https://orcid.org/0009-0003-7434-0139
Figures
Abstract
Plasmodium falciparum erythrocyte membrane protein 1 (PfEMP1), encoded by the hypervariable var gene family, is central to malaria pathogenesis, influencing both disease severity and immune evasion. Classifying var genes into upstream groups (upsA, upsB, upsC, upsE) is important for understanding parasite biology and clinical outcomes, but remains challenging, especially with partial sequences, such as the DBLα tag or RNA-Seq assemblies. We developed upsML, a machine-learning-based classifier trained on 2,530 curated var genes, to accurately assign upstream groups based on sequence features from different partial gene regions. We compared seven methods, including support vector machines, random forests, XGBoost, and HMMER models. Several models in upsML achieve accuracies of 83% for DBLα-tag sequences and 92% for full-length PfEMP1 sequences, thereby significantly outperforming existing tools. Additionally, we developed a model to distinguish internal from subtelomeric var genes, which we applied to a global collection of P. falciparum genomes, revealing a higher frequency of internal var genes in Asia. upsML is available at https://github.com/sii-scRNA-Seq/upsML, providing a robust and efficient resource for large-scale var gene analysis. It can classify var genes from 20 genomes in under one second.
Citation: Pangilinan EA, Quenu M, Claessens A, Otto TD (2026) upsML: A high-accuracy machine learning classifier for predicting Plasmodium falciparum var gene upstream groups. PLoS One 21(4): e0344557. https://doi.org/10.1371/journal.pone.0344557
Editor: Kokouvi Kassegne, Shanghai Jiao Tong University School of Medicine, CHINA
Received: June 23, 2025; Accepted: February 23, 2026; Published: April 16, 2026
Copyright: © 2026 Pangilinan et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: Relevant upsML data for this study are publicly available from the GitHub repository (https://github.com/sii-scRNA-Seq/upsML). All other relevant genome data sources for this study are available within the paper.
Funding: TDO was supported by the ExposUM Institute of the University of Montpellier (grants ANR-21-EXES-0005 and Occitanie Region) and the Wellcome Trust: 104111/Z/14/Z & A. EAP was funded by La Caixa Foundation - Health Research Program (HR20-00635). AC was supported by Fondation pour la Recherche Médicale FRM (ANR 18‐CE15‐0009‐01) and Service Programme d’Excellence I-Site, MUSE Pre-ERC, MOTIVATOR (EQU202303016290). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
Introduction
Malaria remains a major global health burden, with Plasmodium falciparum responsible for the vast majority of the half-million annual deaths. Transmitted by female Anopheles mosquitoes, P. falciparum causes disease during its intraerythrocytic life cycle. While early ring-stage parasites are concealed within infected red blood cells (iRBCs), trophozoite-stage parasites export virulence proteins such as P. falciparum erythrocyte membrane protein 1 (PfEMP1) to the iRBC surface [1]. PfEMP1 enables iRBCs to adhere to endothelial cells via specific binding to host receptors including CD36, EPCR, or ICAM1, allowing parasites to avoid splenic clearance but contributing to microvascular obstruction and inflammation, hallmarks of malaria pathology.
Each parasite expresses only one PfEMP1 variant at a time, due to epigenetically regulated mutually exclusive transcription of var genes, the ~ 60 member gene family that encodes PfEMP1. This controlled expression allows the parasite to switch to a new var gene in subsequent cycles, facilitating immune evasion and persistence during chronic, asymptomatic infections [2].
Despite their extreme sequence diversity, var genes can be grouped into four major upstream sequence (ups) classes: A, B, C, and E, corresponding to the promoter region upstream of exon 1 [3]. These ups groups correlate with gene location and orientation: group A and E var genes are sub-telomeric and transcribed toward the telomere; group B genes are either sub-telomeric or within internal clusters; internal group B genes are also called BC, and group C genes are all internally located. Group E comprises the unusual var gene named var2csa, which has a nearly conserved sequence. Internal var genes are clustered together in six different chromosomal regions that also include other variant surface antigens such as rif and stevor.
PfEMP1 proteins share a conserved domain architecture: an N-terminal segment followed by combinations of Duffy-binding-like (DBL) and cysteine-rich interdomain region (CIDR) domains, ending with a transmembrane domain and acidic terminal sequence (ATS). Different combinations of DBL and CIDR domains can further be classified into domain cassettes (DCs) [4], with different biological properties and consequences for disease aetiology [5,6]. The DBLα-CIDRα “head” structure is nearly universal and biologically critical, as it determines host receptor binding and thus influences disease severity. For instance, group A and some B variants (DC8) bind to EPCR and are associated with cerebral malaria, whereas group B and C variants bind to CD36 and are linked to milder infections [5–8]. VAR2CSA is associated with pregnancy-associated malaria as it is the ligand binding to Chondroitin Sulfate A in the placenta [9].
The DBLα domain is also widely used as a molecular marker (Fig 1). Due to the presence of conserved motifs, degenerate primers can amplify ~450 bp of the DBLα “tag” region across nearly all var genes. This tag exhibits high diversity between parasite clones and provides a scalable tool for population-level studies of var diversity and expression, even in the absence of reference genomes [10]. DBLα tag sequencing has proven to be a cost-effective alternative to whole-genome or transcriptome approaches, capable of tracking population structure and var expression patterns.
Most var genes comprise an upstream regulatory region and two exons. Exon 1 encodes the extracellular domains, including the N-terminal segment (NTS), and a head structure with a DBLα domain and a cysteine-rich interdomain region (CIDRα). Exon 2 encodes the intracellular acidic terminal segment (ATS). Tag DBLα sequences are conserved fragments of the DBLα domain used in PCR amplification. Domain cassettes are conserved DBLα-CIDRα combinations found across strains. Figure adapted from [11].
Recently, Tan and colleagues [12] highlighted the use of DBLα tags as a scalable solution in epidemiological surveillance. Thereafter, a novel tool “cUps” was developed to take a short DBLα ‘tag’ sequence and predict its var gene ups group [13]. The tool uses a database of var gene tag sequences and links those sequences to ups groups. This was achieved using machine learning (ML) Hidden Markov Models (HMM), which were fitted to the different DBLα subclasses and var gene ups groups. The tool has an overall reported accuracy of 74.3% trained on 846 tag sequences [13]. cUps was found to be very accurate at classifying upsA, but lacks accuracy with upsB and upsCs types, consistently.
In light of this, we initially endeavoured to conduct a comparative study of other machine learning techniques to discern the most effective approach for var gene annotation. We developed a new tool, “upsML”, a suite to run different models for different partial var sequences of different lengths and explore alternate algorithms widely used in classification problems:
- Support Vector Machines (SVM): These models aim to find an optimal hyperplane that best separates the data points of the training dataset into different, pre-defined classes or labels. They can take advantage of the kernel tricks, which transform the space of input features into a higher-dimensional space [14].
- Random Forest: an ensemble learning method that combines multiple decision trees. Each tree is trained on a different random subset of the data and features; then, the final classification is based on the majority vote of these trees. It is robust to overfitting and can perform well on imbalanced datasets [15].
- eXtreme Gradient Boosting (XGBoost): a powerful ensemble learning method known for use in imbalanced datasets and sparse training data. It builds decision trees sequentially, where each new tree attempts to correct the errors made by the previous one [16].
We trained these models to link different partial protein sequences encoded by the var gene (tag sequences, domain cassette, the protein sequence encoded by the exon 1, and full-length PfEMP1 sequences) to their upstream group on an existing dataset of 2,530 var gene sequences.
Results
Preparation of var gene database
For this study, var genes from 66 long-read assemblies were collected (see Methods). As the var genes in these assemblies were annotated using different approaches, we re-annotated them as described in the Methods section. On average, 61.7 var genes were annotated per assembly, which aligns with the expected ~60 var genes per isolate (see S1 Table). An additional dataset comprising 43 isolates from The Gambia was also included [17]. Of the 3,409 var genes identified, 2,940 passed our quality control criteria, resulting in an average of 66.8 var genes per isolate. These re-annotated and Gambia sequences were translated and yielded a dataset of 4,357 PfEMP1 sequences. We then removed duplicate PfEMP1 sequences using CD-HIT, resulting in a final dataset of 2,530 unique PfEMP1 sequences (see S2 Table).
Annotation of ups classifications
To classify the upstream sequences of our var genes by ups type, we used previously published annotations [4], which included 172 known 600 bp upstream sequences for upsA, upsB, and upsC types. These reference sequences were aligned against our dataset, resulting in confident mapping of all of them to our var genes. Sequences annotated as upsE, corresponding to the conserved var2csa, were excluded from phylogenetic-based clustering due to their distinct and well-characterised nature.
We then used the annotated sequences from the Rask et al [4] as guides to label a phylogenetic tree constructed from all 600 bp upstream sequences in our dataset (Fig 2). The resulting tree showed distinct and well-supported clusters for upsA (red) and upsC (blue). In contrast, upsB-associated sequences (yellow) formed multiple subclades, reflecting higher diversity. Clades containing upsB sequences from the Rask database or in-between such clades were classified as ‘upsB’.
Manual classification of subtrees into upsA (red), upsB (yellow), and upsC (blue) groups was based on consistency with the reference annotations.
After removing duplicates and excluding the 3D7 reference genes, the final dataset included 468 upsA, 1,516 upsB, 507 upsC, plus an additional 39 upsE sequences (total: 2,530). These formed the training and validation dataset used throughout our study.
Exploration of upsML candidate models
To predict upstream types from PfEMP1 sequences, we tested Support Vector Machines (SVMs), Random Forest (RanFor), and XGBoost (XGB) across four categories of partial protein sequence types: tag, cassette, the protein sequence encoded by the exon 1, hereafter referred to as ‘exon 1’, and full-length PfEMP1 sequences. Each sequence set was filtered to remove highly similar sequences, retaining only the longest representative (S2 Table). Furthermore, we excluded the PfEMP1 sequences from the 3D7 reference genome, as these were later used as a ground truth for testing.
An 80/20 split was used for model training and testing, with 80% of the data allocated for training and 20% reserved for evaluation. We explored amino acid k-mers of length 1, 2, and 4 as input features. Best results were obtained with tetrapeptides, we did all further analysis with that model (Table 1 and S3 Table).
The single amino acid features consistently performed worst, with overall accuracies ranging from 0.70 to 0.79. Increasing the k-mer to 2 improved performance to 0.70–0.86. The highest accuracies were obtained using tetrapeptides, with overall performance ranging from 0.75 to 0.91 (S3 Table). The best-performing method varied by sequence type: SVM Poly and RBF yielded the highest accuracy (0.85) for tag sequences, while SVM Linear performed best for cassettes (0.87), see Table 1. For ‘exon 1’, both SVM Sigmoid and XGBoost achieved the highest accuracy (0.90). For full-length PfEMP1 sequences, SVMs with linear, RBF, and Sigmoid Kernels all achieved an accuracy of 0.92. The differences among models within each sequence type were small (1–2%), which were not statistically significant. Notably, there is a difference in runtime for the models (S4 Table). SVM Linear was consistently over 100 times faster than the others, making it more energy efficient to use. Extending the input sequence length from tags to full-length PfEMP1 sequences increased the accuracy by approximately 2–3% at each step. For instance, the best models achieved 0.85 accuracy for tags, 0.87 for cassettes, 0.90 for ‘exon 1’, and 0.92 for full-length PfEMP1 sequences.
In summary, no single model consistently outperformed the others. However, using tetrapeptides as input features and longer input sequences consistently yielded the best results. Several models perform similarly; therefore, we recommend linear SVM as it has the lowest runtime.
Accuracies vary between upstream types
Next, we were interested in understanding how the tools would perform on the four different ups groups. We generated confusion matrices (Table 2) and compared the specificity and sensitivity of each upstream group across all tetrapeptide models for the four sequence types: tag, cassette, ‘exon 1’, and full-length PfEMP1 sequences (Table 3 and S5 Table).
As expected, upsE classifications were perfect, due to the strain-transcendent nature of the var2csa genes. The upsA group was also classified with high accuracy, except in a few cases involving SVM Poly and Random Forest models. However, 0.39% of upsB sequences were misclassified as upsA when predicting from tag sequences, and 0.35% were misclassified when using full-length PfEMP1 sequences. Most errors occurred in the classification of upsC. Approximately 25–75% of upsC sequences were misclassified as upsB, depending on the model and sequence type. This may be due to the higher prevalence of upsB in P. falciparum genomes, which made them over-represented in the training datasets. Random Forest and XGBoost are often reported to perform better on unbalanced datasets. However, in this case, we did not observe any substantial advantage over the simpler SVM models.
These misclassifications were further examined through per-group specificity and sensitivity analysis (Table 3). Across all models, upsB predictions tended to yield more false positives (i.e., lower specificity), while upsC predictions produced more false negatives (i.e., lower sensitivity). The low specificity of upsB was due to a substantial proportion of upsC sequences being misclassified as upsB. Conversely, upsC predictions were highly specific (specificities above 0.92), indicating that when a sequence was predicted as upsC, it was almost always correct. However, sensitivity was lower, meaning many true upsC sequences were not detected and instead classified as upsB.
Notably, the prediction accuracy using full-length PfEMP1 sequences reached 92%, compared to 90% using ‘exon 1’ only (Table 2). When including exon 2, upsC classification improved, while upsB classification slightly decreased: 78 upsC sequences were correctly classified from full-length PfEMP1 sequences compared to 64 from ‘exon 1’. This suggests a possible linkage between ups type and exon 2 in internal upsC var gene clusters.
Overall, longer sequence inputs yielded better classification performance. SVM Poly and XGBoost offered the best trade-off between accuracy and sensitivity. However, for tag sequences, misclassification between upsB and upsC still occurred in approximately 15% of cases.
Comparison of upsML with cUps
Currently, no tool predicts the upstream type from a full-length PfEMP1 sequences, or other partial sequences, apart from tag sequences. The most recent method, cUps [13], uses tag sequences as input and applies HMM-based classification, trained on 846 tags with a leave-one-out cross-validation (LOOCV) accuracy of 0.74 (Table 4). To enable a direct comparison, we retrained our classifiers using the same 846-tag dataset and implemented LOOCV. Using tetrapeptide features, our SVM models (RBF and Sigmoid kernels) achieved an accuracy of 0.84, which is 10 percentage points higher than cUps. This performance is comparable to that obtained with our larger training set (~2500 sequences), prompting us to assess the uniqueness of the cUps dataset. In applying CD-HIT with parameters similar to those used on our unique tag set, we found that only 700 of the 846 sequences were unique, indicating 17% redundancy. This helps explain the similarity in performance despite the smaller dataset.
In a complementary analysis, we evaluated cUps using our dataset, with an 80/20 training/testing split (as in Tables 1–3). For ups types A, B, and C, cUps achieved sensitivities of 1.0, 0.5, and 0.42, respectively. In comparison, upsML achieved sensitivities of 1.0, 0.9, and 0.3. The specificities were 0.91, 0.83, and 0.67 for cUps, versus 0.99, 0.65, and 0.96 for upsML. Although individual metrics appear comparable, the overall accuracy of cUps was 0.59, significantly lower than upsML’s 0.83. This discrepancy could be due to cUps performing better on the underrepresented upsC class, which contributes less to the overall accuracy due to class imbalance.
Validation on reference 3D7 genome
As further validation, we tested upsML on the well-annotated P. falciparum 3D7 reference genome (Fig 3). This serves as a strong baseline due to the high confidence in the annotated ups group labels. As expected, all models accurately classified upsA and upsE sequences. Most misclassifications occurred between upsB and upsC, consistent with previous results.
“Known” indicates the ground truth. The models used here are the highest accuracy models for each category, except for the Tag SVM linear, as this was included as the fastest model.
For example, using the SVM linear model on tag sequences, three upsB sequences were misclassified as upsC, and four upsC sequences as upsB. Interestingly, several upsC genes, Pf3D7_0412700, Pf3D7_0420900, and Pf3D7_0712000, were classified differently depending on whether the tag or the full sequence was used. Only one sequence, Pf3D7_0461300, was misclassified with both sequence types. This suggests that combining tag- and full-sequence models may improve overall classification performance.
From a biological perspective, such discrepancies may reflect recent recombination events between var gene groups in 3D7 that differ from those in the training set. Additionally, we observed that upsB classifications were nearly perfect, with errors limited to the so-called BC-type genes – internal upsB variants. In these cases, upsML misclassified three internal upsB sequences as upsC. These results led us to hypothesise that it may be possible to develop a classifier to distinguish var genes based on chromosomal location, specifically, internal var genes versus subtelomeric ones.
A novel “positional” classifier for genome location
To further investigate classification errors (Table 2), we found that 90% of misclassifications occurred between upsB and upsC genes located within internal var gene clusters. This supports our hypothesis from the 3D7 analysis: that recombination events are more frequent within internal clusters [11], and thus these sequences may form a biologically meaningful composite class (Fig 3).
Based on this rationale, we re-annotated 1,980 of the 2,530 genes into four categories: upsA, upsE, upsB_subtelomeric, and upsBC_internal, as detailed in the Methods. We then ran all tetrapeptide-based models on this subset (Table 5). Overall, the classification performance for tags remained limited, three percentage points lower than with the first model (Table 1). This slightly worse result could be attributed to the reduced training set size (~75%).
The best overall tag accuracy reached 0.78, with sensitivity for distinguishing B_subtelomeric from BC_internal ranging from 0.60 to 0.83 (S6 and S7 Tables). Using ‘exon 1’ sequences improved sensitivity to 0.67–0.85 and specificity to 0.80–0.91. Notably, full-length PfEMP1 sequences yielded results up to 10 percentage points higher, but these are less relevant given that long-read assemblies are still required for complete var reconstruction and would indicate location.
Importantly, while accuracy for tag sequences is lower compared to the original classifier (Table 1), classification of upsA and upsE remains near-perfect. We then compared performance across sequence types (Table 5, S6 and S7 Tables). For tag sequences, although the overall accuracy dropped, classification between B_subtelomeric and BC_internal was more balanced than in the previous model. For instance, the polynomial SVM yielded sensitivity and specificity between 0.58 and 0.88. Most predictions were correct, contrasting with earlier results where misclassification between upsB and upsC was frequent.
As expected, performance improved with ‘exon 1’ sequences, where over 75% of sequences per class were correctly classified. While full-length PfEMP1 sequences yielded the highest performance (>85% accuracy), ‘exon 1’ might be the most practical input type. Here, overall accuracy reached 0.84, with B_subtelomeric sequences classified at >83% sensitivity and 0.90 specificity using the linear SVM model.
In summary, although overall accuracy is comparable to the previous model, classification is more balanced across classes. Moreover, this analysis introduces a novel classifier for predicting the chromosomal position of var genes.
Regional tetrapeptide contributions reveal internal recombination signals
When testing upsML on the 3D7 reference genome, we observed that certain sequences were classified differently depending on whether the full-length PfEMP1 sequence or only the tag sequence was used (Fig 3). For example, Pf3D7_0420900 has an upstream region classified as upsC, and the full-length model correctly assigned it as such. However, when only the tag sequence was used, the tag model classified it as upsB. A notable advantage of using SVM models is the interpretability of their decision function: the contribution weights of individual tetrapeptides can be extracted. This allows us to explore which regions of a gene – and which specific tetrapeptides – are most informative for classification into ups groups.
In Fig 4, we visualise the tetrapeptide decision weights across the sequence of Pf3D7_0420900. In the top panel, the decision weights fluctuate along the length of the gene, suggesting no single dominant signal, indicative of a complex or mosaic sequence. To interpret this more systematically, we calculated the cumulative sum of the positional weights (CS) score across the gene. For Pf3D7_0420900, the cumulative score strongly supports upsC. However, at the position corresponding to the tag (approximately amino acid position 184), there is a distinct dip in the cumulative score. The tag signal (green line) in this model is strongly associated with upsB, explaining the discrepancy in classification between the tag and full-length models. This suggests that the tag region is more similar to upsB, whereas the remainder of the gene aligns more closely with upsC.
Top panel: Decision weights for each tetrapeptide along the sequence of Pf3D7_0420900, generated by the SVM classifier. Middle and bottom panels: cumulative sum (CS) of the positional weights decision scores for two genes, Pf3D7_0420900 (middle) and Pf3D7_0412700 (bottom). Red lines represent cumulative scores from the full-length gene; green lines represent scores from the tag sequence. Positive values indicate stronger support for upsC, while negative values indicate stronger support for upsB. Localised shifts suggest complex sequence composition and potential ancient recombination.
A second example is shown in the bottom panel of Fig 4, where both the tag and full-length sequence of Pf3D7_0412700 are classified as upsC. The cumulative sum of the positional weights score supports this, although the initial portion of the tag shows some signal associated with upsB, again suggesting a mixed origin.
These regional fluctuations in signal, where different tetrapeptides contribute positively or negatively to distinct classes, may be indicative of historical recombination events. The ability to track such regional contributions across var genes provides new insights into their complex evolutionary history.
Application of upsML to global P. falciparum genome data
Finally, we applied upsML to the varDB dataset, which contains approximately 2,500 P. falciparum genomes. Most var gene sequences in this dataset are near-complete exon 1 sequences. We utilised the linear SVM model for ‘exon 1’ classification, as it is the most computationally efficient, capable of classifying 100 sequences in approximately 5–15 seconds, while maintaining comparable accuracy to more complex models.
Both classification models, the ups type and genomic location (internal vs subtelomeric), were applied. The resulting annotations have been uploaded to the varDB GitHub repository (https://github.com/ThomasDOtto/varDB/tree/master/Datasets/upsML).
To minimise biases arising from uneven sampling across regions, we analysed a normalised subset of the dataset consisting of 660 isolates from eleven countries (60 isolates per region), with each isolate containing between 52 and 65 var genes, except for the Senegal dataset (average of 48) (Table 6). While the primary goal of this analysis was to demonstrate the scalability and speed of upsML for rapid ‘exon 1’ annotation, taking less than one second to annotate 1,200 var genes, it also provides an overview of geographical patterns in ups types. In line with expectations, upsA and upsE were the most confidently classified types, but we did not observe clear geographical structuring for upsA.
In contrast, our positional classifier revealed a consistent enrichment of internal CB-type var genes in isolates from the Asia regions (Table 6, top). Frequencies were highest in Cambodia and Thailand (33.9% each) and lowest in Kenya and Mali (26.7% and 26.8%, respectively). To better understand those differences in all samples, we grouped the countries into East Africa (Kenya, Malawi, Congo – median 28.1%), West Africa (Senegal, The Gambia, Ghana, Mali – median 28.8%), and Asia (Thailand, Cambodia, Vietnam, Laos – median 32.6%, S1 Fig). Differences in internal var gene frequency were significant between Asia and the two African regions, p-values 3.4e-12 (West Africa) and 2.1e-09 (East Africa), but not within Africa (p-value = 0.04 – Welch Two Sample t-test). In contrast, we did not observe a similarly clear difference for upsC (Table 6, bottom), likely reflecting the known difficulty of distinguishing upsB from upsC based on sequence features alone. Thus, internal var genes are significantly more frequent in Asia.
While these observations highlight the utility of our positional classifier for large-scale comparative analyses, they must be interpreted with caution. The underlying varDB assemblies vary in sequencing technology, read length, sampling time, and assembly quality, which may differentially affect the recovery of specific var gene classes.
To draw further robust biological conclusions regarding var gene distributions, more controlled datasets processed under uniform protocols should be used. For example, large-scale datasets such as Pf8 [18] or RNA-Seq-derived contigs from clinical isolates stratified by disease phenotype [19] would provide a stronger basis for assessing associations between ups types, genomic location, and disease outcomes.
Discussion
Understanding the ups type of var genes is valuable for interpreting their functional properties and roles in Plasmodium falciparum biology. However, obtaining upstream sequence information is often challenging. This is especially true for tag sequences or RNA-Seq data, where the upstream region may not be captured, and even in whole-genome Illumina datasets, low-complexity and repetitive features of ups sequences hinder assembly [20].
To address this, we developed upsML, a machine learning–based tool that predicts ups types based on sequence features. We evaluated multiple classifiers using different input types. This is a non-trivial task, as var genes are known to recombine frequently, especially within internal gene clusters [11,21,22]. Overall, larger k-mers generated better results, and the SVM models were slightly better than the other methods. Interestingly, the SVM model using only 1-mers already provides informative classification performance, suggesting that simple amino acid composition carries a signal for ups type discrimination. Future work using larger datasets could explore methods that incorporate positional information, such as recurrent neural networks or other deep learning approaches.
The tool performs well for upsA and upsE genes, which are distinct and less prone to recombination. In contrast, classification of upsB and upsC types is more challenging due to frequent recombination, particularly in internal clusters. Consequently, sensitivity and specificity for these types can be as low as 0.27. Because of class imbalance, a B call is correct 93% of the time, whereas 73% of C types are misclassified as B. These metrics improve with sequence length: overall accuracy rises from 0.82 (tags) to 0.90 (‘exon 1’) and 0.92 (full-length PfEMP1 sequences). Interestingly, the inclusion of exon 2 shifts the sensitivity and specificity between upsB and upsC. While not formally tested, this suggests that exon 2 may be more tightly linked to the upstream region, potentially because recombination events are more likely to involve exon 1.
To further investigate classification errors, we applied an explainable AI approach by analysing the contribution of individual tetrapeptides to upsB or upsC classification in a Pf3D7 var gene that showed discordant results between that tag and PfEMP1 sequences. For this gene, the tag sequence showed a strong upsB signal, whereas the remainder of the gene clearly leaned towards upsC. Interestingly, local drops in the cumulative classification signal aligned with potential recombination breakpoints, highlighting the mosaic structure of the gene. The mosaic nature of var genes has been frequently described [11,23,24], but our ability to observe it using tetrapeptides linked to upstream sequence types offers a novel approach to identifying sequence features associated with recombination events.
We also assessed whether increasing the amount of training data would improve performance. While this is likely, the recombinogenic nature of var genes imposes an inherent ceiling on classification accuracy. In comparing models, we found that performance was generally consistent across algorithms. However, for tag sequences, some models offered marginally higher accuracy at high computational cost – up to 100-fold slower than others. Given this, we opted for the fastest-performing model (linear SVM) for most input types and the polynomial SVM for tags. With trained models, annotating 1,200 var genes from 20 genomes takes approximately one second using linear SVM and 100 seconds using the polynomial model.
We benchmarked upsML against cUps, an HMM-based tool that utilises tag sequences. On their dataset (using leave-one-out cross-validation), upsML outperformed cUps by 10 percentage points (0.83 vs. 0.74 accuracy, Table 4). Notably, 17% of sequences in the cUps dataset were duplicates, suggesting the true accuracy is likely lower. When tested on our dataset, cUps’ overall accuracy was lower, although it showed improved performance on upsC types. The predominance of upsB sequences per genome obscures this distinction.
Given the difficulty in distinguishing upsB and upsC, and the importance of chromosomal positioning in var gene biology, we developed an additional classifier to distinguish subtelomeric from internal var genes (Table 5). For tag sequences, accuracy reached 78%, and for full-length genes, up to 92%. This again suggests that exon 2 may be a strong indicator of chromosomal location [4]. Because most studies rely on tag sequences or partially assembled genomes, this tool enables reasonably accurate inference of genomic position in approximately 78–84% of cases. Applying this positional classifier to a larger global var gene dataset [20] (Table 6), we observed that the relative abundance of internally located var genes varies geographically. A major contrast between Asian and African parasite populations is transmission intensity, with higher transmission in central and eastern Africa and lower transmission in Asia. In low-transmission areas, parasites are expected to persist longer within individual hosts, favouring chronic infections. Nobel and colleagues [25] predicted that chronicity would be associated with increased expression of upsC (internal) var genes, and our analysis supports this prediction. On the other hand, because those numbers are relative, isolates in Africa may have more subtelomeric var genes, perhaps due to different clinical phenotypes or multiple infections, to maintain higher diversity. More broadly, we note that upsML can be used to generate novel hypotheses and help interpret epidemiological differences in var gene architecture.
More generally, despite the extreme polymorphism and recombination of var genes, our work demonstrates that machine learning methods can be applied effectively to classify these genes. Future work could focus on predicting functional features, such as domain cassettes, and linking these to clinical phenotypes.
In conclusion, upsML offers a robust and efficient tool for annotating var genes by their ups type and genomic location. It holds promise for enabling deeper insights into var gene expression patterns in natural infections and facilitating large-scale comparative studies.
Methods
All custom programs were written in Python and can be found on our GitHub site (https://github.com/sii-scRNA-Seq/upsML). The models used in this analysis are also available on our GitHub.
Preparation of var gene database
A database of var genes was extracted from long-read assemblies and used as training data for models. For this, 23 Plasmodium falciparum long read assemblies were recovered from literature and publicly available datasets [26–28]. Genome sequences were downloaded from NCBI, and assemblies were annotated using AUGUSTUS [29]. Var gene sequences were identified through a BLAST search on the list of protein-coding sequences. To this dataset, we added 2,835 var gene sequences from 43 isolates from The Gambia, which were extracted from high-quality PacBio HiFi long-read assemblies. The genomes were annotated through Companion v2 [30]. Var genes were retrieved from all assemblies by screening for genes annotated as potentially coding PfEMP1 by the Companion pipeline and filtering for nucleotide lengths greater than 2500 bp. All identified var genes were passed through pre-established HMM models [4] to define their domain composition. Copies of var2csa were identified by extracting var genes that had a DBLpam domain. Those genes were then extracted from the dataset and annotated as upsE.
Annotation of ups classifications
To annotate and link var genes to their upstream groups (ups), we used a combination of cluster-based analyses and phylogenetic trees. Var genes of known ups groups were taken from published datasets [4], and we mapped the ups annotations for those genes using SAMtools v1.18 [31] and blastn v2.14.0 [32] onto our sequences.
Then, to cluster all var genes into ups groups, we used phylogenetic tree clustering. First, the 500 bp upstream sequences of all var genes were extracted, then aligned with MAFFT v7.520 [33] and a phylogenetic tree was generated using FastTree 2v.1.11 [34]. The var genes of the known ups group were then used to identify and label phylogenetic clades in the tree. Sequences clustering with var genes representative of ups group A, B and C were assigned the corresponding annotation. To generate our final training set, we added the upsE sequences and excluded 3D7 reference sequences.
After assigning the ups types, we obtained a final set of 780 upsA, 2545 upsB, and 861 upsC var genes. Additionally, 110 var2csa sequences from the long read assemblies were then included in the dataset as upsE (S1 Table). To enhance the diversity found in the training datasets, highly similar sequences were filtered out. CD-HIT v4.8.1 [35] was used to cluster sequences at a 98% identity threshold within each gene section set. This retained the unique representatives with the longest sequence in each cluster. A final diverse dataset of 2530 var genes was obtained. This resulted in 468 upsA, 1,516 upsB, 507 upsC, and 39 upsE.
Extraction of partial var sequences
To prepare training datasets for ML models, var genes were translated to protein sequences. We generated four types of datasets (Fig 1) by pulling out partial sequences corresponding to:
- Tags: a sequence of ~150 amino acids located in the DBLα domain that can be amplified through degenerated PCR primers in field isolates. The N-terminal sequence always starts with “LARSFADIG” [36].
- Cassettes defined the DBLα-CIDRα cassette pair.
- ‘Exon 1’: A sequence that contains the NTS domain and is at least 3.5 kb long, but does not contain either a var intron nor the ATS domain.
- Full-length PfEMP1: The complete coding sequence of the var gene, including both exons.
This was performed through custom Python scripts.
Amongst the var genes, some sequences lacked one or more of these components, but to maximise the information gathered from these var genes and gather the largest training datasets, all identified gene sections were used. Additionally, var genes of the canonical Pf3D7 reference genome were removed from this dataset, as it was used for model verification. After excluding identical var genes, a final diverse dataset of 2,530 full-length PfEMP1 sequences, 2,379 ‘exon 1’s, 1,995 domain cassettes, and 1,970 tags was then used to train the machine learning models (S2 Table).
Creating and testing machine learning models
To classify var gene sequences, amino acid sequences of each partial sequence type were decomposed into their raw peptides (k-mer = 1), dipeptides (k-mer = 2), and tetrapeptides (k-mer = 4) count tables. Those were then used as features for six different machine learning algorithms: Random Forest, XGBoost, and SVMs with four different kernels (Linear, polynomial, Radial Basis Function and Sigmoid). All implementations were done in Python; the scikit-learn library v1.3.2 [37] was used for all SVM and random forest models, while xgboost v2.1.1 was used for XGBoost. The hyperparameters of the models were optimised through the Python library optuna v4.1.0. The complete var genes datasets were split into 80% training, 20% test datasets and overall accuracies, precisions, sensitivities, and specificities were calculated and reported for each model using the Python libraries scikit-learn 1.3.2, Biopython 1.83, Pandas 2.0.3, and NumPy 1.24.4. [38].
Comparison to cUps
To ensure a fair comparison with cUps, we first downloaded their original reference dataset of 846 DBLα tag sequences 11. We then applied leave-one-out cross-validation (LOOCV) using the same dataset on the upsML models. This allowed a direct comparison of classification accuracy under the same evaluation strategy reported in the original cUps study 11. We then trained it with the same input as upsML for a comparison with our data.
Validation on reference 3D7 genome
We used the upsML models to annotate the ups type of the 61 3D7 var genes [39], with the tag sequences isolated from the full sequences using the previously mentioned method. The occurrence heatmap was done in Python with pandas 2.0.3, Seaborn 0.13.2, and Matplotlib 3.7.5.
Var gene localisation model
Refined models with added information about var gene localisation were trained in a similar method. To do this, labels of the var sequences were changed to upsA, upsE, upsB_subtelomeric and upsBC_internal. We previously [27] manually annotated the chromosomal position of the var genes. For the remaining genes, we verify their position on chromosomes known to harbour internal var genes and annotate the var genes accordingly. From the 2,545 upsB dataset, 1,016 subtelomeric and 297 internal sequences were successfully identified. This resulted in an annotated dataset of 780 upsA, 110 upsE, 1,016 upsB_subtelomeric, and 1,158 upsBC_internal full-length PfEMP1 sequences.
To produce the models using this alternate dataset, the gene regions were extracted as previously mentioned. This led to final unique datasets of 1,550 tags, 1,541 cassettes, 1,890 ‘exon 1”s, and 1,980 full-length PfEMP1 sequences.
Explainable AI
Feature importances were estimated for the tetrapeptide linear SVM model by fitting models that classify var sequences into the two classes, var group B/C or internal/external var genes. We used the weights of the SVM models, i.e., the coordinates of the vector orthogonal to the hyperplane separating the two classes [36], as direct measures of each feature’s (amino acid k-mer sequence) contribution to classify a PfEMP1 sequence into one class or the other. This feature vector was defined in 01_feature_importance.py by the Python calculate_tetrapeptide_counts function. Here is the pseudo code:
Input: sequence S
Output: tetrapeptide count vector V
- If necessary, translate S to amino acid sequence.
- Slide a window of size 4 along S to extract all overlapping tetrapeptides.
- Count occurrences of each tetrapeptide.
- Output counts in a fixed order defined by ALL_TETRAPEPTIDES.
The coef_ attribute of the scikit-learn LinearSVC class, defined in the documentation as the “weights assigned to the features”, was used to extract weights from linear SVM models. In these code snippets from 01_feature_importance.py, clf.coef_ stores the coefficients of the linear model learned by LinearSVC:
Input: feature matrix X, label vector y, regularisation parameter C
Output: feature weight vector W
- Initialise a linear Support Vector Machine classifier with parameter C.
- Train the classifier on X and y.
- Extract the learned feature weight vector W from the trained model.
- Return W.
Using a sliding window, we then mapped the relative feature importances of each tetrapeptide along full PfEMP1 sequences. This allowed us to identify regions of PfEMP1 sequences associated with different classes (Fig 4a). By summing up the contributions of tetrapeptides along the protein sequences, we could also illustrate how tetrapeptide features progressively influenced the classification outcome from the start to the end of the sequence. The was done using the following function from 02_explainable_ai_plot.py that turns tetrapeptide SVM coefficients into a per-position “decision weight” signal across the sequence:
Input: protein sequence S, tetrapeptide weight mapping Wdict
Output: positional weight profile P along S
- Convert sequence S to uppercase.
- Initialise positional score array P of length equal to S with zeros.
- For each position i from 1 to length(S) − 3:
- Extract tetrapeptide T = substring of S from i to i + 3.
- Retrieve weight w = Wdict[T]; if T is not present, set w = 0.
- Determine centre position c = i + 1 (central residue of tetrapeptide).
- Add w to P[c].
- Return P.
We calculated the cumulative sum (CS) across the entire PfEMP1 sequences (red line in Figs 4b and 4c) to capture the overall classification bias implied by the linear SVM model along the protein. A subregional CS (green line in Fig 4b and 4c) was also calculated to quantify the local contribution of features within shorter DBLa tag sequences. This calculation used the profile (the vector of per-position linear SVM decision weights) and the subregion (the amino-acid interval defining the DBLα tag). This was implemented in 02_explainable_ai_plot.py as:
Input: positional weight profile P along sequence S, subregion interval [a, b]
Output: full-sequence cumulative trace Cfull, subregional cumulative trace Csub
- Compute cumulative sum of P over the full sequence to obtain Cfull.
- Extract subregional profile segment Psub = P[a to b].
- Compute cumulative sum of Psub to obtain Csub.
- Return Cfull and Csub.
3K genomes
In 2018, we published the var genes of over 2500 field isolates [20]. To attribute their ups type, the full and normalised datasets of var genes as amino acids was downloaded from https://github.com/ThomasDOtto/varDB/tree/master/Datasets/upsML. The upsML SVM linear models were used to attribute the ups and localisation types, with the results uploaded to varDB GitHub.
Supporting information
S1 Table. Complete Training Set of Full Var Genes and ups Assignments used.
https://doi.org/10.1371/journal.pone.0344557.s001
(PDF)
S2 Table. All and Unique Training Sets of Isolated Var Genes Components Regions.
https://doi.org/10.1371/journal.pone.0344557.s002
(PDF)
S3 Table. Overall accuracy of upsML, including shorter peptide models.
https://doi.org/10.1371/journal.pone.0344557.s003
(PDF)
S4 Table. upsML Model Speeds when training then classifying 60 3D7 complete PfEMP1 sequences.
https://doi.org/10.1371/journal.pone.0344557.s004
(PDF)
S5 Table. Specificity and Sensitivity of Tetrapeptide Models for first model.
Sensitivity refers to the proportion of true positives, while specificity refers to the proportion of true negatives.
https://doi.org/10.1371/journal.pone.0344557.s005
(PDF)
S6 Table. Confusion Matrixes of Internal/ Subtelomeric Models.
https://doi.org/10.1371/journal.pone.0344557.s006
(PDF)
S7 Table. Sensitivity and Specificity of Internal/Subtelomeric Models.
Split of results for the different sequence types.
https://doi.org/10.1371/journal.pone.0344557.s007
(PDF)
S1 Fig. Relative number of internal var genes per region.
Data from Table 6, of 660 field isolates, from 11 regions, classified with upsML. The genes were grouped as follows: Asia: Thailand, Cambodia, Vietnam and Laos (median 32.6); East Africa: Mali, Kenya, Malawi & Congo (median 28.1); West Africa: Senegal, The Gambia, Mali and Ghana (median 28.8). Isolates with fewer than 30 var genes in the assembly were ignored. The differences in abundance between Asia and the two African regions were significant, as indicated by a Welch Two-Sample t-test (both <1e-08).
https://doi.org/10.1371/journal.pone.0344557.s008
(PDF)
References
- 1. Hadjimichael E, Deitsch KW. Variable surface antigen expression, virulence, and persistent infection by Plasmodium falciparum malaria parasites. Microbiol Mol Biol Rev. 2025;89(1):e0011423. pmid:39807932
- 2. Zhang X, Deitsch KW. The mystery of persistent, asymptomatic Plasmodium falciparum infections. Curr Opin Microbiol. 2022;70:102231. pmid:36327690
- 3. Lavstsen T, Salanti A, Jensen ATR, Arnot DE, Theander TG. Sub-grouping of Plasmodium falciparum 3D7 var genes based on sequence analysis of coding and non-coding regions. Malar J. 2003;2:27. pmid:14565852
- 4. Rask TS, Hansen DA, Theander TG, Gorm Pedersen A, Lavstsen T. Plasmodium falciparum erythrocyte membrane protein 1 diversity in seven genomes--divide and conquer. PLoS Comput Biol. 2010;6(9):e1000933. pmid:20862303
- 5. Lavstsen T, Turner L, Saguti F, Magistrado P, Rask TS, Jespersen JS, et al. Plasmodium falciparum erythrocyte membrane protein 1 domain cassettes 8 and 13 are associated with severe malaria in children. Proc Natl Acad Sci U S A. 2012;109(26):E1791-800. pmid:22619319
- 6. Claessens A, Adams Y, Ghumra A, Lindergard G, Buchan CC, Andisi C, et al. A subset of group A-like var genes encodes the malaria parasite ligands for binding to human brain endothelial cells. Proc Natl Acad Sci U S A. 2012;109(26):E1772-81. pmid:22619330
- 7. Tonkin-Hill GQ, Trianty L, Noviyanti R, Nguyen HHT, Sebayang BF, Lampah DA, et al. The Plasmodium falciparum transcriptome in severe malaria reveals altered expression of genes involved in important processes including surface antigen-encoding var genes. PLoS Biol. 2018;16(3):e2004328. pmid:29529020
- 8. Warimwe GM, Fegan G, Musyoki JN, Newton CRJC, Opiyo M, Githinji G, et al. Prognostic indicators of life-threatening malaria are associated with distinct parasite variant antigen profiles. Sci Transl Med. 2012;4(129):129ra45. pmid:22496547
- 9. Salanti A, Staalsoe T, Lavstsen T, Jensen ATR, Sowa MPK, Arnot DE, et al. Selective upregulation of a single distinctly structured var gene in chondroitin sulphate A-adhering Plasmodium falciparum involved in pregnancy-associated malaria. Mol Microbiol. 2003;49(1):179–91. pmid:12823820
- 10. Tonkin-Hill G, Ruybal-Pesántez S, Tiedje KE, Rougeron V, Duffy MF, Zakeri S, et al. Evolutionary analyses of the major variant surface antigen-encoding genes reveal population structure of Plasmodium falciparum within and between continents. PLoS Genet. 2021;17(2):e1009269. pmid:33630855
- 11. Claessens A, Hamilton WL, Kekre M, Otto TD, Faizullabhoy A, Rayner JC, et al. Generation of antigenic diversity in Plasmodium falciparum by structured rearrangement of Var genes during mitosis. PLoS Genet. 2014;10(12):e1004812. pmid:25521112
- 12. Tan MH, Shim H, Chan Y-B, Day KP. Unravelling var complexity: Relationship between DBLα types and var genes in Plasmodium falciparum. Front Parasitol. 2022;1:1006341. pmid:36998722
- 13. Tan MH, Tiedje KE, Feng Q, Zhan Q, Pascual M, Shim H, et al. A paradoxical population structure of var DBLalpha types in Africa. bioRxiv. 2023. pmid:37986738
- 14. Burges CJC. A Tutorial on Support Vector Machines for Pattern Recognition. Data Mining and Knowledge Discovery. 1998;2(2):121–67.
- 15. Breiman L. Random Forests. Machine Learning. 2001;45(1):5–32.
- 16. Chen T, Guestrin C. XGBoost: A Scalable Tree Boosting System. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. San Francisco, California, USA. 2016. 785–94.
- 17. Nyarko PB, Quenu M, Guery M-A, Cohen C, Hamilton WL, Girgis ST. De novo assembly of complete Plasmodium falciparum isolate genomes using PacBio HiFi sequencing technology. bioRxiv. 2025.
- 18. Abdel Hamid MM, Abdelraheem MH, Acheampong DO, Adam I, Aide P, Ajibaye O, et al. Pf8: an open dataset of Plasmodium falciparum genome variation in 33,325 worldwide samples. Wellcome Open Res. 2025;10:325.
- 19. Andradi-Brown C, Wichers-Misterek JS, von Thien H, Höppner YD, Scholz JAM, Hansson H. Elife. Elife. 2024;12:RP87726.
- 20. Otto TD, Assefa SA, Böhme U, Sanders M, Kwiatkowski D, Pf3k consortium. Evolutionary analysis of the most polymorphic gene family in falciparum malaria. Wellcome Open Res. 2019;4(193).
- 21. Sander AF, Lavstsen T, Rask TS, Lisby M, Salanti A, Fordyce SL, et al. DNA secondary structures are associated with recombination in major Plasmodium falciparum variable surface antigen gene families. Nucleic Acids Res. 2014;42(4):2270–81. pmid:24253306
- 22. Bopp SER, Manary MJ, Bright AT, Johnston GL, Dharia NV, Luna FL, et al. Mitotic evolution of Plasmodium falciparum shows a stable core genome but recombination in antigen families. PLoS Genet. 2013;9(2):e1003293. pmid:23408914
- 23. Bull PC, Buckee CO, Kyes S, Kortok MM, Thathy V, Guyah B, et al. Plasmodium falciparum antigenic variation. Mapping mosaic var gene sequences onto a network of shared, highly polymorphic sequence blocks. Mol Microbiol. 2008;68(6):1519–34. pmid:18433451
- 24. Zilversmit MM, Chase EK, Chen DS, Awadalla P, Day KP, McVean G. Hypervariable antigen genes in malaria have ancient roots. BMC Evol Biol. 2013;13:110. pmid:23725540
- 25. Noble R, Christodoulou Z, Kyes S, Pinches R, Newbold CI, Recker M. The antigenic switching network of Plasmodium falciparum and its implications for the immuno-epidemiology of malaria. Elife. 2013;2:e01074. pmid:24062941
- 26. Ruiz JL, Reimering S, Escobar-Prieto JD, Brancucci NMB, Echeverry DF, Abdi AI. Briefings in bioinformatics. 2023;24(4). pmid:37406192
- 27. Otto TD, Böhme U, Sanders M, Reid A, Bruske EI, Duffy CW, et al. Long read assemblies of geographically dispersed Plasmodium falciparum isolates reveal highly structured subtelomeres. Wellcome Open Res. 2018;3:52. pmid:29862326
- 28. Moser KA, Drábek EF, Dwivedi A, Stucke EM, Crabtree J, Dara A, et al. Strains used in whole organism Plasmodium falciparum vaccine trials differ in genome structure, sequence, and immunogenic potential. Genome Med. 2020;12(1):6. pmid:31915075
- 29. Stanke M, Keller O, Gunduz I, Hayes A, Waack S, Morgenstern B. AUGUSTUS: ab initio prediction of alternative transcripts. Nucleic Acids Res. 2006;34(Web Server issue):W435-9. pmid:16845043
- 30. Haese-Hill W, Crouch K, Otto TD. Annotation and visualization of parasite, fungi and arthropod genomes with Companion. Nucleic Acids Res. 2024;52(W1):W39–44. pmid:38752499
- 31. Li H, Handsaker B, Wysoker A, Fennell T, Ruan J, Homer N, et al. The Sequence Alignment/Map format and SAMtools. Bioinformatics. 2009;25(16):2078–9. pmid:19505943
- 32. Camacho C, Coulouris G, Avagyan V, Ma N, Papadopoulos J, Bealer K, et al. BLAST+: architecture and applications. BMC Bioinformatics. 2009;10:421. pmid:20003500
- 33. Katoh K, Standley DM. MAFFT multiple sequence alignment software version 7: improvements in performance and usability. Mol Biol Evol. 2013;30(4):772–80. pmid:23329690
- 34. Price MN, Dehal PS, Arkin AP. FastTree 2--approximately maximum-likelihood trees for large alignments. PLoS One. 2010;5(3):e9490. pmid:20224823
- 35. Fu L, Niu B, Zhu Z, Wu S, Li W. CD-HIT: accelerated for clustering the next-generation sequencing data. Bioinformatics. 2012;28(23):3150–2. pmid:23060610
- 36. Taylor HM, Kyes SA, Harris D, Kriek N, Newbold CI. A study of var gene transcription in vitro using universal var gene primers. Mol Biochem Parasitol. 2000;105(1):13–23. pmid:10613695
- 37. Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O. Scikit-learn: Machine Learning in Python. J Mach Learn Res. 2011;12:2825–30.
- 38.
Cristianini N, Shawe-Taylor J. An introduction to support vector machines and other kernel-based learning methods. Cambridge: Cambridge University Press. 2000.
- 39. Böhme U, Otto TD, Sanders M, Newbold CI, Berriman M. Progression of the canonical reference malaria parasite genome from 2002-2019. Wellcome Open Res. 2019;4:58. pmid:31080894