Machine Learning Approach Identifies miRNA Biomarkers for Post Surgical Patient Stratification in Prostate Cancer

ABSTRACT

Introduction

Effective management of post‐prostate cancer is hindered by the limitations of current prognostic tools in accurately assessing disease aggressiveness. Radical prostatectomy remains a standard treatment, but some patients develop biochemical recurrence and metastasis, underscoring the need for improved postsurgical prognostic tools.

Methods

This investigation involved sequencing data derived from 38 matched prostate cancer patients who had undergone RP. Initial statistical analysis helped identify the most significant miRNAs, which were further subjected to unsupervised clustering and stepwise selection. A linear discriminant analysis (LDA) model was then trained and tested using a miRNA combination method to pinpoint biomarkers predictive of metastasis.

Results

Out of 1123 miRNAs initially identified, 519 were selected as high‐confidence candidates. Parametric analysis of these miRNAs discerned 41 that effectively distinguished between patients who developed metastasis postoperatively and those who did not. Utilizing LDA, this study harnessed 41 miRNAs in a combinatorial approach, identifying eight key miRNAs (hsa‐miR‐106b‐3p, hsa‐miR‐769‐5p, hsa‐miR‐182‐5p, hsa‐miR‐194‐5p, hsa‐miR‐345‐5p, hsa‐miR‐183‐3p, hsa‐miR‐200a‐3p, hsa‐miR‐301a‐3p) that collectively stratified the metastatic group from control with up to 91% accuracy. This model's effectiveness was supported by a receiver operating characteristic analysis, demonstrating an area under the curve of 80% or higher for the best miRNA combinations. Notably, the performance of this eight‐miRNA panel was consistent with CAPRA‐based risk stratification.

Conclusion

Our study presents a miRNA‐based machine learning model that distinguishes metastatic from non‐metastatic prostate cancer patients following surgery. The panel's alignment with CAPRA underscores its clinical relevance and highlights its potential for integration into future clinical frameworks.

1. Introduction

The management of prostate cancer, characterized by a spectrum from indolent to aggressively metastatic forms, presents a significant clinical challenge [1, 2]. The ability to distinguish between these forms is crucial, as it dictates the therapeutic approach and significantly impacts patient quality of life due to the side effects associated with prostate cancer treatments [3].

Radical prostatectomy (RP) is a standard treatment for localized prostate cancer; however, a subset of patients experience biochemical recurrence and eventual metastasis, highlighting the need for better prognostic tools to guide postsurgical management [4, 5]. While clinical parameters such as Gleason score, PSA kinetics, and pathological staging provide some prognostic insight, they often fail to reliably distinguish between patients at low versus high risk of metastasis, leading to overtreatment or delayed therapeutic intervention [6, 7, 8]. To address these gaps, tissue‐based biomarkers such as the Decipher genomic classifier, Ki‐67, and PTEN loss have been explored for their ability to predict post‐RP progression. However, their predictive accuracy remains variable, and no single marker has been universally adopted in clinical practice [9].

Emerging evidence suggests that micro‐RNA (miRNAs) may serve as promising biomarkers for post‐RP metastasis risk stratification, offering a more refined molecular signature of disease progression [10]. Unlike traditional biomarkers, miRNAs play a direct regulatory role in tumor progression, metastasis, and treatment resistance, making them potential candidates for both prognosis and therapeutic targeting. Given their stability in formalin‐fixed paraffin‐embedded (FFPE) tissues, blood, and urine, miRNAs also present a minimally invasive and highly scalable approach for risk assessment. This growing body of research underscores the need for systematic miRNA profiling to identify robust predictive markers that can enhance post‐RP risk stratification and improve clinical decision‐making in prostate cancer [11, 12, 13]. These insights support the potential of miRNAs as molecular prognostic tools that may complement or enhance existing biomarkers. However, no miRNA panels have yet transitioned into clinical practice for prostate cancer prognosis, reflecting a gap in effectively harnessing this molecular data for patient stratification.

To address this, we apply machine learning approaches to miRNA profiles to identify key discriminative biomarkers for metastatic risk. The study utilizes advanced computational techniques to analyze high‐throughput sequencing data, allowing for a data‐driven approach to patient stratification.

2. Materials and Methods

2.1. Subject Selection

In summary, this study utilized an ethics board‐approved database of 585 patients treated with RP for localized prostate cancer between 1990 and 2000. Patients who developed metastases were matched to those without evidence of disease recurrence based on Gleason score and follow‐up duration. Additional matching criteria, such as pathological stage and PSA levels, were excluded due to limited match availability [14].

For each subject, FFPE tissue from the RP specimen was identified. All slides were re‐reviewed by a pathologist, and a representative slide was selected for each specimen, with cancerous areas clearly marked. The corresponding tumor block from which the slide originated was then retrieved. In each case, the paraffin block chosen for analysis contained the largest tumor focus, as observed on the histologic slide. Tissue was macro‐dissected, and cores were extracted from the area with the highest tumor‐to‐stroma ratio to maximize tumor content. Specifically, two FFPE cores (1 mm diameter, up to 3 mm length after trimming) were used for total RNA extraction. For Gleason score matching, we considered cases with heterogeneous Gleason patterns and selected the highest Gleason score region within the block to ensure consistency in tumor grading across samples.

2.2. Sample Preparation

In summary, total RNA was extracted from two FFPE tissue cores per sample, processed using the Recover All Total Nucleic Acid Isolation Kit (ThermoFisher Scientific). RNA concentration and integrity were verified with a Qubit fluorometer and Agilent Bioanalyzer. Small RNA was enriched from 1 μg of total RNA using the Magnetic Beads Cleanup Module (ThermoFisher Scientific), with quality and quantity assessments performed on an Agilent Bioanalyzer. miRNA isolation was conducted for QPCR analysis of the miRNAs identified in this study. For a detailed isolation protocol, a similar isolation method was used in this study [14].

2.3. QPCR

2.3.1. miRCURY LNA miRNA Custom PCR Panel

The miRCURY LNA miRNA Custom PCR panel was obtained from Qiagen. This panel targets the following miRNAs: hsa‐miR‐106‐3p, hsa‐miR‐765‐5p, hsa‐miR‐182‐5p, hsa‐miR‐194‐5p, hsa‐miR‐345‐5p, hsa‐miR‐183‐3p, hsa‐miR‐200a‐3p, hsa‐miR‐301a‐3p, and hsa‐miR‐221‐3p. It includes the most stably expressed miRNAs as the reference miRNAs: hsa‐miR‐28‐5p, hsa‐miR‐23b‐3p, hsa‐miR‐28‐3p, hsa‐miR‐361‐3p, and hsa‐miR‐152‐3p. Additionally, UniSp3 was used as an inter‐plate calibrator and UniSp6 as a reverse transcription (RT) control. RNA Preparation and Reverse Transcription: Total RNA (10 ng) was reverse‐transcribed in a 10 µL reaction using the miRCURY LNA RT Kit (Qiagen). PCR Amplification: Quantitative PCR was performed in a total volume of 10 µL. Each reaction mixture contained 5 µL of 2× miRCURY SYBR Green Master Mix (Qiagen), which includes ROX reference dye, 4 µL of diluted cDNA template, and 0.5 µL of RNase‐free water. PCR conditions: an initial denaturation step at 95°C for 2 min, followed by 40 cycles of 95°C for 10 s and 60°C for 30 s. Melting curve analysis was performed by gradually increasing the temperature from 60°C to 95°C. Data Analysis: Raw Ct (cycle threshold) values were exported to Qiagen's web‐based analysis software, GeneGlobe where the analysis was conducted by normalizing the reference miRNAs.

2.4. Alignment and Quantification

The alignment and quantification of miRNA reads were conducted using the COMPSRA suite, which aligns reads to the hg38 genome and generates count data specific to miRNA species [15]. This processing yielded raw counts for each miRNA across all samples. Processed counts were filtered to ensure cleaner starting data. Only miRNAs present in all samples and with a minimum of 10 reads in at least 50% of samples from each group (metastatic and non‐metastatic) were retained. Variance among samples within each group was calculated using coefficient of variation to control for excessive within‐group variability. miRNAs with high variance within either the metastatic or non‐metastatic groups were excluded to maintain consistency across samples. The resulting filtered matrix was normalized using the transcripts per million (TPM) method, allowing for accurate cross‐sample comparisons of miRNA expression levels.

2.5. Parametric Analysis

The data set for analysis consisted of the 41 miRNAs normalized by TPM. To ensure the appropriateness of a T‐test, the assumptions of normality and equal variances were tested using Q–Q plots. An independent two‐sample T‐test was performed to assess the significance of differences between the metastatic and non‐metastatic groups. This analysis was conducted using Python and the Spicy package. The T‐test was two‐tailed and used a significance level of α = 0.01 and a fold change cut off of 1.5‐fold.

2.6. Distance‐Matrix and Hierarchical Analysis

The data set for hierarchical clustering analysis utilized the 41 normalized miRNAs that showed a clear stratification between 18 metastatic and 18 matched non‐metastatic samples. A distance matrix was computed to quantify the similarity (or dissimilarity) between samples. The Manhattan distance metric was used, calculated as the sum of the absolute differences for each feature. Python script was employed for this calculation, and the resulting distance matrix was imported into the Orange suite [16]. Hierarchical clustering was performed utilizing ward linkage to group samples based on similarity. To visualize clustering results, a dendrogram was generated using the Orange Suite. The height at each node represents the distance between clusters.

2.7. Principal component analysis (PCA) Analysis

The data set used for PCA analysis consisted of 41 normalized miRNAs that showed a clear stratification between 18 metastatic and 18 matched non‐metastatic samples. PCA was conducted to reduce dimensionality and identify the primary sources of variation in the data set. The analysis was performed using Python and plotted using the Orange package, where the singular value decomposition method was applied to calculate the principal components. The first two principal components explained 62% of the total variance. These components were plotted to visualize the separation among metastatic and non‐metastatic patients. The PCA results were visualized using the Orange suite. Scree plots were generated to evaluate the distribution of variance across principal components and to interpret the relationships among variables and samples.

2.8. Linear discriminant analysis (LDA) Computation

LDA was employed to evaluate the predictive power of different combinations of miRNA markers for distinguishing between metastatic and non‐metastatic samples in the data set. The analysis was conducted using Python 3.8 with several key packages. Data manipulation and processing were handled using pandas, while numpy supported numerical operations and array handling. The itertools package, specifically the combinations function, was used to generate all possible combinations of miRNA markers for feature selection. scikit‐learn provided essential machine learning tools, LDA for implementing the LDA model across different marker combinations, and accuracy score. Together, these packages enabled an exhaustive combination analysis, with results stored and ranked by accuracy. Data Preparation: Labels were binarized to differentiate metastatic (1) and non‐metastatic samples (0). The predictor variables (miRNAs) were selected from an external list. Combination Analysis: To identify optimal marker combinations, an exhaustive approach was taken, evaluating all possible combinations of miRNA markers from a single marker up to the full set. For each combination size, features were selected, and LDA models were trained and tested. Training and Testing: For each feature combination, the data set was split into training and test sets with a 70/30 (Training/Testing) split. Each model was trained using the selected marker combination, and accuracy was evaluated on the test set to assess model performance. Results Storage: Accuracy scores for each combination were stored and ranked to isolate best combining miRNAs.

2.9. Clinical Risk Assessment

Cancer of the Prostate Risk Assessment (CAPRA) scores were computed using preoperative clinical parameters, including PSA, Gleason score, clinical T stage, and lymph node involvement. PSA was categorized as <6 ng/mL (0 points), 6–10 ng/mL (1 point), 11–20 ng/mL (2 points), and >20 ng/mL (3 points). Gleason scores were scored as ≤6 (0 points), 7 (2 points), and ≥8 (3 points). Tumor stage was classified as T1–T2 (0 points) or T3 (1 point). Lymph node involvement was scored as 0 points (negative) or 1 point (positive). The CAPRA score was calculated by summing these individual components, resulting in a possible score range of 0–8. Patients were stratified into extreme risk (CAPRA < 3 or ≥ 6) and moderate risk (CAPRA 3–5) groups for subsequent analyses. Model performance was assessed using Stratified K‑Fold cross‑validation (k = 5) to maintain class balance in each fold. For each combination and group, predicted probabilities from all test folds were pooled to generate receiver operating characteristic (ROC) curves. Area under the ROC curve (AUC) was calculated to summarize discriminatory performance.

3. Results

3.1. Parametric Analysis to Identify a Core Set of miRNAs

For this study, we utilized miRNA sequencing data previously generated by our lab [14]. One thousand one hundred twenty three miRNAs were initially identified across all samples (Figure 1A, Table S1). Following preprocessing, which involved filtering for low read counts, detecting and removing outliers, and adjusting for imbalances in miRNA counts, 519 miRNA targets were identified. This final data set met our quality and reliability criteria. Low‐expression miRNAs were filtered based on metastasis status, excluding those where >50% of TPM values were below 10 in both groups. Correlation plots were used to assess expression differences, revealing one patient with extreme variation across miRNAs, who was subsequently excluded from analysis. The resulting miRNA data set was normalized using TPM reads.

FIGURE 1.

Workflow and differential expression analysis of miRNAs: (A) Schematic representation of the miRNA analysis workflow starting from the total miRNAs identified (1123) through stringent filtration to a final set of 519 miRNAs. The process concludes with the normalization of the miRNA data, preparing it for further analysis. (B) Volcano plot illustrating the differential expression of miRNAs, with log2 fold change on the x‐axis and −log10 p‐value on the y‐axis. Significantly upregulated miRNAs are highlighted in red, significantly downregulated miRNAs are in blue, and nonsignificant changes are shown in gray. Key miRNAs are labeled for further emphasis on their potential biological importance. [Color figure can be viewed at wileyonlinelibrary.com]

To explore the differences in miRNA expression levels between control and metastatic patient groups, we initially employed a parametric analysis of the miRNAs. Considering the possibility of unequal variances between the two groups, Welch's t‐test was utilized. This choice of test, which does not assume homogeneity of variances, allowed us to conduct a two‐tailed analysis.

Our analysis revealed significant expression changes in many miRNAs that have a fold change greater than 1.5 and Welch's t‐test for unequal variances p‐value of less than 0.05 (Figure 1B). A few miRNAs were downregulated in the metastatic patient group, among which are miR‐222‐3p and miR‐221‐3p. However, many of the significant miRNAs were found to be upregulated in the metastatic cohort. We restricted our downstream analysis to a stricter statistical cut of p < 0.01 which gave us a total of 41 miRNAs for downstream analysis (Table S2).

3.2. Unsupervised Stratification of Patients

Our initial effort was centered on characterizing the patient population to ascertain their alignment with our predefined data set criteria of 41 miRNAs. To enhance patient stratification, we employed unsupervised clustering techniques, utilizing Manhattan distance calculations to group similar miRNA expression profiles between patients.

The distance matrix, based on Manhattan distance, reveals that non‐metastatic patients (top‐left cluster) exhibit lower pairwise distances among themselves (indicated by darker purple shades), suggesting greater similarity in their selected miRNA expression profiles (Figure 2A). In contrast, the metastatic patients (bottom right cluster) exhibit more dispersed distance values, suggesting greater heterogeneity in miRNA expression within this group. However, it is important to note that distance in the clustering analysis does not equate to overall sample variability but rather reflects patterns of similarity based on selected features. The distinct separation between the two groups suggests that the identified miRNA signatures are effectively associated with metastatic status.

FIGURE 2.

Comprehensive analysis of patient samples based on discriminant miRNA expression: (A) Heatmap displays the pairwise distance matrix among patient samples, segmented into non‐metastatic controls (“NOMET”) and metastatic (“MET”) groups. Distances were calculated based on discriminant miRNA expression from parametric analysis. The color intensity varies from purple (closer distance, high similarity) to yellow (greater distance, low similarity), illustrating the degree of expression dissimilarity between the samples. (B) Dendrogram resulting from hierarchical clustering based on the calculated distances from the discriminant miRNA list. The clustering delineates the two principal groups: “NOMET” (blue) and “MET” (red). (C) Principal component analysis plot of the patients, based on the same discriminant miRNA list, showing the distribution along the first two principal components that capture the majority of variance in the data. Samples are color‐coded by patient group, with notable outliers labeled to emphasize samples with significant deviation from typical group patterns. [Color figure can be viewed at wileyonlinelibrary.com]

This analytical approach enabled us to validate and refine our patient stratification, ensuring it was based on the initial miRNA expression profiles identified. Recently, such statistical stratification techniques have been successfully utilized, demonstrating their efficacy in accurately categorizing patient groups based on biological data [17]. The clustering effectively demonstrated distinct expression profiles, separating metastatic from non‐metastatic patients with high consistency. A dendrogram from this analysis showed that the patients mostly clustered according to their expected groups, with an overall clustering accuracy of 83.3% (Figure 2B). This clustering was further visualized through a heatmap, displaying a clear distinction in miRNA expression patterns between the two groups (Figure S1). However, there were a couple of patients who veered away from their expected grouping. Upon further examination of their expression profiles, there was a consistent pattern of these outlier patients grouping with their unexpected group.

To explore the underlying structure of the miRNA expression data, dimensional reduction was conducted, and PCA was performed to better stratify the patients. This dimensionality reduction technique allowed us to reduce the data into two principal components (PC1 and PC2), which together captured more than 62% of the total variation present amongst the patients (Figure S2). The resulting scatter plot shows distinct clustering, with non‐metastatic patients tending to group towards the negative values of PC1, while metastatic patients are more dispersed along both axes (Figure 2C). This greater dispersion in the metastatic group is consistent with the pattern observed in the Manhattan distance map (Figure 2A), suggesting a broader range of miRNA expression profiles within this group. A few outlier patients in each group appeared in regions typically associated with the opposite classification, suggesting distinct expression profiles or potential histological misclassification. These patients consistently showed discordant placement in both dimensionality reduction and Manhattan distance clustering (Figure 2A,C). Overall, the observed separation between groups indicates that the selected miRNA panel captures underlying differences associated with metastatic classification.

3.3. Unsupervised Identification of miRNA Clusters

To discern the molecular underpinnings distinguishing metastatic from non‐metastatic prostate cancer cases, we performed unsupervised clustering on the top 41 differentially expressed miRNAs. This analysis utilized a Manhattan distance algorithm to organize the miRNAs into four major groups. The resulting dendrogram from this clustering provided a visual representation of the miRNA groupings, showing distinct expression patterns across patient samples (Figure 3A).

FIGURE 3.

Hierarchical clustering of miRNAs based on patient discrimination. (A) Dendrogram showing the hierarchical clustering of all analyzed miRNAs into four major groups based on their expression profiles in non‐metastatic (“nomet”) and metastatic (“met”) patient samples. These groups are color‐coded from blue (Group 1) to orange (Group 4). (B) Heatmaps and associated dendrograms for three sub‐groups within Cluster 1, demonstrating variations in miRNA expression patterns between “met” and “nomet” conditions. Each subgroup highlights distinct expression profiles within the larger cluster. (C) Similar to Cluster 1, this panel presents the heatmaps and dendrograms for two sub‐groups within Cluster 2, detailing the differential expression of miRNAs between patient conditions within this specific cluster. (D) This panel shows the heatmap and dendrogram for miRNAs in Cluster 3, indicating a unique set of expression patterns distinct from the other clusters. (E) Heatmap and dendrogram illustrating the expression profile of miRNAs in Cluster 4. Table of miRNA Composition (Bottom, right). The accompanying table lists the specific miRNAs associated with each group and sub‐group as shown in the heatmaps, providing a reference for the miRNAs driving the clustering patterns observed in the heatmaps. [Color figure can be viewed at wileyonlinelibrary.com]

Further exploration of these clusters was conducted and the resulting heatmap allowed for an immediate grasp of the expression dynamics demonstrating the distinct molecular signatures between the metastatic and non‐metastatic groups. The miRNAs within each cluster showed varying degrees of expression, with Cluster 1 containing three subclusters, each representing a range from low to medium expression levels (Figure 3B). Interestingly, the three subclusters within Cluster 1 demonstrated a marked difference in expression between the metastatic and non‐metastatic groups. Cluster 2 has two subclusters and demonstrated a clear distinction between metastatic and non‐metastatic patient cohorts (Figure 3C). In contrast, Cluster 3 did not exhibit a drastic difference between metastatic and non‐metastatic groups, indicating a more subtle role in stratifying patients (Figure 3D). Cluster 4 was particularly notable for containing the highest expressing miRNAs across the entire cohort, with some, like miR‐222‐3p and miR‐221‐3p, exhibiting statistically significant repression in metastatic patients (Figure 3E).

To validate these findings, miRNA qPCR was conducted, confirming the differential expression observed in the sequencing analysis (Figure 4). Key miRNAs such as miR‐221‐3p and miR‐222‐3p, found to be downregulated, and miR‐183, miR‐182‐5p, and miR‐106b‐5p, upregulated in metastatic cases, were specifically highlighted. This validation process reinforces the reliability of our clustering approach.

FIGURE 4.

Volcano plot of quantitative PCR validation of miRNA expression: Volcano plot showcasing the quantitative PCR results validating the expression patterns of selected miRNAs derived from the final parametric analysis. The x‐axis represents the log2 fold change in expression between case and control samples, while the y‐axis indicates the −log10 p‐value, reflecting the statistical significance of each comparison. MiRNAs with significant upregulation are highlighted in red, while those with significant downregulation are shown in blue. Key miRNAs, such as miR‐145‐5p, miR‐141‐3p, and pre‐miR‐183, are labeled directly on the plot to denote their particular relevance in the study. This visualization aids in identifying the miRNAs most differentially expressed between the two groups, with dotted lines marking the thresholds for statistical significance (p‐value < 0.05) and notable fold change. [Color figure can be viewed at wileyonlinelibrary.com]

3.4. Utilizing Machine Learning to Identify Predictive miRNAs in a Stepwise Manner

Building a reliable predictive model involves selecting a combination of several miRNAs that work together to classify patients effectively. This provides redundancy in the analysis, allowing buffering against noise and variability in expression levels, thus increasing the model's reliability and reducing the risk of overfitting to any particular miRNA signal. Machine learning algorithms are particularly suited for this task, and the linear classifier, LDA was chosen as the primary algorithm for classifying metastatic and non‐metastatic patients based on their miRNA expression profiles.

To explore the predictive power of miRNA expression within each cluster, an LDA model was trained on the miRNA clusters generated from the unsupervised patient analysis. The accuracy scores for each cluster are summarized in Table 1. LDA predictive accuracy on the entire miRNA content of primary clusters revealed poor performance, with overall accuracy scores ranging from 0.36 to 0.64. However, further subdivision of the clusters (e.g., Clusters 1‐1, 1‐2, and 1‐3) showed modest improvements in accuracy, with scores reaching up to 0.73 in Cluster 1‐3.

TABLE 1.

LDA accuracy scores for each major groups and sub‐groups of miRNAs.

Cluster name	LDA accuracy score
Cluster 1	0.55
Cluster 1, Sub‐Group 1	0.64
Cluster 1, Sub‐Group 2	0.64
Cluster 1, Sub‐Group 3	0.73
Cluster 2	0.36
Cluster 2, Sub‐Group 1	0.45
Cluster 2, Sub‐Group 2	0.55
Cluster 3	0.64
Cluster 4	0.55

Abbreviation: Linear discriminant analysis.

Despite these improvements in subcluster accuracy, the overall predictive power of the LDA model remained low. These results are consistent with the observation that expression profiles within each cluster were relatively similar. In addition, some miRNAs may complement each other effectively, while others may negatively impact the predictive accuracy.

3.5. Comprehensive miRNA Combination Analysis

We evaluated whether combining miRNAs with distinct expression patterns within each cluster would improve classification performance. Consequently, we conducted an extensive analysis, systematically training and testing all possible miRNA combinations within each primary cluster (Clusters 1–4) to determine the optimal set that maximized the LDA accuracy score.

The results of the best combinations are presented in Table 2. Within each cluster, specific miRNA combinations notably improved the model's accuracy. For Clusters 1, 3, and 4, the best combinations achieved an accuracy score of 0.82, while Cluster 2 reached a slightly lower accuracy score of 0.73. However, cluster 3 only resulted in one miRNA capable of predicting with high certainty. Notably, certain miRNAs recurred as key contributors to the predictive model within each cluster. These results indicate that the strategic combination of miRNAs within clusters substantially improved classification accuracy, reaching approximately 80% predictive power within each cluster.

TABLE 2.

LDA accuracy score of best combining miRNAs from each group.

Cluster	LDA score	Recurring miRNAs
Cluster 1	0.82	hsa‐let‐7a‐3p
		hsa‐miR‐194‐5p
		hsa‐miR‐33b‐5p
		hsa‐miR‐183‐3p
		hsa‐miR‐106b‐3p
		hsa‐miR‐33a‐5p
		hsa‐miR‐219a‐5p
		hsa‐miR‐301a‐3p
Cluster 2	0.73	hsa‐miR‐183‐5p
		hsa‐miR‐182‐5p
		hsa‐miR‐769‐5p
		hsa‐miR‐345‐5p
		hsa‐miR‐98‐5p
Cluster 3	0.82	hsa‐miR‐339‐5p
Cluster 4	0.82	hsa‐miR‐15a‐5p
		hsa‐miR‐221‐3p
		hsa‐miR‐106b‐5p
		hsa‐miR‐200a‐3p
		hsa‐miR‐16‐5p

Abbreviation: Linear discriminant analysis.

3.6. Final miRNA Panel

Certain miRNAs appeared in multiple high‐performing combinations within each cluster, suggesting they contributed consistently to classification accuracy. By selecting these miRNAs, we ensured that the model was built on features that not only performed well individually but also synergized with other miRNAs, enhancing the overall accuracy and stability of the model. This approach identified miRNAs that consistently contributed to accurate classification across multiple combinations (Figure 5A).

FIGURE 5.

Machine learning analysis of miRNAs for predictive accuracy: (A) Flowchart illustrating the method used to process the initial set of 41 miRNAs through various stages of unsupervised clustering and linear discriminant analysis (LDA) training rounds. This strategic approach segmented the miRNAs into manageable groups, optimizing computational resource allocation and improving efficiency, culminating in a final set of 8 miRNAs with a predictive accuracy of 91%. (B) Table displaying the 8 final miRNAs identified through the analysis process, their occurrences in the best combining analysis data set, and their respective clusters. This panel highlights the miRNAs that contributed most significantly to the final LDA model's 91% accuracy score. (C) Matrix showing various combinations of the final 8 miRNAs and how each combination affects the model's accuracy. Each row represents a different combination of miRNAs, with green squares indicating inclusion in the model and gray squares showing exclusion. The consistency across combinations demonstrates the reliability and redundancy of the model. [Color figure can be viewed at wileyonlinelibrary.com]

Using these miRNAs we iterated through all sets of combinations, by training and evaluating each set. We isolated a final of eight miRNAs that combined in five unique combinations, all achieving a LDA model predictive accuracy of 91% (Figure 5B). The combinations ranged between three miRNAs to six miRNAs (Figure5C).

Violin plot analysis showed that all eight miRNAs exhibited distinct mean differences between metastatic and control groups (Figure S3). Previous studies have reported that all eight miRNAs are associated with cancer‐related pathways, as summarized in Table 3.

TABLE 3.

Final eight miRNA and their biological roles.

microRNA	Pathways involved	Cancer involvement
hsa‐miR‐106b‐3p	Cell cycle regulation, apoptosis	Prostate cancer, breast cancer, colorectal cancer
hsa‐miR‐769‐5p	Cell migration and invasion	Lung cancer, prostate cancer
hsa‐miR‐182‐5p	Cellular stress response, apoptosis	Prostate cancer, breast cancer, melanoma
hsa‐miR‐194‐5p	Epithelial to mesenchymal transition	Colorectal cancer, liver cancer, prostate cancer
hsa‐miR‐345‐5p	Cell proliferation, apoptosis	Prostate cancer, breast cancer
hsa‐miR‐183‐3p	Cell migration, invasion	Prostate cancer, lung cancer, ovarian cancer
hsa‐miR‐200a‐3p	TGF‐beta signaling, EMT suppression	Breast cancer, ovarian cancer, prostate cancer
hsa‐miR‐301a‐3p	NF‐kB signaling, cell survival	Prostate cancer, pancreatic cancer, lung cancer

Receiver operator characteristic (ROC) curve analysis was conducted to validate the efficacy of this panel. ROC analysis on all 8 miRNAs showed a poor area under the curve (AUC) of 0.53, indicating limited classification performance for the full panel (Figure S4). ROC analysis on each unique miRNA combinations demonstrated an area under the curve (AUC) of 80% or greater. The top‐performing miRNA combinations, particularly combinations one and two, achieved the highest AUC values, suggesting strong classification performance.

To evaluate how well our miRNA‑based model aligned with clinical risk classification, we stratified patients using the CAPRA (Cancer of the Prostate Risk Assessment) score. CAPRA was derived by assigning weighted points to PSA, Gleason score, clinical T stage, and lymph node status, and summing these to obtain a total score (Figure 6A). Patients were grouped into Extreme (CAPRA < 3 or >5) and Moderate (CAPRA 3–5) predictive categories, and each predefined miRNA combination was assessed within these strata (Figure 6B). To summarize group‑level differences visually, pooled ROC curves for each CAPRA category were averaged, producing representative curves for Extreme and Moderate groups. These results indicate that the miRNA combinations show markedly stronger predictive performance in patients within the extreme (average AUC ~ 0.8) CAPRA risk category compared to the moderate group (average AUC ~ 0.54), with a statistically significant difference by paired t‑test (p < 0.001) and a trend toward significance by non‑parametric testing (p = 0.0625, Wilcoxon) (Figure 6C). Overall, the CAPRA classification showed strong concordance with the miRNA‑based predictions: the Extreme group achieved an average accuracy of ~0.8, while the Moderate group averaged ~0.6. Among the combinations, Combinations 1 and 5 demonstrated the best performance in the Extreme group, whereas Combinations 1 and 2 were superior in the Moderate group (Figure 6D). AUC analysis confirmed these trends, with all combinations performing well in the Extreme group, and worse (AUC ≤ 0.5) in the Moderate group.

FIGURE 6.

CAPRA‐based clinical risk stratification and comparison with miRNA‐based predictions: (A) Scoring criteria for CAPRA components (PSA, Gleason score, tumor stage, and lymph node status) used to assign risk points. (B) Patients were grouped into extreme (CAPRA scores 1–2 or 6–7) and moderate (CAPRA scores 3–5) predictive categories based on cumulative CAPRA scores. (C) Stratified five‐fold cross‐validation (k = 5) was performed to evaluate model performance, with receiver operating characteristic (ROC) curves plotted for extreme (AUC = 0.82) and moderate (AUC = 0.54) risk groups. (D) Accuracy and AUC values for individual miRNA combinations are shown for each risk group, highlighting higher predictive performance in the extreme risk group compared to the moderate risk group. [Color figure can be viewed at wileyonlinelibrary.com]

4. Discussion

FFPE samples contain valuable biological information regarding disease but extracting biological molecules like RNA and DNA from these samples is challenging due to fragmentation [18]. The short nature of the miRNAs (~20–23 bp) protects them from excessive degradation by formalin treatments, providing an advantage and paving the way to isolate reliable biological information from the FFPE samples [19]. Building predictive models using this information will promote better patient stratification. In this study, miRNAs were isolated from FFPE samples of prostate cancer patients with varying degrees of metastatic progression, assessed by their Gleason scores and metastatic recurrence following prostatectomy.

The primary aim of this study was to identify a panel of miRNAs that optimally predicts the metastatic potential in postoperative prostate cancer patients. We achieved this by conducting a combinatorial analysis of miRNAs, using the LDA machine learning algorithm. Due to the computational intensity of analyzing all 41 miRNAs simultaneously, we developed a strategic, stepwise approach to narrow down the predictive miRNAs. Starting with the full set of 41 miRNAs, we performed unsupervised clustering based on metastatic versus non‐metastatic patient groups. This initial clustering divided the miRNAs into four clusters, each representing similar expression profiles. To enhance precision, we further subdivided these clusters into seven smaller sub‐clusters, resulting in groups of manageable sizes for the LDA analysis (Figure 5A).

Our initial LDA analysis on these sub‐clusters was limited by similar expression patterns across sub‐clusters. However, this analysis allowed us to identify key miRNAs by examining the frequency of their successful combinations with others in high‐performing LDA combinations. By focusing on these frequently co‐occurring miRNAs, we refined our selection through additional LDA testing within each primary cluster to isolate those miRNAs that combined best within each major group. In the final phase, we analyzed the top‐performing miRNAs from each major cluster and conducted a third LDA analysis. This step identified 8 miRNAs, from which we extracted five unique, predictive combinations, forming our final miRNA panel. Each incremental step in training and testing improved the model's predictive accuracy, validating our approach and supporting the selected miRNA panel as a reliable tool for patient stratification.

A limitation of this study is the small sample size and limited matching, which may affect the generalizability of our findings. While LDA is well‐suited for smaller datasets, a larger, more diverse cohort with better‐matched clinical variables would strengthen the model. Future studies incorporating independent validation cohorts and additional clinical factors will be essential to further refine and validate these findings.

Unlike the prior analysis, which used bootstrapping and univariate screening, our approach integrates unsupervised clustering, LDA training, and combinatorial LDA testing to refine and validate miRNA candidates. By first identifying natural groupings before applying LDA, we ensured selection was data‐driven rather than threshold‐dependent. Instead of a single risk score, we identified multiple combinatorial miRNA panels, improving clinical flexibility.

The 41 discriminant miRNAs were also utilized to cluster patients using unsupervised distance methods. This method allowed us to further characterize the patient population and identify any patients who showed a discrepancy in the characterization. Indeed, for a handful of patients, their miRNA profiles closely matched the opposite group (metastatic or non‐metastatic) with very high accuracy. This suggests that the diagnosis of these patients using biochemical or histological methods may not be fully accurate. This specifically point us to re‐evaluate our histological and biochemical characterization of the metastatic status of the patients in conjunction with miRNA sequencing data. This highlights an important role for miRNAs in FFPE samples to properly characterize the patient status.

Unsupervised clustering also helped to strategically group miRNAs into clusters based on their expression profiles. This allowed for the identification of miRNAs that are co‐expressed in a patient cohort based on their metastatic status and this on its own is also a useful as a predictive tool. Indeed, this is a powerful approach and using this, miRNA biomarkers have be successfully identified [20, 21]. However, machine learning algorithms are particularly suited for this task, as they can identify complex patterns within the data.

LDA is ideal for binary classification tasks. LDA provides interpretable results and is well‐suited for this study, as it maximizes the separation between the two classes (metastatic and non‐metastatic patients). Additionally, the simplicity of LDA reduces the risk of overfitting, which can be a concern when working with smaller datasets like ours, consisting of 18 metastatic and 18 non‐metastatic patients [22]. By assuming linear separability and focusing on maximizing variance between classes, LDA ensures a balance between model simplicity and performance. Cross‐validation confirmed that our model performed well, further validating our decision to use LDA as the optimal approach for this study. More flexible models like Random Forest and Neural Networks risk overfitting with limited samples, and Neural Networks, though powerful, function as “black boxes,” limiting interpretability, an essential factor in clinical applications. SVM can be highly effective; however, it demands careful kernel selection and parameter tuning, which can introduce additional complexity and reduce generalizability in small datasets.

As the identified miRNAs are upregulated in metastatic prostate cancer, they may be therapeutically targeted using antagomiRs, LNA inhibitors, or miRNA sponges to suppress their oncogenic effects. While these strategies show promise, challenges such as efficient delivery and off‐target effects must be addressed for clinical application.

Digital twin models in medicine are emerging as powerful tools that integrate diverse, continuously updated patient data to simulate disease progression and inform clinical decisions (Figure S5) [23, 24]. While our study does not yet meet the criteria of a digital twin because of the absence of real‐time, multimodal data integration, it aligns with early efforts in that direction [25]. At this stage, we view our approach as a foundational component within a broader vision of individualized risk assessment, acknowledging that significant methodological and infrastructural developments will be needed to achieve full digital twin capability. Ethical and societal considerations will also be crucial as such models evolve [26].

Ethics Statement

A well‐established, Sunnybrook Health Sciences institutional research ethics board approved database of 585 patients was utilized. The research was conducted in accordance with the ethical principles outlined in the Declaration of Helsinki. To mitigate potential selection bias in using historical samples (1990–2000), we examined demographics and healthcare access to ensure a representative study population. Consistent inclusion criteria were applied, and all patient data remained deidentified, adhering to ethical guidelines for retrospective research. These measures strengthen the integrity and reliability of our findings. Patients were treated with RP for clinically localized prostate cancer between 1990 and 2000 at Sunnybrook Health Sciences Center, a single, tertiary care hospital. All patients included in this study provided informed consent for the use of their data, which were collected through Sunnybrook Health Sciences under its established guidelines for data sharing and research use.

Consent

Patients were treated with radical prostatectomy for clinically localized prostate cancer between 1990 and 2000 at Sunnybrook Health Sciences Centre, a single, tertiary care hospital. All patients included in this study provided informed consent for the use of their data, which were collected through Sunnybrook Health Sciences under its established guidelines for data sharing and research use.

Conflicts of Interest

The authors declare no conflicts of interest.

Permission to Reproduce Material From Other Sources

No previously published material requiring permission for reproduction was included in this manuscript.

Supporting information

Supplementary Figure 1: Heatmap of miRNA Expression Across Patient Groups.

Supplementary Figure 2: Scree Plot of Principal Component Analysis.

Supplementary Figure 3: Violin Plots of miRNA Expression by Patient Group.

Supplementary Figure 4: Receiver Operating Characteristic (ROC) Analysis for miRNA Combinations.

Supplementary Figure 5: Implementation of a Digital Twin in Clinical Decision‐Making.

Supplementary Table 1: Normalized transcripts per million values of 1123 miRNAs after filtration.

Supplementary Table 2: 42 miRNAs resulting from parametric analysis and filtration.

Supplementary Table 3: Patient clinical characteristics.

Data Availability Statement

Access to the data may be requested from Dr. Arun Seth (arun.seth@utoronto.ca), subject to approval by the appropriate governing bodies.