp (mean; 1.67, range 1–5). The lowest MOI values were detected by ama1-D2 and D3, the microsatellites and glurp. Several statistically significant (adj. p-values < 0.05) pairwise differences between markers were found (Supplementary Table 7). Msp1, Poly-α, and csp had the highest MOI values among their respective marker type and were thus selected for analysis of MOI in the 34 day 0/day X sample pairs (Supplementary Fig. 2). For day 0 samples, all three markers detected the same number of clones (i.e., same MOI) in 14.7% (5/34) of the samples. Poly-α and csp detected the same number of clones in 50% (17/34) of the samples, while msp1 and Poly-α and msp1 and csp had lower concordance rates (23.5% (8/34) and 29.4% (10/34), respectively). For day X samples, the concordance between the three markers was higher. Indeed, all three found the same MOI in 35.3% (12/34) of the samples, and all other marker combinations had similar concordance rates (msp1 and Poly-α: 52.9% (18/34), msp1 and csp: 52.9% (18/34), Poly-α and csp: 50% (17/34)). Sample-level MOI data for all day 0/day X samples is shown in Supplementary Table 8. Fig. 3MOI detected by each marker in all sites and stratified by site (Masaka, Rukara, and Bugarama). The violin plots represent the distribution of MOI values. The black dots represent the mean MOI (value indicated next to each violin), while the colored dots represent the MOI of each sample. Markers: msp1, msp2, glurp, PfPK2, Poly-α, TA40, TA81, cpmp, ama1-D2, ama1-D3, cpp and csp. For length polymorphic markers, 256 samples were analyzed. For AmpSeq markers, 68 samples were analyzed, and only 4 of them were from Masaka. MOI detected by each marker in all sites and stratified by site (Masaka, Rukara, and Bugarama). The violin plots represent the distribution of MOI values. The black dots represent the mean MOI (value indicated next to each violin), while the colored dots represent the MOI of each sample. Markers: msp1, msp2, glurp, PfPK2, Poly-α, TA40, TA81, cpmp, ama1-D2, ama1-D3, cpp and csp. For length polymorphic markers, 256 samples were analyzed. For AmpSeq markers, 68 samples were analyzed, and only 4 of them were from Masaka. Classification of recurrences The total number of recrudescences varied among the different marker combinations and decision algorithms, as shown in Fig. 4. One sample pair (01/03) was excluded from the analysis because it was negative for cpp. The WHO algorithm classified the least amount of recrudescences compared to the other algorithms. Indeed, it classified on average 6.4 recrudescences (range: 5–9). The 2/3 algorithm classified the highest (mean: 10.5) and least consistent amount of recrudescences, varying from 7 with ama1-D3/cpmp/cpp to 19 with msp1/msp2/TA81 (range: 7–19). The Bayesian algorithm found on average 8.8 recrudescences (range: 6.2–10.7). The three algorithms showed the most consistent results with the SNP-rich markers combinations, especially with the ama1-D3/cpmp/cpp, ama1-D2/cpmp/cpp, ama1-D2/ama1-D3/csp and ama1-D2/ama1-D3/cpmp combinations (Table 2). The ama1-D2/ama1-D3/cpmp combination provided the most consistent results independent of the algorithm used; therefore, it was selected as the “reference method” for all further analyses. Fig. 4Number of recrudescences found with each marker combination and algorithm. Number of recrudescences is shown on the y-axis. Marker combinations (x-axis): msp1/msp2/glurp, msp1/msp2/Poly-α, msp1/msp2/TA81, msp1/msp2/TA40, msp1/msp2/PfPK2, cpmp/ama1-D3/csp, cpmp/ama1-D2/csp, cpmp/ama1-D3/cpp, cpmp/ama1-D2/cpp, csp/ama1-D3/cpp, csp/ama1-D2/cpp, ama1-D2/ama1-D3/csp, ama1-D2/ama1-D3/cpmp (highlighted, chosen as “reference method”), ama1-D2/ama1-D3/cpp, and cpmp/cpp/csp. Classification algorithms: WHO, 2/3, Bayesian. For the Bayesian algorithm, the number of recrudescences corresponds to the sum of all posterior probabilities of recrudescence. Number of recrudescences found with each marker combination and algorithm. Number of recrudescences is shown on the y-axis. Marker combinations (x-axis): msp1/msp2/glurp, msp1/msp2/Poly-α, msp1/msp2/TA81, msp1/msp2/TA40, msp1/msp2/PfPK2, cpmp/ama1-D3/csp, cpmp/ama1-D2/csp, cpmp/ama1-D3/cpp, cpmp/ama1-D2/cpp, csp/ama1-D3/cpp, csp/ama1-D2/cpp, ama1-D2/ama1-D3/csp, ama1-D2/ama1-D3/cpmp (highlighted, chosen as “reference method”), ama1-D2/ama1-D3/cpp, and cpmp/cpp/csp. Classification algorithms: WHO, 2/3, Bayesian. For the Bayesian algorithm, the number of recrudescences corresponds to the sum of all posterior probabilities of recrudescence. Table 2Number of recrudescences with each algorithm, mean, standard deviation. The ama1-D2/ama1-D3/cpmp combination (bold) was chosen as “reference method”.WHO algorithm2/3 algorithmBayesian algorithmMeanStandard deviation msp1/msp2/glurp 7118.38.771.66 msp1/msp2/Poly-α 711109.331.70 msp1/msp2/TA81 71910.712.235.02 msp1/msp2/TA40 7129.89.62.05 msp1/msp2/PfPK2 91510.211.42.59 cpmp/ama1-D3/csp 698.77.91.35 cpmp/ama1-D2/csp 699.58.171.55 ama1-D2/ama1-D3/csp 799.38.431.02 ama1-D2/ama1-D3/cpmp 798.18.040.82 cpmp/cpp/csp 598.37.431.75 ama1-D2/ama1-D3/cpp 6106.27.41.84 ama1-D2/cpmp/cpp 688.67.531.11 ama1-D2/cpp/csp 5108.27.732.07 ama1-D3/cpmp/cpp 678.57.161.03 ama1-D3/cpp/csp 597.97.31.69 Table 2Number of recrudescences with each algorithm, mean, standard deviation. The ama1-D2/ama1-D3/cpmp combination (bold) was chosen as “reference method”.WHO algorithm2/3 algorithmBayesian algorithmMeanStandard deviation msp1/msp2/glurp 7118.38.771.66 msp1/msp2/Poly-α 711109.331.70 msp1/msp2/TA81 71910.712.235.02 msp1/msp2/TA40 7129.89.62.05 msp1/msp2/PfPK2 91510.211.42.59 cpmp/ama1-D3/csp 698.77.91.35 cpmp/ama1-D2/csp 699.58.171.55 ama1-D2/ama1-D3/csp 799.38.431.02 ama1-D2/ama1-D3/cpmp 798.18.040.82 cpmp/cpp/csp 598.37.431.75 ama1-D2/ama1-D3/cpp 6106.27.41.84 ama1-D2/cpmp/cpp 688.67.531.11 ama1-D2/cpp/csp 5108.27.732.07 ama1-D3/cpmp/cpp 678.57.161.03 ama1-D3/cpp/csp 597.97.31.69 Number of recrudescences with each algorithm, mean, standard deviation. The ama1-D2/ama1-D3/cpmp combination (bold) was chosen as “reference method”. Variations in outcome between marker combinations were largely driven by a subset of recurrences, as shown in Fig. 5. Indeed, 81.8% (27/33), 54.5% (18/33) and 72.7% (24/33) of the recurrences were assigned the same classification (recrudescence (R), in red, or new infection (NI), in blue) with all marker combinations by the WHO, 2/3, and Bayesian algorithm, respectively. The remaining recurrences were classified differently depending on the marker combination. Concordance between the different algorithms was observed in 39.4% (13/33) of the sample pairs independently of the marker combination used. When replacing glurp with each of the four microsatellites and applying the WHO algorithm, the same number of recrudescences (seven) was found, except when using the microsatellite PfPK2 which detected nine recrudescences. Interestingly, the recrudescences were not found in the same sample pairs for all marker combinations. For 87.9% (29/33) of the sample pairs the results were concordant, but 4 sample pairs gave discordant results depending on the marker combination. The discordant cases are indicated in bold in Fig. 5. Extensive length polymorphic and SNP-rich markers data for each sample is presented in Supplementary Table 9. Fig. 5Recrudescence vs. new infection classification of all sample pairs with the different algorithms. Sample IDs are indicated on the y-axis. Marker combinations (x-axis): msp1/msp2/glurp, msp1/msp2/Poly-α, msp1/msp2/TA81, msp1/msp2/TA40, msp1/msp2/PfPK2, cpmp/ama1-D3/csp, cpmp/ama1-D2/csp, cpmp/ama1-D3/cpp, cpmp/ama1-D2/cpp, csp/ama1-D3/cpp, csp/ama1-D2/cpp, ama1-D2/ama1-D3/csp, ama1-D2/ama1-D3/cpmp (highlighted, chosen as “reference method”), ama1-D2/ama1-D3/cpp, and cpmp/cpp/csp. Classification algorithms: WHO, 2/3, Bayesian. Previously published data from5 was added as “7 M” (for 7 microsatellites). Outcomes are colored in a gradient from light blue (prob. of R: 0) to dark blue (prob. of R: 0.49) for new infections, and from light pink (prob. of R: 0.5) to dark red (prob. of R: 1) for recrudescences. Bold: samples which showed discordant R vs. NI results when replacing glurp with a microsatellite. Underlined: samples which showed discordant R vs. NI results depending on the method used (msp1/msp2/glurp with WHO algorithm, ama1-D2/ama1-D3/cpmp with WHO algorithm, 7 microsatellites with Bayesian algorithm5. Recrudescence vs. new infection classification of all sample pairs with the different algorithms. Sample IDs are indicated on the y-axis. Marker combinations (x-axis): msp1/msp2/glurp, msp1/msp2/Poly-α, msp1/msp2/TA81, msp1/msp2/TA40, msp1/msp2/PfPK2, cpmp/ama1-D3/csp, cpmp/ama1-D2/csp, cpmp/ama1-D3/cpp, cpmp/ama1-D2/cpp, csp/ama1-D3/cpp, csp/ama1-D2/cpp, ama1-D2/ama1-D3/csp, ama1-D2/ama1-D3/cpmp (highlighted, chosen as “reference method”), ama1-D2/ama1-D3/cpp, and cpmp/cpp/csp. Classification algorithms: WHO, 2/3, Bayesian. Previously published data from5 was added as “7 M” (for 7 microsatellites). Outcomes are colored in a gradient from light blue (prob. of R: 0) to dark blue (prob. of R: 0.49) for new infections, and from light pink (prob. of R: 0.5) to dark red (prob. of R: 1) for recrudescences. Bold: samples which showed discordant R vs. NI results when replacing glurp with a microsatellite. Underlined: samples which showed discordant R vs. NI results depending on the method used (msp1/msp2/glurp with WHO algorithm, ama1-D2/ama1-D3/cpmp with WHO algorithm, 7 microsatellites with Bayesian algorithm5. Recrudescence vs. new infection: comparison with previous results The results obtained during this study were compared to those of a study published previously5, which employed the 7 microsatellites assay analyzed with the Bayesian algorithm approach. With the “reference method” ama1-D2/ama1-D3/cpmp analyzed with the WHO algorithm, 7 recrudescences were detected. The msp1/msp2/glurp combination with the WHO algorithm and the 7 microsatellites analyzed with the Bayesian algorithm also detected 7 and 6.54 (7, rounded) recrudescences, respectively. In this case, too, recrudescences were not found in the same sample pairs. For 84.4% (28/33) of the sample pairs, the results were concordant between the different methods used, but 5 pairs gave discordant results, as shown underlined in Fig. 5. Identity-by-descent (IBD) analysis Identity-by-descent analysis of paired samples was conducted on SNP-rich markers data and revealed statistically significant high genetic relatedness (IBD = 1) for 5/33 sample pairs (Table 3). All these pairs were also classified as a recrudescence by the WHO and 2/3 algorithms and had a Bayesian posterior probability of recrudescence > 0.962 for the ama1-D2/ama1-D3/cpmp combination. Two other sample pairs (01/21 and 02/35) had statistically significant moderate genetic relatedness (IBD > 0.5, where a 0.5 value is expected for full siblings34). Both were classified as recrudescence by the WHO, 2/3, and the Bayesian algorithms (Table 3). Table 3Comparison of IBD values with WHO and Bayesian classification algorithms. IBD analysis was performed on cpmp, ama1-D2, ama1-D3, csp, and cpp. Results are compared to those obtained with WHO, 2/3, and Bayesian algorithms for the ama1-D2/ama1-D3/cpmp combination. For the Bayesian algorithm, a sample pair is considered a recrudescence when the posterior probability of recrudescence is > 0.5. R: recrudescence, NI: new infection.Sample pairIBD (95% CI)p-value (IBD)WHO and 2/3 algorithmBayesian algorithm01/16D001/16D211 (0.66–1.00)2.67·10− 8R1.00001/21D001/21D210.665 (0.14–0.98)2.58·10− 3R0.99601/34D001/34D211 (0.62–1.00)6.74·10− 6R0.98902/25D002/25D280.778 (0.31–0.99)5.82·10− 5R0.77502/30D002/30D211 (0.66–1.00)2.42·10− 8R0.96203/02D003/02D211 (0.65–1.00)1.38·10− 7R0.99903/23D003/23D211 (0.65–1.00)2.79·10− 8R0.999 Table 3Comparison of IBD values with WHO and Bayesian classification algorithms. IBD analysis was performed on cpmp, ama1-D2, ama1-D3, csp, and cpp. Results are compared to those obtained with WHO, 2/3, and Bayesian algorithms for the ama1-D2/ama1-D3/cpmp combination. For the Bayesian algorithm, a sample pair is considered a recrudescence when the posterior probability of recrudescence is > 0.5. R: recrudescence, NI: new infection.Sample pairIBD (95% CI)p-value (IBD)WHO and 2/3 algorithmBayesian algorithm01/16D001/16D211 (0.66–1.00)2.67·10− 8R1.00001/21D001/21D210.665 (0.14–0.98)2.58·10− 3R0.99601/34D001/34D211 (0.62–1.00)6.74·10− 6R0.98902/25D002/25D280.778 (0.31–0.99)5.82·10− 5R0.77502/30D002/30D211 (0.66–1.00)2.42·10− 8R0.96203/02D003/02D211 (0.65–1.00)1.38·10− 7R0.99903/23D003/23D211 (0.65–1.00)2.79·10− 8R0.999 Comparison of IBD values with WHO and Bayesian classification algorithms. IBD analysis was performed on cpmp, ama1-D2, ama1-D3, csp, and cpp. Results are compared to those obtained with WHO, 2/3, and Bayesian algorithms for the ama1-D2/ama1-D3/cpmp combination. For the Bayesian algorithm, a sample pair is considered a recrudescence when the posterior probability of recrudescence is > 0.5. R: recrudescence, NI: new infection. 01/16D0 01/16D21 01/21D0 01/21D21 01/34D0 01/34D21 02/25D0 02/25D28 02/30D0 02/30D21 03/02D0 03/02D21 03/23D0 03/23D21 Discussion Accurately estimating antimalarial drug efficacy through Therapeutic Efficacy Studies (TESs) is crucial to inform policy makers, especially in regions where artemisinin partial resistance has emerged. An overestimation of drug efficacy could lead to delays in recognizing the spread and magnitude of drug resistance, and prevent necessary action to mitigate the impact of reduced drug efficacy, while an underestimation of efficacy could lead to premature changes in treatment policies, possibly leading to lower access to treatment for the transition period and additional costs for malaria programs16. Accurate estimation of drug efficacy requires robust and reproducible PCR correction methods which utilize highly sensitive assays and markers with high genetic diversity and low allelic frequencies. Additionally, it requires a decision algorithm which allows for correct classification of a recurrent infection as a recrudescence or a new infection based on several markers. Previous studies have shown that SNP-rich markers analyzed with AmpSeq and msp1/msp2 analyzed with high-resolution capillary electrophoresis (H-CE) have very high sensitivity in detecting minority clones, while glurp has much lower sensitivity19,20. We had initially assessed the sensitivity of microsatellites in detecting minority clones using single-round PCR and found it to be moderate19. Here, we aimed to assess the sensitivity of a nested PCR assay for four microsatellites, expecting higher sensitivity compared to the single-round PCR. Out of the four microsatellites tested, high sensitivity in detecting minority clones was observed for TA40 and PfPK2. For the latter, the sensitivity was comparable to that of the AmpSeq assay, detecting minority clones in a 1:50:1:50 ratio. However, high variability between replicates was observed, in contrast to the one-round PCR assay, which was previously shown to be highly reproducible with low intra- and inter-assay variability19,28. Low marker diversity and high allelic frequency can lead to an overestimation of recrudescences and thus an underestimation of drug efficacy. Therefore, it is critical to use highly diverse markers with low allelic frequencies. Here, we report high diversity (He > 0.95) and low allelic frequency for msp2/3D7, glurp and all SNP-rich markers, and lower diversity (He < 0.90) for msp2/FC27 and most of the microsatellites, with genotypes harboring higher allelic frequencies, thus possibly introducing a bias towards a recrudescence classification for non-statistical methods like match counting approaches that do not take allele frequency into account. This is in line with previous studies, where msp2/3D719,25,35–37, Poly-α38–41, and cpmp19–21 have been consistently reported to be the most diverse length polymorphic, microsatellite and SNP-rich marker, respectively. MOI data can be used as a proxy for marker diversity and sensitivity, also reflecting transmission intensity in the study areas. Indeed, we found significant differences in MOI between sites, likely reflecting differences in transmission intensity. As expected, MOI was lower in Masaka, where in 2018 malaria incidence was reported to be 100–250 cases per 1000 population, compared to Bugarama and Rukara, where incidence was > 450 cases per 1000 population42,43. Higher MOI values correspond to higher numbers of clones detected in a sample and are therefore associated with higher genetic diversity, higher sensitivity in minority clones’ detection, and higher transmission intensity. Within this dataset, msp1, msp2 and most SNP-rich markers detected high MOI values, confirming their high discriminatory power. The multi-allelic family nature of msp1 and msp2 also contributed to higher MOI values for these markers. In contrast, glurp and microsatellites generally had lower MOI values, reflecting the low sensitivity and/or low diversity of these markers. In our study, ama1-D2 and ama1-D3 also had lower MOI values, but this was likely due to the higher thresholds applied to the reads to discard spurious haplotypes, rather than to low sensitivity and diversity. The WHO recommends replacing glurp with the microsatellite with greatest diversity in the study region16. Our diversity and MOI data suggest that Poly-α, PfPK2 and TA40 may be adequate candidates for this purpose. To evaluate this hypothesis, we replaced glurp with each microsatellite and applied different decision algorithms to all marker combinations. With the WHO algorithm, we found a consistent and lowest number of recrudescences across marker combinations. This approach requires concordant results from all markers to confirm a recrudescence, and it is therefore expected that a low number of them are detected. The 2/3 algorithm classified the highest number of recrudescences, up to 19 in the case of the msp1/msp2/TA81 combination. This algorithm was initially introduced to overcome the limitation of glurp in detecting minority clones and thus in detecting recrudescent infections16, however, it is likely to overestimate recrudescences as previously described19. The Bayesian algorithm detected a higher number of recrudescences than the WHO algorithm, but a lower number than the 2/3 algorithm, for most marker combinations. By providing a posterior probability of recrudescence based on marker diversity, this algorithm may overcome the limitation of match counting algorithms underestimating or overestimating the number of recrudescences. Based on these results, it is not possible to conclude which algorithm is most accurate; however, probabilistic approaches may have the potential to provide more accurate outcomes compared to match counting approaches22,44. Indeed, identity-by-descent (IBD) analysis confirmed significant genetic relatedness of clones in 7 sample pairs and yielded the same results as the Bayesian algorithm. The number of recrudescences detected was most consistent across algorithms for the ama1-D2/ama1-D3/cpmp combination. Due to the minimal variability given by the different algorithms, we used this combination as a “reference method”. It is important to note that consistency across algorithms does not necessarily imply accuracy. Therefore, “reference method” should be interpreted as a comparator rather than a gold standard in this context. The comparison of our results with those published previously5 showed that the msp1/mp2/glurp and ama1-D2/ama1-D3/cpmp analyzed with the WHO algorithm and the 7 microsatellites analyzed with the Bayesian algorithm gave the same results in terms of number of recrudescences detected, but results were discordant for some sample pairs. Discordant results were due to the low genetic diversity and high allelic frequency of some markers, and/or to the low sensitivity in detecting minority clones of other markers. Our study has some limitations. First, the nested PCR assay for microsatellites had low reproducibility, which could have resulted in inaccurate results. We also assessed intra-assay variability but did not assess inter-assay variability. Second, erroneously calling stutter peaks (for length polymorphic markers) or haplotypes with low coverage (for SNP-rich markers) as real clones may have led to wrong estimations of diversity, MOI and recrudescences. This is particularly true for microsatellites, for which data analysis is complex due to the high presence of stutter peaks and small size differences between real clones. Third, only 34 sample pairs were used for the analysis of SNP-rich markers, and only 2 of them were collected in Masaka. The small sample size and low number of recrudescent infections may limit the generalizability of the results. Fourth, we found that the ama1-D2/ama1-D3/cpmp combination gave the most consistent results and used it as “reference method”. However, ama1-D2 and ama1-D3 are two fragments of the same gene (ama1), which could make them more likely to be similar in terms of genetic diversity, and thus lead to biased results. Fifth, to remove the spurious haplotypes detected in the positive controls, which were likely due to contaminations and/or sequencing errors, we applied thresholds that were defined based on the positive controls. These contained laboratory strain mixtures which differ in diversity and complexity from patient samples; therefore, the thresholds may not perform equivalently in the two types of samples. We applied quite stringent thresholds for ama1-D2 and ama1-D3, which have probably led to removing true haplotypes. On the contrary, by not applying any thresholds to csp data, spurious haplotypes may have been incorrectly classified as true haplotypes. Finally, the generalizability of our results may be limited to settings with similar malaria transmission intensity to Rwanda. Indeed, the relatively low discordance between marker combinations and algorithms reported here may be largely due to low transmission levels and the resulting low complexity of infections in this dataset. In settings with higher transmission, discrepancies between methods may be much more significant, as reported by a recent study conducted in Uganda40. The AmpSeq assay has previously proven to be highly sensitive and robust19 and therefore holds great potential for accurate PCR-corrected drug efficacy estimations. However, the importance of bioinformatics and data analysis steps should not be underestimated. Indeed, applying different thresholds at different analysis steps, such as for example the mismatch rate for SNPs calling or the marker-specific threshold to discard spurious haplotypes, may significantly impact the final genotyping outcome. A careful evaluation and standardization of the pipelines and downstream analysis steps would be necessary to produce robust results and allow for comparison between studies conducted across different locations and time periods. In conclusion, our results emphasize the importance of using assays with high sensitivity in detecting minority clones, and markers with high diversity and low allelic frequencies, for accurate PCR correction. Both sensitivity and diversity must be considered when choosing a microsatellite to replace glurp. In our study, replacing glurp with a microsatellite did not significantly change the final number of recrudescent infections, and therefore would not have affected drug efficacy estimations. This was likely due to the low complexity of infections resulting from low transmission intensity in the region. Nonetheless, the SNP-rich markers method yielded the most consistent results in detecting recrudescences, regardless of the decision algorithm used, making it a robust alternative to the currently used length polymorphic markers. Decision algorithms based on probabilistic approaches or genetic relatedness rather than match counting need to be further assessed, as they may provide more accurate efficacy estimations. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1 Supplementary Material 2 Supplementary Material 2