DeepHFFT-m7G: A dual-channel self-attention and hybrid feature fusion framework for RNA m7G modification identification.

N7-Methylguanosine (m7G) is a prevalent RNA modification that has attracted significant attention in recent RNA functional research. Multiple studies have shown that m7G modifications play a crucial role in the initiation and progression of various human diseases. Although multiple deep learning methods have been employed to predict m7G modification sites, their accuracy remains suboptimal. In this work, we propose a novel method called DeepHFFT-m7G, which is based on hybrid feature fusion and a dual-channel self-attention network. This approach aims to efficiently identify m7G methylation sites in RNA sequences. First, we integrate four classical RNA sequence features with embedding vectors based on RNA2Vec. Next, a multi-branch convolutional neural network (CNN) constructs a dual-channel feature extraction module, capturing local sequence features. A Transformer encoding module is then introduced to extract global features. Finally, the sequence embedding is transferred to a multi-layer perceptron (MLP) to achieve efficient m7G site prediction. Independent testing demonstrates that DeepHFFT-m7G achieves AUROC, accuracy, MCC, and specificity scores of 97.53%, 96.92%, 93.93%, and 97.63%, respectively, significantly outperforming existing state-of-the-art (SOTA) methods. Furthermore, comparative experiments and visualization analyses further validate the superiority and robust generalization ability of DeepHFFT-m7G.