DBm5U-Deep:预测m5U位点的多尺度深度学习模型OA

RNA 5-methyluridine(m5U)plays crucial biological roles in organismal growth and development,gene expression regulation,and disease onset.Therefore,accurate identification and prediction of m5U-modified sites are of significant importance.Compared with traditional experimental methods,deep learning technologies enable more efficient and cost-effective m5U site prediction.However,existing approaches still exhibit limitations in feature representation capability and predictive accuracy.To address these challenges,the DBm5U-Deep model is proposed,innovatively constructing a collab-orative multi-scale deep framework.This model enhances sequence representation through hierarchical feature extraction and optimizes final predictions using a weighted averaging strategy.Specifically,ribonucleic acid(RNA)sequences are first segmented into 3-mer fragments and converted into low-dimensional continuous vector representations using the GloVe word vector model.These vectors are then fed into a core architecture comprising a convolutional neural network(CNN),a bidirectional long short-term memory network(BiLSTM),and a multi-attention mechanism to fully capture local features and long-range dependencies.Finally,a fully connected layer performs the classification prediction.In five-fold cross-validation,the DBm5U-Deep model achieves AUC and ACC of 97.26%and 92.87%,respectively.On an independent test set,AUC and ACC reach 97.30%and 93.55%,surpassing the current state-of-the-art model by 0.26 and 1.26 percent-age points,respectively.Experimental results demonstrate that DBm5U-Deep exhibits high accuracy and stability in m5U site prediction,providing an efficient and reliable computational tool for RNA modification functional studies and drug target screening.