Background

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 [4,5,6].

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 [11,12,13], 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 [17,18,19]. 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.

Methods

Deep learning approach

DeepSweep is a deep learning model to detect instances of positive selection. It creates and analyses standardized images of the nearby genomic region around a given SNP. In brief, for each SNP of interest, and across all isolates, DeepSweep selects neighbouring SNPs at regularly spaced intervals, and subsequently sorts the remaining genomic matrix in alignment with the longest common haplotype, grouped for each population and for the reference and alternative alleles. The intuition is that SNPs that have undergone recent selective sweeps have a different haplotype structure resulting in distinct images (Additional file 1: Figure S1).

Model structure

DeepSweep uses a convolutional neural network (CNN) architecture, implemented using the Keras library (version 2.2.4) [20] in Python. The model was based on the AlexNet Classifier architecture, widely used for image analysis [21]. Through optimization, it was aimed to fit the smallest sized model (in terms of number of trainable parameters) that showed good predictive performance with low validation loss and high validation accuracy, but also detected features of interest, avoided overfitting, and minimized computational burden. Informally, overfitting is the training of a model that is too specifically tailored to (artefacts in) the training dataset and does not generalize well to unseen data. Statistically, within the framework of the bias–variance trade-off of a model, overfitting occurs where there is excessive variance resulting from an algorithm modelling the random noise in the training data [22]. The approach optimized over various hyper-parameters, including the number of convolutional layers (ranging from 1 to 5 layers), the number of filters (ranging from 2 to 96) and convolutional field sizes (ranging from 3 × 3 to 40 × 40). Regularization techniques (e.g. dropout [22]) were applied to prevent overfitting and support transferability. The model was trained to reduce binary cross-entropy between actual labels and estimated probabilities on images of known- and non-sweeps. The model structure was validated for 500 epochs. The final model has one convolutional layer, two dense layers, four convolutional filters, and a large convolutional field (40 × 9). The haplo-imaging algorithm and the machine learning analyses (Additional file 1: Figures S1, S2) were conducted in Python (version 2.7). The core packages for the machine learning were SnpEff (for annotating effect size) [23], SnpSift (for filtering VCF files) [24], PyVCF (for adjusting and creating VCF files) [25], SciPy and matplotlib (for image manipulation) and Tensorflow (version 1.15).

Simulated data

Sequence data was generated using SFS_Code software [26], which is a forward population genetic simulator. Simulated data corresponded to four sweep types ((i) recent—strong; (ii) recent—weak; (iii) historic—strong; (iv) partial) and compared to a Wright-Fisher “neutral” setting. The parameter settings are outlined (Additional file 1: Table S1), and lead to plausible scenarios for Plasmodium parasites [10]. For each comparison, 160 simulated datasets (128 training; 32 validation) were generated, each dataset with 4 populations of 100 parasite sequences (50% sweep, 50% neutral) and a locus length of at least 1kbp, where the mutation under selection was in the centre of the region. For the combined analysis of the sweep types, 640 simulated datasets (512 training, 128 validation) were used. These data were subsequently transformed into the aforementioned “haplotype images” that serve as input to the image classifier (Additional file 1: Figure S1). These haplo-images showed qualitatively discernible differences in features, with stronger or more recent sweeps leading to more “block-like” features (Additional file 1: Figure S3). The image classifier was trained on the simulated data, and classification accuracy and reduction of binomial loss were estimated. Simulated data was also used to illustrate the impact of changes in a subset of hyperparameters and confirmed that the final model had low validation loss and high validation accuracy (Additional file 1: Table S2).

Plasmodium sequencing data

Publicly available raw Illumina WGS data for P. falciparum (n = 1125) [27] and P. vivax (n = 368) [28], representing 11 malaria endemic countries (Additional file 1: Table S3; accession numbers in Additional file 1: Tables S4, S5). All samples were assessed by estMOI software [29] as either monoclonal or polyclonal samples with only a major dominant clone, to minimize the effects on analysis of multiplicity of infection. The P. falciparum and P. vivax sequences were mapped to the Pf3D7 (23Mbp) and PvP01 (29Mbp) reference genomes, respectively, using bwa-mem software (version 0.7.12; using default parameter settings) [30]. From the resulting alignments, SNPs and insertions and deletions (indels) were called from the consensus of GATK (version 4.1.4.1) [31] and samtools (version 1.9) [32] software (using default parameter settings), as applied in previous studies [4, 10]. SNPs were retained if they had < 10% missing alleles and a minor allele count greater than 4. The resulting dataset comprised of parasite genomes of P. falciparum (1,125 isolates, 74,757 SNPs) and of P. vivax (368 isolates, 126,596 SNPs). The number of missing values was 1,179,202 (2.9%) for P. vivax and 649,337 (1.2%) for P. falciparum. Missing alleles were imputed using the isolate with the longest shared haplotype around the missing position. An overview of the analytical approach is summarized (Additional file 1: Figure S2). The SnpEff tool (https://pcingola.github.io/SnpEff/) was used to annotate SNP variants and predict their effects on genes.

For DeepSweep model training, the presumed positive examples of positive selection are regions surrounding SNPs that are linked to drug resistance with an established scientific literature. For P. falciparum, these included regions around established SNPs in pfcrt (K76T, I356T; chloroquine), pfdhfr (N51I, C59R, S108N, I164L, S306F)/pfdhps (I431V, S436A, A437G, K540E/N, A581G, S613S) (SP), pfmdr1 (N86Y; mefloquine, chloroquine), and pfkelch13 (F446I, Y493H, P574L, R539T, and C580Y; artemisinin) [27]. For P. vivax, these included regions around some known SNPs in pvdhps (A553G, G383A, S382C/A) / pvdhfr (N50I, F57I/L, S/K58R, T61M, N117T/S) (putative SP) and pvmdr1 (F1076L, Y976F, S698G, S513R; putative chloroquine) [4, 6]. This could be considered a relatively small number of training exemplars, which may lead to an increased risk that the implemented machine learning algorithm overfits due to potential artefacts in the training data. Therefore, for each Plasmodium species, “leave-one-group-out” cross-validation was implemented to understand the influence of individual training genes, where each single gene of the positive training examples was omitted in turn, with the model trained on the remaining genes [33]. The final model was fit on 80% of the data (split by SNPs), with 20% left as a hold-out set. The DeepSweep approach was compared to traditional haplotype-based statistics (iHS [34] and Rsb [35]), as calculated with the REHH package [36].

Results

Simulation study

Across the 4 different types of sweep simulations, the predictive accuracy was highest for more recent strong selection (97.1%), followed by weak selection (96.8%) and historic selection (88.2%) and partial selection (86.7%) (Table 1, Additional file 1: Figure S4). The total sensitivity across all sweeps combined was 89.1%, with a specificity of 93.8%, and an overall classification accuracy of 91.4%. The areas under the ROC curve were high for all simulations involving recent selection (> 0.95; maximum 1), consistent with the high predictive ability of DeepSweep. The simulation results showed the potential utility of the approach when combining data across populations with common sweeps at difference stages.

Table 1 Model performance based on simulated data
Full size table

Plasmodium falciparum DeepSweep analysis

The dataset comprised of 1,125 isolates and 74,757 SNPs. Most of these SNPs are in genic regions (76.5%), with 63.0% non-synonymous amino acid changes. Most SNPs have low minor allele frequencies (SNPs with MAF < 1%: 94.6%) (Additional file 1: Figure S5). The image classifier was trained on regions covering the established resistance SNPs in five genes, and found the models validated well using a leave-one-group-out approach. In particular, the overall accuracy was 83.6% (standard deviation 6.0%), where the performance was lower when pfdhfr was omitted (75.0%) and was higher when pfdhps (92.3%) was left out. One interpretation is that pfdhfr is under stronger selection than pfdhps, which would be consistent with pfdhfr N51I, C59R, S108N, I164L and S306F mutations underpinning key haplotypes underlying SP resistance [37]. The final model was fitted on 80% of the data, with 20% of the data used as a validation set, and demonstrated a strong performance both in terms of classification accuracy and reduction of binomial loss (Additional file 1: Figure S6). The trained classifier was then used to make predictions for the entire dataset of P. falciparum SNPs.

The deep learning model identified 387 SNPs in 160 genes (or ~ 2.9% of genes) as putatively under positive selection pressure in the wider dataset (Fig. 1). Further analysis focused on the subset of 11 genes that have > 6 hits (Table 2; see Additional file 1: Table S6 for the 26 genes with > 3 SNPs). Several peaks were in the vicinity of known drug-resistance genes in the training set, with nearby genes likely to be swept along (e.g. on pfdhfr on chromosome 4, pfmdr1 on chromosome 5, pfcrt on chromosome 7, pfdhps on chromosome 8 and pfkelch13 on chromosome 13). There is an additional peak on chromosome 6 that includes Pk4 (PF3D7_0628200) and the HECT domain (PF3D7_0628100). Transcription of Pk4 has been related to artemisinin-induced latency [38], and the HECT domain is thought to alter quinine and quinidine response, and likely co-selected with pfcrt [39]. There is a small peak on chromosome 10 (PF3D7_1013500) in the close vicinity of the gene encoding the autophagy-related protein 18 (PF3D7_1012900), which has been associated with artemisinin resistance. There is a peak on chromosome 12 (PF3D7_1223500) which has been putatively associated with SP resistance [40]. Smaller peaks were observed on chromosome 14 around PF3D7_1462400, which has been associated with chloroquine resistance [41].

Fig. 1
figure 1

Number of DeepSweep Hits per locus across the 14 chromosomes and the relationship to the number of Rsb (top panel; blue line) and iHS (top panel; orange line) hits. a P. falciparum. Blue line is the running average of Rsb hits (p < 0.0001) over the nearest 100 locations. The orange line is the running average of iHS hits (p < 0.0001) over the nearest 100 locations. The vertical blue lines indicate pfdhfr (Chr. 4: 749,001, pfmdr1 (Chr. 5: 960,020), pfcrt (Chr. 7: 404,770), pfdhps (Chr. 8: 549,408); pfkelch13 (Chr. 13: 1,724,817). The tick-marks on the x-axis are chromosomal mid-points. b P. vivax. Top panel: Blue line is the running average of Rsb hits (p < 0.0001) over the nearest 100 locations. The orange line is the running average of iHS hits (p < 0.0001) over the nearest 100 locations. The vertical blue lines indicate pvdhfr (Chr. 5), pvmdr1 (Chr. 10), pvdhps (Chr. 14). The tick-marks on the x-axis are chromosomal mid-points

Full size image
Table 2 Plasmodium falciparum loci identified by DeepSweep (DS; with > 6 SNPs)
Full size table

Plasmodium vivax DeepSweep analysis

The dataset comprised of 368 isolates and 126,596 SNPs. Most of these SNPs are in genic regions (77.6%), with 42.5% non-synonymous amino acid changes. Many SNPs have low minor allele frequencies (SNPs with MAF < 1%: 77.6%) (Additional file 1: Figure S5). The image classifier was trained on the sixteen SNP mutations in the three genes. Using a leave-one-group-out validation approach, the overall accuracy was 79.7% (standard deviation 17.6%), and the performance was lower when pvmdr1 was omitted (57.1%) and was higher when pvdhfr was left out (100%). This difference is consistent with pvmdr1 residues being strongly associated with chloroquine resistance [5] and, although, pvdhfr may contribute to SP drug resistance, there are very few published studies that associate genotypes of this locus with anti-folate susceptibility phenotypes [6]. As with P. falciparum, the trained model had strong performance both in terms of classification accuracy and reduction of binomial loss (Additional file 1: Figure S6). The model identified 577 hits in 237 genes (or ~ 4.3% of genes) as putatively under positive selection pressure in the wider dataset (Fig. 1). Further analysis focused on the subset of 19 genes that have > 6 hits (Table 3; see Additional file 1: Table S7 for the 35 genes with > 3 SNPs). Several loci are near the training genes (pvdhfr on chromosome 5, pvmdr1 on chromosome 10, pvdhps on chromosome 14). Further, there was a peak around the gene encoding for the multi-drug associated protein 1 (pvmrp1), which is a putative resistance candidate [4]. On chromosome 7, there was a peak around a gene coding for cysteine repeat modular protein 1, which is expressed in both vertebrate and mosquito hosts for host tissue targeting and invasion. This locus has been identified as presenting high population differentiation and under directional selective pressure in South America [4]. Finally, there was a larger region that was identified on chromosome 14, which contains pvdhps and a number of other genes that have been found in other analyses [4].

Table 3 Plasmodium vivax loci identified by DeepSweep (DS; with > 6 SNPs)
Full size table

Comparison with established positive selection approaches

An analysis using the established REHH approach was performed, which involved the calculation of the integrated haplotype score (iHS) within populations and the associated Rsb values between pairs of populations (Additional file 1: Tables S8, S9). Although the REHH and DeepSweep methods have a different ranking of the strongest hits, there was an overall positive correlation between the number of hits from Rsb and DeepSweep (Pearson correlation: P. falciparum 0.49, P. vivax 0.20; Additional file 1: Figure S7). However, DeepSweep detected several novel loci that were not identified by REHH. These included loci on chromosomes 6 (PF3D7_0611800), 8 (PF3D7_0811600) and 14 (PF3D7_1461800) for P. falciparum (Additional file 1: Table S6), and on chromosomes 6 (PIR protein), 7 (cysteine repeat modular protein) and chromosome 14 for P. vivax (Additional file 1: Table S7). PF3D7_0611800 has been linked to increased cytoadherence [42], PF3D7_0811600 has previously been linked to SP resistance [40] and the genes coding for the PIR protein and the cysteine repeat protein have been associated with immune response and host invasion [43, 44]. There were several loci that were detected by EHH methods but not by DeepSweep (Additional file 1: Tables S8, S9). Some of the top hits included genes that are linked to immune response and host invasion (e.g. PF3D7_1133400 AMA1, PF3D7_1335100 MSP7). Other hits are housekeeping genes that are less likely to be under selective pressure (e.g. PF3D7_0731800 (alpha/beta hydrolase), PF3D7_1475900 (KELT protein), PVP01_0202900 (18S) and PVP01_1003700 (PPT)).

Discussion

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).

Conclusions

The DeepSweep approach and the wider application of deep learning using genomic images constitutes a novel approach that shows promising results. It provides a robust, accessible and scalable approach for the identification of genomic regions under positive selection, and could assist with detecting established and new types of drug resistance. Thereby, providing insights into transmission dynamics and informing malaria control decision-making.