m5CStack: An integrated framework for m5C site prediction using multi-feature stacking

RNA 5-methylcytosine (m5C) modification sites are essential for understanding the regulation of RNA functions in various biological processes. However, the vast amount of sequence data generated by modern genomics poses significant challenges for traditional identification methods, which often struggle to meet high-throughput demands. Consequently, computational tools have become indispensable for predicting m5C sites. In this study, we present m5CStack, an advanced ensemble learning framework designed to predict m5C modification sites with high accuracy. m5CStack integrates multiple feature encoding techniques and machine learning models through a stacking architecture to enhance the robustness and reliability of predictions. We evaluate the framework on RNA datasets derived from multiple species, including <i>Homo sapiens</i> (human), <i>Mus musculus</i> (mouse), <i>Drosophila melanogaster</i> (drosophila), and <i>Danio rerio</i> (danio). Experimental results demonstrate that m5CStack significantly outperforms previous prediction methods across a range of metrics, including accuracy, sensitivity, and specificity. Furthermore, SHAP-based feature significance analysis reveals the key contribution of specific features, further improving the interpretability of the model. To improve accessibility, a user-friendly web interface is developed, allowing users to input RNA sequences or upload files for prediction, with results displayed in an intuitive format alongside confidence scores. Overall, this study highlights the potential of m5CStack as a powerful tool for RNA modification profiling, offering new insights into the epigenetic regulation of RNA across species.