Design and implementation of multiplexed amplicon sequencing panels to serve genomic epidemiology of infectious disease: A malaria case study

Multiplexed PCR amplicon sequencing (AmpSeq) is an increasingly popular application for cost‐effective monitoring of threatened species and managed wildlife populations, and shows strong potential for the genomic epidemiology of infectious disease. AmpSeq data from infectious microbes can inform disease control in multiple ways, such as by measuring drug resistance marker prevalence, distinguishing imported from local cases, and determining the effectiveness of therapeutics. We describe the design and comparative evaluation of two new AmpSeq assays for Plasmodium falciparum malaria parasites: a four‐locus panel (“4CAST”) composed of highly diverse antigens, and a 129‐locus panel (“AMPLseq”) composed of drug resistance markers, highly diverse loci for inferring relatedness, and a locus to detect Plasmodium vivax co‐infection. We explore the performance of each panel in various public health use cases with in silico simulations as well as empirical experiments. The 4CAST panel appears highly suitable for evaluating the number of distinct parasite strains within samples (complexity of infection), showing strong performance across a wide range of parasitaemia levels without a DNA pre‐amplification step. For relatedness inference, the larger AMPLseq panel performs similarly to two existing panels of comparable size, despite differences in the data and approach used for designing each panel. Finally, we describe an R package (paneljudge) that facilitates the design and comparative evaluation of genetic panels for relatedness estimation, and we provide general guidance on the design and implementation of AmpSeq panels for the genomic epidemiology of infectious disease.


| INTRODUC TI ON
Genetic data are a valuable resource for understanding microbial ecology and the epidemiology of infectious disease. The value of this data type has been highlighted by the COVID-19 pandemic, for which viral sequence analysis has greatly informed patterns of disease spread and evolution, influencing public health policy decisions around the world (Oude Munnink et al., 2021). Applications of genetic data in epidemiology extend from viral and bacterial outbreak management (Gardy & Loman, 2018) to the study of eukaryotic parasites underlying important diseases such as malaria, leishmaniasis and cryptosporidiosis (Cantacessi et al., 2015;Nader et al., 2019;Neafsey et al., 2021).
Many use cases (applications) of genetic data have been identified for malaria (Dalmat et al., 2019), the leading parasitic killer worldwide (WHO, 2019). These include tracking the spread of drug/ insecticide resistance markers and diagnostic mutations (Chenet et al., 2016;Jacob et al., 2021;Kayiba et al., 2021;Lautu-Gumal et al., 2021;Miotto et al., 2020), assessing disease transmission levels Galinsky et al., 2015), identifying sources of infections and imported cases (Liu et al., 2020;Tessema et al., 2019), and estimating genetic connectivity among different populations . Malaria parasite genetic data have also demonstrated utility in therapeutic efficacy studies, such as for distinguishing reinfections from recrudescent infections potentially indicative of drug inefficacy (Gruenberg et al., 2019;Jones et al., 2021). These applications in the malaria field are served by different types of genetic data produced at varying resolution, technical complexity and cost, ranging from genetic panels that comprise as few as 8-12 polymorphic microsatellites or 24 single nucleotide polymorphisms (SNPs) (Daniels et al., 2008;Yalcindag et al., 2012) to whole genome sequencing (WGS) data (Miotto et al., 2015;Takala-Harrison et al., 2015).
To be scalable and sustainable, genetic data should be produced at the minimum resolution that provides robust support for the intended analysis application. WGS data often provide the most complete population genomic perspective on an organism of interest. However, the cost and technical challenges of generating, storing and interpreting WGS data are impediments to scalability and widespread implementation for organisms with large genomes, or microbes with small genomes in samples dominated by host DNA.
Targeted sequencing approaches that focus deep coverage on select genomic regions of interest using multiplexed PCR amplification (AmpSeq) are finding increased application in areas of conservation biology, fisheries science and evolutionary research (Baetscher et al., 2018;Bybee et al., 2011;Dupuis et al., 2018;Hargrove et al., 2021;Natesh et al., 2019;O'Neill et al., 2013;Schmidt et al., 2020). These approaches can also serve genomic epidemiology of infectious diseases by focusing sequencing coverage on the most informative regions of pathogen genomes.
While only a few AmpSeq protocols for eukaryotic parasites have been published to date, pioneer examples for malaria and trypanosomatid parasites have confirmed the viability of this approach with low-parasitaemia host and vector samples, where parasite DNA comprises a very small fraction of the total sample (Jacob et al., 2021;Ruybal-Pesántez et al., 2021;Schwabl et al., 2020;Tessema et al., 2020). Furthermore, one recent study has confirmed the value of designing amplicons to capture multi-SNP Plasmodium falciparum "microhaplotypes," which exhibit polyallelic rather than biallelic diversity to facilitate relatedness inference . New relatedness-based analytical approaches for genomic epidemiology are currently developing for malaria parasites and other sexually recombining pathogens (Henden et al., 2018;Schaffner et al., 2018).
The use of genomic data for estimation of recent common ancestry shared by pairs or clusters of parasites or mosquitoes has shown strong potential to provide epidemiologically useful insights over small geographical distances (10s to 100s of kilometres) and short timescales (weeks to months) relative to traditional population genetic parameters of population diversity and divergence Taylor et al., 2017). While many analyses of recent co-infection. We explore the performance of each panel in various public health use cases with in silico simulations as well as empirical experiments. The 4CAST panel appears highly suitable for evaluating the number of distinct parasite strains within samples (complexity of infection), showing strong performance across a wide range of parasitaemia levels without a DNA pre-amplification step. For relatedness inference, the larger AMPLseq panel performs similarly to two existing panels of comparable size, despite differences in the data and approach used for designing each panel.
Finally, we describe an R package (paneljudge) that facilitates the design and comparative evaluation of genetic panels for relatedness estimation, and we provide general guidance on the design and implementation of AmpSeq panels for the genomic epidemiology of infectious disease.

K E Y W O R D S
amplicon sequencing, epidemiology, genome, genotyping, malaria, relatedness common ancestry in malaria parasites to date have used WGS data, targeted genotyping of as few as 200 biallelic SNPs or 100 polyallelic loci (e.g., microsatellites or microhaplotypes) may also be used to infer relatedness with necessary precision , making AmpSeq an excellent candidate for relatedness estimation.
However, there remains uncertainty in the molecular epidemiology field as to the suitability of existing panels for profiling pathogen populations in specific geographical locations that did not inform the original panel designs, and it is unclear which protocol features are most conducive to implementation in both high-and low-resource settings. Should each disease field adopt a common multiplexed amplicon protocol and panel, or should bespoke panels be implemented regionally to address genetically distinct pathogen populations and specific use cases?
To address these questions, here we describe the design and comparative evaluation of two new multiplexed amplicon assays for P. falciparum malaria parasites: a four-locus panel composed of highly diverse loci, useful for estimating the number of genetically distinct strains within an infection (complexity of infection; COI) as well as distinguishing between continuing and newly acquired infections in any geographical setting; and a 129-locus panel composed of drug resistance markers and many diverse loci for relatedness inference initially designed for application in South America (a region that did not inform previously published panel designs). Both assays use nonproprietary reagents (including standard PCR oligos) to maximize accessibility and affordability in malaria-endemic settings. The panels are supported by new open-source bioinformatic analysis pipelines to facilitate widespread use. We also show that the core sets of multiplexed PCR oligos can flexibly accommodate various new targets not included in the original designs, allowing for panel customization towards detecting locally relevant resistance markers, polymorphic loci and co-infecting parasite species. We use WGS data to explore the degree to which our newly described and previously published genotyping panels can serve studies in diverse geographies vs. the alternative of customizing panels with targets that are locally informative but not globally useful. We suggest there is value in genotyping panels that can be flexibly adapted to incorporate informative targets from pathogen populations of interest. The analyses and resources described here clarify the rapidly diversifying options for targeted microbial sequencing ( Figure 1) by providing tools and guidance for the comparative evaluation and refinement of AmpSeq approaches.

| Panel designs
We developed a small multiplex of four highly polymorphic antigenic loci, dubbed "4CAST": CSP, AMA1, SERA2 and TRAP ( Figure 2). All four amplicons use previously published primer sequences (Miller et al., 2017;Neafsey et al., 2015), as no modification was required for successful multiplexing.
In designing the larger multiplexed amplicon panel we call "AMPLseq" (short for "Assorted Mix of Plasmodium Loci"), we first built a large pool of candidate loci, anticipating significant attrition of candidates due to primer incompatibility. We prioritized four classes of loci: loci within antigens of interest (Helb et al., 2015), loci with high population diversity for relatedness inference , loci included in the SpotMalaria v1 panel (Chang et al., 2019;Jacob et al., 2021) and known antimalarial drug resistance markers. We contracted the services of GTseek LLC (https://gtseek.com) to design multiplexed oligo panels according to the criteria previously described for the Genotyping-in-Thousands by F I G U R E 1 Amplicon sequencing and other genotyping approaches for genomic epidemiology of infectious diseases. Schematic of three common approaches for molecular surveillance data generation. Genomic DNA can be extracted from clinical samples and then processed using any of the three methods shown: SNP barcoding, amplicon sequencing or whole genome sequencing (WGS). Our two amplicon panels, AMPLseq and 4CAST, are shown with representations of their loci and amplification. Pre-amplification (selective whole genome amplification), which increases the ratio of parasite to human DNA in samples, is generally recommended for WGS and some amplicon sequencing panels (AMPLseq, but not 4CAST). SNP barcoding provides data in the form of variant calls at each SNP site; amplicon sequencing provides extremely deep coverage at select, small regions of the genome; and WGS generally provides shallower coverage of the entire genome could be added to the panel without compromising amplification of the other loci. We also successfully added primers amplifying known markers of drug resistance within the genes dhfr, dhps, mdr1 and kelch13 (Table S1). Furthermore, we added previously described primers targeting a region within Pvdhfr (Lefterova et al., 2015) to identify Plasmodium falciparum/P. vivax co-infections undetected in preliminary screening by microscopy or rapid diagnostic test (RDT). The final panel contains this single P. vivax locus and 128 P. falciparum loci (Figure 2), with a median length across all amplicons of 276 bp ( Figure S1).

| Panel protocols
To create the primer pool used in 4CAST PCR1, we combined 100 µm of each 4CAST primer ( conditions are provided in Protocol S1. We combined PCR2 products in equal volumes and subjected the resultant library to doublesided size selection using Agencourt AMPure XP beads (Beckman Coulter). We verified size selection via Agilent Bioanalyzer 2100 and sequenced the selected library at 6 pm with >10% PhiX in pairedend, 500-cycle format using MiSeq Reagent Kit v2 (Protocol S1).
We followed a similar nested PCR and pooled clean-up procedure for AMPLseq library construction. Primer sequences, input volumes and concentrations are listed in Table S2 and PCR conditions and size selection steps are described in Protocol S2. As detailed therein, AMPLseq library construction differs to 4CAST library construction in a few minor aspects. For example, primer input quantities vary slightly (800 pmol ± 33%) to account for amplification rate differences among loci. PCR1 products are diluted prior to PCR2 and only single-sided (left-tailed) bead-based size selection is used to enhance yield. Sequencing also occurs via paired-end, 500-cycle MiSeq but with a higher final library loading concentration (12 pm) and a lower fraction of PhiX (8%).

| Mock samples
We generated mock samples from parasite lines 3D7 and Dd2, cultured at 3% haematocrit in commercially obtained red blood cells as previously described (Trager & Jensen, 1976 1:10 dilution of the 10,000 3D7 parasites/µl control with 10 ng/ µl human gDNA. We also generated a 10,000 parasites/µl positive control as described above but using Dd2 instead of 3D7 strain gDNA. We generated mixed-strain control templates by combining the 10,000 3D7 parasites/µl control with this 10,000 Dd2 parasites/µl control at 1:1, 3:1 and 10:1 ratios (respectively). We serially diluted the 1:1 ratio to concentrations of 1000, 100 and 10 parasites/µl and diluted the 3:1 and 10:1 ratios to concentrations of 1000 and 100 parasites/µl using 10 ng/µl human gDNA diluent as before. We also applied selective whole genome amplification (sWGA) to all above control templates representing ≤1000 parasites/µl. The 50 µl sWGA reaction followed Oyola et al. (2016) with the exception of fixing template input volume to 10 µl. We purified sWGA products with Agencourt AMPure XP beads (Beckman Coulter) on the KingFisher Flex (Protocol S3) and verified amplification success via a NanoDrop (ThermoFisher Scientific).

| Clinical samples
We tested the panels on clinical dried blood spot (DBS) samples from Mali and Guyana. Tran and colleagues collected samples in Kalifabougou, Mali, between 2011 and 2013 as previously described (Tran et al., 2013). We aligned reads to the P. falciparum v3 reference genome assembly using bwa-mem (Li, 2013)

| Amplicon data analysis
We developed the application AmpSeQC (Supporting Information S2) to assess sequencing quality and yield ( Figure S2). We also used AmpSeQC for P. vivax detection by concatenating the P. falciparum 3D7 and P. vivax PvP01 reference genomes during the bwa-mem alignment step. For in-depth assessment of P. falciparum sequence variation, we processed paired-end Illumina sequencing data in the form of FASTQ files using a custom analysis pipeline (Supporting Information S2) that utilizes the Divisive Amplicon Denoising Algorithm (dada2) tool designed by Callahan et al. (2016) to obtain microhaplotypes ( Figure S2). We mapped microhaplotypes obtained from dada2 against a custom-built database of 3D7 and Dd2 reference sequences for each amplicon locus and filtered microhaplotypes based on edit distance, length and chimeric identification using a custom R script (Supporting Information S2). We summarized observed sequence polymorphism into a concise format by converting individual microhaplotypes into pseudo-CIGAR strings using a custom python script. Microhaplotypes were discarded if supported by fewer than 10 read-pairs or by less than 1% total read-pairs within a locus, or if they exhibited other error features (Supporting

Information S3).
We analysed native and pre-amplified mock samples to determine precision and sensitivity of the dada2 pipeline and filters. We defined a true positive (TP) as a microhaplotype with a pseudo-CIGAR string identical to the reference strain (either 3D7 or Dd2
Paragon HeOME v1, designed using the CleanPlex algorithm ity at different spatial scales. Primers also target various drug resistance-associated loci and mitochondrial sequences with conserved primer binding sites among Plasmodium spp. Library construction requires sWGA prior to PCR1 but no special processing between PCR1 and PCR2.
We also compared our amplicon panels to a molecular barcode assay containing 24 SNPs (Daniels et al., 2008). The SNPs targeted by this Taqman qPCR-based assay were chosen principally for their high average MAF (>0.35) across parasite sample collections from Thailand and Senegal, with further filtering to remove tightly linked SNPs and to minimize the generation of identical barcodes for closely related strains (Daniels et al., 2008).

| paneljudge and in silico data simulations
We used WGS data to simulate genotypic panel data for simulations.
This publication uses data from the MalariaGEN Plasmodium falciparum Community Project as described online pending publication and public release of data set Pf7 (https://www.malar iagen.net/ resou rce/34). Specifically, we used genomic data from monoclonal samples collected in Mali, Malawi, Senegal and Thailand (Zhu et al., 2019), and from Colombia and Venezuela (accession numbers in Table S3). We also used previously published monoclonal genomic data from Guyana (SRA BioProject PRJNA543530; Mathieu et al., 2020) and French Guiana (SRA BioProject PRJNA242182; Pelleau et al., 2015). We used the scikit-allel library (Miles et al., 2020) "read_ vcf," "is_het," "haploidify_samples" and "distinct_frequencies" functions to process the data and estimate microhaplotype frequency and diversity metrics (Supporting Information S1).
We assessed the performance of different panels for relatedness inference using simulated data. We simulated data on pairs of haploid genotypes (equivalent to pairs of monoclonal malaria samples) using paneljudge, an R package that we built to simulate data under a hidden Markov model (HMM) , which is the same HMM used in the single-population implementation of hmmIBD (Schaffner et al., 2018) (Danecek et al., 2011). We counted the number of distinct microhaplotypes observed at each locus per sample and estimated COI as the maximum number of distinct microhaplotypes observed at any locus within a sample.
To evaluate panel performance for geographical attribution, we identified microhaplotypes at loci as described above. We used the microhaplotype sequences themselves and visualized these data using the Rtsne package (Krijthe, 2015), with 5000 iterations, ϴ of 0.0 and perplexity parameter of 10. We performed principal component analyses (PCAs) using the "prcomp" function in base R version 4.1.2 (R Core Team, 2021).

| 4CAST and AMPLseq validation
We validated assay precision (defined as TP/(TP + FP)), sensitivity (defined as TP/(TP + FN)), and depth of coverage using 3D7 mock samples representing parasitaemia levels of between 10 and 10,000 parasites/µl in 10 ng/µl human DNA. Following automated and manual filtration steps (Supporting Information S3), both 4CAST and AMPLseq generated 3D7 microhaplotype calls with 100% precision for all parasitaemia levels assessed, both with and without preamplification by sWGA. 4CAST achieved high sensitivity and depth without preliminary sWGA, generating a median of 43 read-pairs per locus from native templates representing 10 parasites/µl (Figure 3a).
Median depth increased to 443 and 1312.5 read-pairs per locus for native templates representing 100 and 1000 parasites/µl, respectively. Read-pair counts were also evenly distributed among 4CAST loci using native DNA (Figure 3a).
Unlike 4CAST, AMPLseq required sWGA for 3D7 mock samples representing 10 and 100 parasites/µl (Figure 3b). Following sWGA on mock samples representing 10 parasites/µl, the assay generated ≥10 read-pairs at a median of 126 loci, with a median of 465 readpairs after excluding loci with fewer than 10 reads. Values were statistically similar for pre-amplified samples representing 100 parasites/µl and increased to 692 read-pairs for pre-amplified samples representing 1000 parasites/µl ( Figure S3). Positions on the y-axis indicate median read-pair support across replicate samples. Low sensitivity observed using native templates (grey) representing 10 and 100 3D7 parasites/µl suggests that clinical samples should generally be pre-amplified with sWGA (results at right). (c) Ratio of 4CAST read-pairs from microhaplotypes assigned to 3D7 (x-axis) or Dd2 (y-axis) from 3D7 + Dd2 mock mixtures with strain ratios of 1:1 (tan), 3:1 (pink) and 10:1 (dark red), respectively. All samples contained 1000 parasites per µl in total, that is across both strains. Dashed lines represent the expected ratio, and each point represents a 4CAST locus per sample (n = 4 per condition). (d) AMPLseq read-pair ratios observed in native mock mixtures (1000 parasites/µl) plotted as above for 4CAST We also validated the sensitivity of 4CAST and AMPLseq for genotyping polyclonal infections by using mock samples containing both 3D7 and Dd2 templates (likewise in 10 ng/µl human DNA).
4CAST showed high sensitivity for Dd2 without the need for sWGA. At 1000 parasites/µl, the assay detected Dd2-specific microhaplotypes at each of its four loci in all 1:1, 3:1 and 10:1 mixture replicates (Figure 3c). At 100 parasites/µl, median Dd2 sensitivity remained 100% at 1:1 and 3:1 ratios but was slightly lower (94%) at 10:1. At 10 parasites/µl, 1:1 ratios yielded a median of three target loci for 3D7 and a median of two target loci for Dd2; median sensitivity in these samples rose to 3.5 and 3 loci (respectively) following pre-amplification with sWGA, but this led to unbalanced read-pair support between the two strains ( Figure S4A), possibly due to differential sWGA success on low-quality Dd2 vs. higher-quality 3D7 templates. 4CAST read-pair ratios generated from native templates, by contrast, showed a strong correlation with input ratios at 100 parasites/µl ( Figure S4A) and 1000 parasites/µl (Figure 3c). Ratios were less informative at 10 parasites/µl ( Figure S4A).
AMPLseq was also successful in detecting Dd2-specific microhaplotypes, but only at a maximum of 77 of 83 dimorphic loci (in the 1:1 ratio at 10,000 parasites/µl). With sWGA, Dd2-specific sequences were detected at a minimum of two dimorphic loci for all three input ratios (1:1, 3:1, 10:1) and parasitaemia levels (≥10 parasites/µl) assessed. Like with 4CAST, however, the use of sWGA decorrelated read-pair ratios from input ratios ( Figure S4B). This decorrelation was not observed with native templates (Figure 3d) at 1000 and 10,000 parasites/µl for which AMPLseq achieved high read-pair support without the use of sWGA.
We also tested both panels on gDNA extracted from dried blood spots collected by the Guyana Ministry of Health in 2020 from individuals diagnosed as P. falciparum-positive via microscopy or RDT.
Ten Guyanese samples were tested with both panels, and an additional six were tested with AMPLseq. Using 4CAST, we observed coverage across all loci in all samples, with a median depth per locus of 1162 read-pairs without sWGA (Figure 3a). Using AMPLseq (with sWGA), we observed a median of 122 loci with ≥10 read-pairs and a median depth of 298 read-pairs per covered locus (Figure 3b).
Additionally, we tested both panels on gDNA extracted from 16 dried blood spot samples collected in Mali in 2011 (Tran et al., 2013) and subsequently stored at room temperature for 10 years.
Using 4CAST (without sWGA), we observed a median depth of 407 read-pairs per locus (Figure 3a). Using AMPLseq (with sWGA), we observed a median of 75 loci with ≥10 read-pairs and a median depth of 112 read-pairs per covered locus (Figure 3b).

| Evaluation of panel performance for relatedness
We used the R package paneljudge to assess in silico the impact of choosing a specific genotyping panel for relatedness inference.
Considering the choice of panel, we evaluated relatedness estimation from data simulated on our 4CAST and AMPLseq panels, the SpotMalaria v2 (Jacob et al., 2021) and Paragon HeOME v1  amplicon panels, and a barcode of 24 SNPs (Daniels et al., 2008). When data were simulated using microhaplotype frequency estimates of Senegalese parasites, we found that almost all estimates of unrelated or clonal pairs were correctly classified, regardless of the panel (Figure 4a). All three large panels also performed similarly well in accurately identifying related (

| Geographical attribution
We again engineered amplicon data in silico to evaluate the relative signal in genotyping panels for geographical attribution of samples. We subsampled WGS variant calls, called microhaplotypes, and evaluated these data using PCA ( Figure S5) and t-SNE visualizations ( Figure 5). We found that all three larger panels (AMPLseq,

| COI estimation
We evaluated COI estimation based on 4CAST as opposed to the single locus AMA1, which is commonly used for this purpose, alone or together with a single additional locus (Lerch et al., 2017;Miller et al., 2017;Nelson et al., 2019). We engineered in silico samples with COI ranging from 2 to 10 (100 engineered samples per COI level) and used the maximum number of unique microhaplotypes present at any locus as a simple data summary method to estimate COI.
4CAST provided more accurate estimates of COI than AMA1 alone in these simulated data, especially at simulated COI levels between 5 and 7. (Figure 6a). Estimation improved at engineered COI = 8 using AMPLseq ( Figure S7), but to reap the full benefit of the larger panel in practice will require an inferential approach that accounts for both

F I G U R E 4
In silico relatedness estimation comparisons among panels and empirical AMPLseq validation against WGS. (a) Evaluation of relatedness estimation from data simulated on genotyping panels using the paneljudge R package. Pairs of haploid genotypes were simulated at each locus of a panel, using microhaplotype frequencies estimated from a given parasite population (Colombia, Thailand or Senegal, as shown in columns from left to right). Genotype pairs were simulated at three levels of relatedness: unrelated (relatedness = 0.01), related (relatedness = 0.50) and clonal (relatedness = 0.99), as shown in the rows from top to bottom. Relatedness estimates of these pairs were classified using their 95% confidence intervals (LCI = lower limit of the 95% confidence interval, UCI = upper limit of the 95% confidence interval). Estimates could be correctly classified, misclassified or unclassified, as described in the grey box. Each bar represents the proportion of simulations per condition (n = 400) classified in each category. Bars that are filled with a colour represent correctly classified simulations, bars that are hashed represent misclassified simulations and bars that are filled with white represent simulations that were unable to be classified.   AMPLseq also identified COI = 2 for C6-GUY but without consistent support (two loci presenting two alleles in replicate 1 and six loci presenting two alleles in replicate 2). Six additional Guyanese samples were assayed by AMPLseq and one was classified as COI = 2.
This sample (A5-GUY) gave a stronger minor variant signal in both AMPLseq (15 loci presenting two alleles in both replicates) and WGS data (10.9% SNP heterozygosity; Figure S8).

| Longitudinal sampling: distinguishing continuing vs. newly acquired infections
We used 4CAST to examine longitudinal samples that were likely to be diverse and polyclonal. We sequenced samples from two asymp-  Figure S10), in which we detected a series of polyclonal infections, with some strains sustained over many time points ( Figure   S10A), and a series of distinct monoclonal infections ( Figure S10B).
In all cases, the individuals were asymptomatic and did not receive antimalarial treatment between visits; however, these simple examples demonstrate the clarity that 4CAST can bring to tracking infection turnover in longitudinal studies and suggests its potential in distinguishing recrudescence vs. reinfection in therapeutic efficacy studies.

| P. falciparum and P. vivax coinfection detection
To test the ability of AMPLseq to detect P. vivax co-infections via co-amplification of PvDHFR, two additional Guyanese blood spot samples that had been diagnosed as P. vivax-only (G4G430) and P.
vivax + P. falciparum co-infection (G4G180) via microscopy were included in the sample set. These samples did not undergo sWGA.
PvDHFR was detected at high depth in both samples (1068-1822 read-pairs for G4G430 and 234-560 read-pairs for G4G180) ( Figure 9). Only G4G180 also showed read-pair support at P. falciparum loci (>10 read-pairs at 100-115 loci). PvDHFR was not detected in any native or pre-amplified 3D7 or mixed-strain (3D7 + Dd2) templates. This demonstrates high specificity of both PvDHFR and P. falciparum AMPLseq primers to their intended target species without any apparent amplification inhibition by the presence of congeneric

DNA.
PvDHFR was also detected at low levels (

| DISCUSS ION
The utility of AmpSeq for molecular surveillance of infectious diseases is evidenced by the growing number of protocols recently published or under development for Plasmodium and other pathogen taxa (Aydemir et al., 2018;Fola et al., 2020;Jacob et al., 2021;Mitchell et al., 2021;Moser et al., 2021;Ruybal-Pesántez et al., 2021;Schwabl et al., 2020;Tessema et al., 2020). Considering the first of these criteria, intended use case, our investigations above suggest a straightforward mapping of panels by size and feature to use case. The small 4CAST panel is well suited to COI estimation (Figure 6), and profiles four highly diverse antigens for the same effort and cost traditionally used to profile a single locus. Because of the very high diversity of the loci in the 4CAST panel in most parasite populations, this panel is also well suited to any application requiring genetic delineation of distinct parasite lineages ( Figure 7). In therapeutic efficacy studies, for example, it is essential to determine whether subjects who become parasitaemic following drug treatment are exhibiting a recrudescence of an incompletely cleared strain from the initial infection (which could indicate treatment failure), or if they have become reinfected with a distinct parasite strain subsequent to treatment. We suggest that the 4CAST panel would be significantly more informative than traditional genotyping approaches used in therapeutic efficacy studies, such as profiling length polymorphisms or allele-specific amplification in the msp1/msp2/glurp genes (Reeder & Marshall, 1994;Snounou, 2002), especially if coupled with an inferential approach that accounts for some chance sharing of alleles dependent on their population frequencies. 4CAST is also more cost-effective than independent monoplex amplification and Illumina sequencing of individual loci (Early et al., 2019;Gruenberg et al., 2019;Lerch et al., 2017).
Our work demonstrates that the AMPLseq panel performs comparably to two existing multiplexed amplicon sequencing panels of similar size (Jacob et al., 2021;Tessema et al., 2020) for any use case reliant on estimation of parasite relatedness (Figure 4), despite different design criteria and data sets that informed the panels.
Potential public health use cases that employ relatedness information include measuring the connectivity of parasites between locations to define units of control, and monitoring changes in the level of transmission Daniels et al., 2015;Knudson et al., 2020). The AMPLseq panel and its peers are well suited to detecting imported vs. local infections given their capacity to distinguish parasites from distinct countries, as long as population genetic differentiation is sufficiently high ( Figure 5). Finally, the larger panels offer the capacity to monitor genetic markers associated with drug resistance (Figure 8) or, in some panels, detect co-infection with other Plasmodium species (Figure 9).
The second panel selection criterion, protocol complexity and compatibility with available instruments, should be prefaced with a reminder that all of these protocols employ nested PCRs as the fundamental mechanism to produce sequencing libraries targeting small genomic regions of interest. Equipped with a few key instruments, most laboratories with access to pre-and post-PCR hood space are probably capable of the nested PCR library construction approach.
Key instruments needed are a centrifuge, thermocycler, vortexer, magnetic rack, fragment sizer (e.g., Bioanalyzer or TapeStation) and DNA quantitation device (e.g., Qubit fluorometer and/or degraded sample collections unless lower batch sizes ( Figure   S9C) or larger (more expensive) sequencing instruments are utilized.
The impact of DNA integrity on panel performance should be further assessed in future work. Additional performance assessment using parasitaemias formulated prior to DNA extraction (e.g., diluting ring-stage parasites) would also enhance future sensitivity tests. GTseek, we successfully added 4CAST, drug resistance and PvDHFR loci to the GTseek target set without any primer modifications. The AMPLseq panel is thus probably receptive to further augmentation.
As the AMPLseq and 4CAST protocols utilize unmodified, commercially available oligos as primers, further customization should be feasible in any setting. However, note that not all targets are amenable to incorporation into the multiplex, as we failed despite multiple attempts to include amplicons targeting the crt gene associated with chloroquine resistance (Fidock et al., 2000), or the hrp2/3 genes, which can contain deletions that lead to false-negative diagnosis via rapid diagnostic test (Gamboa et al., 2010). studies and surveillance efforts led by different groups. This latter factor, which we term portability of analyses, has the potential to provide regional or global insight through syntheses across studies.
However, the portability of certain analyses is hampered by ascertainment bias, an inherent limitation of any targeted sequencing approach for analyses based on the genotypic state of select loci in different countries. That is, a panel designed based on observations of genetic diversity through WGS in countries A and B may not provide a fair means of comparing diversity in countries C vs. D, if diversity there is distributed differently in the genome than in countries A and B. WGS is the ultimate tool for avoiding this bias. However, the problems of comparing populations profiled with different panels may be mitigated by comparing inferred relatedness levels within populations rather than actual genotypic diversity measures . Overlap of loci among panels would further facilitate direct estimation of relatedness between samples included in different studies, as in Carrasquilla et al. (2022), where the importance of confidence intervals around relatedness estimates is highlighted.
The AMPLseq panel we describe here contains a significant number (n = 47) of targets from the SpotMalaria panel, and we expect that future P. falciparum panel designs will also tend to exhibit some degree of overlap with other panels, both by deliberate design and through blind convergence based on key genomic features, such as high diversity and sequence amenability to PCR primer design.
As molecular surveillance efforts for malaria parasites and other organisms are more widely adopted and become increasingly diverse, it will be essential for the community to develop standardized approaches for the design, validation, interpretation and sharing of targeted amplicon sequencing data. The paneljudge R package described here provides an excellent means to comparatively evaluate existing and hypothetical panel performance via data collected from previous population genomic surveys, and the bioinformatic analysis pipelines we have developed are suitable for interpreting Illumina data from diverse targets and panels in different organisms. We anticipate the growth of this field and the development of new analytical tools to extract even more epidemiological and ecological insight from increasingly large AmpSeq data sets for diverse taxa.

CO N FLI C T O F I NTE R E S T S
The authors declare no conflict of interest.

ACK N OWLED G M ENTS
This project was funded in whole or in part with federal funds from

O PEN R E S E A RCH BA D G E S
This article has earned an Open Data Badge for making publicly available the digitally-shareable data necessary to reproduce the reported results. The data is available at http://www.ncbi.nlm.nih.
gov/sra Benefits Generated: A research collaboration was developed with scientists from the countries providing genetic samples, and all principal collaborators are included as co-authors. The results of this research have been shared with the provider communities and the broader scientific community (see above), and the research addresses a priority concern, in this case the public health concern of malaria. More broadly, our group is committed to international scientific partnerships, as well as institutional capacity building.