Integrating DNA methylation and clinical features to predict prostate cancer prognosis

This study investigates the relationship between DNA methylation patterns and clinicopathological characteristics in prostate cancer. We performed targeted next-generation sequencing (NGS) on methylation sites linked to prostate cancer and calculated the corresponding methylation rates. Clinical samples were divided into training and prediction sets. In the training set, we utilized unsupervised clustering and Support Vector Machine (SVM) modeling to distinguish between prostate cancer and non-cancer samples. For predictions, K-Nearest Neighbors (KNN) assessed sample similarity, while SVM facilitated classification. Bayesian methods integrated probabilities to predict cancer status and cluster assignments. To address uncertainties identified by Uniform Manifold Approximation and Projection (UMAP), we validated our results using Random Forest Support Vector Machine (RFSVM), which highlighted significant methylation sites for SVM training. Depth-correction methods were applied to mitigate variations in sequencing depth.

In the training dataset, the leave-one-out cross-validation (LOOCV) prediction accuracy of RFSVM was 0.85 (AUC: 0.91); for RFSVM-depth, it was 0.83 (AUC: 0.93). The LOOCV prediction accuracy for Bayesian SVM (BSVM) was 0.87 (AUC: 0.94), decreasing to 0.83 (AUC: 0.91) with depth correction. In the test dataset, our Bayesian prediction achieved an accuracy of 0.8793 (sensitivity: 0.8182, specificity: 0.9167), which improved to 0.9655 with depth correction (sensitivity: 1.0, specificity: 0.9444). RFSVM demonstrated an accuracy of 0.8621 (sensitivity: 0.8182, specificity: 0.8889), dropping to 0.7759 with depth correction. Among 58 samples, predictions showed 67% complete consistency, with 84% consistency in methylation rates and 78% in specific methylation sites. Figure 6 presents a comprehensive comparison of the performance of RFSVM, BSVM, XBSVM, and RFESVM on training and test datasets, including feature selection and prediction consistency. Confusion matrices, ROC curves, and Venn diagrams were used to detail the feature importance and prediction consistency of each method.

These findings highlight the importance of analyzing individual methylation sites and broader methylation patterns in understanding the role of DNA methylation in prostate cancer, providing valuable insights into the effects of data preprocessing and feature selection.