Using deep learning to identify recent positive selection in malaria parasite sequence data

Malaria, caused by Plasmodium parasites, is a major global health burden, with an estimated 229 million cases and 409,000 deaths in 2019 alone [1]. Plasmodium falciparum causes almost half of all malaria cases, and the majority of deaths are children in sub-Saharan Africa; Plasmodium vivax accounts for 65% of malaria cases in Asia and South America [1]. Malaria control involves a combination of case management using diagnosis and treatment, and prevention using insecticide-treated nets, indoor residual spraying, and intermittent preventive treatment.

Resistance to anti-malarial medicines is a threat to the global efforts to control and eliminate malaria. Resistance originates from Plasmodium genetic mutations that increase in frequency over time and “sweep” through populations. During the past fifty years, several first-line treatments for P. falciparum malaria, including chloroquine and sulfadoxine-pyrimethamine (SP), have been rolled-out and then subsequently replaced due to the emergence of resistance. Recently, resistance to artemisinin has been reported in the form of delayed parasite clearance in Southeast Asia, posing a threat to the current first-line artemisinin-based combination therapy [2, 3]. For P. vivax, the spread of resistance to chloroquine, primaquine, mefloquine, and SP has been reported in various regions of the world [4, 5]. The underlying mutations causing resistance for P. vivax are less well defined than for P. falciparum [4‐6].

Protecting and monitoring the efficacy of antimalarial treatments is a top priority for malaria endemic countries. There is a need to not only continuously monitor for drug resistance, which includes clinical reporting, but also to screen the parasite genome for known resistance mutations (e.g. in P. falciparum: pfcrt (PF3D7_0709000), pfdhfr (PF3D7_0417200), pfdhps (PF3D7_0810800), pfmdr1 (PF3D7_0523000), and pfkelch13 (PF3D7_1343700) [3]) and to identify potentially novel loci under putative positive selection. These insights are being facilitated by the characterization of genomic variation using whole-genome sequencing (WGS) across many Plasmodium isolates, and the subsequent application of statistical and population genomics methods to detect sweeps. In particular, sweeps can be identified through statistical approaches considering population differentiation, site-frequency spectra, or linkage disequilibrium and extended haplotype homozygosity (e.g. the within population integrated haplotype score (iHS), and the between population ratio (Rsb)) [7]. Whilst these methods have been developed for the human genome [8], they have been applied to Plasmodium and identified known genetic mutations contributing to drug resistance [9, 10]. Recently tools have been developed for the efficient computation of these statistics from WGS libraries, such as REHH, SweeD and OmegaPlus [11‐13], but they require parameter optimization and their results are sensitive to the SNPs included, population definition, and to the statistical significance thresholds used to make inferences.

In recent years, researchers have explored the possibility of augmenting traditional approaches to the detection of selective sweeps with machine learning methods [14]. To date, sweep detection algorithms have been applied to pre-calculated population genetic statistics (e.g. Tajima’s D, Fay and Wu’s H) [7]. Gradient boosted decision trees and random tree classifiers have been trained on simulated data and applied to human 1000 Genomes Project data [15]. However, these methods do not solve the challenge of defining and calculating the population genetic statistics used as predictors of selection, a task which can be complex and time-consuming, especially when there are multiple sub-populations for cross-comparison. Deep (machine) learning methods may provide a viable alternative, and allow algorithms to learn through a hierarchy of features, where their definition and relationships can be inferred by the algorithm rather than externally defined [16]. The application of neural networks and deep learning has been explored within population genetics [17‐19]. More generally, these methods are gaining traction in healthcare and biomedical settings, where enormous amounts of data are being generated, which contain extremely valuable signals and information, at a pace far surpassing what “traditional” methods of analysis can process [19].

The detection of recent positive selection seems amenable to deep learning approaches, where learning to recognize features in raw SNP data, such as the length and shape of shared haplotypes in genes with known sweeps within and between populations, may help to identify sweeps across the genome. The work presented applies a deep learning image-classification approach, which does not require prior extraction or selection of population-genetic statistics, to classify selective sweeps from “haplotypic” images. Using large P. falciparum (n = 1125) and P. vivax (n = 368) WGS datasets, partitioned into training and validation sets, the analysis shows that a deep learning approach (called “DeepSweep”) calibrates well with other haplotype-based methods and other studies, and has the potential to detect novel signatures of positive selection.

The application of whole genome sequencing (WGS) is gaining traction across malaria endemic countries. With the resulting development of Plasmodium parasite genomic databases (“big data”), there is an opportunity for the implementation of machine learning methods to inform disease control. The detection of genomic signatures of selective sweeps resulting from the spread of mutations associated with anti-malarial drug resistance is one application of WGS data. This work presents a supervised (deep) learning approach (DeepSweep), which after being trained on haplotypic “images” of established drug resistance genes in P. falciparum and P. vivax parasites, resulted in the identification of loci known to be under recent positive selection. Whilst the strength of sweep signals per locus found by DeepSweep correlated with established EHH methods (e.g. between population Rsb), the machine learning approach has the advantage of not requiring a rigid definition and calculation of population-genetic statistics, incorporating information within and across populations, and relatively lower requirements for the pre-processing of raw SNP data. Like other machine learning approaches, it has the potential to scale up to large numbers of samples, and is parallelizable across genomic regions, thereby making it a potentially useful “big data” tool. In the absence of sufficient computational power, it is possible to develop sampling strategies that can select the subset of the data and samples that contain the highest density of information relevant to DeepSweep. Different model structures were assessed, but performance could be improved by further fine tuning of model hyperparameters (e.g. the number and size of the convolutional filters).

DeepSweep detected a set of loci not detected by the EHH methods, potentially because a deep learning approach can holistically incorporate information from the raw SNP data, which could be fragmented across separate populations and genomic windows, for the calculation of population-genetic statistics. Indeed, the simulation study demonstrated the potential of including haplo-images with not only single, but multiple populations, to allow the algorithm to take advantage of features that are common across regions and be robust to different stages of the sweeps. However, DeepSweep does require “representative” positive training examples, and in the context applied, assumes that the training drug resistance related loci have undergone or are undergoing selective sweeps in some of the populations. This assumption is not unrealistic given that some antimalarial drugs have been rolled out in different populations at different times resulting in differential stages of selective sweeps [40]. The DeepSweep and EHH approaches, as well as alternative methods (e.g. HaploPS [45]), can be considered complementary and could be run in parallel. However, as these approaches will increasingly use WGS, there are general challenges that affect variant-calling and ascertainment (e.g. extreme genome GC content), which can impact on the density and accuracy of genomic variant inputs, as well as the final population genomic analysis. Typically, WGS analysis leads to a dense set of well supported variants in robust genomic regions, with the application of calling algorithms incorporating information on known high quality polymorphisms [6]. Further, highly variable or problematic regions, such as var genes in P. falciparum, are typically removed from analysis [46]. In general, DeepSweep appeared to perform well across different GC content settings (P. falciparum 19%, P. vivax 58%), as well as in a simulated data setting which did not impose any constraint on GC content. However, in general, it is important to evaluate the quality of genomic variants used in an analysis. A further consideration is that most approaches use haplotype data, which in the human context require phasing from genotypes. Whilst the Plasmodium life cycle involves haploid asexual stages, complex clinical infections can complicate and confound population genetic analyses, and therefore analysis was restricted to infections with a dominant clone. However, it may be possible to extend DeepSweep to process individual parasite sequences for samples with multiplicity of infection. Irrespective, any novel loci identified should be confirmed through functional work [47]. Further, complementary methods that look at isolate relatedness, as determined by identity by descent (e.g. IsoRelate [48]), could also be implemented. New loci detected by DeepSweep that were not identified by other methods (e.g. on chromosomes 6, 8 and 14 for P. falciparum and on chromosomes 6, 7 and 14 for P. vivax) provide interesting candidates for confirmation studies.

A potential future opportunity is to apply models across species, for example, to detect P. falciparum loci after being trained on P. vivax signatures, and vice-versa. Such an application could assist to detect regions where drug resistance loci are unknown or less established, such as P. vivax. However, the impacts of differences in sample size and degree of polymorphism between species need to be considered. Relatedly, “real data” was used for training, but an alternative may be to use coalescent or forward-in-time simulation to create positive and negative labelled exemplars. However, there is a risk that images might not be representative of actual selective sweeps in nature. The deep learning algorithm has applications beyond positive selection, including for other evolutionary signatures (e.g. balancing selection) or application to other organisms (e.g. mosquitoes and humans).