Development and validation of a miRNA-based prognostic model for high-grade serous ovarian cancer: a retrospective cohort study

Summary

Background

Ovarian cancer, particularly high-grade serous ovarian cancer (HGSOC), is the most lethal gynecological malignancy, mainly due to late-stage diagnosis and limited prognostic biomarkers. Current clinical markers, such as CA125, have limited prognostic accuracy for risk stratification. MicroRNAs (miRNAs) have emerged as promising biomarkers due to roles in tumor biology and stability in biofluids. This study aimed to identify and validate prognostic miRNA biomarkers in HGSOC.

Methods

A machine learning pipeline was implemented to develop a prognostic model using miRNA data. Candidate miRNAs were identified through feature selection and differential expression analyses. Recursive Feature Elimination determined the optimal predictor set among miRNAs combined with age, stage, and MUC16. Model development used a hold-out split, with hyperparameter optimization under 5-fold cross-validation within training. Performance was evaluated using area under the receiver operating characteristic curve (AUC), recall, and balanced accuracy. SHAP analysis assessed feature contributions, while enrichment analyses characterized miRNA–mRNA interactions and pathways.

Findings

The final model, incorporating 9 miRNAs with clinical variables, achieved an AUC of 0.762 [95% CI: 0.621–0.903], exceeding previously reported signatures. Key miRNAs, including hsa-miR-205-5p and hsa-miR-150-5p, were associated with angiogenesis, invasion, and chemoresistance pathways. In RT-qPCR validation, discriminative performance decreased; however, the continuous risk score remained independently associated with survival, achieving a C-index of 0.85 in multivariable analysis.

Interpretation

We present an interpretable miRNA-based prognostic model for HGSOC integrating molecular features. Although ROC-based discrimination was limited in external validation, survival analyses supported independent prognostic value, with the continuous risk score significantly associated with survival. Continuous and classification-based stratification identified survival groups, supporting clinical relevance of the model and identified miRNA signature.

Funding

CNPq; FAPERJ; Brazilian Ministry of Health (INCA/MS).

Research in context.

Evidence before this study

We searched PubMed to identify studies evaluating the prognostic value of microRNAs (miRNAs) in high-grade serous ovarian cancer (HGSOC), with a focus on those applying artificial intelligence (AI) methods. The search covered articles published up to 2025 with no language restrictions, using combinations of (“miRNA” OR “microRNA”) AND (“ovarian cancer” OR “epithelial ovarian cancer” OR “HGSOC”) AND (“machine learning” OR “artificial intelligence” OR “predict”) AND (“prognosis” OR “relapse”), yielding 40 relevant studies. Of these, eight were review articles broadly discussing the role of miRNAs in ovarian cancer, particularly the hsa-miR 200 family. Most original studies relied on Cox regression or RT-qPCR to evaluate the prognostic value of individual or small sets of miRNAs, typically combined with clinical variables. Recent examples include exosomal miRNA signatures for plasma-based detection (AUCs 0.77–0.94), but these focus primarily on diagnosis rather than HGSOC-specific prognosis. Only about 15% of studies used machine learning (ML) approaches to model outcomes like relapse or survival, and seven focused specifically on HGSOC. Among existing tools, the MiROvaR signature, a 35-miRNA based model for epithelial ovarian cancer, showed initial promise but reduced performance in HGSOC-only cohorts (AUC = 0.58 in TCGA), indicating limited applicability to this more aggressive subtype. More recently, the TMErisk model (2023) integrates tumor microenvironment features via ML (10 algorithms tested on TCGA/GEO data), outperforming prior signatures for HGSOC prognosis prediction and linking low-risk groups to BRCA1 mutations and enhanced immune responses. In parallel, clinical diagnostic tests such as ROMA, Ova1®, and Overa® use serum biomarkers to aid in the detection of epithelial ovarian cancer. However, these tools have notable limitations: ROMA is affected by menopausal status, while Ova1® and Overa® show high sensitivity but low specificity, leading to frequent false positives. Together, these findings highlight a significant gap in accurate, HGSOC-specific, AI-driven prognostic tools that can robustly guide risk stratification and treatment decisions in clinical practice.

Added value of this study

In this study, we developed and validated a ML–based prognostic model for HGSOC that integrates miRNA expression profiles with clinical variables. Our workflow combined differential expression analysis, feature selection, and explainable AI to derive a robust panel of 9 miRNAs that, when modeled together, served as powerful prognostic biomarkers. Using this panel, we successfully stratified patients into subgroups with significantly different overall survival outcomes, enabling more personalized prognostic insights for women with HGSOC. A major barrier to the clinical adoption of AI models is their “black-box” nature, which often limits interpretability. Here, we addressed this challenge by incorporating explainable AI methods that elucidated the contribution of each feature to the model’s predictions. Furthermore, we biologically characterized the identified miRNAs, organizing them into regulons associated with key signaling pathways and tumor progression processes. Finally, the model’s generalizability and clinical relevance were confirmed through external validation in real-world clinical samples from a Brazilian cohort, reinforcing its potential as a reliable prognostic tool for HGSOC.

Implications of all the available evidence

Our study addresses a critical unmet need in ovarian cancer research by providing a clinically interpretable, AI-driven prognostic tool specifically tailored for HGSOC, a subtype for which existing models and serum-based tests have shown limited performance. By integrating ML, explainable AI, and biological interpretation, we overcame the “black-box” barrier that has hindered clinical adoption of AI models. The model achieved high predictive accuracy (AUC = 0.762), effectively stratifying patients into distinct prognostic groups with translational relevance. Validation in both the TCGA and an independent Brazilian clinical cohort demonstrates its robustness and generalizability across populations and healthcare systems, bridging the gap between research and clinical applicability. The integration of clinical variables enhances its capacity for individualized risk assessment, while the identification of key miRNAs within plasma and extracellular vesicles supports their feasibility as liquid biopsy biomarkers, enabling non invasive monitoring of prognosis and relapse. Together, these findings mark a paradigm shift from static, exploratory biomarker studies toward dynamic, interpretable, and clinically actionable AI tools in oncology. The proposed model has the potential to guide treatment decisions, inform follow-up strategies, and improve patient outcomes by embedding molecular and AI-based risk stratification into routine clinical workflows. Prospective validation through multi-institutional and longitudinal studies will be an essential next step toward regulatory approval and implementation in precision oncology.

Introduction

Ovarian cancer (OC) remains a major global health challenge, with 324,603 new cases and 206,956 deaths reported worldwide in 2022.¹ In Brazil, 7,310 new cases are expected annually (2023–2025), with 4,444 deaths recorded in 2023.² High mortality is largely due to nonspecific symptoms and late diagnosis, only 25% of cases are detected early³ resulting in a five-year survival rate of ≤40%.⁴ Epithelial OC accounts for over 90% of cases, with high-grade serous ovarian cancer (HGSOC) being the most prevalent and aggressive subtype.⁵

The standard first-line treatment consists of cytoreductive surgery followed by platinum-based chemotherapy, most commonly carboplatin and paclitaxel.^6,7 This regimen remains the standard of care worldwide, especially in public healthcare systems such as those in Brazil, India, South Africa, and Mexico, where access to targeted therapies is often limited by cost and availability.8, 9, 10, 11 Although most patients initially respond, up to 70% with advanced-stage disease relapse within two years.^12,13 The introduction of PARP inhibitors (PARPi) has improved outcomes for patients with homologous recombination deficiency (HRD). Approximately half of HGSOC cases exhibit HRD and^14,15 typically respond well to platinum-based chemotherapy, benefitting further from PARPi maintenance therapy.¹⁶ Nevertheless, access to PARPi and other novel therapies remains limited in many low- and middle-income settings, while effective treatment options for HR-proficient patients are still scarce.¹⁷ In this context, reliable biomarkers capable of predicting prognosis and treatment response are essential to optimize therapeutic decisions and improve clinical outcomes.

Currently used biomarkers, mucin 16 (MUC16, also known as carbohydrate antigen 125 or CA-125) and human epididymis protein 4 (HE4), support diagnosis and prognosis but lack sufficient sensitivity and specificity, particularly for early detection and reliable risk stratification.¹⁸ Beyond BRCA and HRD testing, no molecular markers are routinely applied in clinical practice to predict long-term outcomes or guide therapy. This gap underscores the need for integrative approaches that combine molecular and clinical features to enhance prognostic accuracy and enable more personalized treatment strategies.¹⁹

Among molecular regulators, microRNAs (miRNAs) stand out as promising biomarkers. By fine-tuning gene expression at the post-transcriptional level, they regulate key oncogenic pathways including proliferation, apoptosis, DNA repair, and metastasis.^20,21 Their remarkable stability in biological fluids supports their potential as non-invasive biomarkers. Profiling miRNA expression deepens understanding of tumor biology and provides new opportunities for precision oncology.

Machine learning (ML) methods further advance this field by enabling the analysis of high-dimensional biological data and uncovering complex, non-linear relationships between molecular features and clinical outcomes.^22,23 In ovarian cancer, where prognostic biomarkers remain limited and often inaccessible, ML offers a powerful framework for developing accurate, reproducible, and scalable tools for patient stratification and treatment guidance. Here, we developed an interpretable predictive model for HGSOC that integrates miRNA profiles with age and MUC16 levels. The model achieved high accuracy, strong generalization, and consistent performance when validated in an independent external cohort, demonstrating its robustness and potential for clinical application across diverse healthcare systems.

Methods

Data sources and study participants

This study leveraged public datasets for model discovery and an independent institutional cohort for external validation.

Discovery cohort

The Cancer Genome Atlas - Ovarian Cancer (TCGA-OV) miRNA expression data with batch correction were obtained from the Gene Expression Omnibus (GEO) under accession GSE164767.²⁴ Only canonical miRNAs (22 nucleotides) with a median expression >15 Reads Per Million mapped reads (RPM) were included. TCGA mRNA, clinical, and survival data were downloaded using the Bioconductor package TCGAbiolinks (v2.28.3).²⁵ Additional clinical data on platinum response and TP53/BRCA1/2 mutation status were retrieved from cBioPortal’s “Ovarian Serous Cystadenocarcinoma (TCGA, Nature 2011)” cohort.²⁶ Survival data were obtained from the PanCancer Atlas.²⁷ The study included only patients with complete, matched data for miRNA expression and relevant clinical-pathological variables.

Raw RNA-Seq FASTQ files from TCGA and Genotype-Tissue Expression (GTEx) were accessed using NCBI database of Genotypes and Phenotypes (dbGap) (accesion numbers phs000178.v11.p8.c1 and phs000424.v11.p2, respectively).

For differential expression analyses, preprocessed microarray datasets were retrieved via GEOquery (v2.68). Three independent datasets were included: GSE103708,²⁸ which contains exosomal and intracellular miRNA profiles from three normal human ovarian surface epithelial (HOSE) cell lines and four HGSOC cell lines; GSE83693, containing eight primary HGSOC tumors and four normal ovarian tissues²⁹; GSE47841, including 12 HGSOC tumors and nine normal samples.³⁰ These datasets provided complementary information on tumor versus normal expression patterns and exosomal miRNA signatures, supporting the identification of biologically relevant features for model construction. A complete overview of datasets and their characteristics is presented in Supplementary Table S1.

To assess the detectability of the selected miRNAs in circulating samples, serum miRNA expression data were obtained from GEO (GSE106817).²⁸ This dataset includes 4,046 samples from non-cancer controls, ovarian cancer patients, borderline ovarian tumors, benign ovarian diseases, and other solid cancers. Samples classified as others_OV were excluded from the analysis. Preprocessed log₂-normalized microarray expression values were obtained directly from GEO. Expression levels of the miRNAs included in our panel were extracted for downstream analyses, and group-wise comparisons were visualized using box plots. Additional cohort characteristics are provided in Supplementary Table S2b.

Studies were eligible if they included primary epithelial ovarian cancer samples with available miRNA expression data and adequate clinical annotation. For differential expression analyses, only datasets with both tumor and normal ovarian tissues were included. Datasets based exclusively on cell lines, non-human samples, incomplete annotations, or very small sample sizes were excluded.

Experimental analysis - validation cohort

Ethics statement

This retrospective study was approved by the Ethics in Human Research Committee of the Brazilian National Cancer Institute (INCA) (CAAE 52409221.4.0000.5274) in Rio de Janeiro, Brazil, and was conducted under all applicable regulations concerning the involvement of human study participants.

Sample collection and study design

This retrospective study included 40 HGSOC patients diagnosed between 2017 and 2024 at INCA. Fresh frozen tumor samples were obtained from the National Tumor Bank. Inclusion criteria comprised patients without a history of other primary malignancies and RNA samples with at least 70% tumor cellularity available for analysis. Samples were collected during cytoreductive surgery: prior to systemic therapy in patients receiving adjuvant chemotherapy and during debulking surgery after five or six cycles of cisplatin–paclitaxel chemotherapy in patients treated with neoadjuvant chemotherapy (NACT). The patients’ clinical data are summarized in Supplementary Table S3.

RT-qPCR analysis of miRNAs

Total RNA was extracted from tumor tissues using TRIzol reagent (Invitrogen) following the manufacturer’s protocol. RNA quality and concentration were assessed using a NanoDrop ND-1000 spectrophotometer (Thermo Fisher) by measuring the 260/280 absorbance. For miRNA expression, 10 ng of total RNA was reverse-transcribed using the miRCURY LNA RT Kit (Qiagen, Hilden, Germany). qPCR was performed in duplicate using the miRCURY LNA SYBR Green PCR Kit (Qiagen) on a ViiA™ 7 Real-Time PCR System (Thermo Fisher). U6 snRNA was used for normalization, and relative expression was calculated by the 2ˆ-ΔCt method. Primer sequences used are listed in Supplementary Table S4.

Outcome definition

Patients from TCGA (discovery) and INCA (external validation) cohorts were stratified by overall survival (OS). Consistent with prior studies investigating prognostic signatures in HGSOC,³¹ prognosis was defined using a 3-year OS cutoff, classifying patients into a good prognosis group (OS > 3 years) and a poor prognosis group (OS ≤ 3 years). A substantial proportion of deaths occur within the first 3 years after diagnosis, reflecting a period of high-risk recurrence and mortality³² (Fig. 1a).

Fig. 1.

Integrative pipeline for development of the miRNA-based prognostic model. Overview of the analytical workflow integrating miRNA-seq data from tumor samples with clinical survival data from TCGA. (a) Patients were stratified into good and poor prognosis groups using a 3-year overall survival threshold. (b) Differentially expressed miRNAs (DEmiRs) were identified from GEO datasets by comparing tumor and normal samples. Data preprocessing included expression filtering, removal of highly correlated features, and selection based on variability. Feature selection was performed using multiple complementary approaches. (c) Selected features were used to train supervised ML models with class balancing and hyperparameter tuning. Model performance was evaluated using AUROC, sensitivity, accuracy, and F1-score. Model interpretability was achieved through the SHAP (SHapley Additive exPlanations) framework, providing an explainable artificial intelligence perspective.

Predictor definition and feature selection

Differentially expressed miRNAs (DEmiRs) were identified using the LIMMA package (v4.0.1) across the three GEO datasets (GSE103708, GSE83693 and GSE47841). Significance was determined by an FDR-adjusted <0.01 and |log2FC| ≥ 1. Only DEmiRs in at least two datasets and present in the tumor tissue datasets (GSE83693 and GSE47841) were retained. For Venn diagram visualization, the suffixes “-3p” and “-5p” were temporarily removed from miRNA identifiers to maximize overlap across datasets reporting precursor miRNAs. This step was applied only for visualization purposes; all downstream statistical and ML analyses retained the 3p and 5p isoforms. Visualization was performed using the EnhancedVolcano package(v1.18).

For model development, using the training set (90%), redundant and non-informative miRNAs were removed. Highly correlated miRNAs (Pearson R² > 0.90) were removed to minimize redundancy, and zero-variance features were excluded through variance filtering. Three feature selection methods identified potential prognostic miRNAs: Fast Correlation-Based Filter (FCBF, v2.8.0)³³ with minimum_su = 0.01, univariate Cox regression (survival package v3.5)³⁴ selecting miRNAs with p < 0.05 over 5-year OS; and Least Absolute Shrinkage and Selection Operator (LASSO) regression (caret v6.0)³⁵ with 50-fold cross-validation selecting miRNAs with non-zero coefficients. Only miRNAs identified by at least two of these methods (FCBF, Cox and LASSO) and DEmiR analysis were used for downstream analysis (Fig. 1b).

Machine learning model: development and performance

The dataset was randomly partitioned into a training set (90%) and an internal hold-out test set (10%), ensuring that the test set remained completely isolated from model development. Model training and hyperparameter optimization were performed within the training set using 5-fold stratified cross-validation, with the AUC used as the primary optimization metric. Several ML classifiers were evaluated, including Tabular Prior-data Fitted Network (TabPFN, v1),³⁶ Logistic Regression (LR), Random Forest (RF), Extra Trees, XGBoost, LightGBM, Support Vector Machines (SVM), and a Multi-Layer Perceptron (MLP).

Predicted probabilities were calibrated using CalibratedClassifierCV, and the optimal classification threshold was determined using out-of-fold predictions from the training set, selecting the threshold that maximized balanced accuracy. Model performance was evaluated using AUC, accuracy, balanced accuracy, recall, specificity, F1-score, Negative Predictive Value (NPV), Brier score, and Expected Calibration Error (ECE). Confidence intervals were estimated using nonparametric resampling with 2000 iterations. To enhance interpretability, the contribution of each miRNA to model predictions was quantified using SHapley Additive exPlanations (SHAP) (v0.36.0).³⁷

Following the initial definition of candidate features, Recursive Feature Elimination (RFE) was applied within the training pipeline to evaluate models containing between three and thirteen predictors. For the integrated model, miRNA expression, MUC16 gene expression, and patient age were jointly considered, and RFE was used to identify the most informative subset of predictors. All analyses were performed in Python 3.11 using scikit-learn (v1.2.2),³⁸ and, where applicable, the original libraries for XGBoost(v.0.90)³⁹ and LightGBM(v.2.3.1).⁴⁰ Visualizations were generated with Seaborn and Matplotlib (Fig. 1c).

In the discovery cohort, MUC16 values were represented as log₂-normalized gene expression levels derived from RNA-seq data. Missing data were addressed through predefined inclusion criteria established prior to analysis. Only samples with complete survival information and miRNA expression data were included. Cases with missing clinical follow-up or incomplete miRNA measurements were excluded from the analyses.

Cross-platform miRNA profiling and standardized modeling pipeline

In the discovery phase, miRNA expression was profiled using RNA sequencing, while external validation was performed using RT-qPCR, a targeted and clinically applicable platform. In the RT-qPCR cohort, miRNA levels were calculated as relative expression values prior to model application.

To facilitate comparability between platforms, data preprocessing was incorporated into the ML pipeline during model training. Hyperparameter optimization (Grid Search) evaluated alternative preprocessing strategies, including StandardScaler (Z-score normalization), RobustScaler, and a passthrough option (no scaling). Scaling parameters were estimated exclusively from the training data and subsequently applied unchanged to the external validation cohort, preventing information leakage and ensuring consistent model application across datasets. This strategy allows predictions to rely on relative expression patterns rather than platform-specific signal intensities, facilitating model transfer between RNA-seq and RT-qPCR measurements.

For the TabPFN model, external feature scaling was not applied because the algorithm internally performs feature normalization as part of its inference procedure.³⁶

Functional analysis and biological interpretation

Analysis of RNA-seq and differentially expressed genes

To explore potential mRNA targets regulated by the key miRNAs identified, we performed differential expression analysis using RNA-seq data (FASTQ files) from TCGA-OV cohort and normal ovarian tissue from the GTEx database.

Quality assessment of raw reads was conducted using FASTQC with default parameters, followed by adapter trimming and quality filtering using Trimmomatic (v0.39)⁴¹ with default settings, including the removal of adapters, sliding window trimming (4:15), and a minimum read length of 36 bp. High-quality reads were aligned to the human reference genome GRCh38 (Ensembl release 83) using the STAR aligner (v2.6.0c)⁴² with default parameters optimized for paired-end reads, allowing up to two mismatches per read. Gene-level quantification was performed using HTSeq-count (v2.0.4)⁴³ with the “union” mode to assign reads to genes based on GTF annotation (GRCh38 - version 83).

Differential gene expression analysis between tumor and normal samples was performed using the DESeq2 package (v1.40.2)⁴⁴ with adjusted p-values < 0.001 considered statistically significant.

Identification of target genes for relevant miRNAs

Putative target genes of miRNAs selected through the ML analysis were identified using mirDIP³³ and miRTarBase.³⁴ For each database, the list of relevant miRNAs from the ML pipeline was used as input. Only high-confidence interactions (confidence score ≥0.9) were retained from mirDIP, and experimentally validated targets were selected from miRTarBase. To ensure reliability, only mRNA targets common to both databases were considered. These were further filtered by statistically significant negative correlations (Pearson r < 0, p ≤ 0.05) between miRNA and mRNA expression in TCGA-OV samples, calculated using the Hmisc package (v4.7).³⁵ Finally, only mRNAs differentially expressed in HGSOC were retained. The resulting miRNA–mRNA regulatory network was visualized and analyzed using the igraph (v1.5.0) and ggraph (v2.1) R packages.

Analysis of enriched pathways

The top 5000 miRNA target genes ranked by log-fold change were subjected to functional enrichment analysis using the Reactome (2024) database in Enrichr (v3.0). Enrichment significance was assessed by a hypergeometric test with Benjamini–Hochberg correction (adjusted p ≤ 0.05). Redundant pathways were removed using Jaccard similarity filtering (Jaccard ≥0.80), retaining up to two representative pathways per group.

Single-cell target mRNA analysis

To identify cell types enriched for each miRNA’s target genes, differentially expressed targets were analyzed using Seurat’s AddModuleScore function (v4.3)⁴⁵ in R. This method computes module scores by comparing gene set expression with random control genes, enabling the assessment of gene activity across single cells and cell types. We applied this method to single-cell data from Guimarães et al. (2024),⁴⁶ focusing on HGSOC subpopulations.

Statistical and survival analyses

Comparisons of mean scores and clinical variables were performed using the Kruskal–Wallis test, followed by pairwise Wilcoxon rank-sum tests. Survival differences were visualized with Kaplan–Meier curves (survival v3.5; survminer v0.4.9) and assessed by the log-rank test. The surv_cutpoint function was used to determine optimal expression cutoffs that maximized survival separation. Relative risks were estimated through univariate and multivariate Cox proportional hazards models. Clinical survival data were obtained from two cohorts: TCGA (overall survival, OS; and progression-free interval, PFI, from PanCancer Atlas) and INCA (OS: diagnosis to death/last follow-up; PFS: last chemotherapy or medical record to recurrence, with censoring). HRD status was determined using the genomic scar score from the TCGA PanCancer Atlas using XenaBrowser platform. This score sums loss of heterozygosity (LOH), telomeric allelic imbalance (TAI), and large-scale state transitions (LST). Samples with scores ≥42 were classified as HRD-positive, following established thresholds.^47,48

The risk score was derived from the combination of the selected miRNAs using univariate Cox regression coefficients as weights (predict(cox_model, type = “lp”)). The prognostic independence of the selected miRNA panel was evaluated using multivariate Cox regression, incorporating clinical covariates (HRD status, tumor stage, and age) and visualized in a heatmap and florest plot from survminer (v0.4.9) package. Variables with p ≤ 0.05 were considered independently prognostic.

Role of the funding source

This work was supported by the National Council for Scientific and Technological Development (CNPq), Grant No. 09/2022 Research Productivity Fellowship; the Rio de Janeiro State Research Foundation (FAPERJ), Grant No. E-26/201.322/2022 Young Scientist of Our State Program; and a scholarship from the Brazilian Ministry of Health. The funding sources had no role in the study design; data collection, analysis, or interpretation; manuscript writing; or the decision to submit the article for publication. The authors were not paid by any pharmaceutical company or other private agency and declare no involvement of commercial entities. The corresponding author confirms that all authors had full access to the data.

Results

Cohort description and group definition

Clinicopathological and miRNA expression data from the TCGA HGSOC cohort were integrated to develop a ML–based prognostic model (Fig. 1a–c). Prognosis was defined by overall survival (OS) using a 3-year threshold, classifying patients into good prognosis (OS > 3 years; n = 209, 60.9%) and poor prognosis (OS < 3 years and vital status of death; n = 134, 39.1%) groups (HR [Poor vs Good] = 2.5 × 10¹⁰, 95% CI: 0–Inf). Of the 445 patients available, 78 were excluded due to incomplete survival data or because they did not have the cut-off criteria, resulting in a final cohort of 343 patients. Thus, the median OS was 4.7 years (IQR: 3.7–5.0) in the good prognosis group and 1.7 years (IQR: 0.99–2.42) in the poor prognosis group (p < 0.0001; Fig. 2a). Similarly, the median progression-free interval (PFI) (HR [Poor vs Good] = 0.26, 95% CI:0.2–0.34) was 1.8 years (IQR: 1.23–3.63) versus 0.9 years (IQR: 0.55–1.22) (p < 0.0001; Fig. 2b).

Fig. 2.

Prognostic group stratification and development of the miRNA-based prognostic signature in HGSOC. (a) Kaplan–Meier overall survival and (b) progression-free interval curves for patients with HGSOC (TCGA-OV cohort), stratified into good- (n = 209) and poor-prognosis (n = 134) groups using a three-year overall survival cutoff (log-rank test, p < 0.0001). (c) Selection of miRNAs across feature-selection methods. miRNAs were identified by each method: Fast Correlation-Based Filter (FCBF), univariate Cox regression, Least Absolute Shrinkage and Selection Operator (LASSO) regression, and differential expression (DEmiRs) and by their intersections. Vertical bars indicate the number of miRNAs shared between combinations of methods, while the horizontal bars show the total number identified by each method. (d) Correlation heatmap of the 47 selected miRNAs. The matrix displays pairwise Pearson correlation coefficients, each dot size represents the number of genes associated with each pathway, while dot color reflects the statistical significance (adjusted p-value) (red denotes positive and blue denotes negative correlations), highlighting co-expression patterns within the selected signature.

Prognostic groups also showed significant associations with clinical features, patients with good prognosis were younger at diagnosis (p < 0.0001) and exhibited higher tumor mutation burden, greater platinum sensitivity, better treatment response, and higher HRD scores (all p < 0.05; Supplementary Table S2).

Selection of miRNAs associated with prognosis

To establish a robust miRNA panel, we analyzed publicly available datasets profiling miRNA expression in both tumor and normal ovarian samples, including primary tissues and cell lines. Differential expression analyses across three independent studies identified 737 and 1301 DEmiRNAs from exosomal and total RNA data,28, 29, 30 respectively, and 49 and 279 DEmiRNAs from two additional datasets containing paired tumor and normal tissues. miRNAs consistently dysregulated in at least two datasets spanning both tissue and cell line sources were retained as candidate biomarkers (Supplementary Tables S1a,b, and S5).

To further refine this list, we applied three complementary feature selection methods, such as FCBF, univariate Cox regression, and LASSO regression to reduce dimensionality and identify prognostically relevant miRNAs. These approaches selected 6, 15, and 80 candidates, respectively. Integrating these results with the differential expression analysis yielded 49 miRNAs consistently identified by at least two of the four methods from an initial set of ∼1500 candidates (Fig. 2c and d and Supplementary Table S6), highlighting their potential prognostic value.

Comparison of machine learning models for prognostic classification

Several ML algorithms were evaluated for prognostic classification using the expression profiles of the 49 selected miRNAs. Among them, the Logistic Regression (LR) classifier achieved the best performance (Supplementary Table S7). Trained with cross-validation, the LR model (Model 1) reached an AUC of 0.718 [95% CI: 0.569–0.867], balanced accuracy of 0.643 [95% CI: 0.484–0.802], recall of 0.714 [95% CI: 0.565–0.864], and F1-score of 0.606 [95% CI: 0.444–0.768], demonstrating stronger predictive power (Supplementary Fig. S2a).

To interpret the contribution of individual miRNAs, we applied SHAP, which identified members of the miR-200 family, miR-143, and miR-106b as key contributors, all consistently deregulated across tumor tissue and exosomal datasets (Supplementary Fig. S1a and b). Model optimization through iterative evaluation of miRNA subsets (Supplementary Fig. S2b) identified a 10-miRNA panel that maximized predictive accuracy (AUC: 0.704 [95% CI: 0.553–0.855]) (Model 2; Supplementary Table S7). The comparative analysis of other algorithms is presented in Supplementary Table S8.

Integrating clinical variables, age and MUC16 levels into the 10-miRNA model further improved performance (Model 3 - AUC = 0.741 [95% CI: 0.596–0.887], balanced accuracy = 0.702 [95% CI: 0.551–0.854], recall = 0.786 [95% CI: 0.650–0.922], F1-score = 0.667 [95% CI: 0.510–0.823]) (Fig. 3a; Supplementary Fig. S2c and Supplementary Table S7). In contrast, adding HRD status did not yield additional gains in predictive accuracy or AUC (Supplementary Fig. S3a–c). To assess the incremental value of this integration, we compared ROC curves for each of the biomarkers (10 miRNAs) individually; none achieved an AUC above 0.60 (Fig. 3b), underscoring the advantage of combining features within a multivariate model. Finally, Kaplan–Meier survival analyses confirmed the prognostic relevance of most panel components: eight of the 10 miRNAs were significantly associated with overall survival, except hsa-miR-151a-3p and hsa-miR-106b-3p (Supplementary Table S9 and Supplementary Fig. S4). Collectively, these results demonstrate that while individual miRNAs hold prognostic value, their predictive strength is markedly enhanced when integrated into a combined multivariate model.

Fig. 3.

Prognostic performance and feature importance of the miRNA-based model in HGSOC. (a) ROC curves comparing predictive performance: 10 miRNAs (AUC = 0.704), 10 miRNAs + clinical data (AUC = 0.741), and the RFE-optimized 9-miRNA + clinical model (AUC = 0.762). (b) ROC curves for individual miRNAs showing sensitivity-specificity performance. (c) Confusion matrix summarizing model predictions for good- and poor-prognosis groups. (d) TabPFN model calibration plot (reliability analysis). The LOESS curve (blue) illustrates the relationship between predicted probabilities and observed outcomes (Brier score: 0.199; ECE: 0.042). (e) Kaplan–Meier analysis of overall survival based on model-predicted prognosis, with optimal cut-points defined by the surv_cutpoint R function (log-rank test, p < 0.013). (f) SHAP summary plot illustrating the 12 most influential features contributing to model predictions. (g) Multivariate Cox proportional hazards analysis for up to 5-year overall survival, incorporating the 10-miRNA-derived risk score with clinical and other molecular covariates. (h) Distribution of patient risk scores generated by the model alongside a heatmap of normalized expression levels for the 9 miRNAs. Colors represent scaled expression values (log2 + 1), illustrating distinct molecular profiles between low- and high-risk patient groups. MUC16 values correspond to log₂-normalized mRNA expression. Patients were classified as early stage (I–II) or advanced stage (III–IV) and according to the age at diagnosis (>50 or <50 years).

TabPFN model excelled in predicting HGSOC prognosis

Model 4, a transformer-based probabilistic foundation model (TabPFN) for tabular data designed to small datasets,³⁶ accurately classified patients according to prognostic outcome (Supplementary Table S7, Fig. 3a), despite using a reduced number of features identified by RFE (AUC = 0.762 [95% CI: 0.621–0.903], Balanced Accuracy = 0.714 [95% CI: 0.565–0.864] and Recall = 0.857 [95% CI: 0.741–0.973]). The model correctly identified 12 of 14 patients with unfavorable prognosis and 12 of 21 with favorable prognosis (Fig. 3c). The calibration curve shows good agreement between predicted probabilities and observed outcomes, with the LOESS trend closely following the ideal diagonal line. The Brier score (0.199) and Expected Calibration Error (ECE = 0.042) indicate that the model provides reasonably accurate and well-calibrated probability estimates on the test dataset (Fig. 3d). Kaplan–Meier analysis based on model-derived classifications showed a highly significant difference in overall survival between groups (log-rank, p = 0.013; HR [Poor vs Good]: 3.22 (95% CI: 1.23–8.43); Fig. 3e). Model explainability analysis using SHAP revealed the relative contribution of each variable, highlighting the top 9 miRNAs together with age, figo stage and MUC16 as the most influential predictors (Fig. 3f).

Multivariate Cox regression confirmed that the 9-miRNA panel is an independent prognostic factor, maintaining predictive power after adjustment for age, tumor stage, MUC16 levels, HRD status, and molecular subtype (Global p-value (log-rank) < 0.001; Fig. 3g). The prognosis score showed a hazard ratio (HR = 2.38; p-value < 0.001). Risk scores derived from the 9 miRNAs revealed distinct expression patterns across the prognostic spectrum (Fig. 3h). Notably, hsa-miR-150-5p, hsa-miR-143-3p, and hsa-miR-23a-3p, displayed clear differential expression between low- and high-risk groups, reinforcing their contribution to patient stratification.

Molecular and clinical correlates of the miRNA signature

To further explore clinical relevance, we examined associations between the 10 miRNAs and major molecular and clinical features. Significant correlations were observed with TP53 mutation status (hsa-miR-150-5p, hsa-miR-23a-3p, hsa-miR-151a-3p) and HRD status (hsa-miR-23a-3p, hsa-miR-200a-5p, hsa-miR-151a-3p and hsa-mir-143-3p) (Supplementary Fig. S5a and b and Supplementary Table S9).

Finally, analysis across the four TCGA mRNA-based molecular subtypes of HGSOC (differentiated, immunoreactive, mesenchymal, and proliferative) showed that hsa-miR-485-3p (p < 0.0001), hsa-miR-143-3p (p = 0.0007), hsa-miR-150-5p (p = 0.0009), hsa-miR-149-5p (p = 0.0009), and hsa-miR-106b-3p (p = 0.0093) were significantly upregulated in the mesenchymal and proliferative subtypes, both associated with poorer prognosis (Supplementary Fig. S5c).

Functional insights into miRNAs

To explore the biological roles of the 10 prognostic miRNAs in ovarian cancer, we identified their putative mRNA targets by integrating tumor transcriptomes with normal ovarian tissue data. Differential expression analysis revealed 3,311 upregulated and 1,720 downregulated genes in tumors (Fig. 4a; Supplementary Table S10). Among the upregulated genes were TPD52, KIF11, and FOXM1; downregulated genes included GCKR and AASS. Using mirDIP and miRTarBase, we identified 4,711 predicted or validated targets for the 10 miRNAs, each regulating 30–100 genes. Cross-referencing these with differentially expressed genes yielded 1,173 dysregulated targets (Fig. 4b), such as FOXM1, MKI67, MYB, MMP9, and E2F2 regulated by hsa-miR-149-5p (Supplementary Table S11).

Fig. 4.

Biological characterization of miRNAs identified as prognostic biomarkers. (a) Heatmap with hierarchical clustering of differentially expressed genes (DEGs) between tumor and normal ovarian samples. The matrix displays log-normalized expression values for DEGs identified with an adjusted p < 0.01. Hierarchical clustering was performed using Pearson correlation. Colors represent scaled expression values (Log10 across genes). For visualization, raw gene counts were transformed using variance-stabilizing transformation implemented in DESeq2 to generate heatmaps. (b) miRNA–mRNA interaction network showing predicted interactions between the 10 selected miRNAs (central nodes with white borders) and their differentially expressed target genes (colored peripheral nodes) in ovarian cancer. Node color represents gene expression changes (log2foldchange), and edge thickness indicates interaction confidence. (c) Pathway enrichment analysis of miRNA target genes. The scatter plot depicts significantly enriched biological pathways for the predicted targets of each differentially expressed miRNA. Dot size represents the number of genes associated with each pathway, and color intensity reflects the adjusted p-value (FDR-adjusted). (d) Box plots represent the Log2 expression levels of 10 candidate miRNAs in Non-cancer Controls, Ovarian Cancer, Borderline Ovarian Tumor, Benign Ovarian Disease, and Other Solid Cancers. Global differences were assessed using the Kruskal–Wallis test, followed by pairwise Wilcoxon tests. (e) Expression levels of the 10 miRNAs in exosomal and total-cell fractions from ovarian cancer cell lines. Boxplots display log₂-normalized expression levels in exosomes (tumor_exo) and whole cells (tumor_cell). Comparisons were performed using the Wilcoxon test. ∗p < 0.05; ∗∗p < 0.01; ∗∗∗p < 0.001; ∗∗∗∗p < 0.0001.

Pathway enrichment analysis revealed significant associations with cancer-related processes (Fig. 4c). For instance, hsa-miR-485-3p (182 targets) was enriched in programmed cell death and interferon α/β signaling, while hsa-miR-205-5p (207 targets) was linked to extracellular matrix organization and PI3K/AKT signaling (Supplementary Table S12).

In serum samples, all analyzed miRNAs were detectable, with mean expression values ranging from 2.05 to 4.60. Comparative analysis between non-cancer controls and ovarian cancer samples revealed that 8 of the 10 miRNAs were significantly more expressed in tumor samples, with the exception of hsa-miR-187-3p and hsa-miR-485-3p (Fig. 4d). We further assessed miRNA expression across different ovarian cancer stages and observed no statistically significant differences, suggesting stage-independent detectability and supporting the potential applicability of the model across patients with varying disease stages (Supplementary Fig. S6). To assess translational potential, we compared exosomal and intracellular expression in cell lines and in serum of ovarian cancer patients. The hsa-miR-150-5p showed significantly higher exosomal expression (p = 0.02) and hsa-miR-205-5p showed a similar trend, suggesting that some prognostic miRNAs are secreted through extracellular vesicles and may serve as non-invasive biomarkers (Fig. 4e).

To infer the functional context of the 10-miRNA panel, we analyzed the expression of their predicted mRNA targets using single-cell RNA sequencing (scRNA-seq) data from HGSOC samples (Supplementary Fig. S7a). Since miRNAs repress target gene expression, identifying where these targets are most highly expressed can indicate the cell types in which the miRNAs exert greater regulatory effects. Targets of hsa-miR-187-3p, hsa-miR-205-5p, hsa-miR-200a-5p, and hsa-miR-151a-3p were enriched in stromal and mesenchymal cells; hsa-miR-485-3p, hsa-miR-143-3p, hsa-miR-23a-3p, and hsa-miR-150-5p in malignant epithelial cells; hsa-miR-149-5p; and hsa-miR-106b-3p in myeloid cells (Supplementary Fig. S7b).

Based on these findings, we further investigated the biological significance of the two miRNAs hsa-mir-150-5p and hsa-mir-205-5p in HGSOC. hsa-miR-205-5p correlated with venous (p = 0.022; Fig. 5a) and lymphatic invasion (p = 0.038; Fig. 5b), with its targets enriched in endothelial cells (Fig. 5c). Among the top predicted targets KLF15, SOX12, and PARD6B (Fig. 5d), the latter exhibited a strong inverse correlation with hsa-miR-205-5p (r² = −0.23, p < 0.0001; Fig. 5e), was markedly upregulated in tumors (Fig. 5f). Furthermore, low hsa-miR-205-5p level correlated with reduced overall survival (Fig. 5g), high level of the target PARD6B correlated with worse outcome (Fig. 5h). The hsa-miR-150-5p is related to higher tumor mutational burden (p = 0.019; Fig. 5i), TP53 mutations (p = 0.028; Fig. 5j), and enrichment in the immunoreactive subtype relative to the proliferative subtype (Fig. 5k). Its predicted targets (KCNF1, FGFR3, and GATA4) were mainly expressed in malignant cells (Fig. 5l,m). KCNF1 was inversely correlated with hsa-miR-150-5p (r² = −0.25, p < 0.0001; Fig. 5n), and upregulated in tumors (Fig. 5o). The hsa-miR-150-5p was associated with poorer outcomes (Fig. 5p), and the KCNF1 downregulation was correlated with poor prognosis (Fig. 5q), supporting a tumor-suppressive role for hsa-miR-150-5p via repression of oncogenic targets such as KCNF1.

Fig. 5.

Prognostic relevance and functional characterization of hsa-miR-205-5p and hsa-miR-150-5p. (a and b) Boxplots showing expression levels of hsa-miR-205-5p in samples with venous (a) and lymphatic (b) invasion. (c) Violin plot displaying the inferred activity score of hsa-miR-205-5p obtained from single-cell RNA-seq analysis. (d) miRNA–mRNA interaction network centered on hsa-miR-205-5p. Node color indicates gene expression changes (red = upregulated, blue = downregulated in tumors), and edge thickness represents interaction confidence. (e) Correlation between PARD6B and hsa-miR-205-5p expression (r² = −0.23, p < 0.0001). (f) Boxplot showing PARD6B overexpression in tumor compared with normal ovarian tissues. (g and h) Kaplan–Meier survival curves of TCGA-OV patients stratified by high vs. low expression of hsa-miR-205-5p (log-rank test, p = 0.005) (g) and PARD6B (log-rank test, p = 0.016) (h). (i–k) Boxplots showing higher hsa-miR-150-5p expression according to (i) tumor mutational burden (TMB) status, (j) TP53 mutation status, and (k) molecular subtype. (l) Violin plot displaying the inferred activity score of hsa-miR-150-5p from single-cell RNA-seq data. (m) miRNA–mRNA interaction network centered on hsa-miR-150-5p. Node color and edge thickness are defined as in panel D. (n) Correlation between KCNF1 and hsa-miR-150-5p (r² = −0.25, p < 0.0001). (o) Boxplot showing significant overexpression of KCNF1 in tumor tissues compared with normal ovarian tissues. (p and q) Kaplan–Meier survival curves for TCGA-OV patients stratified by high vs. low expression of hsa-miR-150-5p (p = 0.0052) (p) and KCNF1 (log-rank test, p = 0.009) (q). Statistical significance was assessed using the Kruskal–Wallis test followed by pairwise Wilcoxon tests. ns, not significant. ∗p < 0.05; ∗∗p < 0.01; ∗∗∗p < 0.001; ∗∗∗∗p < 0.0001.

Collectively, these analyses provide biological interpretability to the predictive model, reinforcing that the identified miRNAs are not only statistically robust but also functionally relevant. By linking miRNA expression to cellular context, target regulation, and clinical outcomes, this integrative framework enhances confidence in the model’s predictions and supports its attractiveness for translational and clinical application.

External validation cohort

To evaluate the generalizability and clinical applicability of the 9-miRNA prognostic model (Model 4), we performed an external validation using an independent cohort of Brazilian patients treated at the National Cancer Institute (INCA) (Supplementary Tables S3 and S13). Unlike the discovery and test cohort, miRNA expression in this set was measured by RT-qPCR, a more accessible and cost-effective technique, underscoring the potential for clinical translation of the model. MUC16 expression levels were also replaced by CA125 protein levels (U/mL). Notably, hsa-miR-106-5p and hsa-miR-149-5p remained significantly associated with overall survival in the INCA cohort (Supplementary Fig. S8), reinforcing their biological relevance and contribution to the robustness of the 9-miRNA prognostic signature.

As the prognostic model was trained on adjuvant-treated tumors. Primary external validation was performed in this subgroup (n = 31). The binary classification model for 3-year overall survival correctly classified 23 of 31 patients, yielding an overall accuracy of 74.2%, with high specificity for favorable outcomes (21/25, 84%) but low sensitivity for poor prognosis (2/6, 33.3%), yielding a balanced accuracy of 0.587 [95% CI: 0.4000–0.7933], with an AUC of 0.48 [95% CI: 0.1660–0.7940] (Supplementary Fig. S9a and b). When the analysis was restricted to late-stage adjuvant-treated tumors (stage III–IV, n = 23), which are more clinically aligned with the training population, performance improved modestly, with an overall accuracy of 78.2%, a balanced accuracy of 0.671 [95% CI: 0.4211–0.9474], and an AUC of 0.58 [95% CI 0.1337–1.0000] (Fig. 6a and b). In contrast, performance declined further in the neoadjuvant chemotherapy (NACT) cohort (n = 9), with an overall accuracy of 55.6% (Supplementary Fig. S9c and d), indicating reduced generalizability across treatment contexts and highlighting the impact of biological and clinical differences relative to the training population.

Fig. 6.

External validation of the miRNA-based prognostic model in the INCA cohort (model 4 - TabPFN). (a) Confusion matrix restricted to late-stage (III–IV) adjuvant-treated patients (n = 23). (b) ROC curve for the stage III–IV subgroup. (c) Forest plot of multivariable Cox proportional hazards analysis. The miRNA risk score remained an independent prognostic factor (HR = 2.62, 95% CI: 1.21–5.60; log-rank p = 0.014) after adjustment for tumor stage, age, and CA125 levels (C-index = 0.85). (d) Distribution of continuous risk scores and clinical features (status, age, and CA125) alongside a heatmap of the 9-miRNA expression profiles across patients. (e) Kaplan–Meier survival curves stratified by the prognosis group predicted by the model (log-rank, p = 0.019). (f) Kaplan–Meier survival curves stratified by risk groups derived from the COXPH model (log-rank, p = 0.043). AUC [95% CI], balanced accuracy [95% CI], accuracy [95% CI], sensitivity [95% CI], specificity [95% CI], and negative predictive value (NPV) [95% CI].

Despite the limited performance of the binary framework, the continuous 9-miRNA-based risk score demonstrated robust prognostic value. In multivariable Cox proportional hazards analysis, the risk score remained independently associated with overall survival after adjustment for tumor stage, age, and CA125 levels (HR = 2.62, 95% CI: 1.21–5.60; log-rank, p = 0.014; Fig. 6e). The full multivariable model showed strong discriminative performance, with a concordance index of 0.85. The distribution of the risk score revealed substantial inter-patient variability, supporting its capacity to capture underlying heterogeneity in disease progression (Fig. 6f). Kaplan–Meier analysis further supported the prognostic relevance of the signature. Stratification based on the continuous risk score using predefined cutoffs showed significant differences in overall survival between groups (log-rank, p = 0.019; HR [Poor vs Good] = 10.31 (95% CI: 0.93–114.85) Fig. 6g). Notably, stratification based on the binary classification provided even clearer separation of survival curves (log-rank, p = 0.043; HR [Low vs High Risk] = 0.23 (95% CI: 0.05–1.08) Fig. 6h), suggesting that the discretized model effectively captures clinically relevant risk groups in this cohort.

Discussion

In this study, we developed and validated an explainable ML model integrating molecular and clinical features to predict overall survival in patients with HGSOC. The resulting 9-miRNA signature demonstrated the ability to stratify patients by risk and remained an independent prognostic factor after adjustment for established clinical variables.

Importantly, external validation in an independent Brazilian cohort demonstrated that a model trained using RNA-Seq data can be successfully transferred to a clinically accessible RT-qPCR platform, supporting both its biological robustness and technical portability across measurement platforms. Notably, incorporating data scaling within the training pipeline enabled predictions to rely on relative expression patterns rather than platform-specific intensities, reducing the impact of technical variability and facilitating cross-platform applicability.⁴⁹

Although complete measurement of all miRNAs in the panel is required for prediction, which may pose challenges in cases of missing data or technical variability, the relatively small number of markers enhances the feasibility of future clinical implementation.

Because the primary model was developed using a 3-year overall survival threshold to identify patients at risk of early adverse outcomes, survival was initially treated as a binary endpoint. Although this approach simplifies prediction, it reduces the information contained in continuous survival times. To address this limitation, we performed complementary time-to-event analyses using Cox proportional hazards regression, which confirmed that the model-derived risk score remained significantly associated with overall survival, supporting the robustness and clinical relevance of the miRNA signature.

In the external validation cohort, we observed a discrepancy between the performance of the binary classification model and the continuous risk score. Although the classification model showed limited discriminative ability when assessed by AUC, survival-based analyses demonstrated that the continuous risk score retained strong prognostic value and remained independently associated with overall survival after adjustment for clinical variables. Notably, stratification based on the binary classification provided improved separation of survival curves compared with predefined cutoffs of the continuous score, suggesting that discretization may better capture clinically meaningful risk groups. From a clinical perspective, this is particularly relevant, as treatment decisions are typically guided by categorical risk stratification rather than continuous measures. The reduced performance observed in the NACT cohort likely reflects underlying biological and clinical differences relative to the training population, highlighting the impact of treatment-related heterogeneity on model generalizability.

Together, these findings indicate that, while continuous modeling captures the full spectrum of risk, its translation into discrete groups may enhance interpretability and support clinical decision-making, particularly in heterogeneous or modestly sized cohorts.

Our findings highlight the added value of miRNA-based molecular features in improving prognostic precision beyond conventional markers such as age and CA-125. Although CA-125 remains widely used in ovarian cancer management, its limited sensitivity in early stages and poor specificity underscore the need for complementary biomarkers that improve early detection and risk stratification. Existing multimarker algorithms such as ROMA, Ova1®, and Overa® offer incremental improvements but are influenced by menopausal status or exhibit relatively high false-positive rates.50, 51, 52 These limitations underscore the need for novel molecular tools capable of achieving more balanced predictive performance.

The integration of high-dimensional transcriptomic data with ML enabled the identification of a concise and biologically meaningful miRNA panel. While the initial modeling framework included a larger set of candidate miRNAs, feature selection guided by SHAP explainability analysis, clinical correlation and RFE reduced the model to a compact 9-miRNA signature without loss of predictive accuracy. Notably, several miRNAs in the panel exhibited higher expression in ovarian cancer samples compared with non-cancer controls, supporting their biological relevance in disease progression.

The explainable framework also enabled the identification of key miRNAs driving model predictions and linking their contributions to cancer-related pathways. Among the most influential features were hsa-miR-205-5p and hsa-miR-150-5p, both previously implicated in ovarian cancer biology. miR-205 has been associated with epithelial plasticity, invasion, and activation of PI3K–AKT signaling, promoting tumor progression and extracellular-matrix remodeling in ovarian and other epithelial cancers.53, 54, 55 Several circulating miRNAs, particularly members of the miR-200 family, are associated with aggressive disease phenotypes and oncogenic processes, including epithelial–mesenchymal transition, invasion, and chemoresistance.56, 57, 58 In contrast, miR-150-5p has been reported to exert tumor-suppressive functions, regulating proliferation, migration, and immune-related pathways in ovarian cancer and other malignancies.^59,60 Consistent with these observations, miR-205-5p expression was associated with vascular invasion, whereas miR-150-5p expression correlated with improved clinical outcomes and enrichment in the immunoreactive molecular subtype of HGSOC. Linking model predictions to biologically plausible mechanisms strengthens confidence in the model and supports its interpretability for clinical applications.

Comparative analyses further emphasize the robustness of our model. Unlike the MiROvaR 35-miRNA signature, whose performance declined when restricted to HGSOC cases (AUC = 0.58 ± 0.02),⁶¹ our Model 4 (9 miRNAs plus three clinical variables) maintained stable discrimination within this histological subtype. The use of the TabPFN architecture likely contributed to this stability by enabling strong predictive performance even in relatively small clinical datasets.³⁶

Consistent with observations in cross-cohort validation studies,^55,61,62 model performance was attenuated in the external dataset. Notably, restricting the analysis to stage III–IV tumors, more closely reflecting the stage distribution of the discovery cohort, resulted in improved performance, underscoring the influence of cohort composition on predictive behavior. Several factors likely contributed to this reduction in performance. HGSOC is a biologically and clinically heterogeneous disease,⁶³ which can limit the transferability of molecular predictors across populations. In addition, the Brazilian cohort differed from the TCGA dataset⁶⁴ in key aspects, including a higher proportion of early-stage cases and differences in age distribution, both of which may influence the relationship between molecular features and clinical outcomes.

From a translational perspective, this miRNA panel may also be suitable for liquid biopsy–based applications. Circulating miRNAs are highly stable in biological fluids and have been reported as independent prognostic indicators in ovarian cancer.^56,65 The detectability of all miRNAs included in our panel in publicly available serum datasets, including early and late stage tumor samples, further supports the feasibility of developing minimally invasive assays for disease detection, monitoring and risk stratification.⁶⁶ In addition, implementation using RT-qPCR or digital PCR technologies could facilitate adoption in clinical laboratories and enable scalable clinical deployment.67, 68, 69, 70

This study has several limitations. First, the external validation cohort remains relatively small, which results in wider confidence intervals and reduced stability of classification-based performance metrics. Second, the model was trained using RNA-seq data from a predominantly advanced-stage North American cohort, whereas external validation was performed in a Brazilian cohort using RT-qPCR, introducing both biological and technical variability. These differences represent a domain shift that likely contributed to the reduced performance observed in binary classification metrics.

At the same time, this cross-platform validation represents a stringent test of model robustness. By applying scaling parameters derived from the training dataset to the external validation cohort, predictions relied on relative expression patterns rather than absolute platform-specific values, minimizing technical bias. The preservation of prognostic association under these conditions supports the biological relevance of the miRNA signature and reflects the challenges expected in real-world clinical implementation.

In summary, we present an interpretable miRNA-based prognostic model for HGSOC that integrates molecular and clinical features within an explainable ML framework. The use of a concise miRNA panel and validation using RT-qPCR measurements highlights the potential feasibility of translating this approach into clinically applicable assays. Future studies involving larger multi-center cohorts and prospective evaluation will be essential to determine the clinical utility of this model for risk stratification and personalized management of ovarian cancer.

Contributors

All authors contributed to the drafting, reviewing, and approval of the final manuscript. C.Е.Т. developed and implemented the ML models, performed bioinformatics analyses, and contributed to data interpretation. C.B.P.C. assisted with the characterization of the Brazilian cohort. C.E.T., N.T., G.B., A.S., L.M., and D.J.G.P. collected, processed, and prepared the biological samples and clinicopathological data for analysis. Validation experiments using RT-qPCR were conducted by C.E.T., N.T., G.B., A.S. Additional bioinformatic analysis was done by H.Z. J.B.P.V. contributed to the validation efforts and provided critical discussions. M.B. and A.D.P.C.F. supervised the development and implementation of ML models. M.B. conceived and designed the study, secured funding, and provided supervision.

Data sharing statement

Data related to gene expression and survival analysis were obtained from The Cancer Genome Atlas (TCGA) public repository through the GDC portal (https://portal.gdc.cancer.gov/) and dbGaP (phs000178.v11.p8). Data for the analysis of differentially expressed miRNAs were obtained from the Gene Expression Omnibus (GEO) (GSE103708, GSE83693, GSE47841 and GSE106817). All data used for model training were obtained from the Gene Expression Omnibus (GEO) (GSE164767). The code for reproducing the figures in this study is available at the GitHub: https://github.com/bioinformatics-inca/prognosis_Predictor_miRNA.

AI statement

The author(s) used ChatGPT to improve text readability and subsequently reviewed and edited the content, taking full responsibility for the publication.

Declaration of interests

C.E.T., N.T., G.B., A.S., A.D.P.C.F. and M.B. are inventors on a pending patent application filed by the Brazilian National Cancer Institute and registered at the Brazilian National Institute of Industrial Property (INPI), related to the miRNA-based prognostic model for ovarian cancer described in this study (BR 10 2025 006150 3). All other authors declare no competing interests.