Integrative transcriptomic and machine learning approaches to decipher mitochondrial gene regulation in severe Plasmodium vivax malaria Authors: Priyanka Roy, Yash Aggarwal, Sanjay Kumar Kochar, Dhanpat Kumar Kochar, Ashis Das Published: 2025 ABSTRACT AbstractMitochondria in Plasmodium vivax are functionally vital despite possessing a highly reduced genome and differing substantially from the human organelle. Beyond their classical role in energy production, they dynamically coordinate processes like pyrimidine biosynthesis and heme metabolism, adapting their functions across the intra-erythrocytic development cycle (IDC). Their unique architecture and stage-specific roles enable the parasite to fine-tune mitochondrial gene expression, involving both protein-coding sense transcripts and long non-coding natural antisense transcripts (NATs). This study unveils an unprecedented regulatory complexity by integrating transcriptomic profiling with advanced machine learning to decode the role of mitochondrial sense and natural antisense transcripts (NATs) in severe P. vivax malaria. We reveal distinct, clinically relevant expression signatures, where NATs emerge not as transcriptional by-products but as potent regulators tightly linked to mitochondrial pathways and translational machinery. This dual-layered transcriptomic landscape reflects an intricate molecular strategy by which the parasite fine-tunes mitochondrial function to survive under severe disease conditions. Importantly, while these findings illuminate novel regulatory mechanisms and position mitochondrial NATs as promising targets for antimalarial drug development, they represent preliminary insights derived from a limited clinical cohort and should not be interpreted as definitive clinical indicators. Validation in larger and diverse patient populations is essential to confirm their broader biological and clinical relevance. However, these results serve as indicators for potential innovative therapeutic interventions aimed at disrupting parasite bioenergetics and regulatory networks.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12936-025-05725-8.Keywords: Mitochondria, Plasmodium vivax, Transcriptome, Machine Learning, NATs Introduction Malaria, a significant global health challenge, is caused by protozoan parasites of the genus Plasmodium, with Plasmodium vivax being one of the most widespread and prevalent species. According to the World Health Organization (WHO) 2024 report, an estimated 263 million malaria cases were recorded globally in 2023. Strikingly, Plasmodium vivax represented approximately 48.2% of malaria infections in the South-East Asia region, with India emerging as a major contributor, accounting for nearly 25% of the global P. vivax burden [1], highlighting its substantial burden on public health. Historically, P. vivax was thought to cause benign tertian malaria, however many studies have reported that it causes significant severity and morbidity, where changes in gene expression plays a major role on the underlying mechanisms contributing to this shift. Transcriptional analyses have shown that variations in gene expression profiles can affect the parasite's ability to invade erythrocytes and modulate host immune responses [2, 3]. Most eukaryotic genomes produce non-coding RNAs (ncRNAs), which integrate complex suites of gene activity and contribute to the complexity of gene regulation [4]. The lncRNAs are a subclass of non-coding RNAs (ncRNAs) that are longer than 200nt and are typically distinguished by their inability to code for proteins. They can be categorized as intergenic, intronic, sense lncRNAs, transcribed from the sense strand of protein-coding genes, and Natural Antisense Transcripts (NATs), transcribed from the antisense strand of protein-coding genes, based on the genomic area of transcription [5]. The lncRNAs can modulate gene expression at numerous levels, including transcription, post-transcription, RNA turnover and protein translation [6, 7]. The presence of ncRNA in Plasmodium has the potential to regulate key events in the complex life cycle and modulate the virulence ability of the parasite, by controlling transcriptional and post-transcriptional processes [8]. Patankar et al., 2001 and Militello et al., 2005 reported the presence of antisense transcripts in Plasmodium parasite and their role in gene regulation [9, 10]. Additionally, the presence of antisense transcripts in gametocytes and ookinetes implies that these antisense RNAs could be essential for the control of gene expression and parasite development [11]. Studies from our laboratory have showed the presence of NATs using a custom designed strand specific whole genome microarray in P. vivax from patients showing complicated and uncomplicated disease manifestations [12]. In Plasmodium the role of the mitochondria is particularly fascinating because its function varies depending on the stage of the parasite’s lifecycle. It has a minimalistic genome encoding only three proteins cytc oxidase subunit 1 (cox1) and subunit 3 (cox3), members of the cytc oxidase complex (complex IV), and cytb (cob), a member of cytochrome bc1 complex. Despite the reduced genome, Plasmodium mitochondria play crucial roles in metabolism, including pyrimidine biosynthesis and electron transport chain in the asexual blood stages [13, 14]. Apart from this, there is a functional citric acid cycle within the mitochondrion which is responsible for the de novo biosynthesis of heme. Despite the widespread haemoglobin breakdown in the digestive vacuole, heme production in Plasmodium is divided between the mitochondrion, the apicoplast, and the cytosol [13–16]. Despite mitochondria having their own genome with transcriptional and translational machinery, majority of the proteins found in mitochondria are encoded in the nucleus due to a huge endosymbiotic gene transfer from the mitochondrial to the nuclear genome [17]. There are possible energy requirement differences between the asexual intra-erythrocytic developmental stages and the sexual stages such as gametocytes. In asexual stages, no cristae are visible in the mitochondrial membrane while in the mitochondria of gametocytes, crista-like structures accumulate in the respiratory chain complexes [18, 19]. Another interesting observation remains that there is a significantly lower level of Complex V components in the asexual stages, compared to the sexual stages of the parasites [19]. Even though the Plasmodium mitochondria has limited number of metabolic process and pathways, mitochondrial gene products and their expression are functionally vital, with stage-specific remodelling evident between asexual and gametocyte stages, highlighting their fundamental contribution to parasite viability and transmission potential. A wide range of drugs are designed to target mitochondrial pathways, but the emergence of resistance has become a significant concern during their development. The relaxed selective pressure for mitochondrial electron transport chain function during the blood stage, along with mitochondrial genome polyploidy seems to accelerate the emergence of atovaquone-resistance mutations [20]. Based on the data in this study, this is the first comprehensive evidence of mitochondrial transcripts from P. vivax clinical isolates with differing disease manifestations showing transcripts in both sense and NATs orientation. Through integrative transcriptomic profiling and machine learning models, we uncovered key insights into mitochondrial gene regulation mediated by NATs, particularly in the context of severe disease manifestation. Notably, we observed distinct differential expression patterns of NATs across clinical conditions, with a strong association to mitochondrial pathways and the translational machinery. These findings suggest that NATs are not random transcriptional by-products but potentially function as regulatory elements that fine-tune mitochondrial gene expression during pathogenesis. Altogether, in correlation with the gene expression signatures, the mitochondrial NATs in Plasmodium may provide alternative targets for antimalarial drug development and disease intervention strategies. Materials and methods Curating gene list with potential mitochondrial localization To employ a systematic approach to compile a comprehensive list of P. vivax mitochondrial genes, a multi-layered bioinformatic strategy was employed, integrating web-based prediction tools with functional annotations. Given the absence of a universally conserved mitochondrial localization signal in Plasmodium species and the limitations of individual prediction tools each trained on distinct datasets and algorithms multiple complementary platforms were utilized, following the recommendations of Martelli et al. [21]. The four primary tools included: TargetP, which identifies mitochondrial transit peptides (mTPs) characterized by enrichment of arginine, leucine, and serine residues critical for matrix targeting [22]; Cello2GO, which leverages BLAST-based homology to assign subcellular localization [23]; MitoFates, which predicts mitochondrial pre-sequences using classical targeting features and motif analyses [24]; and Mitoprot II, a tool that employs discriminant analysis of amino acid composition, charge, hydrophobicity, and amphiphilic potential to identify mitochondrially imported proteins [25]. Additionally, the analysis was strengthened by incorporating genes annotated with mitochondrial targeting domains in PlasmoDB [26], alongside orthologous genes that have been a part of the Plasmodium falciparum mitochondrial proteome [27]. This combined approach ensured the inclusion of both predicted and empirically confirmed mitochondrial candidates. To ensure prediction robustness, a stringent cutoff score of > 0.6 was applied for each tool, and duplicate predictions across different platforms were consolidated. This filtering yielded a final non-redundant list of unique nuclear-encoded, mitochondria-targeted candidates in P. vivax. Blood sample collection and processing Venous blood samples were obtained from patients diagnosed with Plasmodium vivax infection at S.P. Medical College, Bikaner, India. To enrich for infected erythrocytes, the collected blood was processed using a Histopaque density gradient. All patient samples were collected by a dedicated team of clinicians under ethical approval, with informed consent obtained in accordance with hospital guidelines. The presence of P. vivax infection was initially confirmed by slide microscopy and rapid diagnostic tests (RDTs), followed by PCR-based validation to ensure mono-infection [28, 29]. The cohort consist of 14 P. vivax patients (5 uncomplicated (UNC) and 9 complicated with hepatic dysfunction (HD) collected between 2008 and 2010. The patient’s diagnosis was confirmed as hepatic dysfunction as per WHO guidelines [30] by the team of clinicians. Additionally, no samples processed showed any co-morbidities. Detailed clinical characteristics of the enrolled patients are presented in Table S1. For transcriptomic profiling, total RNA from each sample was individually hybridized onto custom-designed 8 × 60 K single-color microarrays. Microarray hybridization Total RNA was extracted from patient blood samples following established protocols [12]. The concentration and purity of RNA were assessed using a NanoDrop 2000 spectrophotometer (Thermo Scientific) and Qubit fluorometric assay (Thermo Scientific, USA). RNA integrity was evaluated on an Agilent TapeStation system to ensure high-quality input for downstream analysis. Microarray labeling and hybridization were carried out at the Agilent-certified facility of Genotypic Technology Pvt. Ltd., Bengaluru, India. RNA samples were processed using the Agilent Quick-Amp Labeling Kit (p/n 5190-0442). In this process, total RNA was reverse transcribed at 40 °C using an oligo-dT primer linked to a T7 polymerase promoter to synthesize first-strand cDNA. This was followed by second-strand synthesis to generate double-stranded cDNA, which served as a template for in vitro transcription. During in vitro transcription, cyanine-3 (Cy3) labeled cRNA was produced using Agilent Cy3 CTP. All enzymatic reactions, including cDNA synthesis and transcription, were performed at 40 °C. The labeled cRNA was purified using Qiagen RNeasy columns (Cat. No. 74106), and quality metrics yield and specific activity were confirmed using the NanoDrop ND-1000. Prior to hybridization, labeled cRNA was fragmented at 60 °C and applied to Agilent Custom P. vivax 8 × 60 K Gene Expression Microarrays. Hybridization was carried out using the Agilent Gene Expression Hybridization Kit (Part No. 5190-0404) in SureHyb chambers at 65 °C for 16 h. Post-hybridization washes were performed using Agilent Gene Expression Wash Buffers (Part No. 5188-5327). Microarray slides were scanned with the Agilent Microarray Scanner (Model G2600D), and raw data were extracted using Agilent Feature Extraction software from images for further analysis. Microarray design A custom 60 K sense-antisense microarray was meticulously designed to enable comprehensive transcriptomic profiling of Plasmodium vivax, with a particular emphasis on mitochondrial gene regulation. The array was constructed using genomic annotations from the P. vivax Sal-1 reference genome PlasmoDB v47 (www.plasmodb.org), incorporating both nuclear and mitochondrial open reading frames (ORFs). Each ORF was represented by five distinct oligonucleotide probes (60–132 mers) in both sense and antisense orientations. Additionally, probes targeting plasmodial mitochondrial sequences were cross-verified against the Homo sapiens mitochondrial genome and the probes for nuclear encoded transcripts were also cross-verified against the Homo sapiens nuclear genome; probes with significant matches were removed, any cross-hybridizing probes were also excluded. The array was fabricated on the Agilent microarray platform, ensuring optimal hybridization performance and data quality. Probes from our previously validated 15 K and 244 K microarray [3, 12] were selectively integrated into the current 60 K design to enhance coverage and reliability. In this study, we specifically report transcripts, detected through hybridization, that are either encoded by the P. vivax mitochondrial genome or are predicted to be mitochondrial-targeted. Microarray data analysis Feature-extracted raw image data, filtered for signals at least two-fold above background intensity, were processed using Agilent GeneSpring GX Software, following the workflow described in Boopathi et al., 2016. Data normalization was performed using the 75th percentile shift method to ensure consistency across samples and reduce technical variability. To enhance strict stringency, only probes detecting transcripts in at least 60% of samples were retained for further analysis. Average fold expression values were calculated for each gene across all complicated cases and differential gene expression was calculated for the severe disease manifestation (HD vs UNC). Genes exhibiting a fold change ≥ 1.5(equivalent to a log2 fold change of + 0.6 for upregulation or − 0.6 for downregulation were considered as differentially expressed, and their statistical significance was determined using a student t-test (p ≤ 0.05), with multiple testing correction applied via the Benjamini–Hochberg false discovery rate (FDR) implemented through the limma and genefilter R packages in Bioconductor. Machine learning Preprocessing and feature selection The differentially expressed genes were statistically ranked by adjusted p-values and used to construct training datasets for machine learning. To systematically evaluate predictive performance and feature stability, the differentially expressed genes were ranked by adjusted p-values and progressively reduced in subsets from top 92, 80, 70, 60, 50, 40, 30, and 20 transcripts for sense transcripts, and 77, 70, 60, 50, 40, 30, and 20 transcripts for NATs. This stepwise reduction enabled us to examine how the number of features influenced model robustness and interpretability. Our machine learning models, trained on these transcriptomic signatures, were further interrogated to uncover potential regulatory transcripts in sense and NATs that may act as drivers involved in the modulation of mitochondrial gene expression. The resulting models not only demonstrated high predictive performance but also guided downstream analyses to explore enriched pathways and biological functions. This integrative approach helps direct, preliminary studies, for identifying potential molecules and improving clinical stratification of malaria patients. Unlike simple accuracy, these metrics offer nuanced insights, ensuring the model’s effectiveness is not masked by skewed class ratios. A high F1-Score, MCC, and AUC collectively indicate the model’s strength in accurately predicting both complicated and uncomplicated cases, providing a holistic assessment of its predictive capabilities across different class distributions. Model training We employed three machine learning algorithms, including XGBoost [31] with parameters n_estimators = 100,max_depth = 2,learning_rate = 0.05,subsample = 0.8,colsample_bytree = 0.8,reg_alpha = 0.1,reg_lambda = 1, random forest [32] with parameters for max_depth = 3, min_samples_split = 5, n_estimators = 100, logistic regression [33] with parameters max_iter = 1000 to test and predict key regulators of disease severity in case of mitochondrial transcripts from sense and NATs. The Leave-One-Out (LOOCV) cross-validation technique was utilized [34]. Due to the limited number of samples, the dataset was not further split into independent training and testing sets, as this would have resulted in very small partitions. Instead, we employed Leave-One-Out Cross-Validation (LOOCV) to address critical challenges in evaluating machine learning models trained on limited and high-dimensional biological data. For imbalanced datasets, standard cross-validation may create folds with zero minority-class samples, skewing performance metrics. Small datasets with high dimensionality are prone to overfitting. LOOCV’s exhaustive resampling helps detect overfitting by testing the model’s consistency across all possible training configurations. In standard k-fold cross-validation, the dataset is partitioned into k subsets, with each subset serving as the test set once. LOOCV represents the extreme case of k-fold CV, where k equals the total number of samples. LOOCV guarantees each test set contains exactly one sample, preserving class representation and providing a more reliable estimate of model sensitivity/specificity. For each iteration of LOOCV, N-1 samples are used to train the model while the remaining single sample serves as the test set. This process is repeated N times, ensuring every sample is used as the test set exactly once. Mathematically, the LOOCV estimate of the mean squared error (MSE) is given by: where N denotes the number of iterations, and the term in the brackets denotes the model trained on all samples except the i-th observation. LOOCV helps to ensure no data wastage, which is ideal for precious samples, and provides stability by removing randomness from fold generation, which is crucial for publishing reproducible results. Computational cost for employing this technique increases linearly with the sample size, making it prohibitive for large datasets. In our study, LOOCV ensured rigorous validation without sacrificing statistical robustness. Evaluation We use several indexes, including F1-Score, Matthews Correlation Coefficient (MCC), and Area Under the Curve (AUC), to evaluate the performance of models, each with its specific formula, significance, and application, particularly in scenarios with imbalanced data (non-uniform between uncomplicated and complicated samples). The metric of F1-Score is the harmonic mean of precision and recall, calculated as F1-Score provides a balanced assessment of a model’s ability to identify both true positives and true negatives. A score close to 1 indicates strong performance, particularly vital in datasets with unequal class distributions. MCC is calculated using the formula: where TP, TN, FP, and FN represent true positives, true negatives, false positives, and false negatives, respectively. It offers a comprehensive evaluation that considers all elements of the matrix (TP, TN, FP, FN). It ranges from − 1 (inverse prediction) to + 1 (perfect prediction). MCC is especially useful in imbalanced datasets for its balanced insight. AUC (Area Under the Curve) quantifies a model’s ability to distinguish between classes, independent of class imbalance, AUC ranges from 0 to 1. To gain deeper insights into model interpretability, we employed SHAP (SHapley Additive exPlanations, v0.47.2) to unravel the contribution of individual features to the prediction outcome. By computing Shapley values, we quantified the impact of each feature on the model’s decision-making process, enabling a precise, feature-level understanding of predictive influence [35]. Genes were ranked by mean absolute SHAP values across LOOCV folds. To strengthen prioritization, we also examined stability across models and folds, and cross-validated biological relevance using gene set enrichment analysis. Only genes meeting both statistical and biological criteria were emphasized. Permutation testing for model validation To rigorously assess whether our machine learning models captured genuine biological signal rather than chance correlations, we employed permutation testing as a negative control. In this approach, the class labels (uncomplicated vs complicated) were randomly shuffled while preserving the gene expression structure, thereby destroying any true association between features and labels. The models (Random Forest, Logistic Regression, XGBoost) were retrained and evaluated on these permuted datasets across 200, 500 and 1000 iterations and performance metrics were recalculated. The distribution of scores obtained from permuted datasets represents the null distribution the level of performance expected if the models were learning noise only. By comparing the observed performance on the real data to this null distribution, we calculated empirical p-values, defined as the proportion of permuted scores greater than or equal to the observed score. This approach directly tests whether the observed model performance is statistically greater than chance. Gene set enrichment and pathway analysis To uncover the biological meaning of the machine learning prioritized signatures, we carried out gene set enrichment analysis (GSEA) on the final optimized subsets of transcripts i.e. 80 sense transcripts and 70 NATs that achieved the best balance between predictive accuracy and interpretability. GSEA included Gene Ontology (GO) term prediction and pathway enrichment using the PlasmoDB resource [26], which provided a user-friendly interface for exploring parasite biology. Analyses were performed with a significance threshold of p = 0.05, and pathway information was obtained from KEGG [36] and MetaCyc [37]. Focusing enrichment only on the optimized transcript sets avoided the noise and overfitting associated with very large or arbitrary gene lists, ensuring that the functional interpretations reflected the most robust biological signals. This strategy allowed us to directly link model-selected features to meaningful pathways relevant to regulatory network in the mitochondria. Results Our previous studies have also documented P. vivax infections presenting with severe complications such as cerebral malaria, acute respiratory distress syndrome, jaundice, acute renal failure, etc., in adult patients [12]. Building on these findings, the current study focuses on 14 adult cases of P. vivax malaria, encompassing both complicated and uncomplicated manifestations, classified in accordance with WHO guidelines [32]. Detailed clinical characteristics of these patients are provided in Table S1. Mitochondrial associated genes A curated list of 636 mitochondrial-associated genes was compiled, encompassing 19 genes encoded by the mitochondrial genome and 617 nuclear-encoded genes predicted to target the mitochondria. These predictions were derived using a suite of tools including TargetP [22], Cello2GO [23], MitoFate [24], MitoprotII [25], PlasmoDB [26] and orthologous genes of the Plasmodium falciparum mitochondrial proteome [27] (Supplementary Data-2). To ensure the robustness of the gene selection, a stringent criterion was applied with a cutoff score > 0.6 for each tool. The comprehensive list provided in (Supplementary Data-2) comprises computationally predicted nuclear-encoded mitochondrial-targeted transcripts along with core mitochondrial transcripts, forming the basis for the transcriptomic analyses in this study. Differential gene expression analysis in clinical isolates To assess differential expression between HD and UNC groups, we applied a fold-change threshold of ≥ 1.5 (corresponding to log₂ fold change ≥ + 0.6 for upregulation or ≤ − 0.6 for downregulation) and a significance cutoff of p ≤ 0.05, with multiple-testing correction using the Benjamini–Hochberg false discovery rate method. Of the 636 mitochondrial transcripts curated, 626 sense transcripts and 581 NAT transcripts were retained for analysis (Supplementary Data-3). Among these, 92 sense and 77 NATs were differentially expressed respectively (Supplementary Data-3). Of the 92 differentially expressed sense transcripts, 66 (71.7%) were upregulated and 26 (28.3%) were downregulated. In contrast, among the 77 differentially expressed NATs, 27 (35.1%) were upregulated, whereas 50 (64.9%) were downregulated (Fig. 1). Differential expression distribution of mitochondrial sense and NATs transcripts in complicated P. vivax malaria Evaluating machine learning sensitivity across varying thresholds of significant gene expression To explore the predictive potential of mitochondrial transcripts, our initial analysis focused on evaluating features derived from statistically significant differentially expressed genes (DEGs) in both sense and NATs. These features were assessed for their utility in machine learning aimed at identifying key markers or genes in mitochondrial regulation. From the differential expression analysis, 92 sense and 77 NATs were statistically ranked by adjusted p-values (Supplementary Data-4), and subsequently prioritized for downstream predictive modelling and pathway enrichment. To gauge predictive strength, we systematically evaluated subsets of these DEGs ranging from the top 92 to 20 for sense transcripts, and from 77 to 20 for NATs (Supplementary Data-4), and then used three different machine learning algorithms. Remarkably, performance metrics including F1, MCC, sensitivity, and specificity consistently approached or exceeded 90% for subsets ranging from the top 80 to 20 in sense, and 70 to 20 in NATs. These findings underscore the robustness of these gene sets in predictive modelling and highlight their potential biological relevance. To further assess the stability of these models, we computed standard deviations (averaged across all tested gene sets (92 to 20 for sense transcripts and 77 to 20 for NATs) across LOOCV iterations and repeated the process with random shuffling both for sense and NATs (Tables 1, 2). Based on the comparative evaluation, the top 80 genes in sense and top 70 in NATs offered the most reliable prediction performance across all algorithms. Given its consistently high performance and stability across feature subsets, random forest was selected as the primary model for downstream interpretability analyses, including SHAP-based feature attribution and for further gene set enrichment analysis to enable a deeper, systems-level understanding of the mitochondrial regulatory landscape potentially governed by these transcripts. Table 1Machine-learning model performance across progressively reduced subsets of the 92 differentially expressed sense transcriptsPerformance metrics (accuracy, precision, sensitivity, specificity, F1-score, MCC, and AUC) for Random Forest, Logistic Regression, and XGBoost models trained using systematically reduced gene subsets derived from the 92 differentially expressed sense transcripts statistically prioritized by adjusted p-values Machine-learning model performance across progressively reduced subsets of the 92 differentially expressed sense transcripts Performance metrics (accuracy, precision, sensitivity, specificity, F1-score, MCC, and AUC) for Random Forest, Logistic Regression, and XGBoost models trained using systematically reduced gene subsets derived from the 92 differentially expressed sense transcripts statistically prioritized by adjusted p-values Table 2Machine-learning model performance across progressively reduced subsets of the 77 differentially expressed NATs transcriptsPerformance metrics (accuracy, precision, sensitivity, specificity, F1-score, MCC, and AUC) for Random Forest, Logistic Regression, and XGBoost models trained using systematically reduced gene subsets derived from the 77 differentially expressed sense transcripts statistically prioritized by adjusted p-values Machine-learning model performance across progressively reduced subsets of the 77 differentially expressed NATs transcripts Performance metrics (accuracy, precision, sensitivity, specificity, F1-score, MCC, and AUC) for Random Forest, Logistic Regression, and XGBoost models trained using systematically reduced gene subsets derived from the 77 differentially expressed sense transcripts statistically prioritized by adjusted p-values To establish a robust baseline, 40 genes were randomly selected from the full set of mitochondrial sense and NATs transcripts and used as machine learning features (Supplementary Data-4). This control experiment was designed to validate the predictive strength of statistically derived features. The outcome clearly demonstrated that without prior feature selection, the machine learning models failed to yield reliable predictive performance highlighting the critical role of informed, data-driven feature selection in enhancing model accuracy and biological interpretability. Comparison of the 15 K P. vivax microarray data for assessing the performance of machine-learning models A brief comparison was performed using the expression dataset available for the 8 × 15 K P. vivax array [3], covering both sense and NAT transcripts. This array was hybridized with samples from 14 patients: 5 with uncomplicated vivax malaria and 9 with complicated vivax malaria. It contains multiple probes for both sense and NATs, representing 4180 sense genes and 2260 NATs in P. vivax Sal-I [3]. To identify differential gene expression patterns between complicated and uncomplicated infections, we applied a fold-change cutoff of ≥ 1.5 and a significance threshold of p ≤ 0.05, with Benjamini–Hochberg correction for multiple testing. Of the 294 sense and 186 mitochondrial NATs transcripts analysed, 71 sense transcripts and 47 NATs met the criteria for differential expression. The differentially expressed genes in sense and NATs were statistically ranked by adjusted p-values and subsequently prioritized for downstream predictive modelling (Supplementary Data–5). We then systematically evaluated subsets of these DEGs from the top 71 to 20 for sense transcripts and from 47 to 20 for NATs (Supplementary Data–5) using three different machine learning algorithms. Notably, performance metrics such as MCC, sensitivity, and specificity remained consistently high across these gene subsets, indicating the robustness of the selected features for predictive modelling. To further examine model stability, we calculated standard deviations across LOOCV iterations and repeated the analyses with randomly shuffled labels for both sense and NATs (Tables 3 and 4). Based on this comparative assessment, the top 60 sense genes and top 30 NATs provided the most reliable classification performance across all algorithms. Table 3Comparison of machine-learning model performance on the 15 K P. vivax microarray dataset for senseThis table summarizes the performance of Random Forest, Logistic Regression, and XGBoost models evaluated across systematically reduced gene subsets derived from the 71 differentially expressed sense transcripts. Genes were ranked and prioritized based on adjusted p-values, and model performance was assessed Comparison of machine-learning model performance on the 15 K P. vivax microarray dataset for sense This table summarizes the performance of Random Forest, Logistic Regression, and XGBoost models evaluated across systematically reduced gene subsets derived from the 71 differentially expressed sense transcripts. Genes were ranked and prioritized based on adjusted p-values, and model performance was assessed Table 4Comparison of machine-learning model performance on the 15 K P. vivax microarray dataset for NATsThis table summarizes the performance of Random Forest, Logistic Regression, and XGBoost models evaluated across systematically reduced gene subsets derived from the 47 differentially expressed NATs transcripts. Genes were ranked and prioritized based on adjusted p-values, and model performance was assessed Comparison of machine-learning model performance on the 15 K P. vivax microarray dataset for NATs This table summarizes the performance of Random Forest, Logistic Regression, and XGBoost models evaluated across systematically reduced gene subsets derived from the 47 differentially expressed NATs transcripts. Genes were ranked and prioritized based on adjusted p-values, and model performance was assessed To establish a reliable baseline, we also selected 40 genes at random from mitochondrial sense and NAT transcripts and used them as machine-learning features (Supplementary Data–5). As expected, models built on randomly chosen genes failed to achieve stable or meaningful predictive performance, underscoring the importance of statistically driven feature selection for improving model accuracy. Evaluation of predictive performance by permutation testing Across all tested models and gene sets, the permutation testing consistently yielded low p-values (< 0.05) for accuracy, F1 score, MCC, and AUC, indicating that the observed performance was highly unlikely to arise by chance (Table 5, 6, Table S2-S5). For the sense gene datasets, the top-performing 80 gene set with Random Forest attained an observed MCC score of 0.866, with the corresponding null distribution producing a markedly lower mean score (p = 0.01). Likewise, for the NATs 70 gene set, Random Forest achieved an observed MCC score of 0.866, while the mean permuted F1 score was substantially lower, yielding an empirical p-value of 0.005. Similar results were obtained for Logistic Regression and XGBoost. The consistent statistical separation between observed and null distributions across multiple independent models (200, 500 and 1000 permutations) strengthens confidence that the classifiers are capturing biologically relevant signal rather than random structure. Table 5Permutation-based evaluation of predictive performance across reduced subsets of the 92 differentially expressed sense transcriptsThis table presents the results of permutation testing performed for all progressively reduced sense gene subsets evaluated using Random Forest, Logistic Regression, and XGBoost models. For each subset, empirical p-values were calculated across key performance metrics (accuracy, precision, sensitivity, specificity, F1-score, MCC, and AUC) using 200 iterations Permutation-based evaluation of predictive performance across reduced subsets of the 92 differentially expressed sense transcripts This table presents the results of permutation testing performed for all progressively reduced sense gene subsets evaluated using Random Forest, Logistic Regression, and XGBoost models. For each subset, empirical p-values were calculated across key performance metrics (accuracy, precision, sensitivity, specificity, F1-score, MCC, and AUC) using 200 iterations Table 6Permutation-based evaluation of predictive performance across reduced subsets of the 77 differentially expressed NATs transcriptsThis table presents the results of permutation testing performed for all progressively reduced NATs gene subsets evaluated using Random Forest, Logistic Regression, and XGBoost models. For each subset, empirical p-values were calculated across key performance metrics (accuracy, precision, sensitivity, specificity, F1-score, MCC, and AUC) using 200 iterations Permutation-based evaluation of predictive performance across reduced subsets of the 77 differentially expressed NATs transcripts This table presents the results of permutation testing performed for all progressively reduced NATs gene subsets evaluated using Random Forest, Logistic Regression, and XGBoost models. For each subset, empirical p-values were calculated across key performance metrics (accuracy, precision, sensitivity, specificity, F1-score, MCC, and AUC) using 200 iterations Gene ontology (GO) enrichment analysis of top-performing predictive gene set Gene Ontology (GO) analysis was employed to characterize the biological attributes of the genes. For sense transcripts, the top 80 genes were stratified into 57 upregulated and 23 downregulated genes, and GO enrichment was performed on each group independently to capture biological processes specifically associated with up and downregulation (Supplementary Data-6). The GO enrichment analysis revealed significant functional clustering exclusively within the upregulated gene set, while the 23 downregulated genes showed no statistically significant enrichment. The upregulated genes demonstrated a highly coordinated mitochondrial gene regulatory program centred on interconnected metabolic networks essential for P. vivax adaptation within host erythrocytes. The most prominent enrichments included amino acid catabolism (GO:0006546, GO:0009071), protein synthesis machinery (GO:0006412), and thiamine diphosphate metabolism (GO:0042357). The simultaneous activation of the glycine cleavage system, thiamine biosynthesis, and mitochondrial translation machinery establishes a metabolic interdependency network wherein glycine catabolism generates one-carbon units for nucleotide biosynthesis while requiring thiamine diphosphate as an essential cofactor for TCA cycle enzymes. Additionally, endocytic pathways, including clathrin-dependent and receptor-mediated endocytosis, were enriched, that could possibly suggest coordinated mechanisms for nutrient acquisition. In other systems, studies have shown that membrane-impermeant lipid probes can traffic directly from the plasma membrane to mitochondria, bypassing intermediate compartments [38]. In parallel, clathrin-mediated endocytosis has been linked to multiple cellular functions, including regulating surface protein expression, sensing environmental cues, facilitating nutrient uptake, controlling signalling pathway activation, retrieving proteins following vesicle fusion, and directing membrane components for lysosomal degradation [39].The predominance of upregulated endocytic components could therefore reflect an adaptive response in which enhanced endocytic uptake may support a metabolic focusing strategy, potentially channelling resources toward mitochondria. However, this proposed link between endocytic activity and mitochondria warrants further investigation and validation. In NATs, the top 70 predictive genes were classified into 23 upregulated and 47 downregulated genes (Supplementary Data-6). Enrichment analysis of upregulated transcripts revealed associations with mitochondrial support functions, including prolyl-tRNA aminoacylation (GO:0006433), glycine decarboxylation via the glycine cleavage system (GO:0019464), and thiamine diphosphate metabolic and biosynthetic processes (GO:0042357 and GO:0009229). These findings suggest that such NATs may be related to the regulation of organellar translation and mitochondrial cofactor networks, but the directionality or functional impact cannot be directly inferred. Additional enrichment in translational termination (GO:0006415) and complex disassembly (GO:0043624 and GO:0032984) points to possible roles in mitochondrial ribonucleoprotein turnover. For downregulated NATs, enrichment was observed for biological processes connected to mitochondrial function, RNA processing, and protein regulation. Processes such as heme metabolic and biosynthetic pathways (GO:0046160 and GO:0006784), respiratory chain complex IV assembly (GO:0008535), dolichol and polyprenol biosynthesis/metabolism (GO:0019408, GO:0016094 and GO:0016539), mRNA 3'-end processing, polyadenylation, and cleavage (GO:0031124, GO:0006378 and GO:0006379) and translation (GO:0006412) were associated with the set of downregulated NATs. While these patterns provide insight into potential involvement of NATs in mitochondrial gene regulation and cellular resource management, upregulation or downregulation of a NAT may not uniformly equate with activation or suppression of the corresponding biological pathway; rather, these results indicate complex regulatory relationships. Pathway enrichment analysis of top-performing predictive gene set In addition to the Gene Ontology, the pathway analysis was also performed to assess the functional relevance, the top-performing genes were systematically mapped to KEGG [36] and MetaCyc [37] pathways to identify activated and suppressed processes across classes. Pathway analysis of up and downregulated genes in both sense and NATs was performed using PlasmoDB [26]. For sense transcripts, 57 upregulated and 23 downregulated genes were mapped to 6 MetaCyc and 1 KEGG pathway while for NATs, 23 upregulated and 47 downregulated genes were mapped to 15 MetaCyc and 5 KEGG pathways (Supplementary Data-7). Among the upregulated sense transcripts, significant enrichment was observed in thiamine-tRNA regulatory network essential for mitochondrial bioenergetics. The most significant enrichment was thiamine diphosphate salvage III accompanied by multiple thiamine biosynthetic pathways and tRNA splicing. This coordinated upregulation establishes dual thiamine acquisition strategies (salvage and de novo biosynthesis) that ensure continuous supply of thiamine diphosphate (ThDP) cofactor for critical mitochondrial dehydrogenases in the TCA cycle. In parallel, enrichment of the tRNA splicing pathway points to elevated processing and maturation of mitochondrial and cytosolic tRNAs, reinforcing translational capacity and mitochondrial gene expression. Together, these changes, highlight an upregulation of cofactor metabolism and translational machinery, indicative of heightened mitochondrial activity. In contrast, among the downregulated sense transcripts, the only significantly enriched pathway was aminobenzoate degradation. Aminobenzoate is a precursor in folate biosynthesis, which contributes one-carbon units critical for nucleotide synthesis and mitochondrial redox balance. Its suppression suggests a reduced flux toward folate turnover and one-carbon metabolism, potentially restricting cofactor recycling and limiting mitochondrial biosynthetic capacity. This downregulation may reflect an adaptive strategy of the parasite to selectively repress certain mitochondrial metabolic routes while prioritizing cofactor synthesis and translational processes. In the case of NATs, pathway enrichment analysis showed that upregulated NATs were associated with thiamine diphosphate (TPP) metabolism, including both salvage and biosynthetic pathways. As TPP acts as a cofactor for mitochondrial enzymes involved in the TCA cycle, this association may reflect a potential for modulating cofactor availability to support energy metabolism; however, the impact of NAT upregulation on these pathways requires further investigation. Enrichment in L-cysteine biosynthesis and xenobiotic degradation pathways suggests possible links to redox buffering and metabolic adaptability. Overall, these associations indicate that upregulated NATs may contribute to mitochondrial function by supporting cofactor supply, redox stability, and nutrient utilization, but their precise regulatory effects on gene expression and pathway activity remain to be established. For downregulated NATs, pathway analysis highlighted associations with mitochondrial energy metabolism, including enzymes implicated in pyruvate decarboxylation to acetyl-CoA, valine/leucine/isoleucine degradation, and propanoate metabolism. This pattern may suggest reduced anaplerotic input to the TCA cycle and diminished flux through central carbon metabolism, as indicated by enrichment for citrate cycle and pyruvate metabolism pathways among suppressed genes. The parallel association with glycolysis/gluconeogenesis points toward a more global restriction of energy-yielding pathways. Additionally, connections to reduced dolichol and polyisoprenoid biosynthesis and aminoacyl-tRNA biosynthesis imply possible impacts on protein glycosylation and translation, processes that are closely coupled to mitochondrial activity. Importantly, these findings are based on transcript-level associations, and do not directly demonstrate suppression or activation of metabolic functions, highlighting the need for further functional validation. Together, while sense transcripts mapped largely to thiamine diphosphate salvage and amino acid metabolism, NATs displayed a more profound regulatory footprint by modulating both anabolic (cofactor biosynthesis, cysteine metabolism) and catabolic (pyruvate decarboxylation, TCA cycle, glycolysis) pathways. This dual influence underscores the pivotal role of NATs in fine-tuning a regulatory influence on mitochondrial function in Plasmodium vivax, balancing energy generation and metabolic flexibility that may critically determine parasite survival and pathogenicity. Feature importance analysis To identify the key contributors among the top performing genes, SHAP analysis was applied to the best-fitting Random Forest model to interpret feature importance. The top-ranking genes in sense and NATs transcripts (Fig. 2A–B) demonstrated strong predictive relevance, from the overall 80 genes selected for sense and 70 genes for NATs, providing the most reliable prediction performance across all algorithms tested. A. SHAP-based feature importance for the top predictive mitochondrial sense transcripts. SHAP analysis was applied to the best-performing Random Forest model to interpret feature importance across the top 80 selected sense transcripts, which provided the best predictive performance and model interpretability. B. SHAP-based feature importance for the top predictive mitochondrial NATs. SHAP analysis was applied to the best-performing Random Forest model to interpret feature importance across the top 70 selected NATs transcripts, which provided the best predictive performance and model interpretability Genes with SHAP values were considered to have a greater influence on the model’s predictions and are potential candidates for involvement in mitochondrial regulatory processes. Notably, genes prioritized by feature importance overlapped with enriched pathways related to polyisoprenoid pathway, amino acid catabolism, mitochondrial translation, and central carbon metabolism, thereby linking computational prioritization with biologically meaningful processes. To further investigate the directionality of feature contributions, SHAP scatter density plots were generated separately for sense and NATs (Fig. 3A–B). In these plots, each dot corresponds to a sample, with the x-axis representing the SHAP value (impact on model prediction) and the color indicating relative expression (red = high expression, blue = low expression). A. SHAP scatter density plot for the top predictive mitochondrial sense transcripts. Scatter density SHAP analysis of the top 80 selected sense transcripts identified by the best-performing Random Forest model. Each dot represents a sample, with dot colour indicating the feature value (red = high expression, blue = low expression), and the x-axis representing the SHAP value, reflecting each gene’s impact on the prediction outcome. B. SHAP scatter density plot for the top predictive mitochondrial NATs transcripts. Scatter density SHAP analysis of the top 70 selected sense transcripts identified by the best-performing Random Forest model. Each dot represents a sample, with dot colour indicating the feature value (red = high expression, blue = low expression), and the x-axis representing the SHAP value, reflecting each gene’s impact on the prediction outcome For sense, transcripts, the feature importance analysis identified a subset of mitochondrial-associated genes as prominent model predictors (Table 7) [14, 19, 40–46], each representing a distinct aspect of mitochondrial physiology critical for Plasmodium. The emergence of genes involved in oxidative phosphorylation, energy metabolism, and mitochondrial ribosomal function among top features underscores the parasite's reliance on robust mitochondrial activity for bioenergetic maintenance, protein synthesis, and metabolic adaptation. This could potentially suggest that variations in these mitochondrial processes may be associated with changes in parasite growth and viability in clinical setting. Table 7SHAP-based feature importance analysis identified for subsets of mitochondrial-associated transcripts in sense SHAP-based feature importance analysis identified for subsets of mitochondrial-associated transcripts in sense On the other hand, NATs were also represented among top-ranked features (Table 8) [19, 44, 47–53], with their association pointing towards nuanced regulation of mitochondrial activity particularly in ATP synthesis, cofactor assembly, membrane transport, and proteostasis. The marked presence of NATs in these key regulatory roles suggests an additional layer of control, enabling the parasite to finely adjust mitochondrial function to diverse physiological demands. These findings suggest that NATs may represent an advanced layer of gene regulation within Plasmodium mitochondria, with the potential to become particularly significant in physiologically demanding or severe conditions. Table 8SHAP-based feature importance analysis identified for subsets of mitochondrial-associated transcripts in NATs SHAP-based feature importance analysis identified for subsets of mitochondrial-associated transcripts in NATs Collectively, these scatter density plots complement the feature ranking by providing critical insights into the directionality of gene contributions, revealing that transcript abundance in both sense and NATs, with NATs demonstrating a stronger regulatory footprint. Importantly, SHAP helped us to distinguish which genes were truly influential in driving the model’s predictions, rather than being statistical noise. In this case, SHAP facilitated the prioritisation of candidate genes that may play regulatory roles in mitochondrial pathways, especially among NATs, thereby generating biologically meaningful insights that could guide future functional validation and mechanistic studies. Discussion The study delivers an integrative framework combining transcriptomic profiling and machine learning to unravel mitochondrial gene regulation in severe P. vivax malaria, with a particular emphasis on the biological significance of NATs. These findings present a paradigm-shifting view into how mitochondrial dynamics are orchestrated in the parasite and highlight the critical, multifaceted roles of both sense and NATs transcripts. Our study incorporated transcriptomic data to identify key genes of interest, which were subsequently employed as training sets for model verification. The underlying rationale was that if the model achieved robust predictive accuracy based on these features, it would not only provide predictive power but also strengthen the biological plausibility of the findings. In practice, ML facilitated dimensionality reduction by refining a broad set of statistically significant transcripts into a smaller, reproducible subset of genes. Thus, ML served a major role for enhancing predictive validation thereby reducing potential false positives and uncovering mechanistic leads for future investigation. This integrative framework provides a more interpretable approach for extracting biologically meaningful insights from complex transcriptomic datasets. The predominance of upregulated sense mitochondrial transcripts in complicated malaria reflects a possible environment driven modulation of mitochondrial metabolic programs under host-imposed stress. Enrichment of pathways mainly in the sense transcripts were amino acid catabolism, mitochondrial translation machinery, and thiamine diphosphate metabolism strongly suggests a strategic metabolic focusing wherein P. vivax optimizes energy production and biosynthetic processes critical for survival [14, 54, 55]. The glycine cleavage system and thiamine biosynthesis form an interdependent metabolic network providing crucial one-carbon units and cofactors necessary for mitochondrial enzymatic reactions, highlighting a tightly coordinated metabolic adaptation [54, 55]. Such reprogramming likely supports increased mitochondrial bioenergetic demands during the erythrocytic stage, consistent with reported roles of the Plasmodium mitochondrion in maintaining parasite viability [13, 25]. In addition to this, the downregulation of sense transcripts hints towards a nuanced reallocation of mitochondrial resources favouring energy and cofactor production over certain anabolic pathways, likely as a cost-saving measure under infection-induced physiological stress. This aligns with the concept of metabolic focusing where parasites selectively activate essential pathways while suppressing non-critical routes to optimize survival against host immune responses [56], or severe disease pressure. The landscape of NATs indicates a nuanced regulatory role. Upregulated NATs were associated with enrichment in biological processes relevant to mitochondrial translational capacity and cofactor biosynthesis, such as prolyl-tRNA aminoacylation, thiamine diphosphate metabolism, glycine decarboxylation, and translational termination [23, 24, 37]. These associations may reflect the potential for NATs to modulate cofactor supply necessary for mitochondrial protein translation and broader metabolic activity [14]. Conversely, downregulated NATs appeared linked to pathways, including pyruvate metabolism, TCA cycle, glycolysis, and respiratory chain complex IV assembly, suggesting possible involvement in the fine-tuning of energy-generating processes. This pattern may represent a metabolic trade-off aimed at limiting mitochondrial activity and conserving resources under specific physiological conditions [41, 57], though the functional impact of NAT downregulation requires further clarification. Additionally, reduced NATs expression in dolichol and polyisoprenoid biosynthesis may be related to changes in protein glycosylation and mitochondrial protein import, processes crucial for organelle integrity [25]. The co-occurrence of downregulation in RNA processing and polyadenylation pathways points to additional layers of translational control. Overall, these observations suggest that NATs may act as dynamic regulators responsive to physiological states [11, 58], though functional validation will be needed to unravel their precise effects. Interestingly, analysis of concordant expression of the NATs and sense transcripts revealed striking findings. Notably, PVX_099630 (thiamine pyrophosphokinase, putative), a pivotal enzyme in the thiamine diphosphate salvage III pathway, exhibited simultaneous upregulation of both sense transcripts and NATs, underscoring a potential regulatory coordination. This could mechanistically indicate that the NATs act as molecular buffers that prevent protein levels from rising proportionally to mRNA levels. The coordinate upregulation during severe malaria suggests this as a possible stress adaptation mechanism. The parasite enhances metabolic capacity for survival while preventing potentially lethal metabolic imbalances could compromise cellular integrity. A similar phenomenon is observed in case of PVX_092250 (glycine cleavage system H protein, putative), which has shown significant association with GO terms suggesting different catabolic processes (Glycine catabolic process, serine family amino acid catabolic process) amongst others, where both the sense transcripts and NATs are upregulated in severe malaria. The glycine cleavage system and amino acid catabolic pathways identified are metabolically powerful. Uncontrolled activity could lead to excessive NADH production disrupting cellular redox balance or amino acid depletion respectively. This has the potential of affecting other cellular processes [59]. In both the above-mentioned cases, the NATs could be providing a mechanism which would prevent cellular damage from uncontrolled enzyme production during severe malaria and could be part of a sophisticated survival strategy. The use of LOOCV ensured that each sample was evaluated exactly once, thus preventing performance estimates from depending on a single arbitrary split. Nevertheless, given the limited dataset size, variance across folds remained high, as reflected in the observed standard deviations, an expected limitation of small clinical cohorts. This design helps detect overfitting to specific individuals by repeatedly testing generalization across all samples. While LOOCV reduces bias in small datasets, it does not resolve high variance. To strengthen confidence in the predictive performance, we complemented LOOCV with rigorous permutation testing, which established that the observed accuracy was significantly greater than chance expectations, addressing concerns about overfitting in small clinical cohorts [32]. This addresses the concern that randomly selected genes or latent co-expression modules might drive spurious results. Importantly, the low p-values across most metrics and gene sets confirm that the predictive performance reported here reflects true signal in the gene expression data rather than overfitting to noise. Despite these constraints, the use of machine learning classifiers with LOOCV and SHAP interpretation delivers validation of transcriptomic signatures, confirming the predictive power and biological significance of mitochondrial sense and NATs transcripts [31, 32, 35]. The consistently high F1 scores, MCC, and AUC metrics highlight the reproducibility and reliability of identified features. Notably, SHAP values emphasize NAT-associated transcripts as having stronger and more stable contributions, underscoring their integral role in mitochondrial gene regulation [60]. In addition, the comparative analysis of 8 × 15 K array demonstrated consistent trends across the gene subsets which was also observed in our study. When these differentially expressed genes were assessed using machine learning, all three algorithms showed stable performance, with random forest producing the most robust and consistent predictive results. The reliable performance across multiple feature subsets, together with the poor predictive behaviour of randomly selected genes, highlights the relevance of these statistically prioritized transcripts. Collectively, these findings indicate that reproducibility across datasets further reinforces the strength of our modelling approach. This study elucidates NATs as master regulators of mitochondrial function in P. vivax, capable of orchestrating both activation and suppression of critical metabolic and translational pathways. By modulating cofactor biosynthesis, translation, and energy metabolism, NATs afford the parasite metabolic plasticity to adapt to the hostile host environment during severe malaria [61, 62]. This refined regulatory control extends beyond mitochondrial gene expression, potentially affecting pathways critical for parasite development and drug resistance [20]. The identification of NATs as key modulators of mitochondrial regulators of both metabolic pathways and translational control suggests sophisticated regulatory networks that could be utilized for drug development. The robust predictive performance of our machine learning models alongside transcriptomic data demonstrates mitochondrial NATs as potential molecular players in clinical applications. Despite the robustness of our integrative approach, certain limitations warrant acknowledgment, first, the clinical cohort size was modest, reflecting the inherent challenges of obtaining severe P. vivax samples. Although the use of LOOCV maximized data utilization and reduced bias relative to a single train test split, it cannot fully eliminate the risk of overfitting or guarantee generalizability, given biological variability across patients. In small datasets where the number of features vastly exceeds the number of samples, model performance estimates may exhibit high variance even when average metrics appear robust. However, studies have also documented that train/test split approaches produce robust and unbiased performance estimates regardless of sample size [63]. In addition, multiple microarray probes for each NAT were designed with a 3′ bias, meaning the 5′ regions may not always be captured, and complete NAT sequences may may not be always detectable. While RNA-Seq offers greater potential for resolving full-length NATs, this still depends on transcript integrity and the proportion of full-length molecules present in the pool. Accordingly, while the SHAP rankings and predictive accuracies observed here are compelling, they should be regarded as preliminary rather than definitive clinical predictors. The positioning of the Plasmodium mitochondrion as a potential target is supported by our findings, yet validation in larger and geographically diverse patient cohorts is essential to confirm these regulatory signatures. Importantly, the unique mitochondrial architecture and NATs mediated regulatory mechanisms uncovered in this study highlight multiple potential avenues ranging from targeting parasite-specific metabolic enzymes to disrupting NATs driven regulation. Future directions should integrate multi-tool mitochondrial annotation, interpretable machine learning models in expanded cohorts, and experimental validation of mitochondrial targeted transcripts, and RT-qPCR validation of key NATs will be performed as samples become available in future work. Such approaches will not only refine predictive gene signatures and strengthen mechanistic inference but also accelerate the translation of transcriptomic insights into actionable targets for combating malaria. Supplementary Information Additional file 1. Additional file 2. Additional file 3. Additional file 4. Additional file 5. Additional file 6. Additional file 7. Acknowledgements We thank all the patients and technical workers for participating in and supporting this project. We thank Birla Institute of Technology & Science, Pilani (BITS Pilani), Pilani campus, Pilani, Rajasthan, India and SP Medical College, Bikaner, Rajasthan, India for providing the necessary facilities required for this work. We also thank Genotypic Technology Pvt. Ltd., Bangalore, India, for the microarray hybridization and consultation regarding data analysis. We further acknowledge Dr. Gitanjali Yadav (National Institute of Plant Genome Research, New Delhi, India) for kindly reading the manuscript and providing valuable inputs. Author contributions AD conceived the study and networked to facilitate the outcome. PR was involved in performing the experimentation, analysis, and data interpretation, designing the figures, and drafting the manuscript under the constant guidance of AD. YA was involved in designing, training, and evaluating the machine learning models used in this study. DKK and SKK supervised sample collection and clinical characterization. All authors read and approved the manuscript. Funding This work is funded by the Indian Council of Medical Research (ICMR) under the Extramural Ad-hoc Scheme (PID: 2019-1121, a multicentric project with AD as PI). Availability of data and materials All microarray results have been deposited in the Gene Expression Omnibus database (GSE 300144). All the datasets underlying in this study will be shared on reasonable request to the corresponding author. All data generated are provided in the main text and supplementary data files. All patient meta-data included in this article are included in Table S1. The codes are https://github.com/MPSB-Lab/ML-Microarray Reprint requests should be addressed to AD. All Venn diagram illustrations have been made using Venny [64]. Heatmaps were generated using the Seaborn and Matplotlib libraries, while SHAP plots were created using the SHAP and Matplotlib libraries in python v3.11.13. Declarations Ethics approval and consent to participate Informed consent was obtained from all individual participants in the study according to the SP Medical College Hospital's guidelines (No. F (Acad)SPMC/2003/2395). Permission to use these samples for further studies was given through IERB approval No. F29(Acad)SPMC/2020/3151 dated 05.09.2020. Informed consent was obtained from all individual participants included in the study. We sincerely thank all the patients for their voluntary informed consent and participation. Competing interests The authors declare no competing interests. Footnotes Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Additional file 1. Additional file 2. Additional file 3. Additional file 4. Additional file 5. Additional file 6. Additional file 7. Data Availability Statement All microarray results have been deposited in the Gene Expression Omnibus database (GSE 300144). All the datasets underlying in this study will be shared on reasonable request to the corresponding author. All data generated are provided in the main text and supplementary data files. All patient meta-data included in this article are included in Table S1. The codes are https://github.com/MPSB-Lab/ML-Microarray Reprint requests should be addressed to AD. All Venn diagram illustrations have been made using Venny [64]. Heatmaps were generated using the Seaborn and Matplotlib libraries, while SHAP plots were created using the SHAP and Matplotlib libraries in python v3.11.13.