Glioma identification from microRNA biomarkers using machine learning

Article type: Research Article

Keywords: brain cancer, glioma, machine learning, meningioma, MicroRNA biomarkers

Authors: Rakesh Kanth Andugala, Alyson Cieslik, Maria Braoudaki, Iosif Mporas

Affiliations: School of Physics, Engineering and Computer Science, University of Hertfordshire, Hatfield, United Kingdom; School of Health, Medicine and Life Sciences, University of Hertfordshire, Hatfield, United Kingdom

License: Copyright © 2026 Andugala, Cieslik, Braoudaki and Mporas. CC BY 4.0 This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

Article links: DOI: 10.3389/fsysb.2026.1771910  |  PubMed: 42344946  |  PMC: PMC13286781

Relevance: Moderate: mentioned 3+ times in text

Full text: PDF (4.6 MB)

Introduction

Gliomas are a heterogeneous group of primary brain tumours arising from glial cells and are classified by the World Health Organization (WHO) according to their histopathological and molecular characteristics. (ref. Louis et al., 2021). These tumours present symptoms, including headaches, seizures, and cognitive deficits, driven by their infiltrative growth (ref. IJzerman-Korevaar et al., 2018). Despite aggressive treatments, such as surgical resection, radiotherapy, and chemotherapy, patient prognosis remains poor, underscoring the need for early and accurate diagnosis. Conventional diagnostic methods, including magnetic resonance imaging (MRI) and histopathological biopsy, face limitations (ref. Li K. et al., 2024). MRI lacks specificity to differentiate gliomas from other brain lesions, while biopsy, although definitive, is invasive, costly and carries procedural risks that can often delay therapeutic decision-making. Recent studies highlight microRNAs (miRNAs), small non-coding RNAs that regulate gene expression, as promising minimally invasive biomarkers for glioma diagnostics (ref. Zhou et al., 2018). MiRNAs exhibit tumour-specific expression profiles in biofluids (ref. Anelli et al., 2021), such as blood (serum or plasma), saliva, urine, and cerebrospinal fluid, offering high sensitivity and specificity in distinguishing gliomas from other brain lesions or other neurological conditions. However, developing reliable computational models using miRNA expression data is challenging due to their high dimensionality, the typically small sample sizes in available training datasets, class imbalances, and variability in data collection and processing (ref. Muhamed Ali et al., 2018). To address these issues, this article presents a machine learning pipeline to identify miRNA biomarkers for glioma classification, using online miRNA expression datasets from Gene Expression Omnibus (GEO) (ref. Edgar et al., 2002). Our pipeline employs feature selection and class balancing with Synthetic Minority Over-sampling Technique (SMOTE) to identify miRNA biomarkers by employing both binary (glioma vs controls) and multi-class (glioma vs meningioma vs controls) classification, using several machine learning (ML) algorithms. The miRNAs found to be top ranked in glioma identification were then analysed and compared with other biomarkers known from the literature.

Related works

miRNAs in glioma classification

Previous research on glioma identification from molecular data and previous studies on cancer classification using miRNAs. Specifically, regarding glioma classification using omics data, (ref. Si et al., 2022), glioma was distinguished from Primary Central Nervous System Lymphoma (PCNSL) by analysing serum miRNA expression profiles from blood serum using Least Absolute Shrinkage and Selection Operator (LASSO) and Support Vector Machine – Recursive Feature Elimination (SVM-RFE), identifying ten diagnostic miRNAs. In (ref. Li et al., 2023), glioma was distinguished from controls based on the within-sample relative expression orderings of miRNAs, reporting 100% accuracy. In (ref. Slimene et al., 2022), glioma was classified among four other cancer types (Breast invasive carcinoma (BRCA), Stomach and Esophageal carcinoma (STES), Pan-kidney cohort (KIPAN), and Ovarian Serous Cystadenocarcinoma (OV)) by analysing miRNA expression profiles (Reads Per Million (RPM) normalised) from tissue samples, using DeepInsight (ref. Sharma et al., 2019) for feature extraction and ResNet50 (ref. Koonce and Koonce, 2021) for classification, achieving 92.9% prediction accuracy. In (ref. Ni et al., 2024), low-grade glioma (LGG) was classified from five other cancer types (including lung adenocarcinoma (LUAD), kidney pan-cancer (KIPAN), thyroid carcinoma (THCA), sarcoma (SARC), and esophageal carcinoma (ESCA)), based on miRNA expression profiles from human tissue samples, MiRS-HF deep learning framework with two graph convolutional networks (LAGCN and GCNCC) and feature weighting, achieving accuracy of 95.6% for KIPAN, 94.8% for ESCA, 77.9% for SARC, 72.8% for THCA, 63.8% for LGG and 63.2% for LUAD and identifying 20 miRNA biomarkers. In (ref. Kaczmarek et al., 2021), LGG was classified from 12 other cancer types (Bladder urothelial (BLCA); Breast invasive (BRCA); Cervical/endocervical (CESC); Head and neck squamous cell (HNSC); KIPAN; LGG; Liver hepatocellular (LIHC); LUAD; Lung squamous cell (LUSC); Prostate (PRAD); STES; THCA), by analysing patient-matched miRNA and mRNA expression data from The Cancer Genome Atlas (TCGA) (ref. Tomczak et al., 2015) tissue samples, using the Graph Transformer Network (GTN) to model miRNA-mRNA interactions, achieving 93.56% classification accuracy. In (ref. Wang T. et al., 2021), LGG (Grade2,3) by analysing miRNA, mRNA, and DNA methylation data from tissue samples, using the proposed Multi-Omics Graph convolutional NETworks (MOGONET) deep learning framework combining graph convolutional networks and View Correlation Discovery Network (VCDN), achieving ROSMAP dataset (of Alzheimer’s disease (AD) patients vs. normal control (NC)) (accuracy (ACC): 0.815, F1 score (F1): 0.821, area under the receiver operating characteristic curve (AUC): 0.874), LGG (ACC: 0.816, F1: 0.814, AUC: 0.840), BRCA (ACC: 0.829, F1 weighted: 0.825, F1 macro: 0.774), and KIPAN (matched adaptive group-regularized ridge regression (GRridge) (ref. Van De Wiel et al., 2016) in AUC, surpassed in ACC and F1. In (ref. Waspada et al., 2017), miRNA expression data (1881 features) from 22 cancer tissue types and normal samples were classified using supervised machine learning methods, decision tree, naïve Bayes, neural network (Artificial Neural Network), and deep learning (Deep Neural Network). Deep learning achieved 91.49% accuracy in the multi-class scenario. Binary classifications (breast or cervical cancer vs. normal) yielded up to 100% accuracy with deep learning and neural networks. In (ref. Pyman et al., 2018), glioma was distinguished from 26 other cancer types and non-neoplastic samples, using miRNA expression profiles from human tissues, employing stacked autoencoders and a multi-layer perceptron, achieving 94.73% for multi class classification accuracy. In (ref. Laplante and Akhloufi, 2020), glioma was classified from 20 anatomical sites (grouped from 27 cancer location) using miRNA stem-loop expression data from TCGA (ref. Tomczak et al., 2015) human tissue samples, employing a deep neural network and an overall F1 score of 96.88%, with 18 of 20 classes having an F1 score greater than 90%. In (ref. Noviandy et al., 2023), LGG and glioblastoma (GB were classified by analysing molecular components, genetic mutations, and clinical features from TCGA (ref. Tomczak et al., 2015) human tissue samples, using ensemble machine learning with Genetic Algorithm Logistic Regression (GA-LR) (ref. Hosseini, 2020), Boruta (ref. Kursa and Rudnicki, 2010), and CatBoost Feature Importance (CFI) (ref. Dhananjay and Sivaraman, 2021) feature selection, achieving 86% F1-score. In (ref. Lokeswaran et al., 2024), classification of five brain tumour types (medulloblastoma, glioma, rhabdoid tumour, primitive neuro-ectodermal tumour (PNET), and normal) was presented, using gene expression data from human tissue samples, with Principal Component Analysis (PCA) and Pearson correlation feature selection, achieving 90% multiclass classification accuracy with LR. In (ref. Oberoi et al., 2023), LGG, GB, and their subtypes (oligodendroglioma, astrocytoma, glioblastoma, and mixed glioma) were classified by analysing gene mutation and clinical data from human tissue samples, achieving 90.70% accuracy for grade classification using an ensemble of Logistic Regression (LR), Support Vector Machine (SVM), and K-Nearest Neighbors (KNN) with Mutual Information feature selection and 84.88% accuracy for subtype classification using a Bagging Classifier with Mutual Information feature selection. In (ref. Wang et al., 2023), classification of LGG into grade 2 and grade 3 was performed by analysing multimodal data (miRNA, mRNA, and DNA methylation) from human tissue samples, utilising a proposed deep learning framework, Dual Trustworthy Mechanism for Classification (DTMC), reporting 89.8% AUC.

miRNAs in other cancer types classification

Beyond brain cancer, miRNAs have also been utilized to distinguish other malignancies. For instance, in (ref. Kaczmarek et al., 2022) neoplastic from non-neoplastic breast and skin tissue samples were classified using miRNA expression profiles, while AUC of the Deep Cancer Classifier (DCC) (ref. Pyman et al., 2018) was 99.9% for breast cancer and 97.1% for skin cancer. In (ref. Pal and Rami, 2024), classification between normal and breast cancer was performed on miRNA expression patterns from human blood samples, achieving up to 100% accuracy with Artificial Neural Networks (ANN) and Minimum Redundancy Maximum Relevance (mRMR). In (ref. Slimene et al., 2021) breast invasive carcinoma was distinguished from control using miRNA RPM expression data from human tissue samples, with SVM achieving 0.99 accuracy, 1.0 sensitivity, and 0.99 specificity. In (ref. Mehejabin et al., 2020) breast cancer samples were classified from normal ones using miRNA expression data from human tissue samples, with SVM achieving 99.23% accuracy. In (ref. Li et al., 2022) a prognostic risk model classifying cervical cancer patients into high or low risk categories was constructed based on miRNA expression and clinical data from TCGA (ref. Tomczak et al., 2015) human tissue samples, using Random Forest (RF) for feature selection and multivariate Cox regression, achieving 0.94 AUC on 5-year survival prediction. In (ref. Kantamneni et al., 2024) subtypes of lung cancer (adenocarcinoma, squamous cell carcinom a, and Small Cell Lung Cancer (SCLC)) along with healthy controls were classified based on miRNA expression profiles from GEO (ref. Edgar et al., 2002), using a late fusion of FNN and DNN (Feedforward Neural Network and Deep Neural Network) model, achieving 90.7% accuracy. In (ref. Andreini et al., 2022) breast cancer classification from healthy controls was performed based on miRNA expression data from human tissue samples, using a two-stage SVM and Random Forest (RF) machine learning approach, achieving 0.9901 accuracy.

Methodology

The methodology followed for glioma identification is presented in this section, including a description of the datasets used and the processing steps of the employed architecture, which is illustrated in Figure 1.

Datasets and preprocessing

The miRNA expression data were acquired from four GEO (ref. Edgar et al., 2002) datasets: GSE112264 (ref. Urabe et al., 2019), GSE113486 (ref. Usuba et al., 2019), GSE113740 (ref. Yamamoto et al., 2020), and GSE139031 (ref. Ohno et al., 2019). All cohorts were derived from human serum samples and processed using the same microarray platform (Toray 3D-Gene Human miRNA V21, GPL21263), which ensured consistency across the integrated data. Therefore, explicit batch-effect correction methods (e.g., ComBat (ref. Johnson et al., 2007) or limma removeBatchEffect (ref. Ritchie et al., 2015)) were not applied to avoid potential over-correction, which could inadvertently remove biologically meaningful variation.

To further mitigate potential inter-study variability and batch effects, we evaluated the impact of different preprocessing strategies, specifically comparing Min-Max scaling and z-score standardization. We implemented z-score standardization as it provided better stability across datasets. Figure 2 illustrates the UMAP visualization before and after z-score normalisation.

This processed dataset supported two experimental configurations: a binary classification setup involving 552 samples (285 glioma and 267 non-cancer controls), and a multi-class setup (glioma vs. meningioma vs. non-cancer controls) comprising 569 samples (285 glioma, 17 meningioma, and 267 non-cancer controls). Each sample consists of 2,565 miRNA values, with the distribution across datasets and classes detailed in Table 1.

An independent external validation cohort, GSE211692 (ref. Matsuzaki et al., 2023), was additionally obtained from GEO. This dataset contains human serum miRNA expression profiles generated on the GPL21263 platform. After label mapping and alignment with the trained feature space, the external cohort included 5,884 samples in total, 241 glioma samples and 5,643 non-cancer control samples. The GSE211692 dataset was used only for independent validation testing.

Feature selection

Five feature ranking algorithms were employed to perform dimensionality reduction. These are (i) the Least Absolute Shrinkage and Selection Operator (LASSO) (ref. Tibshirani, 1996) based on L1 regularisation, (ii) the Minimum Redundancy Maximum Relevance (mRMR) (ref. Peng et al., 2005), (iii) the ReliefF (ref. Kononenko, 1994), (iv) the Recursive Feature Elimination (RFE) (ref. Guyon et al., 2002) and (v) the Random Forest (RF) (ref. Breiman, 2001). To prevent information leakage and overfitting, feature selection was performed independently within each iteration of the cross-validation process. Features were selected solely based on the training folds, ensuring that the test fold remained completely unseen during the dimensionality reduction phase.

Machine learning algorithms

Six machine learning algorithms (ML) were used to train glioma identification models. These are the Logistic Regression (LR) (ref. Hastie et al., 2009), the K-Nearest Neighbours (KNN) (ref. Cover and Hart, 1967), the Decision Tree (DT) (ref. Quinlan, 1986), the Random Forest (RF) (ref. Breiman, 2001), the Support Vector Machine (SVM) (ref. Cortes and Vapnik, 1995) with linear kernel, and the Extreme Gradient Boosting (XGB) (ref. Chen et al., 2015). Each model was trained on various miRNA feature subsets obtained after the feature selection process (tested dimensionalities: 5, 10, 20, 50, 100, 250, 500, 750, 850, 1,000, 1,100, 1,500, 2,000, 2,565), and hyperparameters were optimized using grid search following a 5-fold stratified cross-validation protocol.

Class imbalance handling

Class imbalances were addressed using the Synthetic Minority Oversampling Technique (SMOTE) (ref. Chawla et al., 2002). SMOTE was used to generate synthetic samples for the minority class only for the training subsets, while the test subsets remained unchanged with the original samples only.

However, it is important to note that while SMOTE addresses numerical imbalance, it generates synthetic samples by interpolating between neighbouring minority-class samples in feature space. For very small classes such as meningioma $(n=17)$, this does not introduce new biological variation and may smooth or blur genuine within-class heterogeneity (e.g., potential substructure or outliers), which can affect biological interpretability. Consequently, this scarcity remains a primary limitation that may impact the model’s ability to generalize complex biological features compared to more prevalent classes.

Performance evaluation

Five metrics of performance were estimated across all evaluated glioma identification models, namely, the classification Accuracy, the F1-score, the Precision, the Recall, and the Area Under the Receiver Operating Characteristic Curve (AUC).

Results

The methodology for glioma identification presented in Section 3 was evaluated using two experimental setups, binary classification (glioma vs. controls) and multi-class classification (glioma vs. meningioma vs. controls), with and without SMOTE oversampling.

Binary classification

The binary classification (glioma vs. controls) results without SMOTE data augmentation are tabulated in Table 2. The best performing feature selection (FS) algorithm and the corresponding number of selected miRNA features are shown in the second column and ‘*’ stands for all five feature ranking algorithms.

As can be seen in Table 2, all evaluated ML algorithms achieved their best performance when using the top 10 ranked miRNA features from LASSO, ReliefF, mRMR, RFE or RF importance. Among them, hsa-miR-663a was commonly selected across all five feature selection methods. Several others, such as hsa-miR-125a-3p (ReliefF, mRMR, RFE, RF importance), hsa-miR-6784-5p (LASSO, mRMR, RFE, RF importance), and hsa-miR-873-3p (LASSO, ReliefF, mRMR, RFE), were shared among the 4 FS methods, while the remaining miRNAs differed across the different evaluated algorithms. Compared with the discussion section (top 100 features), hsa-miR-663a and hsa-miR-6784-5p were identified in both the top 10 and top 100 feature analyses. Except DT, all other evaluated ML algorithms classified correctly all miRNA samples.

The binary classification results using SMOTE data augmentation are tabulated in Table 3. As can be seen in Table 3, applying SMOTE did not change the classification performance, as the results remain identical to those obtained without SMOTE. This was expected since in the binary classification setup there are 285 glioma samples and 267 control samples, i.e., the dataset is already relatively balanced.

Multi-class classification

The multi-class classification (glioma vs. meningioma vs. controls) results without SMOTE data augmentation are tabulated in Table 4. As can be seen in Table 4, the classification results without SMOTE demonstrate consistently high performance across all models, with KNN and XGB classifying all samples correctly. However, the limited meningioma sample size $(n=17)$ should be also considered in the interpretation of the results due to the potential risk of overfitting. KNN performed best using the 100 top ranked miRNAs as ranked by RF importance, while XGB performed best using the 500 top ranked miRNAs as ranked by LASSO. When meningioma samples are included in the multi-class setup, to maintain high classification accuracy 100 miRNA features are required compared to the 10 features in the case of binary classification setup.

The multiclass classification results using SMOTE data augmentation are tabulated in Table 5. As can be seen in Table 5, applying SMOTE improves classification performance across four out of six models. Compared to Table 4 (without SMOTE), where KNN achieved perfect accuracy with 100 miRNAs and XGB required 500 miRNAs, the use of SMOTE enables both KNN and XGB to reach perfect performance using 100 miRNAs. The top 100 miRNAs selected for KNN using RF feature selection were the same in both with and without SMOTE setups. Moreover, DT, RF, and SVM, which previously required 850–1,000 miRNAs, achieved absolute performance with only 100 features when using SMOTE. This demonstrates the positive effect of SMOTE in improving class balance in ML model training, resulting in reducing the number of required miRNAs to maximise classification performance. Crucially, while SMOTE balances the data numerically, it does not introduce new biological variation; thus, the scarcity of original samples remains a constraint for broader generalization.

For clinical translation, we also considered the trade-off between predictive performance and assay practicality. In our experiments, a panel size of 20 miRNAs provided a good balance between performance and feasibility for qPCR/ddPCR-based workflows, achieving an F1-Score of 0.98, while reducing the panel to 5 miRNAs led to a notable performance drop (F1-Score of 0.88).

In addition to the AUC scores reported in Tables 2–5, the corresponding Receiver Operating Characteristic (ROC) curves for the best performing ML + FS in binary and multiclass classification, without and with SMOTE, are shown in Figure 3.

Discussion

Further to the classification analysis for glioma identification presented in Section 4, we conducted additional analysis over the five different feature ranking algorithms to investigate miRNAs that could serve as potential biomarkers for glioma. In specific, the top 100 miRNA features of the binary setup (glioma vs. controls) from each of the LASSO, ReliefF, MRMR, RFE, and RF feature ranking algorithms were compared to identify common ones among them. The Venn diagram showing the number of common miRNAs is shown in Figure 4. Seven common miRNAs were found, namely, the miR-125a-3p, miR-4276, miR-4648, miR-4763-3p, miR-663a, miR-6784-5p and miR-873-3p.

Figure 5 shows the expression levels of the seven selected miRNAs in controls, glioma, and meningioma. The box plots indicate that miR-125a-3p, miR-4276, miR-4648, miR-6784-5p, and miR-873-3p are downregulated in both glioma and meningioma compared with controls, suggesting potential tumour suppressor roles, whereas miR-663a and miR-4763-3p are upregulated in both tumour types, indicating possible oncogenic properties.

To assess the robustness and generalizability of the 7 miRNAs, validation on the independent dataset GSE211692 was performed. The four datasets (GSE112264, GSE113486, GSE113740 and GSE139031) were used for training, without SMOTE (Table 6) and with SMOTE augmentation (Table 7). For reference purposes, except the best ML model using only the 7 miRNA signatures, in Tables 6, 7 we also report the performance of best ML + FS from Section 4 for each ML algorithm on the validation dataset. Despite the substantial class imbalance in the external dataset, the selected miRNA signatures maintained consistently high discriminative ability across classifiers, with AUC values remaining very high and the 7-miRNA model achieving high performance in both settings (LR: 99.68% accuracy without SMOTE in Table 6 and 99.68% with SMOTE in Table 7). Although a slight reduction in some metrics, particularly glioma recall and F1-score, was observed during the independent validation when compared with the evaluation results, the results indicate the robustness and the generalizability of the 7 miRNA signatures when tested on independent data. It is worth noting that the selected miRNA signatures are aligned to or competing in performance the other data-driven ML + FS models while using 3 less miRNAs, and without the requirement for feature selection training.

Further to the seven miRNAs identified as potential biomarkers, linkage with both these 7 miRNAs and additional circulating miRNAs previously reported as glioma-related diagnostic or prognostic biomarkers was performed, to place our findings within the broader biomarker context. In specific, miR-328 was previously identified as part of a panel of miRNAs able to diagnose GB with 91.7% accuracy (ref. Ebrahimkhani et al., 2018). In addition, low expression levels were shown to be associated with poor prognosis in GB patients (ref. Wu et al., 2012). It has also been identified as a diagnostic biomarker of non-small cell lung cancer (ref. Ulivi et al., 2013). Using a bioinformatics approach, miR-4763-3p was identified as a biomarker of diffuse glioma when part of an index composed of three miRNAs (ref. Ohno et al., 2019). miR-873 has previously been shown to be able to serve as a diagnostic biomarker for glioma, with low expression levels linked to poor prognosis (ref. Li et al., 2020). Serum levels of miR-769-3p were able to differentiate glioma from controls with an AUC of 0.925 (ref. Wang et al., 2020). In (ref. Yerukala Sathipati et al., 2022) it was also identified it as part of a signature of 14 diagnostic miRNA biomarkers for bladder urothelial carcinoma, further validating its utility as a cancer biomarker. As previously mentioned, (ref. Li et al., 2023) created a model that utilised the REOs of five pairs of miRNAs to diagnose glioma from non-cancer. Three of the miRNAs included in these pairs, miR-125a-3p, miR-1914-5p and miR-1225-3p, were also identified in this study, supporting the potential utility of these miRNAs as diagnostic biomarkers for glioma. In addition, miR-125a-3p has been identified as a biomarker in pancreatic, colon, prostate and gastric cancer (ref. Salehi et al., 2022; ref. Wang et al., 2017; ref. Damodaran et al., 2021; ref. Li et al., 2025), and miR-1914-5p was previously identified as a diagnostic marker of ovarian cancer using a machine learning approach (ref. Hamidi et al., 2023). Our study also revealed that miR-663a could serve as a potential biomarker for glioma, which, to the best of our knowledge, has not been previously reported. miR-663a has been shown to act as a tumour suppressor in GB. It has been found to be downregulated in GB tissue with overexpression being associated with inhibition of proliferation, migration and invasion (ref. Wang L. et al., 2021). Collectively, it has the potential to act as a biomarker for GB. However, this association has only been reported by (ref. Wang L. et al., 2021) and requires further research to validate findings. Other studies have suggested this miRNA as a biomarker for the diagnosis of other types of cancer, including osteosarcoma and lung adenocarcinoma (ref. Huang et al., 2019; ref. Li Y. et al., 2024), providing a basis for its utility as a biomarker. In addition, miR-6784-5p, miR-4276 and miR-4648, which were ranked in the top 100 by each of the five feature selection methods in this study, have not previously been reported as biomarkers for glioma, therefore, they represent potentially novel biomarkers that require further research to validate their clinical relevance. (ref. Chen and Dhahbi, 2022) used ML to determine a panel of five miRNAs that could differentiate between cancer and non-cancer samples. miR-663a and miR-6784-5p were two of the miRNAs identified in this panel. Machine learning approaches also predicted miR-6784-5p to be a biomarker in ovarian and gastric cancer (ref. Hamidi et al., 2023; ref. Wang L. et al., 2021; ref. Huang et al., 2019; ref. Li Y. et al., 2024; ref. Chen and Dhahbi, 2022; ref. Wang et al., 2025). miR-4276 was identified as a circulating diagnostic biomarker of gastric cancer using bioinformatics and validated in human blood samples (ref. Li et al., 2025; ref. Yoshikawa et al., 2023) determined that the upregulation of miR-4648 was a biomarker for the recurrence of colorectal cancer, but to the best of our knowledge, there are no reports on its role in glioma. In addition, a machine learning model identified miR-4648 as a diagnostic biomarker of pancreatic cancer (ref. Chi et al., 2023). Although these miRNAs have not been associated with glioma before, the results from this study and the fact that they have been shown to be biomarkers for other cancers provide a basis for future research into their clinical validity as glioma biomarkers. It is important to distinguish between diagnostic biomarkers and mechanistic drivers. Diagnostic biomarkers are used primarily to indicate the presence of disease and support detection and patient stratification based on consistent statistical associations. In contrast, mechanistic drivers actively contribute to disease biology and may represent therapeutic targets; however, establishing such roles requires functional validation beyond correlative analyses. Therefore, this study focuses on the diagnostic value of circulating miRNAs for glioma identification, while acknowledging that leading candidates warrant further mechanistic investigation.

From a clinical translation perspective, the identified miRNA panels may be amenable to established RT-qPCR and droplet digital PCR (ddPCR) workflows. This would typically involve standardized steps for biofluid collection and handling, RNA extraction, reverse transcription, and amplification using miRNA-specific primers (e.g., LNA-based assays) with suitable endogenous or exogenous controls for normalization. Prior work has demonstrated that miRNA signatures discovered using high-throughput profiling can be transitioned to targeted qPCR/ddPCR validation with acceptable analytical reproducibility for panels of modest size, supporting feasibility for future assay development and independent clinical validation (ref. Mustafov et al., 2025; ref. Yu et al., 2025).

To provide additional biological context for the potentially novel biomarkers (miR-663a, miR-6784-5p, miR-4276, and miR-4648), we explored predicted targets using miRDB and investigated enriched pathways using miRPath v4.0 (KEGG). Table 8 lists the top 10 predicted gene targets of each miRNA from miRDB. Pathway enrichment suggested that miR-663a, miR-4276 and miR-4648 are involved in axon guidance and nervous system development, while miR-4648 was additionally predicted to be involved in brain development. These predictions motivate follow-up validation (e.g., cross-checking targets in TargetScan and verifying downstream pathway perturbations in glioma). In addition to KEGG pathway context, functional enrichment analysis highlighted biological processes linked to transcriptional regulation, nervous system development, phosphorylation-related signalling, and axon guidance (Tables 9–12). Notably, the enrichment of axon guidance and nervous system development is consistent with the neural origin of glioma. Across all four GEO datasets, miR-6784-5p, miR-4276 and miR-4648 showed consistent downregulation in glioma, while miR-663a was consistently upregulated (Table 13); all comparisons were statistically significant after multiple testing correction (adjusted $p\le 4.71\times 10^{-4}$ in all datasets). Similar analysis of the top 20 miRNAs of each of the five feature ranking algorithms resulted in 2 common miRNAs, miR-633a and miR-125a-3p. In the case of multiclass classification (glioma vs. meningioma vs. controls) none of the common miRNAs (miR-125a-3p, miR-4648, miR-4763-3p, miR-6786-3p, miR-6799-5p, miR-92b-5p) from the top 100 of each feature ranking algorithm are known clinically relevant biomarkers for meningioma or have any known association with meningioma.

Limitations

A limitation of this study concerns the metadata for dataset GSE113740. Review of the available annotations confirms that the 25 tumour samples are labelled only as “disease status: glioma,” with both pathological grade and pathological T stage listed as “uncertain.” This precludes reliable subtype assignment (e.g., glioblastoma versus IDH-mutant and 1p/19q-codeleted oligodendroglioma). Accordingly, these samples were analysed as a single glioma group.

We acknowledge that this introduces biological heterogeneity, as glioma subtypes differ substantially in molecular features, clinical behaviour, and prognosis. Pooling samples may therefore dilute subtype-specific miRNA signals, particularly between highly aggressive entities such as glioblastoma and more favourable tumours such as 1p/19q-codeleted oligodendrogliomas. Despite this limitation, our miRNA panels consistently discriminated glioma from controlsacross datasets (AUC $>0.95$), supporting their utility for broad glioma detection. Future work will prioritise molecularly stratified cohorts aligned with the WHO 2021 integrated histo-molecular classification (e.g., TCGA-GBMLGG) to enable refined subtype-level analyses.

In addition, age, sex, inflammation, or comorbidities need to further be investigated as other key factors that may affect circulating miRNA profiles. In specific, age-related differences in plasma miRNA expression have been previously reported, with brain-enriched miRNAs such as miR-134 showing peak levels in mid-life, with sex-specific patterns (ref. Sheinerman et al., 2018; ref. La Grotta et al., 2025). Sex-related dimorphism has also been observed, including female-specific downregulation of miR-4297 in glioma (ref. Xu et al., 2022). In addition, systemic inflammation can increase circulating levels of miRNAs such as miR-21, miR-155, and miR-122, while comorbidities associated with physiological stress or organ dysfunction (e.g., frailty or renal impairment) have been linked to elevated miR-23a-5p and miR-26a-5p (ref. Vibo et al., 2024; ref. Xu et al., 2022). Due to the lack of relevant meta-data information on the evaluated datasets the above factors have not been considered in the present study.

Finally, the multi-class classification included a limited number of meningioma samples (n = 17). Although SMOTE was employed to address this class imbalance, synthetic oversampling cannot fully replicate the true biological diversity and complexity of real clinical samples. This limitation may also affect the stability and biological interpretability of meningioma-related feature rankings. Therefore, the performance metrics for this specific class should be interpreted cautiously and further validated using larger cohorts. Moreover, independent validation of the multiclass classification setup with meningioma class was not feasible because the independent dataset GSE211692 does not have meningioma samples, and thus the generalizability of the selected miRNA signatures when meningioma samples are included could not be validated.

Conclusion

This study identified a panel of key miRNA signatures that can classify glioma subtypes and assessed their predictive accuracy using multiple machine learning models. Leveraging GEO blood serum miRNA expression data, the proposed pipeline integrated pre-processing, feature selection, ML classifiers, and SMOTE-based imbalance handling. The selected miRNAs enabled near-perfect binary classification (F1-score, AUC = 1.00) and robust multi-class performance (F1-scores 0.97–0.99, AUC 0.99), with KNN consistently achieving the highest scores. Several miRNAs identified in this study, including miR-328, miR-4763-3p, miR-873, miR-769-3p, miR-125a-3p, miR-1914-5p, and miR-1225-3p, have been previously reported as glioma biomarkers, supporting the clinical relevance of our findings. Importantly, miR-663a, miR-6784-5p, miR-4276, and miR-4648 ranked as top 100 by each of five feature selection techniques are potential novel biomarkers for glioma, with no prior reports linking them to the disease. While these miRNAs have been implicated in other cancer types, their identification here provides a foundation for future glioma-specific biomarker validation. Our findings demonstrate that targeted miRNA selection combined with optimised predictive modelling can achieve high diagnostic accuracy with compact feature sets, supporting feasibility for clinical deployment. Limitations include the small meningioma sample size and possible batch effects, which warrant validation in larger, diverse cohorts. Future work should explore multi-omics integration, deep learning–based models, and clinical translation into standardised, non-invasive diagnostic protocols.

References

1. P. Andreini, S. Bonechi, M. Bianchini, F. Geraci. Microrna signature for interpretable breast cancer classification with subtype clue.. J. Comput. Math. Data Sci., 2022. [DOI]
2. L. Anelli, A. Zagaria, G. Specchia, P. Musto, F. Albano. Dysregulation of Mirna in leukemia: exploiting mirna expression profiles as biomarkers.. Int. Journal Molecular Sciences, 2021. [DOI | PubMed]
3. L. Breiman. Random forests.. Mach. Learning, 2001. [DOI]
4. N. V. Chawla, K. W. Bowyer, L. O. Hall, W. P. Kegelmeyer. Smote: synthetic minority over-sampling technique.. J. Artificial Intelligence Research, 2002. [DOI]
5. J. W. Chen, J. Dhahbi. Identification of four serum mirnas as potential markers to screen for thirteen cancer types.. PLoS One, 2022. [DOI | PubMed]
6. T. Chen, T. He, M. Benesty, V. Khotilovich, Y. Tang, H. Cho. Xgboost: extreme gradient boosting.. R. Package Version 0 (, 2015. [DOI]
7. H. Chi, H. Chen, R. Wang, J. Zhang, L. Jiang, S. Zhang. Proposing new early detection indicators for pancreatic cancer: combining machine learning and neural networks for serum mirna-based diagnostic model.. Front. Oncol., 2023. [DOI | PubMed]
8. C. Cortes, V. Vapnik. Support-vector networks.. Mach. Learning, 1995. [DOI]
9. T. Cover, P. Hart. Nearest neighbor pattern classification.. IEEE Transactions Information Theory, 1967. [DOI]
10. M. Damodaran, M. Chinambedu Dandapani, SimonDuraiRaj, SandhyaSundaram, S. VenkatRamanan, I. Ramachandran. Differentially expressed mir-20, mir-21, mir-100, mir-125a and mir-146a as a potential biomarker for prostate cancer.. Mol. Biol. Rep., 2021. [DOI | PubMed]
11. B. Dhananjay, J. Sivaraman. Analysis and classification of heart rate using catboost feature ranking model.. Biomed. Signal Process. Control, 2021. [DOI]
12. S. Ebrahimkhani, F. Vafaee, S. Hallal, H. Wei, M. Y. T. Lee, P. E. Young. Deep sequencing of circulating exosomal microrna allows non-invasive glioblastoma diagnosis.. NPJ Precision Oncology, 2018. [DOI | PubMed]
13. R. Edgar, M. Domrachev, A. E. Lash. Gene expression omnibus: ncbi gene expression and hybridization array data repository.. Nucleic Acids Research, 2002. [DOI | PubMed]
14. I. Guyon, J. Weston, S. Barnhill, V. Vapnik. Gene selection for cancer classification using support vector machines.. Mach. Learning, 2002. [DOI]
15. F. Hamidi, N. Gilani, R. Arabi Belaghi, H. Yaghoobi, E. Babaei, P. Sarbakhsh. Identifying potential circulating mirna biomarkers for the diagnosis and prediction of ovarian cancer using machine-learning approach: application of boruta.. Front. Digital Health, 2023. [DOI | PubMed]
16. T. Hastie, R. Tibshirani, J. H. Friedman. The Elements of Statistical Learning., 2009
17. S. Hosseini. A new machine learning method consisting of ga-lr and ann for attack detection.. Wirel. Netw., 2020. [DOI]
18. C. Huang, Y. Sun, S. Ma, A. S. Vadamootoo, L. Wang, C. Jin. Identification of circulating mir-663a as a potential biomarker for diagnosing osteosarcoma.. Pathology-Research Pract., 2019. [DOI | PubMed]
19. M. Ijzerman-Korevaar, T. J. Snijders, A. de Graeff, S. C. Teunissen, F. Y. de Vos. Prevalence of symptoms in glioma patients throughout the disease trajectory: a systematic review.. J. Neuro-Oncology, 2018. [DOI | PubMed]
20. W. E. Johnson, C. Li, A. Rabinovic. Adjusting batch effects in microarray expression data using empirical bayes methods.. Biostatistics, 2007. [DOI | PubMed]
21. E. Kaczmarek, A. Jamzad, T. Imtiaz, J. Nanayakkara, N. Renwick, P. Mousavi. Multi-omic graph transformers for cancer classification and interpretation,” in. Pacific Symposium on Biocomputing 2022 (, 2021
22. E. Kaczmarek, B. Pyman, J. Nanayakkara, T. Tuschl, K. Tyryshkin, N. Renwick. Discriminating neoplastic from nonneoplastic tissues using an mirna-based deep cancer classifier.. Am. J. Pathology, 2022. [DOI | PubMed]
23. S. Kantamneni, P. Hegde, N. Patil. Utilizing deep learning methods for cancer detection through analysis of microrna expression profiles,” in. 2024 15th International Conference on Computing Communication and Networking Technologies (ICCCNT) (, 2024
24. I. Kononenko. Estimating attributes: analysis and extensions of relief,” in. European Conference on Machine Learning (, 1994
25. B. Koonce, B. Koonce. Resnet 50 [W:]: Convolutional Neural Networks with Swift for Tensorflow., 2021
26. M. B. Kursa, W. R. Rudnicki. Feature selection with the boruta package.. J. Statistical Software, 2010. [DOI]
27. R. La Grotta, P. Crocco, S. Dato, G. Passarino, G. Rose. Circulating micrornas mir-21-5p, mir-23a-3p and mir-26a-5p reflect clinical and molecular features of aging.. Sci. Rep., 2025. [DOI | PubMed]
28. J.-F. Laplante, M. A. Akhloufi. Predicting cancer types from mirna stem-loops using deep learning,” in. 2020 42nd Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC) (, 2020
29. K. Li, Q. Zhang, D. Niu, H. Xing. Mining mirnas’ expressions in glioma based on geo database and their effects on biological functions.. BioMed Res. Int., 2020. [DOI | PubMed]
30. L. Li, L. Guo, Y. Zhang, Y. Su. Machine learning-based construction of a mirnas prognostic risk assessment model of cervical cancer,” in. 2022 41st Chinese Control Conference (CCC) (, 2022
31. H. Li, L. Ma, F. Luo, W. Liu, N. Li, T. Hu. Construct of qualitative diagnostic biomarkers specific for glioma by pairing serum micrornas.. BMC Genomics, 2023. [DOI | PubMed]
32. K. Li, Q. Zhu, J. Yang, Y. Zheng, S. Du, M. Song. Imaging and liquid biopsy for distinguishing true progression from pseudoprogression in gliomas, current advances and challenges.. Acad. Radiol., 2024a. [DOI | PubMed]
33. Y. Li, N.-D. Yu, X.-L. Ye, M.-C. Jiang, X.-Q. Chen. Construction of lung cancer serum markers based on relieff feature selection.. Comput. Methods Biomechanics Biomed. Eng., 2024b. [DOI | PubMed]
34. X. Li, Q. Wang, Z. Xu, X. Yang, R. Zhao, H. Wang. Screening and evaluation of specific blood mirnas as potential biomarkers in diagnostics of gastric cancer.. Sci. Rep., 2025. [DOI | PubMed]
35. S. Lokeswaran, P. Manikandan, R. Rajakumar, M. Marimuthu. Brain tumor classification from gene expression dataset through supervised machine learning algorithms,” in. 2024 International Conference on Intelligent Algorithms for Computational Intelligence Systems (IACIS) (, 2024
36. D. N. Louis, A. Perry, P. Wesseling, D. J. Brat, I. A. Cree, D. Figarella-Branger. The 2021 who classification of tumors of the central nervous system: a summary.. Neuro-oncology, 2021. [DOI | PubMed]
37. J. Matsuzaki, K. Kato, K. Oono, N. Tsuchiya, K. Sudo, A. Shimomura. Prediction of tissue-of-origin of early stage cancers using serum mirnomes.. JNCI Cancer Spectrum, 2023. [DOI | PubMed]
38. T. Mehejabin, F. Rahman, S. Yeasmin, M. Sarkar. Identification of Most relevant breast cancer mirna using machine learning algorithms,” in. 2020 11th International Conference on Computing, Communication and Networking Technologies (ICCCNT) (, 2020
39. A. Muhamed Ali, H. Zhuang, A. Ibrahim, O. Rehman, M. Huang, A. Wu. A machine learning approach for the classification of kidney cancer subtypes using mirna genome data.. Appl. Sci., 2018. [DOI]
40. D. Mustafov, L. Thomas, S. S. Siddiqui, G. I. Lambrou, M. Braoudaki. New microrna-based therapies reveal common targets in paediatric medulloblastoma and adult glioblastoma.. Sci. Reports, 2025. [DOI]
41. J. Ni, D. Yan, S. Lu, Z. Xie, Y. Liu, X. Zhang. Mirs-hf: a novel deep learning predictor for cancer classification and mirna expression patterns.. IEEE Journal Biomedical Health Informatics., 2024. [DOI]
42. T. R. Noviandy, M. H. Alfanshury, T. F. Abidin, H. Riza. Enhancing glioma grading performance: a comparative study on feature selection techniques and ensemble machine learning,” in. 2023 International Conference on Computer, Control, Informatics and its Applications (IC3INA) (, 2023
43. V. Oberoi, B. S. Negi, D. Chandan, P. Verma, V. Kulkarni. Grade and subtype classification for glioma tumors using clinical and molecular mutations,” in. 2023 International Conference on Modeling, Simulation & Intelligent Computing (MoSICom) (, 2023
44. M. Ohno, J. Matsuzaki, J. Kawauchi, Y. Aoki, J. Miura, S. Takizawa. Assessment of the diagnostic utility of serum microrna classification in patients with diffuse glioma.. JAMA Network Open, 2019. [DOI | PubMed]
45. J. K. Pal, B. R. Rami. Machine learning based identification of candidate mirna biomarkers for micro-invasive breast cancer diagnosis.. bioRxiv,, 2024. [DOI]
46. H. Peng, F. Long, C. Ding. Feature selection based on mutual information criteria of max-dependency, max-relevance, and min-redundancy.. IEEE Trans. Pattern Analysis Machine Intelligence, 2005. [DOI | PubMed]
47. B. Pyman, A. Sedghi, S. Azizi, K. Tyryshkin, N. Renwick, P. Mousavi. Exploring microrna regulation of cancer with context-aware deep cancer classifier,” in. BIOCOMPUTING 2019: Proceedings of the Pacific Symposium (World Scientific),, 2018
48. J. R. Quinlan. Induction of decision trees.. Mach. Learning, 1986. [DOI]
49. M. E. Ritchie, B. Phipson, D. Wu, Y. Hu, C. W. Law, W. Shi. Limma powers differential expression analyses for rna-sequencing and microarray studies.. Nucleic Acids Research, 2015. [DOI | PubMed]
50. A. S. Salehi, N. Parsa-Nikoo, F. Roshan-Farzad, R. Shams, M. Fathi, H. A. Aghdaei. Microrna-125a-3p, -4530, and -92a as a potential circulating microrna panel for noninvasive pancreatic cancer diagnosis.. Dis. Markers, 2022. [DOI | PubMed]
51. A. Sharma, E. Vans, D. Shigemizu, K. A. Boroevich, T. Tsunoda. Deepinsight: a methodology to transform a non-image data to an image for convolution neural network architecture.. Sci. Reports, 2019. [DOI | PubMed]
52. K. Sheinerman, V. Tsivinsky, A. Mathur, D. Kessler, B. Shaz, S. Umansky. Age-and sex-dependent changes in levels of circulating brain-enriched micrornas during normal aging.. Aging (Albany NY), 2018. [DOI | PubMed]
53. P.-p. Si, X.-h. Zhou, Z. Qu. Identification of serum mirnas as effective diagnostic biomarkers for distinguishing primary central nervous system lymphoma from glioma.. J. Immunol. Res., 2022. [DOI | PubMed]
54. I. Slimene, I. Messaoudi, A. Elloumi Oueslati, Z. Lachiri. Microrna expression classification for human disease prediction,” in. 2021 18th International Multi-Conference on Systems, Signals & Devices (SSD) (, 2021
55. I. Slimene, I. Messaoudi, A. E. Oueslati, Z. Lachiri. Deep learning-based cancer disease classification through microrna expression,” in. 2022 IEEE Information Technologies and Smart Industrial Systems (ITSIS) (, 2022
56. R. Tibshirani. Regression shrinkage and selection via the lasso.. J. R. Stat. Soc. Ser. B Stat. Methodol., 1996. [DOI]
57. K. Tomczak, P. Czerwińska, M. Wiznerowicz. Review the cancer genome atlas (tcga): an immeasurable source of knowledge.. Contemp. Oncology/Współczesna Onkol., 2015. [DOI]
58. P. Ulivi, G. Foschi, M. Mengozzi, E. Scarpi, R. Silvestrini, D. Amadori. Peripheral blood mir-328 expression as a potential biomarker for the early diagnosis of nsclc.. Int. Journal Molecular Sciences, 2013. [DOI | PubMed]
59. F. Urabe, J. Matsuzaki, Y. Yamamoto, T. Kimura, T. Hara, M. Ichikawa. Large-scale circulating microrna profiling for the liquid biopsy of prostate cancer.. Clin. Cancer Research, 2019. [DOI | PubMed]
60. W. Usuba, F. Urabe, Y. Yamamoto, J. Matsuzaki, H. Sasaki, M. Ichikawa. Circulating mirna panels for specific and early detection in bladder cancer.. Cancer Science, 2019. [DOI | PubMed]
61. M. A. Van De Wiel, T. G. Lien, W. Verlaat, W. N. van Wieringen, S. M. Wilting. Better prediction by use of co-data: adaptive group-regularized ridge regression.. Statistics Medicine, 2016. [DOI | PubMed]
62. R. Vibo, K. Jõgi, A. Remm, A. Rebane, J. Kõrv. Different expression patterns of inflammation-related genes and serum micrornas in young-onset ischemic stroke.. Sci. Rep., 2024. [DOI | PubMed]
63. J. Wang, F. Yan, Q. Zhao, F. Zhan, R. Wang, L. Wang. Circulating exosomal mir-125a-3p as a novel biomarker for early-stage colon cancer.. Sci. Reports, 2017. [DOI | PubMed]
64. K. Wang, S. Yang, Y. Gao, C. Zhang, Q. Sui. Microrna-769-3p inhibits tumor progression in glioma by suppressing zeb2 and inhibiting the wnt/β-catenin signaling pathway.. Oncol. Letters, 2020. [DOI | PubMed]
65. L. Wang, B. Lang, Y. Zhou, J. Ma, K. Hu. Up-regulation of mir-663a inhibits the cancer stem cell-like properties of glioma via repressing the kdm2a-mediated tgf-β/smad signaling pathway.. Cell Cycle, 2021a. [DOI | PubMed]
66. T. Wang, W. Shao, Z. Huang, H. Tang, J. Zhang, Z. Ding. Mogonet integrates multi-omics data using graph convolutional networks allowing patient classification and biomarker identification.. Nat. Communications, 2021b. [DOI]
67. Q. Wang, J. Zhu, C. Pan, J. Shi, C. Meng, H. Guo. Dual trustworthy mechanism for illness classification with multi-modality data,” in. 2023 IEEE International Conference on Data Mining Workshops (ICDMW) (IEEE),, 2023
68. X. Wang, Z. Li, C. Zhang. Integrated analysis of serum and tissue microrna transcriptome for biomarker discovery in gastric cancer.. Environ. Toxicol., 2025. [DOI | PubMed]
69. I. Waspada, A. Wibowo, N. S. Meraz. Supervised machine learning model for microrna expression data in cancer.. J. Ilmu Komput. Dan Inf., 2017. [DOI]
70. Z. Wu, L. Sun, H. Wang, J. Yao, C. Jiang, W. Xu. Mir-328 expression is decreased in high-grade gliomas and is associated with worse survival in primary glioblastoma. 2012
71. W. Xu, L. Huang, B. Xie, B. Yang. Serum microrna-4297 is a sex-specific predictive biomarker of glioma grade and prognosis.. Front. Neurology, 2022. [DOI]
72. Y. Yamamoto, S. Kondo, J. Matsuzaki, M. Esaki, T. Okusaka, K. Shimada. Highly sensitive circulating microrna panel for accurate detection of hepatocellular carcinoma in patients with liver disease.. Hepatol. Communications, 2020. [DOI | PubMed]
73. S. Yerukala Sathipati, M.-J. Tsai, S. K. Shukla, S.-Y. Ho, Y. Liu, A. Beheshti. Microrna signature for estimating the survival time in patients with bladder urothelial carcinoma.. Sci. Reports, 2022. [DOI]
74. Y. Yoshikawa, M. Fukunaga, J. Takahashi, D. Shimizu, T. Masuda, T. Mizushima. Identification of the minimum combination of serum micrornas to predict the recurrence of colorectal cancer cases.. Ann. Surgical Oncology, 2023. [DOI]
75. Y. Yu, W. H. Wong, Y. Hu, Y. Shen, X. Xu, S. Lu. Translation of mirna blood-based discovery to molecular testing for clinical diagnosis of endometriosis.. Women’s Health, 2025. [DOI]
76. Q. Zhou, J. Liu, J. Quan, W. Liu, H. Tan, W. Li. Micrornas as potential biomarkers for the diagnosis of glioma: a systematic review and meta-analysis.. Cancer Science, 2018. [DOI | PubMed]