While most prostate cancers diagnosed at a localized stage have high survival rates, approximately 40% of patients will experience biochemical cancer recurrence (BCR) even after apparently curative surgery. This recurrence is driven by inherent biological factors, microscopic residual disease, and distant micrometastases not detectable at the time of treatment.
DNA methylation is a chemical modification of DNA -- specifically the addition of a methyl group to cytosine bases at CpG sites -- that controls whether genes are turned on or off without changing the underlying DNA sequence. It is one of the primary epigenetic mechanisms regulating gene expression throughout development and disease.
In cancer, two distinct patterns emerge: hypermethylation of tumor suppressor gene promoters silences protective genes, while hypomethylation of oncogene regions activates cancer-driving genes. These methylation changes are detectable early in tumor development, are relatively stable compared to gene expression, and vary systematically across different cancer types and stages.
Prior work has shown that DNA methylation signatures can predict recurrence-free survival in lung and thyroid cancers. However, a quantitative methylation-based recurrence prediction model specifically for prostate cancer had not been established, leaving a clinical gap that this study aimed to fill.
The study analyzed data from 490 prostate cancer patients in The Cancer Genome Atlas (TCGA), a large publicly available cancer genomics database. Each patient had paired DNA methylation data (from 22,602 genes), RNA expression data (from 20,531 genes), and complete clinical information including recurrence status. Patients were split 70/30 into training and testing cohorts.
The researchers performed two parallel genome-wide analyses: identifying differentially methylated genes (DMGs) -- genes showing significantly different methylation levels between patients who recurred and those who did not -- and differentially expressed genes (DEGs) -- genes showing significantly different RNA expression levels between the same groups. This dual analysis enabled identification of genes where methylation was functionally linked to expression changes.
A critical filtering step compared the two gene lists: genes that were both differentially methylated and differentially expressed, and showed a negative correlation between methylation level and expression level (meaning more methylation = less gene activity, consistent with methylation silencing the gene) were prioritized. This correlation approach identifies genes where the epigenetic change is functionally meaningful, not just coincidental.
From 684 DMGs and 691 DEGs, ten genes emerged with significant methylation-expression correlations: TNNI2, SPIN2, COL5A3, RNF169, CCND1, FGFR1, SLC17A2, FAMM71F2, RREB1, and AOX1. These ten genes formed the basis of the predictive model.
A Support Vector Machine (SVM) model was trained using the methylation levels (beta values) of the ten identified genes as input features and recurrence vs. no-recurrence as the binary outcome. SVM is a machine learning method that finds the optimal boundary to separate two classes in high-dimensional feature space, well-suited for small datasets with many variables.
The model produces a continuous recurrence score calculated as a weighted combination of the ten genes' methylation values: RecurrenceScore = 0.46*TNNI2 - 0.59*SPIN2 + 0.15*COL5A3 + 0.35*RNF169 + 0.10*CCND1 + 0.23*FGFR1 + 0.37*SLC17A2 + 0.61*FAMM71F2 + 0.19*RREB1 + 0.72*AOX1. Patients with scores above 0 are classified as high-risk, those at or below 0 as low-risk.
The model achieved an AUC of 0.773 on the ROC curve, indicating good discriminatory ability between patients who will and will not experience recurrence based solely on their methylation patterns. Multivariate Cox regression confirmed the SVM score as an independent predictor of recurrence (hazard ratio 0.45; 95% CI 0.28-0.69; P less than 0.001), meaning it retains predictive power even after adjusting for other clinical variables.
Notably, traditional clinical staging variables including Gleason score, T-stage, and M-stage did not retain independent predictive power in multivariate analysis, while the SVM methylation score and lymph node (N-stage) status remained independently significant. This suggests the methylation model captures biological recurrence risk beyond what standard pathological staging conveys.
Patients were divided into high-risk (SVM score greater than 0) and low-risk (SVM score at or below 0) groups. Kaplan-Meier survival analysis showed that low-risk patients had significantly better recurrence-free survival (RFS) compared to high-risk patients -- the two groups separated clearly over time, confirming the model's ability to stratify patients by recurrence probability.
The separation in outcomes between risk groups held consistently across different pathological subtypes and lymph node stages, suggesting the methylation score captures an underlying biological difference in tumor aggressiveness that is not fully accounted for by standard clinical variables.
Gene Ontology (GO) analysis of the 684 differentially methylated genes revealed enrichment in biological processes including transcription regulation, signal transduction, and immune response. KEGG pathway analysis highlighted involvement in T-cell receptor signaling, cytokine-cytokine receptor interactions, and natural killer cell-mediated cytotoxicity -- suggesting that the methylation changes driving recurrence risk involve alterations in immune surveillance pathways.
To validate the TCGA findings in a controlled biological system, the researchers measured the expression and methylation of all ten genes in two cell lines: PC3 (a prostate cancer cell line) and PNT2 (a non-cancerous normal prostate cell line). Gene expression was measured by quantitative real-time PCR (qRT-PCR), and methylation was assessed by methylation-sensitive PCR.
The expression patterns in cell lines closely matched the TCGA data. Genes that were hypomethylated in the TCGA recurrent group (TNNI2, SPIN2, COL5A3, SLC17A2, AOX1) showed significantly increased expression in the PC3 cancer cells compared to PNT2 cells. Genes hypermethylated in the TCGA data (RNF169, CCND1, FGFR1, FAMM71F2, RREB1) showed increased expression in cancer cells -- consistent with loss of methylation-mediated silencing in cancer.
Methylation-sensitive PCR confirmed the distinct methylation patterns: COL5A3, FGFR1, and SLC17A2 showed clear hypermethylation in PC3 cancer cells, consistent with the bioinformatic analysis. This laboratory confirmation strengthens confidence that the biomarkers identified from TCGA data reflect genuine biological differences between cancer and normal prostate tissue.
Pearson correlation analysis between methylation and expression levels in cell lines showed variable but significant R-squared values for several gene pairs, with both positive and negative correlations observed -- reflecting the complexity of how methylation affects different regulatory regions and genes individually.
CCND1 (Cyclin D1) is one of the most well-characterized of the identified genes -- it drives cell cycle progression and is dysregulated in many cancers. Its methylation-linked upregulation in cancer cells contributes to uncontrolled cell proliferation, a fundamental hallmark of cancer progression and recurrence.
FGFR1 (Fibroblast Growth Factor Receptor 1) is a cell surface receptor that, when activated by growth factors, triggers signaling cascades promoting cell growth and differentiation. FGFR1 alterations have been linked to metastasis in colorectal and renal cancers, and its methylation-driven expression changes in prostate cancer may facilitate invasion and spread.
AOX1 (Aldehyde Oxidase 1) has prior published associations with prostate cancer survival in a genome-wide scan study. Its hypomethylation and elevated expression in cancer cells connects metabolic deregulation to cancer progression. AOX1 has also been implicated in bladder cancer through epigenetic silencing, suggesting shared epigenetic cancer mechanisms across urological malignancies.
The involvement of immune-related pathways (T-cell receptor signaling, NK cell cytotoxicity) in the differentially methylated genes is consistent with the growing understanding that epigenetic alterations in cancer affect not only cancer cells directly but also the tumor immune microenvironment -- potentially explaining why some cancers evade immune surveillance and recur.
This study presents a proof-of-concept for a 10-gene DNA methylation model that predicts prostate cancer recurrence independently of standard clinical staging variables. With AUC 0.773, it outperforms Gleason score and T-stage in multivariate analysis, suggesting that methylation captures complementary biological information about tumor biology.
The practical appeal of methylation-based testing is that DNA methylation patterns are stable and can be measured from routine tissue samples (biopsies or surgical specimens) using established laboratory methods. A test requiring only 8-10 methylation measurements would be far simpler and potentially less costly than broader genomic tests.
Key limitations include the single-institution TCGA dataset, relatively short follow-up for some non-recurrent patients (introducing potential misclassification bias), and the modest number of recurrent cases available. External validation in independent prospective cohorts, particularly with standardized follow-up durations, will be essential before clinical translation. Further mechanistic studies on the cancer-specific roles of poorly characterized genes in this panel are also needed.