Towards an evolutionary baseline model of Plasmodium falciparum for population-genomic inference

Abstract

Malaria has caused over 15.7 million deaths in the 21^st century and was responsible for ~600 thousand deaths globally in 2023 alone. Although many effective antimalarial drugs have been developed and widely adopted to reduce the occurrence and severity of the disease, recurrent resistance to the frontline treatment has been of major concern. Multiple drug resistance alleles at intermediate and high allele frequency have been identified in specific Asian and African populations of P. falciparum, the deadliest malaria parasite. With the improvement in throughput of sequencing technologies and global efforts such as the MalariaGEN project to build genomic surveillance, we now have access to tens of thousands of genomes of P. falciparum from across the world. With this data, it is becoming increasingly possible to employ powerful population genetics approaches to understand the selective pressures and demographic history of the parasite. While several empirically motivated outlier-based approaches have been employed to identify targets of drug resistance, there is a lack of a framework that jointly accounts for the multiple concurrent processes occurring in natural populations of P. falciparum. We argue that a baseline evolutionary model that accounts for simultaneously acting evolutionary processes is needed to understand patterns of genomic variation in P. falciparum populations. Here, we identify key components essential for building such a baseline model for the malaria-causing pathogen. Such effort will be important to develop an appropriate null model to test evolutionary hypotheses using genomic datasets, will provide a path forward to improve the accuracy of inference of evolutionary parameters, and help identify new gene candidates involved in drug resistance.

INTRODUCTION

Plasmodium falciparum, a unicellular protozoan parasite that is responsible for malaria, accounts for over 90% of malaria mortality worldwide (Snow 2015; World Health Organization 2024). It causes the most severe and fatal outcomes among all malaria-causing species, with approximately half of the world’s population at risk of contracting the disease (Centers for Disease Control and Prevention 2024). In the 20th century alone, P. falciparum claimed an estimated 150–300 million lives and continues to kill approximately half a million people per year (Arrow et al. 2004; World Health Organization 2024). Despite continuous efforts to develop effective antimalarial drugs, P. falciparum has repeatedly developed wide-spread resistance to frontline treatments, including chloroquine in the 1960s and sulphadoxine–pyrimethamine in the 1980s (Payne 1987; Basu et al. 2025). Most recently, acquired partial resistance to current frontline artemisinin-based combination therapy (ACT) has emerged, jeopardizing recent advancements in malaria control (Phyo et al. 2016; Rosenthal et al. 2024). Given the severity and widespread impact of P. falciparum transmission, understanding the evolutionary dynamics shaping its populations has become increasingly critical. With the advent of a large amount of publicly available whole-genome population-genetic data from more than 30,000 genomes of P. falciparum (Ahouidi et al. 2021; Abdel Hamid et al. 2023; Abdel Hamid et al. 2025), ~2000 genomes of P. vivax individuals (Adam et al. 2022), and initial effort for P. malariae (Popkin-Hall et al. 2024) and P. ovale (Carey-Ewend et al. 2024), population-genomic inference is likely to become a widespread approach to understand the evolution of drug resistance, adaptation to the host, and key demographic events in Plasmodium (Volkman et al. 2007; Chang et al. 2012; Parobek et al. 2016; Guo et al. 2024).

When a drug-resistance or any other beneficial allele rapidly increases in frequency and/or reaches fixation in the population, it results in molecular signatures at sites linked to the beneficial fixation. Specifically, there is a reduction in genetic diversity (Maynard Smith and Haigh 1974), a skew in the site frequency spectrum (SFS), and a unique spatial pattern of linkage disequilibrium (LD) generated at linked sites, also known as selective sweeps (Stephan 2019). Thus, statistical summaries of DNA sequence variation are commonly used to detect candidates of recent fixation of beneficial mutations from population-genetic data (Nielsen 2005). For instance, outlier-based approaches use loci present in the tails of the distribution of summary statistics, like measures of genetic diversity, population differentiation, skew in the SFS relative to expectations under neutrality (Tajima 1989) and haplotype diversity (e.g., Garud et al. 2015), to identify possible gene candidates that have recently experienced sweeps. However, variability in such statistics across the genome is generated not only by a rapid increase in frequency of beneficial mutations or those under balancing selection, but also due to other non-adaptive processes like variation in mutation or recombination rates and simply due to strong bottlenecks, i.e., the sharp reduction in the size of a population (Teshima et al. 2006).

There have been two opposing schools of thought when accounting for this issue. One school of thought recommends identifying putative candidates of positive selection by simply selecting all loci present in the tails (1% or 5%) of the distributions of summary statistics (e.g., Carlson et al. 2005^; Kelley et al. 2006). This approach has many caveats (Thornton and Jensen 2007), with the two most important ones being that it always yields potential gene candidates even under the absence of any recent selective sweeps in the population (because a fixed fraction of all genes, such as 1% or 5% will always be identified as candidates). Secondly, this approach assumes that loci that have undergone sweeps will be enriched in the tails of the distributions, which is not necessarily the case (as shown by Teshima et al. 2006), and thus the false positive rate is unknown. For instance, when using this approach, severe population bottlenecks can drastically increase the false positive rate when detecting selective sweeps (Crisci et al. 2013), and the effects of purifying selection and recombination rate variation can result in falsely identifying outlier loci even without any beneficial fixations (Johri, Stephan, et al. 2022).

To address these caveats, the second school of thought suggests a model-based approach, wherein putative sweep candidates are identified by first generating an expectation of the distribution of the statistic of interest without any positive selection by explicitly modeling the history of the population of interest (Teshima et al. 2006; Johri, Aquadro, et al. 2022). The caveat of the model-based approach is that model violations or poor model fit may lead to incorrect inferences. We therefore stress the importance of using model-based approaches and of working towards building accurate and realistic models so that we can detect loci under positive or balancing selection with more confidence and with fewer false negatives so that a fuller picture of the genes responsible for adaptation can be obtained (Johri, Aquadro, et al. 2022; Soni et al. 2023; Soni and Jensen 2024). In fact, while a number of outlier-based approaches have been employed in Plasmodium species to identify targets of drug resistance (e.g., Volkman et al. 2007^; Parobek et al. 2016), there is a lack of an evolutionary model that jointly accounts for multiple constantly operating evolutionary processes in their natural populations, shaping molecular variation across the genome.

Accounting for an evolutionary baseline model in unicellular pathogenic species can be uniquely challenging for multiple reasons. (a) Unlike human genomes, which are sparsely populated with functionally important genomic elements, genomes of unicellular species are often characterized by highly streamlined genomes (Lynch and Conery 2003), that are gene-rich, making the effects of selection pervasive across the genome. (b) Unicellular species often exhibit some form of clonal reproduction or self-fertilization (Weedall and Hall 2015), which reduces the effectiveness of recombination in their population, in turn increasing the effects of selection across the genome. (c) Finally, pathogenic organisms experience a highly complex demographic history due to continual bottlenecks during host invasions and strong bouts of selection due to exposure to antibiotics/drugs (Renzette et al. 2013). Thus, in order to model molecular variation in natural populations of pathogenic species, it becomes especially important to jointly account for the effects of multiple evolutionary factors, like their mating system, skewed distribution of offspring, the effects of selection on putatively neutral sites, and complex demography.

We argue that evolutionary baseline models of Plasmodium falciparum must account for such factors, in particular, recurrent bottlenecks, realistic genome architecture (density of coding and intergenic regions), selection against deleterious mutations, along with heterogeneity in mutation and recombination rates across the genome (Johri, Aquadro, et al. 2022). Such baseline models can then be employed to generate realistic expectations and thus to better identify putative sweep candidates (detailed in Box 4 of (Johri, Aquadro, et al. 2022). We here elucidate how such an evolutionary baseline model in P. falciparum may be constructed and which population-genetic considerations might play an important role in evolutionary inference. With the advent of a large amount of publicly available whole-genome population-genetic data of P. falciparum (Ahouidi et al. 2021; Abdel Hamid et al. 2023; Abdel Hamid et al. 2025), there will be many opportunities for performing population-genomic inference. It is therefore imperative that we construct baseline models, which would aid in correctly interpreting evolutionary forces shaping genomic variation in human pathogens.

INFECTION CYCLE

The P. falciparum life cycle consists of two main phases: the human phase (intermediate) and the mosquito phase (definitive). In the salivary gland of the female mosquito, ~10³-10⁴ haploid sporozoites, or the spore-like, motile phase of the parasite, remain in the salivary gland until the next blood meal (reviewed in Kappe et al. 2010^; Churcher et al. 2017^; Graumans et al. 2020^; Andolina et al. 2024). The intermediate phase starts when an infectious female Anopheles mosquito bites a human, sporozoites are injected with the saliva into the dermis of the human host (reviewed in Ejigiri and Sinnis 2009). Only ~20% of sporozoites reach the salivary glands of mosquitoes, and ~100 are transmitted during the blood meal to the dermis of the human host (Rosenberg and Rungsiwongse 1991; Vanderberg and Frevert 2004; Amino et al. 2006; Jin et al. 2007; reviewed in Kappe et al. 2010). The sporozoites remain motile in the dermis and follow a random path until a portion of them encounters a blood vessel, at which point they quickly circulate into the liver. Those that do not reach the bloodstream are either destroyed in the skin or drained by the lymphatic system (Amino et al. 2006; Yamauchi et al. 2007). On average, the majority of sporozoites that reach the liver take between 1 and 3 hours to leave the skin stage (Yamauchi et al. 2007). After reaching the liver, during the hepatocytic (liver cell) stage (see Figure 1), the surviving sporozoites (~ 10 in number) invade hepatocytes and begin their first round of asexual reproduction (Mota et al. 2001; Venugopal et al. 2020). After 13–14 rounds of mitosis, the result is a multi-nucleated schizont (cell with tens of thousands of nuclei) containing thousands of haploid daughter cells called merozoites. After 6–7 days of multiplying in the liver, the schizont bursts and releases up to 90,000 (10³ - 10⁴) merozoites into the bloodstream (Vaughan et al. 2012), where each of the merozoites invades a red blood cell (Venugopal et al. 2020).

Figure 1:

The lifecycle of Plasmodium falciparum. The progression of various life stages is displayed in human (purple) and mosquito (blue) hosts. The inner-most circle shows the stage of the lifecycle and the corresponding ploidy of P. falciparum in grey circles. Note that the gametocyte blood stage and the asexual blood stage may occur concurrently within a human host as parasites can be in different stages at any given time. The yellow ring displays the approximate time spent in the corresponding stage (not drawn to scale). The outermost circle displays recurrent bottlenecks experienced by P. falciparum during the life cycle and the numbers indicate present estimates of the number of cells. The width of the black arrows are correlated to their estimated within-host population sizes.

Within the red blood cells (RBCs), the parasites undergo additional rounds of asexual reproduction over the course of 48-hour cycles (Hawking et al. 1968; Smith et al. 2020), referred to as the asexual erythrocyte (RBC) stage (Figure 1), where they multiply into a schizont containing 32 daughter merozoites after 5 rounds of mitosis (Venugopal et al. 2020). At this point, the schizont bursts and these merozoites rapidly invade new RBCs to initiate another replication cycle (Cowman and Crabb 2006). This exponential process allows the population to reach an approximate count of 10⁹-10¹³ haploid individuals in the bloodstream, increasing the total parasite load by approximately 8 to 10-fold every 48 hours (Arrow et al. 2004b; Kappe et al. 2010; Graumans et al. 2020). It is estimated that ~5 mitotic events occur per 48-hour cycle, with an average of 2–3 of these cycles occurring per human infection (McDew-White et al. 2019; Bucşan and Williamson 2020). With each replication cycle, less than 10% of the asexual parasites within a RBC terminally differentiate into sexually competent cells called gametocytes (Carter et al. 2013; Collins et al. 2018; Reuling et al. 2018; Tadesse et al. 2019). All the gametocytes that emerge from a single schizont are either male or female (Silvestrini et al. 2000; Smith et al. 2000). This process of the development and maturation of gametes (i.e., gametocytogenesis), takes an additional 9–12 days within the human host (Hawking et al. 1968; Vaughan et al. 2012). Gametocytes do not undergo further replication once differentiated, and are in fact the only mode of transmission from a human to a mosquito (Chawla et al. 2021); parasites in the asexual erythrocytic stage die after ingestion by the mosquito (Smith and Jacobs-Lorena 2010).

The definitive mosquito phase (also known as the sporogonic phase) of the life cycle, which lasts from ~10–18 days, begins upon the ingestion of infected blood by a female Anopheles, with approximately 1–10³ P. falciparum gametocytes likely to be ingested (Kappe et al. 2010; Lin et al. 2014; Sato 2021). Once the gametocytes are exposed to the mosquito midgut environment, the female gametocytes differentiate to produce one macrogamete (larger reproductive cell), and the male gametocytes become eight microgametes (smaller reproductive cells) referred to as the gamete stage (Figure 1). The microgamete fertilizes the macrogamete, producing a diploid zygote. The zygote undergoes meiosis and differentiates into an ookinete (Figure 1), which then has four haploid genomes in its nucleus (i.e., has a ploidy of 4N) for ~17–23 hours (Mzilahowa et al. 2007). In other words, the zygotic stage only lasts for a single cell division before becoming a tetraploid ookinete (Mzilahowa et al. 2007). The motile ookinete journeys towards the midgut lining, where ~80% of them are destroyed by the mosquito immune system (Shiao et al. 2006). The motile ookinete forms an oocyst (Figure 1) that is embedded on the outer midgut lining of the mosquito and undergoes several rounds of mitosis (Graumans et al. 2020). Once fully matured (taking 11–12 days post-blood meal), the oocyst ruptures and releases 10³-10⁴ haploid sporozoites (immature sporous form) that invade the salivary gland of the mosquito (Rosenberg and Rungsiwongse 1991; Wang et al. 2018; Musiime et al. 2019). A severe bottleneck occurs at transmission, where only approximately 20% of sporozoites reach the salivary glands where they are stored until less than 1% are transmitted during the infectious bite (Figure 1) when the mosquito takes its next bloodmeal (Rosenberg and Rungsiwongse 1991; Graumans et al. 2020).

Regarding the duration of the parasite life cycle inside the mosquito relative to mosquito feeding behavior and lifespan, it is important to note that once a female mosquito has mated, she immediately seeks a blood meal and rests for approximately three days while the eggs develop (Chavasse 2002). When the eggs have matured, she lays them and seeks her next blood meal. Adult female Anopheles are generally short-lived, and in optimal conditions, can live up to 1–3 weeks on average in this stage (Arrow et al. 2004b; Lambert et al. 2022), and usually tend to take one blood meal per gonadotrophic cycle (i.e. pregnancy), thus limiting the time spent in the mosquito host to approximately 21 days. However, because mosquito species that are significant malaria vectors, such as Anopheles gambiae and Anopheles funestus, frequently take two blood meals per gonadotrophic cycle (Scott and Takken 2012), and the frequency of feeding episodes ranges from 2–4 days, there’s likely some variance in the number of days (~11–15 days) that a parasite spends in the mosquito host (Guelbéogo et al. 2018; Brackney et al. 2021).

A detailed literature survey of previous estimates of the approximate within-host population sizes, the duration of the different stages of their life cycles, and the corresponding ploidy (Table S1, Figure 1), points us to a few key observations. Firstly, P. falciparum seems to live largely in a haploid state (~97% of its life cycle), while it is diploid and tetraploid for only a small fraction (~3%) of the time (Table S2). Thus, heterozygous effects of selected mutations are important for a relatively limited amount of time during their life cycle, and therefore their populations can likely be modeled to be haploid. While population-genetic dynamics are very similar in a haploid and diploid population under neutrality, and/or when selected mutations are semi-dominant (i.e., the two alleles contribute equally to fitness), this is not necessarily the case with recessive/dominant selected mutations. In particular, when present in diploids or polyploids, the effects of partially or fully recessive mutations can be masked. Thus, recessive deleterious mutations can be more effectively purged in haploid populations, reducing the expected genetic load (i.e., reduced mean fitness of the population compared to that of a mutant-free population) that haploid populations carry (reviewed in Otto and Gerstein 2008). In addition, assuming the same number of individuals in a haploid and diploid population, if the rate of new beneficial mutations is low (i.e., Nu_b<<1), the rate of fixation of beneficial mutations in haploid populations can be much larger (or smaller) than the diploid population if the beneficial mutation is recessive (or dominant; Orr and Otto 1994). On the other hand, if beneficial mutations are common (i.e., Nu_b ≥ 1), haploid populations will always fix beneficial mutations at higher rates. Therefore, haploidy and dominance will both be important to consider when accurately modeling the dynamics of selected mutations in malaria populations.

Secondly, there are at least 3 bottleneck events during one lifecycle, with the most severe being ~10⁶-fold at the point of human to mosquito transmission. Moreover, this event follows an extremely rapid exponential growth in the human blood. Recurrent bottlenecks can lead to multiple merger events (Birkner et al. 2009; Tellier and Lemaire 2014), which refers to the coalescence of more than two lineages in the same generation. Moreover, recent experimental evidence has indicated that some strains are more likely to proliferate in the mosquito than others (Li et al. 2019), possibly leading to a wider distribution of offspring numbers than assumed in a Wright-Fisher population (referred to as progeny skew). This suggests that multiple merger events are highly likely during within-host evolution of P. falciparum. In other words, when reconstructing the genealogy backwards in time, multiple lineages are likely to coalesce into one at the same node, suggesting that multiple merger coalescents (Pitman 1999; Möhle and Sagitov 2001; Möhle and Sagitov 2001) might best model their coalescent genealogy as opposed to the Kingman coalescent (Kingman 1982), used most commonly. In general, multiple merger events result in a U-shaped site frequency spectrum (Eldon and Wakeley 2006; Blath et al. 2016), lead to an increase in linkage disequilibrium between alleles (Eldon and Wakeley 2008), and an increase in differentiation between populations (measured by F_ST; Eldon and Wakeley 2009), thus affecting expectations of various population genetics summary statistics. While it is clear that multiple merger events will be common within the mosquito and human hosts, it is unclear whether such events during recurring bottlenecks will affect the genealogy at longer time scales (Tellier and Lemaire 2014; Irwin et al. 2016). If the number of hosts is large and the sampling is geographically scattered and species-wide, the long-term genealogy may converge to the Kingman coalescent (Wakeley and Aliacar 2001; Städler et al. 2009; Heuer and Sturm 2013). However, this is an important unexplored question for the future, which will have consequences for modeling P. falciparum populations accurately, and the answer is likely to depend on the transmission dynamics and the sampling scheme.

MUTATION RATE

As de novo mutations are the primary source of genetic variation, quantifying the rate at which new mutations arise is an essential component in understanding population genetics in Plasmodium populations. While the rate at which new mutations enter P. falciparum populations varies across life cycles and time periods (Table 1), the 48-hour asexual erythrocytic cycles are the most important for replication-based mutations. During the asexual erythrocytic cycles, the haploid parasite rapidly replicates in human red blood cells, reaching extremely high within-host population sizes (Figure 2), which vastly increases the possible points of origin for new mutations to arise. Mutation accumulation (MA) experiments (Kibota and Lynch 1996; reviewed in Halligan and Keightley 2009), where a single individual is propagated in the laboratory at ideally a population size of 1, thus maximizing the effects of genetic drift, are used to estimate the rate of de novo mutations. Such studies in P. falciparum (where lab cultures had small population sizes) have indicated that the rate at which new single base substitutions are observed over the 48-hour mitotically dividing period, the asexual erythrocytic cycle, is between 2.10 × 10⁻¹⁰ and 5.23 × 10⁻¹⁰ per site/ 48 hours (Bopp et al. 2013; Claessens et al. 2014; Hamilton et al. 2016; McDew-White et al. 2019; Table 1). However, strongly deleterious mutations that arise may not be easily observed by MA as they are purged from the population prior to sampling (Eyre-Walker and Keightley 2007). The observed mutation rate of single-base substitutions in these studies was adjusted to account for unobserved mutations of strong deleterious effect based on the expected and observed ratios of nonsynonymous and synonymous base substitutions compared to the reference genome (3D7) to approximate the true de novo mutation rate This mean adjusted base substitution rate accounting for strongly deleterious mutations ranges from 0.854 × 10⁻⁹ (McDew-White et al. 2019) to 1.70 × 10⁻⁹ (Bopp et al. 2013) per site per asexual erythrocytic cycle. Assuming 5 mitotic divisions per 48-hour cycle (McDew-White et al. 2019), we may therefore expect the adjusted per-mitotic division mutation rate to range from 1.71 × 10⁻¹⁰ (McDew-White et al. 2019) to 3.40 × 10⁻¹⁰ (Bopp et al. 2013).

Table 1.

Estimated P. falciparum de novo mutation rates for single bases (per site/generation) and microsatellite repeats (per locus/generation) over specified periods. Here, “Asexual cycle” refers to a 48-hour mitotically dividing period corresponding to the asexual blood stage in Figure 1. All methods were performed on experimental populations except for longitudinal sampling by Chenet et al. (2015), which involved repeated sampling of a wild Colombian population. Confidence intervals (95%) are provided where available.

Annotation	Rate	Period	Method	Reference
Single base substitutions	3.18 × 10−10 (2.44 × 10−10 – 3.92 × 10−10)	Asexual cycle	Mutation accumulation (31 clones)	McDew-White et al. 2019
	5.23 × 10−10 (3.69 × 10−10 – 6.77 × 10−10)	Asexual cycle	Mutation accumulation (3 clones)	Bopp et al. 2013
	2.10 × 10−10 (1.49 × 10−10 – 2.72 × 10−10)	Asexual cycle	Mutation accumulation (37 clones)	(Hamilton et al. 2017)
	4.07 × 10−10 (CI not available)	Asexual cycle	Mutation accumulation (19 clones)	Claessens et al. 2014
Microsatellites	4.45 × 10−7 (3.48 × 10−7 – 5.42 × 10−7)	Asexual cycle	Mutation accumulation (37 clones)	Hamilton et al. 2016
	4.43 × 10−7 (4.06 × 10−7 – 4.81 × 10−7)	Asexual cycle	Mutation accumulation (31 clones)	McDew-White et al. 2019
	1.59 × 10−4 (6.98 × 10−5 – 3.70 × 10−4)	Meiosis	Genetic cross (35 progeny)	Su et al. 1999; Anderson et al. 2000
	0.54 – 3.77 × 10−2 (8.26 × 10−4 – 7.37 × 10−2)	Per month	Longitudinal sampling (79 haplotypes)	Chenet et al. 2015

Figure 2.

Mutation rate heterogeneity across the P. falciparum genome. The histogram displays per-gene synonymous divergence rates ($K_s$) between P. falciparum and P. reichenowi directly estimated for 4931 genome-wide genes (available at Supplementary Data 4 in Otto et al. 2014). The dark blue line represents the beta distribution fitted to the histogram data, linearly rescaled around the mean per site per 48-hour erythrocytic asexual cycle single base substitution rate estimated by McDew-White et al. (2019). The orange line shows the commonly used parameters fitted to a normal distribution for the per site per 48-hour erythrocytic asexual cycle mutation rate estimated by Bopp et al. (2013). Both estimates are adjusted for strongly deleterious and lethal mutations and assume 33 mitotic divisions per generation. Mean values for each distribution are indicated with dashed lines.

Although all three studies that measured mutation rates in in-vitro experimental populations of P. falciparum agree on the final estimated mutation rates, Hamilton et al. (2017) observed fewer indels that are not multiples of three (i.e., indels that result in a frameshift) in coding regions than expected, suggesting that the effect of purifying selection was present in these experimental populations. It is thus highly likely that the estimated rates of de novo mutations are presently underestimated, especially because it is not straightforward to account for lethal mutations in MA experiments. We can also estimate how many mitotic asexual divisions are likely in a single generation (i.e., between their sexual cycles). There are ~3–7 divisions of the zygote in the mosquito gut, ~13–14 divisions in the human liver, and ~14 or more divisions in the human blood, totaling ~30–35 divisions during the life cycle (see Figure 1). Assuming the rate of mutation per cell division estimated by McDew-White et al., and that the parasite undergoes on average 33 cell divisions in a generation, the rate of de novo mutations of single base substitutions would be 5.6 × 10⁻⁹ per site/ generation, somewhat higher than previously believed. Mutation rates in unicellular eukaryotes (~10⁻⁹-10⁻¹¹ per site/generation) are typically similar to those of eubacteria and much lower than in multicellular eukaryotes (~10⁻⁸ – 10⁻⁹ per site/generation; Lynch et al. 2016). While the P. falciparum mutation rate seems to be on the higher end of the range observed in unicellular eukaryotes, with most free-living species like Chlamydomonas reinhardtii (3.8×10⁻¹⁰ per site/gen), Paramecium tetraurelia (1.9×10⁻¹¹ per site/gen) and S. cerevisiae (2.6×10⁻¹⁰ per site/gen) having much lower rates, the vector of African trypanosomiasis, Trypanosoma brucei (1.4×10⁻⁹ per site/gen) and the fungus Neurospora crassa (4.1×10⁻⁹ per site/gen) exhibit higher mutation rates, very similar in magnitude to P. falciparum (Lynch et al. 2016).

Mutation rate exhibits heterogeneity across genomic landscapes, influenced by factors such as nucleotide composition, local sequence context, and methylation status (Nachman and Crowell 2000; Ness et al. 2015; Lynch et al. 2016; Smith et al. 2018). While region-specific rates may not have been directly estimated for P. falciparum, the distribution of mutation rates across the genome is expected to follow rates of fixed differences at putatively neutral synonymous sites (K_s) between closely related species. Thus, by fitting a beta distribution (α = 3.55, β = 56.56; SE = 0.07, 1.18) to the genome-wide per-gene K_s values between P. falciparum and P. reichenowi (closely-related species causing malaria in chimpanzees and gorillas) reported by Otto et al. (2014), we obtain an approximate distribution of the mutation rate across the genome which can be scaled around a mean estimate of choice, such as the adjusted rate of 5.6 × 10⁻⁹ per site/generation estimated above (Figure 2), allowing us to model mutation rate heterogeneity across the Plasmodium genome.

P. falciparum has an extremely high AT content (80.6% of sites genome-wide) compared to other eukaryotes and Plasmodium species (Hamilton et al. 2017) across a highly repetitive genome (Zilversmit et al. 2010). This leads to high genome-wide rates of de novo small tandem repeat indels, i.e., microsatellites, due to replication slippage (Lovett 2004), prevalent in both coding and noncoding regions (Hamilton et al. 2017). Microsatellite indel rates observed per 48-hour erythrocytic asexual cycle from Hamilton et al. (2017) and McDew-White et al. (2019) were highly consistent (Table 1). In addition, studies have estimated microsatellite indel rates observed between meiotic generations (Su et al. 1999; Anderson et al. 2000), as well as per month across a wild Columbian population using a Bayesian phylogenetic approach (Chenet et al. 2015), which could be useful if de novo mutations were likely to accrue as a function of time rather than the number of cell divisions (Gao et al. 2016). However, microsatellite indels that are not multiples of three are disproportionately uncommon in coding regions compared to intron/intergenic regions due to strong purifying selection against frameshift de novo microsatellite mutations. As a result, the values reported in Table 1 represent the lower bounds of the de novo microsatellite mutation rates, while the true rates may be up to 48% higher when accounting for unobserved lethal or strongly deleterious frameshift mutations (Hamilton et al. 2017).

GENOME ARCHITECTURE

The AT-rich 23 Mb P. falciparum genome is dense with coding regions, with exonic sites covering approximately 52% of the genome across over 5268 genes; the mean length of exons, introns and intergenic distance is 949 bp, 179 bp, and 1,694 bp respectively, with an average (and median) of 2.39 (and 1) exons per gene with 47% of genes identified as having no introns (details available in Gardner et al. 2002). In eukaryotes with similar genome-wide coding sequence density, such as Drosophila melanogaster (dos Santos et al. 2015), this would be expected to result in pervasive linked effects of selection (Begun and Aquadro 1992; Charlesworth et al. 1993; Comeron et al. 2012). It is also notable that these estimates exclude any extrachromosomal DNA and the highly variable subtelomeric regions that contain major gene families responsible for immune evasion and virulence, such as Var genes (Otto, Böhme, et al. 2018). Because these regions are difficult to align and find orthologous regions in (Otto, Böhme, et al. 2018), they are often excluded from population genetics studies.

Comparative analysis has uncovered that approximately 60% of P. falciparum genes genome-wide were conserved among all five Plasmodium species studied (Cai et al. 2012); this set of genes represent the core genome, which conserve essential functions across the entire genus. Alternative studies found that 57% of genes on Chromosome 2 had detected homologs in other species (Gardner et al. 1998), and 27% (1407 of 5268) of all genes remain unannotated (Behrens and Spielmann 2024). However, note that due to ascertainment bias, predicted gene sequences from many protists, especially those that have diverged substantially from Opisthokonts (e.g., those belonging to Alveolates, Discobids), lack known orthologs and assigned functions (Burger et al. 2016; Johri et al. 2019; Prieto-Baños et al. 2025). Thus, as expected, when sequence-independent computational algorithms based on geometry were applied to the unannotated set of genes in P. falciparum, similarity to previously known protein domains was identified in 25% of these genes (Behrens and Spielmann 2024), suggesting that these represent putatively functional coding regions.

Due to the high AT-content of the P. falciparum genome, fewer than a third of coding sites are synonymous; we estimate the proportion to be 17.5%. Thus, assuming that mutations at all nonsynonymous sites have some effect on the fitness of the individual, the proportion of coding sites experiencing direct selection would be 82.5%. The proportion of noncoding regions under purifying selection is unclear. However, previous studies (Siepel et al. 2005) that have used phylogenetic methods to estimate this proportion across model organisms found that in the human, C. elegans, D. melanogaster, and S. cerevisiae genomes where ~3%, 30%, 18%, and 70% are annotated as coding, ~ 2%, 11%, 20%, and 6% of their genomes were predicted to be conserved noncoding respectively. As the P. falciparum genome resembles the S. cerevisiae genome in terms of the coding density (~52.6%), total number of genes (~5k vs ~6k in yeast) and the genome size (23 Mb vs 12 Mb in yeast), we assume that ~5–10% of their noncoding regions are conserved. As such, assuming that 82.5% of coding regions in Plasmodium genomes are conserved (43% of genome-wide sites) and ~10% of genomic sites represent conserved noncoding regions, approximately 53% of genome-wide sites likely experience direct selection. Further analysis is needed to more accurately identify and quantify phylogenetically conserved sites with approaches such as phastCons (Siepel et al. 2005) or GERP (Davydov et al. 2010), which in turn would enable analyses to effectively quantify the direct and linked effects of selection genome-wide. There is a striking conservation of the number of chromosomes, number of genes, and synteny across the seven species belonging to the Laverania genus, which includes P. praefalciparum and P. reichnowi, and reasonable genomic synteny extends to P. ovale spp. (Rutledge et al. 2017), as well as to P. vivax, P. knowlesi, and P.y. yoelli that share ~77% of orthologs with P. falciparum (Carlton et al. 2008). With whole genome sequences of 14–15 closely related taxa, such an endeavor seems plausible.

RECOMBINATION AND SELFING RATE

Recombination refers to the shuffling of genetic variation that occurs during meiosis, where genetic material from two homologous chromosomes is exchanged. Recombination results in novel combinations of genetic variants, ultimately bolstering the effectiveness of natural selection. Due to its influential role in shaping patterns of genomic variation, understanding the heterogeneity in recombination rate across the genome and between populations is important to understand the patterns of genetic diversity in a species (Cutter and Payseur 2013; Peñalba and Wolf 2020). P. falciparum undergoes sexual recombination once per life cycle during the gamete stage (Figure 1), between the chromosomes from a haploid female macrogamete and a haploid male microgamete (Mzilahowa et al. 2007). Numerous studies have characterized the patterns of recombination in P. falciparum using microsatellite markers or tiling arrays (Walliker et al. 1987; Walker-Jonah et al. 1992; Kerr et al. 1994; Su et al. 1999; Hayton et al. 2008; Jiang et al. 2011). Recent work by Miles et al. (2016) used short-read whole genome sequencing to achieve a comprehensive understanding of recombination in P. falciparum crosses by sequencing the parents and offspring from three pairs of crosses between diverse lab strains. This study achieved a resolution of ~300 bp for each recombination event using SNPs and indels in the strains, which was an order of magnitude greater than previous work using an SNP array (Jiang et al. 2011). The average recombination rate for all three crosses was 7.4 × 10⁻⁷ recombinations/bp per meiosis, confirming the previous observation of a recombination rate that is an order of magnitude higher than mammals such as rats, mice, and humans (Jensen-Seaman et al. 2004) but smaller than an estimate of 3.5 × 10⁻⁶ in yeast (Ruderfer et al. 2006). Because unicellular eukaryotes generally have small chromosomes and at least one crossover is required per meiosis, this results in a relatively high rate of crossover on average (Lynch 2006). The probability of gene conversion per base pair (per haploid genome) was estimated to be 3.6 × 10⁻⁷ per meiosis, with average tract lengths of 1.4 kb, after adjusting for small gene conversion events that were likely missed. These gene conversion tract lengths are also similar to those of yeast (Mancera et al. 2008), but longer than those of humans (Jeffreys and May 2004). Genome-wide patterns of recombination showed a significantly lower recombination rate near centromeres and subtelomeres, and with higher recombination rates associated with repetitive DNA. Both recombination and gene conversion showed elevated rates in exons, unlike humans, where recombination primarily occurs in hotspots near but not within the coding regions of genes (Myers et al. 2005). The higher GC content of coding regions in P. falciparum may partially explain this preference, but it may also be explained by a methodological bias against finding recombination events in AT-rich noncoding regions. Additionally, a 12-bp motif was found to be significantly enriched for recombination in multiple studies (Jiang et al. 2011; Miles et al. 2016); however, limited data currently prevents the confident classification of any recombination hotspots (Miles et al. 2016). The prdm9 gene, a well-known mediator of recombination hotspots in many species (Baudat et al. 2010) has an ortholog in the P. falciparum genome (Jiang et al. 2011), so future studies using a higher number of crosses or a population genetic method to estimate the recombination rate (reviewed in Peñalba and Wolf 2020) may be able to more confidently identify recombination hotspots in the Plasmodium genome.

While the actual rate of recombination is quite high, the effective rate of recombination is substantially lower in P. falciparum populations. This is mostly because of two reasons: (1) as the female and male gametocytes mature from the same pool of parasite, selfing can occur between genetically identical clones of P. falciparum, making recombination entirely ineffective. (2) Due to the bottlenecks during ingestion by mosquitoes in the P. falciparum lifecycle, only a small number of strains (usually between 1–20; Nkhoma et al. 2020) are available to recombine. Moreover, the differentiation and pairing of gametes occurs within hours of ingestion by a mosquito (Baton and Ranford-Cartwright 2005), so recombination events usually occur between parasites ingested from a single human host rather than strains from multiple blood meals (Camponovo et al. 2023). This is likely to increase the probability of selfing and mating between genetically highly related chromosomes, reducing the “effective” rate of recombination in the population. Note that while the rate of selfing is likely to always be non-zero, it will depend on the number of distinct strains in the mosquito. For instance, assuming random mating, the rate of selfing with 10 distinct haplotypes will be 0.1, while it will be 0.5 with 2 distinct haplotypes. Thus, the effective rate of recombination will mostly depend on the overall level of inbreeding occurring in the population, i.e., $R_{\mathit{eff}}=R\left(I-F_{it}\right)$. Here $R$ is the population-scaled rate of recombination, and $F_{it}$ represents the overall level of inbreeding, which will be influenced by both selfing events and mating between closely related individuals, and is calculated by comparing the heterozygosity of parasite populations within the guts of individual mosquitoes $\left(H_W\right)$ to that of the entire population of parasites sampled from mosquitoes $\left(H_S\right)$ with $F_{it}=I-\frac{H_W}{H_S}$ (Mzilahowa et al. 2007). As expected, the number of genetically distinct malaria strains present within a human host, called the multiplicity of infection (MOI) or the complexity of infection (COI; Viriyakosol et al. 1995^; Snounou and Beck 1998^; Paschalidis et al. 2023), was found to positively correlate with the amount of transmission and therefore to rates of outcrossing (Camponovo et al. 2023) leading to higher nucleotide diversity and less linkage disequilibrium in such populations (Anderson et al. 2000).

Several attempts have been made to estimate the inbreeding rate in P. falciparum populations using either the parasite population in the mosquito or the human host. Obtaining estimates of inbreeding from infected mosquitoes is likely the most direct way. Analysis from the mosquito stage is also most relevant to its effect on recombination, particularly because haploid oocysts attached to the mosquito midgut contain thousands of cells descended from one zygote (a single meiosis), and thus can be used to infer the genotypes of zygotes and infer $F_{it}$ (Ranford-Cartwright et al. 1991). Genotyping of microsatellite loci in oocytes from mosquito midguts in high-transmission areas has yielded a few estimates of inbreeding rate: 0.56–0.6 in Malawi (Mzilahowa et al. 2007), 0.33 in Tanzania (Hill et al. 1995), and 0.4 in Western Kenya (Razakandrainibe et al. 2005). Due to the recent large-scale malaria sequencing efforts from at least 20,000 patient blood samples (Abdel Hamid et al. 2023), inbreeding estimates using Plasmodium samples from infected human patients (commonly referred to as $F_{WS}$) have become prevalent and are largely correlated with estimates using COI, though they use a relatively indirect approach (Auburn et al. 2012; Manske et al. 2012).

Inbreeding rates obtained using parasite populations in the human host yield somewhat higher estimates from ~0.9 in African populations to >0.95 in South America, Southeast Asia, and Oceania (Ahouidi et al. 2021; Abdel Hamid et al. 2023; Abdel Hamid et al. 2025) as summarized in Table S3. However, a more fine-grained analysis of African P. falciparum populations found large amounts of variation in mean $F_{WS}$, with countries with high rates of transmission like Kenya and Malawi (F_WS~0.7) having lower rates of inbreeding than countries with low rates of transmission like Senegal and Ethiopia (F_WS > 0.95; Amambua-Ngwa et al. 2019). Thus, it is possible that differences between the mosquito-based and human-based studies simply reflect the differences in the specific populations used - i.e., rates of inbreeding could vary locally vs globally. However, other methodological differences might also contribute to this discrepancy. For instance, estimates using parasites sampled from humans do not account for the drastic bottleneck experienced by the parasites immediately before recombination, and while mosquito-based studies used a few microsatellite loci, the human-based studies have included data from the whole-genome. In either case, inbreeding rates seem to vary between 0.4–0.95, governed by rates of transmission in the specific local population.

Because the effective rate of recombination in P. falciparum populations depends on the rate of inbreeding, which in turn depends on the rate of transmission, it is likely to also vary across populations. The bite frequency and infection rate vary by geographic region. Across multiple studies, there is a consistent pattern of intense hotspots in rural and peri-urban regions, and lower intensity within city centers (Duchemin et al. 2003; Doumbe-Belisse et al. 2018). This is because the number of infected bites per human per year (EIR) is much lower in indoor (31.14) than outdoor (66.65) settings (Doumbe-Belisse et al. 2018). Similarly, across multiple regions in sub-Saharan Africa, the EIR was found to be 7.1 in the city centers, 45.8 in peri-urban areas, and 167.7 in rural areas (Duchemin et al. 2003). Thus, the likelihood of transmission can vary substantially depending on spatial factors and population density.

High-transmission regions in Africa and South Asia have higher nucleotide diversity, and much faster decay of linkage disequilibrium (LD) over distance relative to low-transmission regions in South America, Southeast Asia, and Oceania (Achidi et al. 2008; Ahouidi et al. 2021; Abdel Hamid et al. 2023; Abdel Hamid et al. 2025). However, the magnitude of the difference in LD decay is moderate across populations, with nucleotide diversity in coding regions varying from a median of 0.00019 in Eastern Southeast Asia to 0.00028 in West Africa, and r² in coding regions decaying to baseline levels within ~2,000–10,000 bp for Asian populations and ~500–1000 bp for African populations (Baudat et al. 2010; Ahouidi et al. 2021). Thus, LD decays in P. falciparum much faster than in human populations (Auton et al. 2015) but is similar to Drosophila populations, where LD decays in ~500–10000 bases . While these estimates are derived from samples pooled from multiple populations collected at different time points, they provide time and space-averaged estimates that may not necessarily correspond to present estimates in local populations. For instance, in regions of varying transmission intensity in Colombia, where clonal genotypes have been observed to persist for up to 8 years (median=537.5 days), LD decay was found to vary drastically between subpopulations, with LD in subpopulations with low transmission decaying to background levels in ~500 kb and relatively higher transmission areas within ~250 kb (Echeverry et al. 2013). Similarly, certain malaria outbreaks in low transmission regions like eastern Panama (Obaldia et al. 2015) and northwest Ecuador (Sáenz et al. 2015) have been composed of only one to three strains, and high levels of monoclonal infections have been identified in regions with low transmission like Cambodia (Parobek et al. 2016) or the highlands of Ethiopia (Holzschuh et al. 2024). The LD in Cambodian populations of Plasmodium has been shown to decay to baseline levels hundreds of base pairs slower than populations in sub-Saharan Africa, supporting reduced recombination in low-transmission regions (Miotto et al. 2013; Samad et al. 2015). Thus, the levels of inbreeding and the effectiveness of recombination varies at the local scale, with regions with high transmission and endemicity generally having more effective recombination than regions with lower or declining malaria transmission (Anderson et al. 2000).

Assuming a Wright-Fisher panmictic population, a previous study estimated the haploid effective population size of P. falciparum in Senegal to be 50,000 (Chang et al. 2012). Assuming the relevant value of the per-site rate of recombination and inbreeding estimates of 0.6 and 0.9 for high vs low transmission populations, we find the theoretical expectation of the LD-decay (Figure 3) to match empirical estimates quite well (see Figure 2c of Ahouidi et al. 2021). We theoretically predict LD to reach baseline levels in ~250 bp in high transmission vs ~1000 bases in low transmission populations. An important implication of vastly different recombination rates across populations is that the expected signatures of recent selective sweeps would likely be highly population-specific, deeming it necessary to generate population-specific expectations when performing selection scans using the P. falciparum genome (see the section Drug resistance and selective sweeps for a more detailed discussion).

Figure 3:

The expected decay of linkage disequilibrium (between hosts) with distance between alleles in a high vs low transmission population of P. falciparum. Here, N=80,000, $r=7.4\times {10}^{-7}$ per site/generation, and $N_{\mathrm{e}}=N/(1+F)$ where $F=0.6$ and 0.9 in high and low transmission populations respectively. Here, linkage disequilibrium is measured using ${\sigma }_d^2$ such that $E\left[{\sigma }_d^2\right]=\frac{10+\rho }{22+13\rho +{\rho }^2}$, where $\rho =2N_{e}rx(1-F)$, and $x$ is the distance between alleles (Ohta and Kimura 1971).

DISTRIBUTION OF FITNESS EFFECTS

New de novo mutations occur every generation (see the section Mutation Rate) and can either be harmful or beneficial for the individual, have the same fitness effect as the wildtype, or may have no effect on fitness at all. The latter category of mutations is called “neutral” mutations, examples of which are many synonymous mutations and those that occur in intergenic regions that serve no function. Most new mutations in functionally important regions tend to be deleterious, while a small minority (less than ~1%) may be beneficial (Eyre-Walker and Keightley 2007; Bank et al. 2014). Although we understand this broad-level classification of mutations, the magnitude of the effect of each mutation on fitness follows a continuous distribution between −1 (signifying lethal) and +1 (signifying the fittest individual). This distribution of fitness effects (DFE) of new mutations is important to understand the expected amount of genetic variation in a population, the amount of genetic load, the rate of Muller’s ratchet, and the effects of selection on linked sites.

Most genes in P. falciparum exhibit low values of divergence at nonsynonymous vs synonymous sites, measured by d_N/d_S, with a median of 0.17 (Jeffares et al. 2007) and a mean d_S value of 0.057 when using P. reichnowii (Prugnolle et al. 2008), consistent with widespread purifying selection. Fast-evolving genes tend to be lowly expressed, suggesting that these are likely under reduced selective constraints, consistent with universal trends. However, some of the P. falciparum coding regions have been found to comprise lowly conserved domains that exhibit higher levels of nonsynonymous substitutions (d_N/d_S ~0.35) and have been predicted to represent low complexity hydrophilic non-globular domains (Wootton 1994; Gardner et al. 1998; Gardner et al. 2011). Gardner et al. (2011) found that for a subset of 40 genes that have conserved orthologs and synteny in the P. falciparum genome, these protein domains of low conservation occupy ~50% of the total coding sequence, suggesting that a substantial number of new mutations in coding regions may be mildly deleterious and/or effectively neutral. Relatedly, the observed amount of polymorphism at nonsynonymous (${\pi }_N$) vs synonymous sites (${\pi }_S$), measured by ${\pi }_N/{\pi }_S$ in P. falciparum populations, was found to be unexpectedly high, with a mean value of ~0.52 and with many genes exhibiting values above 1 (Chang et al. 2012). Similarly, the site frequency spectra of both nonsynonymous and synonymous sites were found to be very similar in P. falciparum populations (Chang et al. 2012), suggesting the presence of many mildly deleterious mutations. However, as the mutation spectrum of de novo mutations is AT-biased (Hamilton et al. 2017), it is unclear what effect that might have on the expected statistics. Alternatively, it is possible that the unique life cycle of the parasite with multiple bottlenecks results in summary statistics at selected and neutral sites being similar (Chang et al. 2013). Thirdly, the global population of P. falciparum is likely far from equilibrium due to the drastic bottleneck during the speciation event that occurred only 10,000 years (or 111 generations) ago (Loy et al. 2017). Because variation at effectively neutral sites takes longer to reach equilibrium levels than at selected sites (Brandvain and Wright 2016), the speciation bottleneck could explain higher diversity at nonsynonymous relative to synonymous sites. Finally, selection on synonymous sites might also yield such observations (Musto et al. 1999; Peixoto et al. 2004; Chan et al. 2017; Sinha and Woodrow 2018). These hypotheses need to be tested methodically to better understand patterns of variation observed in P. falciparum populations.

While there are no experimental estimates of fitness effects of new mutations in P. falciparum yet, it seems like a feasible goal, as both gene editing and in-vitro competition assays have been successfully performed to study drug resistance (e.g. Walliker et al. 2005^; Stokes et al. 2021^; Hagenah et al. 2024). Although such studies need to be performed for more genes, the few mutations evaluated in genes involved in drug resistance show that most mutations are either neutral or deleterious (Stokes et al. 2021), in line with genome-wide expectations. A gene knock-out study (Zhang et al. 2018) conducted in-vitro suggested that approximately 50% of all genes are essential during the asexual erythrocytic stage, suggesting that a substantial fraction of large deletions and frameshift mutations would likely be lethal or strongly deleterious in P. falciparum populations, though this number is probably much larger because a number of non-essential genes are expressed during the sexual phase of the parasite (not captured by their study). Note that due to the high AT content in the P. falciparum genome, the rate of de novo insertions/deletions is relatively high (discussed in section Mutation Rate). Such insertions/deletions in protein-coding regions are likely to lead to nonfunctional, misfolded, or truncated protein products due to frameshifts, especially when their length is not a multiple of 3 and thus might have strongly deleterious effects. This is consistent with an overrepresentation of fixed indels in coding regions whose lengths are multiples of 3 (Jeffares et al. 2007). Insertions/deletions in noncoding regulatory regions, however, might have only mild selective effects. Thus, the DFE that combines the effects of base substitutions and insertions/deletions in functionally important parts of the genome will likely be bimodal, with a peak at strongly deleterious and another at mildly deleterious mutations. Note that while indels are not usually studied often due to computational challenges, they may offer variation that is important for immune evasion in P. falciparum and thus their contribution to evolutionary dynamics might be significant.

As the probability of fixation of very mildly deleterious mutations is appreciable, such mutations can contribute substantially to differences between species and subpopulations. They contribute to outliers when identifying targets of local adaptation via population differentiation tests (Johri, Charlesworth, et al. 2021). Therefore, a null model that incorporates the effect of mildly deleterious mutations will be important for correctly interpreting outliers of selection scans across the P. falciparum genome.

EFFECTS OF SELECTION AT LINKED SITES

Background selection

Deleterious mutations are purged from populations via purifying selection, and this leads to a decrease in diversity at functionally important sites in the genome. However, alleles linked to deleterious alleles can also be removed by purifying selection acting on selected sites. This process is called background selection (Charlesworth et al. 1993). While strongly deleterious mutations get purged rapidly from populations, mildly and moderately deleterious mutations can segregate in populations for a longer period of time before being purged, resulting in a distortion of the gene genealogy itself (Charlesworth 2013). Thus, background selection not only reduces neutral diversity, but can also skew the expected site frequency spectrum at neutral sites linked to selected sites (Charlesworth 2013; Nicolaisen and Desai 2013). This distortion of expected summary statistics can lead to the misinference of demographic history when using methods that assume strict neutrality (Ewing and Jensen 2016; Johri, Riall, et al. 2021). In particular, background selection results in a skew of the site frequency spectrum towards low-frequency alleles, leading to a false inference of recent population growth. Despite the pervasive effects of background selection demonstrated across multiple species (Cutter and Payseur 2013), currently only a minority of methods account for the effects of selection on linked sites while estimating parameters of demography (Sheehan and Song 2016; Johri et al. 2020; Johri et al. 2023). Moreover, a recent simulation study suggests that selection in P. falciparum might bias inferences of demography and population structure (Guo et al. 2024). It is therefore important to understand the extent of background selection that might be expected in P. falciparum populations.

The P. falciparum genome is ~23 Mb, with 14 chromosomes, comprising 52.6% coding, 5.7% intronic, and 53.2% intergenic regions. We estimate that 53% of the P. falciparum genome is likely to be under direct selection (see the section on Genomic Architecture above). As most new mutations at selected sites are deleterious (Eyre-Walker and Keightley 2007), the number of new deleterious single-base pair mutations in a haploid individual per genome per generation (U_base) can be estimated to be 0.53 × μ × 23Mb, where μ is 5.6 × 10⁻⁹ per site/generation, which is 0.068. Because there are a significant number of indels in P. falciparum that are likely to be harmful, we include them in estimating the genome-wide deleterious mutation rate. McDew-White et al. (2019) estimate an average rate of ~4.4 × 10⁻⁷ indels per locus/generation (= 8.8×10⁻⁸ per locus per cell division). Assuming that there are 123,722 loci in the core genome (as found in McDew-White et al. 2019), we estimate that each individual is likely to have ~0.05 new insertions/deletions every generation. Note that some indels will have no effects on the fitness of the individual, especially if they occur in non-coding regions. Assuming that 53% of them are selected against, we get U_indel~0.03. This gives us a total U ~ 0.096, i.e., there are likely 0.10 deleterious mutations per individual/ generation.

Using estimates of cross-over rates for chromosomes (see Figure 3 from Miles et al. 2016), we can estimate the expected nucleotide diversity at neutral sites in the presence of background selection (BGS) relative to strict neutrality, denoted by the symbol B, (Nordborg et al. 1996). Note that a value of B close to one indicates minimal effects of BGS while a value closer to zero suggests a drastic reduction in diversity due to BGS. We find that B for an average-sized chromosome in P. falciparum is only ~0.99 and 0.96 in high (assuming F=0.6) and low transmission (assuming F=0.9) populations, respectively (discussed in the section Recombination). Similar values (B~0.96–0.99) can be estimated for chromosomes of different lengths: chromosome 1, 10, and 14 corresponding to 640 kb (shortest), 1.69 Mb (average), and 3.3 Mb (longest), with the chromosome-wide map length in Morgans (R) being ~0.5, 1, and 2.25, respectively (Table S4) and accounting for selfing. In summary, the reduction in diversity due to BGS alone in P. falciparum genomes is unlikely to be much, far smaller than that expected in D. melanogaster populations (where B~0.5).

Because P. falciparum has many chromosomes (14), strongly deleterious mutations on other chromosomes can also lower diversity at a focal chromosome (Santiago and Caballero 1998; Charlesworth 2012), which is referred to as unlinked BGS. We estimate that these unlinked effects of BGS acting on a chromosome of size 1.6 Mb would result in ~0.96× the nucleotide diversity relative to under neutrality. Assuming multiplicativity of fitness effects, the combined effect of both linked (caused by deleterious mutations on the same chromosome) and unlinked BGS (caused by deleterious mutations on other chromosomes) will result in B ~ 0.93, which we estimate to be similar across all chromosomes (Table S4). Thus, the effects of background selection in lowering nucleotide diversity are expected to be minimal in P. falciparum, which is consistent with very short chromosomes and thus high rates of recombination in this unicellular organism. However, background selection can also skew the site frequency spectrum of neutral alleles (Charlesworth 2013; Nicolaisen and Desai 2013), and the extent of this skew in P. falciparum remains to be tested, as it will be important to determine biases of inference of population history. Note that our calculations assume panmixia and equilibrium, both of which are unlikely to realistically represent malaria populations. Thus, further investigation accounting for the parasite life cycle and host-pathogen dynamics is needed to fully understand the effects of background selection on genomic variation in P. falciparum.

Drug resistance and selective sweeps

Since malaria eradication became a global priority in the 20th century, recurrent emergence of antimalarial-resistant strains against every first-line treatment developed has offset progress in transmission control and mortality rates. Parasites resistant to chloroquine, the first antimalarial widely implemented in the 1950s, have emerged at least four times due to mutations in the P. falciparum gene encoding the chloroquine resistance transporter (pfcrt; Payne 1987^; Wellems and Plowe 2001). Resistant strains first appeared in Southeast Asia, which spread to Africa, and have also been observed in Papua New Guinea and South America (Payne 1987; Wootton et al. 2002). Despite chloroquine being removed as a first-line drug due to its failure, the presence of chloroquine-resistant mutations remains almost entirely fixed in low-transmission populations like Southeast Asia and South America (Bacon et al. 2009; Griffing et al. 2010; Muhamad et al. 2011; Phompradit et al. 2014). On the other hand, the chloroquine-resistant pfcrt genotype became almost undetectable in the high transmission Malawi population within twelve years after its removal as a first-line drug (Kublin et al. 2003; Frosch et al. 2014; Takala-Harrison and Laufer 2015). Resistance to the subsequent first-line antimalarial drug, sulfadoxine-pyrimethamine arises from sequential mutations accumulated in dihydrofolate reductase (pfdhfr) and dihydropteroate synthase (pfdhps; Cowman et al. 1988^; Triglia et al. 1997) genes. These highly resistant haplotypes followed a similar pattern to the widespread accumulation of chloroquine resistance, where the lineage emerged in Southeast Asia, spread to Africa, and emerged independently in South America and Oceania (Mita et al. 2007; Mita et al. 2011). Highly resistant sulfadoxine-pyrimethamine haplotypes remain highly prevalent in Southeast Asia and South America despite their removal as a drug (Mita et al. 2011). Genomic surveillance along the Kenyan–Ugandan border, including new whole-genome sequences from Bungoma, Western Kenya, suggests that East African P. falciparum comprises several subpopulations with distinct structure and diverse ancestral origins, some derived from independent lineages rather than from West or Central Africa (Osborne et al. 2024).

Currently, there is a major concern regarding resistant lineages of the first-line treatment of the 21st century: artemisinin combination therapy (ACT). ACT combines artemisinin and its derivatives, which are short-acting, with a longer-acting drug to ensure full removal of resistant parasites in an infected individual. The first evidence of artemisinin resistance (mediated by SNPs in the pfkelch13 gene) emerged in western Cambodia in 2008 as treatment failed to completely clear parasite infections (Noedl et al. 2008). By 2013, resistance was observed in multiple other Southeast Asian countries, such as Vietnam, Thailand, and Laos, associated with resistance to the long-acting drug piperaquine, with substantial increases in the frequency of resistance alleles from 4% to 63% observed in Cambodia from 2007 to 2013, respectively (Saunders et al. 2014; Amaratunga et al. 2016; Imwong et al. 2017; Thanh et al. 2017; Amato et al. 2018). Despite ACT being an effective treatment, there are numerous signs that resistance is rapidly emerging, facilitated through clonal spread of pfkelch13 mutations (Parobek et al. 2017) that may be maintained by low transmission. Multiple independent pfkelch13 mutations have emerged in African populations as well (e.g., Rosenthal et al. 2024^; Brhane et al. 2025^; Niaré et al. 2025). Pfkelch13 mutations appear to have initially expanded in Uganda at a similar speed to emergence in southeast Asia (Rosenthal et al. 2024). Unlike the spread of chloroquine and sulfadoxine-pyramethimine, differences in the flanking haplotypes suggest that resistance these mutations appear to originate in Africa, instead of having spread from Southeast Asia (Conrad et al. 2023; Niaré et al. 2025). Although there are multiple factors that may contribute to shaping the effects of drug pressure in a geographic region, similarities between the less genetically diverse and highly structured populations in may facilitate the rapid spread and fixation of resistant alleles (Ariey et al. 2003; reviewed in Amato et al. 2018).

With multiple drug resistance alleles segregating in the P. falciparum populations, there is a need to understand the geographic location and time of origin of these mutations, as well as the identification of the targets of adaptation in the genome. As new beneficial mutations increase in frequency and reach fixation, linked alleles hitchhike with the beneficial mutation to result in specific genomic signatures, referred to as a selective sweep. There is a decrease in nucleotide diversity near the beneficial fixation (Maynard Smith and Haigh 1974), a skew in the site frequency spectrum (Braverman et al. 1995), and a spatial pattern in linkage disequilibrium (Kim and Nielsen 2004), all of which are frequently used to detect loci where sweeps may have occurred (Nielsen 2005). While these methods have a reasonably high accuracy when selection is extremely strong and very recent, it becomes exceedingly difficult to detect sweeps that were only moderately strong or if the beneficial fixation occurred ~0.5N generations ago (Przeworski 2002; Soni et al. 2023). It is especially difficult to distinguish between false and true positives if the population has experienced bottlenecks (Teshima et al. 2006; Crisci et al. 2013). Because within-host pathogen populations undergo strong recurrent bottlenecks every generation, it becomes important to account for nonequilibrium demography when attempting to identify loci involved in adaptation in natural populations. We consider some parameters that may specifically affect the detection of selective sweeps in P. falciparum populations below.

Because malaria populations drastically vary in their transmission dynamics across the globe, there is a large difference in the amount of effective recombination in different populations. For instance, African populations on average have much larger effective recombination rates than the South American populations, though transmission is highly heterogeneous at smaller geographic scales (see the section Recombination). This can result in very different lengths of genomic signatures due to sweeps (Figure 4). For instance, we estimate that in high-transmission populations, signatures of sweeps immediately post-fixation can extend to ~6 kb when positive selection is weak (s=0.01) and up to 45 kb when selection is extremely strong (s=0.1). In contrast, in low-transmission populations, signatures of sweeps are expected to extend to much longer regions: ~25 kb and 180 kb when selection is weak vs strong, respectively. Thus, while it might be easier to detect valleys of diversity in low-transmission populations, which would span larger regions of the genome (up to 1/4th of the short chromosomes), high-transmission populations would allow for more precise identification of the target of selection due to smaller signatures around relevant loci, if selection was recent. These signatures of reduced diversity are significantly reduced if the sweep occurred N_e generations ago, which would be ~9000–10,000 years ago, assuming that P. falciparum undergoes 8–9 generations (referring to all divisions in a transmission cycle) per year (reviewed in Ejigiri and Sinnis 2009^; Smith et al. 2014^; Vaughan and Kappe 2017^; Graumans et al. 2020). Note that this is a lot more recent than the speciation event where P. falciparum jumped to the human host and started diverging from P. prefalciparum around 40,000–60,000 years ago (Otto, Gilabert, et al. 2018). Thus, while gene candidates involved in recently evolved drug resistance can be detected successfully, those involved in adaptation to the human host will be difficult to detect.

Figure 4:

Expected signatures of selective sweeps in populations of high transmission (left) vs low transmission (right). Recovery of neutral nucleotide diversity ($\pi$) is shown as a function of the distance from the beneficial fixation. Here, we assume that the haploid Wright-Fisher population size ($N$) without accounting for inbreeding is 80,000 (estimated from (Chang et al. 2012; Otto et al. 2018), Wright’s inbreeding coefficient ($F$) was assumed to be 0.6 in the high-transmission and 0.9 in the low-transmission population. We assume that $N_{\mathrm{e}}=N/(1+F)$, $r_{\mathit{eff}}=r(1-F)$, and that the beneficial mutation is semidominant. The recovery of diversity post-fixation is calculated using equation 13 from Kim and Stephan (2000). Immediately post-fixation, neutral diversity recovers to 95% of its baseline value in ~6 kb ($F=0.6$) and 25 kb ($F=0.9$) when selection is weaker ($s=0.01$) and in ~46 kb ($F=0.6$), and 180 kb ($F=0.9$) when selection is extremely strong ($s=0.1$).

Consistent with variations in effective recombination rates across transmission dynamics, valleys of nucleotide diversity observed near genes responsible for drug-resistance have been reported to span much smaller regions in African populations: 10–13 kb in Ghana (Alam et al. 2011), 16–40 kb in Cameroon (McCollum et al. 2008), 50–70 kb in South Africa (Pearce et al. 2005), and larger regions in populations of moderate to low transmission: 100kb in a population near the Thailand/Myanmar border (Nair 2003), and ~600 kb in Venezuela (McCollum et al. 2007). It would be important to account for this difference in rates of recombination for accurately estimating the selection coefficient of beneficial mutations in different populations.

Another important implication of vastly different recombination rates across populations is that the efficacy of selection, as well as the magnitude of indirect effects of selection, would be different across populations. Because low-transmission populations will have a reduced effective population size, beneficial mutations will be subjected to stronger genetic drift (i.e., there will be more mutations with 2N_es < 1; Eyre-Walker and Keightley 2007). Secondly, populations with low transmission are likely to experience stronger linked effects of selection, more specifically, increased background selection as well as increased interference between selected alleles (Hill and Robertson 1966; Charlesworth and Jensen 2021). Both of these would decrease the probability of fixation of beneficial alleles in low-transmission populations Moreover, as discussed above, the expected genomic size of signatures of recent selective sweeps would likely be highly population-specific. It might therefore be necessary to generate population-specific expectations to perform selection scans using the P. falciparum genome.

The malaria parasite has a unique life cycle (as discussed in the section Infection Cycle), such that the parasite population experiences subsequent rounds of expansions within a host, followed by bottlenecks during transmission to a new host. While the phases of population expansion result in stronger efficacy of purifying and positive selection, bottlenecks down to ~10 parasites greatly reduce the effective population size, increasing the effects of genetic drift. Chang et al. (2013) evaluated the effect of the malaria life cycle on the effects of selection and found that compared to a Wright-Fisher (WF) population under equilibrium, the probability of fixation of beneficial mutations was drastically reduced, because new beneficial mutations are likely to be lost during infection bottlenecks. Moreover, the time to fixation of beneficial mutations was found to be larger in the malaria-specific demography, as opposed to a WF population under equilibrium. The longer time to fixation will allow for more time for linked neutral mutations to escape the haplotype carrying the beneficial mutation, resulting in less reduction of diversity due to sweeps compared to standard expectations. However, note that this study was restricted to understanding the dynamics of selected mutations within hosts. Thus, it is unclear how the effects of the within-host population size changes will affect evolutionary dynamics across the entire population.

Another important factor specific to the malaria life cycle is that meiosis occurs only once per generation (i.e., the full life cycle) while the parasite replicates asexually (i.e., without recombination) during the rest of the cycle. Interestingly, a recent simulation study (Ollivier et al. 2025) found that in organisms that undergo mitotic cell divisions interspersed infrequently with meiosis, hard sweeps can decrease diversity across all chromosomes. This is because during phases of clonal expansion, all chromosomes are fully linked to each other, extending hitchhiking effects genome-wide. This effect also lowers the baseline levels of neutral diversity genome-wide, in turn reducing the valley of diversity observed post-sweep. As P. falciparum undergoes a single meiosis event every ~35 mitotic divisions, which can often be selfing, a similar effect is likely to be observed, again reducing the expected effects of sweeps, possibly reducing the power to detect them. Thirdly, another factor pertinent to malaria populations is that the beneficial effects of mutations may be specific to particular stages of the life cycle. For instance, while drug resistance mutations have benefits within the human host, they are unlikely to be beneficial within the mosquito host, and in fact might even be costly (Segovia et al. 2025). A notable example of this is atovaquone, which selects for mutations in the parasite’s cytochrome b gene that contributes to drug resistance in humans, but are later lethal in the mosquito (Goodman et al. 2016). Similarly, mutations that allow better survival in the mosquito gut might play no role in increasing fitness in the human host. A systematic study about how such rapid temporal fluctuations in fitness will affect fixation probabilities and signatures of sweeps is currently lacking.

Finally, natural populations of the parasite have experienced a complex history with multiple migration and bottleneck events. P. falciparum, which is most closely related to the gorilla vector P. praefalciparum, diverged from their most recent common ancestor 40–60 thousand years ago (Otto, Gilabert, et al. 2018; Galaway et al. 2019). From Africa, the parasite spread globally, mirroring the migration of modern humans. Population structure observed today reflects both this ancient expansion and recent events, with African, South Asian, and South American parasites forming distinct, segregated genetic clusters. For example, Indian isolates segregate from Southeast Asian in PCA space, lying between Bangladesh/Myanmar and Thailand/Cambodia samples. Neighbor-joining trees support the clear separation between African and Asian lineages, and Indian samples occupy an intermediate position (Kumar et al. 2016). Meanwhile, South American isolates form a tight cluster closely related to African parasites, which may suggest the trans-Atlantic slave trade being the primary route of introduction (Michel et al. 2024). Despite such complex demography, inference of selection coefficients in malaria populations has assumed a panmictic population at equilibrium (e.g., Wootton et al. 2002^; Nsanzabana et al. 2010), making these estimates less reliable. A careful study that incorporates the population history along with the specifics of the malaria life cycle is needed to better estimate parameters of selection in malaria populations.

CONCLUSION

An appropriate baseline model is essential to perform evolutionary inference from population genomic data and is especially challenging in pathogenic organisms due to their unique life cycles, complex demographic history with recurrent bottlenecks during transmission, strong bouts of selection against drugs, high-coding density across the genome, and skewed progeny distributions. Here, we have presented key considerations when building such a baseline model for inference in populations of the deadliest malaria parasite, P. falciparum. The complex evolutionary dynamics of P. falciparum require a baseline model that incorporates the effects of simultaneously acting evolutionary processes such as alternating cycles of asexual and sexual reproduction, selfing, transmission bottlenecks, and genome architecture (i.e., heterogeneity in coding density, recombination, and mutation rates). Importantly, an appropriate null model must adopt a framework that accounts for the within-host dynamics at shorter time scales and the global population history of multiple hosts over longer evolutionary times (Figure 5). This could be achieved by modeling a population of multiple hosts as a metapopulation composed of individual infected hosts as demes, with migration between demes representing transmission between hosts. Such an approach could allow for the flexibility to model the specifics of the life cycle, such as within-host clonal expansions and bottlenecks, as well as stochasticity during transmission between hosts (i.e., allowing for extinctions of lineages).

Figure 5:

Depiction of a baseline evolutionary model for P. falciparum incorporating within-host dynamics (depicted by the solid circles that represent the parasite), between-host transmission (modeled as migration), and the long-term demographic history of the host metapopulation (where $G_{\mathit{n}}$ represents the n^th generation).

While the development of theoretical models incorporating such factors should be the ultimate goal, a useful alternative currently is to employ realistic simulations (e.g., SLiM; Haller and Messer 2019) and accompanying simulation-based statistical approaches (e.g., Approximate Bayesian computation; Beaumont et al. 2002) to account for the unique population-genetic parameters of pathogenic species in order to identify gene candidates involved in adaptation. Although simulation-based inference can be computationally very intensive and will likely be very time-consuming for organisms with such complex infection dynamics, at the very least, simulations can be used to generate null expectations of patterns of variation and uncertainties around inferred estimates. The development of such a model will help infer accurate evolutionary history, including patterns of global migration of the parasite, to understand the evolutionary dynamics of loci involved in drug resistance, and to allow hypothesis testing, an essential component of basic science research.