A deep multi-feature learning framework for transcriptome-wide methyluridine site discovery

Abstract

RNA 5-methyluridine (m5U) is an important post-transcriptional modification involved in numerous biological and regulatory processes, highlighting the need for accurate computational identification methods. This study proposes D-m5U, an efficient deep learning framework for transcriptome-wide prediction of RNA m5U modification sites. The proposed framework employs a hybrid feature representation by integrating Pseudo Dinucleotide Composition (PseDNC), Pseudo Trinucleotide Composition (PseTNC), Z curve (12-bit) encoding, and Normalized Moreau Broto (NMB) descriptors. SHAP (Shapley Additive Explanations) based feature selection is then applied to identify the most informative features, followed by classification using a Deep Neural Network (DNN). The proposed model was evaluated on the Full Transcript and Mature mRNA benchmark datasets using 10-fold cross-validation and independent testing. D-m5U achieved accuracies of 93.29% and 96.75% on the Full Transcript and Mature mRNA datasets, respectively, and 94.43% and 96.14% on the corresponding independent test datasets. Compared with existing predictors, D-m5U improved the classification accuracy by 5.31% and 3.37% on the training datasets, and by 4.12% and 3.15% on the independent test datasets. These results demonstrate the robustness and generalization capability of D-m5U and indicate that it provides an effective computational framework for transcriptome-wide RNA m5U site prediction.