Abstract
BACKGROUND: 5-formylcytidine (f5C) is a unique post-transcriptional RNA modification present at the wobble position of mRNAs and tRNAs, that plays a critical role in mitochondrial protein synthesis and is potentially involved in translation regulation. Recent studies have revealed that f5C modifications may promote cancer metastasis by driving mitochondrial mRNA translation. However, existing computational methods for predicting f5C modification sites are scarce and have significant limitations, which has hindered in-depth investigations into their molecular functions and regulatory mechanisms. RESULTS: To address these limitations, we developed Fusion_f5C-Pred, an innovative dual-branch deep learning framework that integrates both sequence and structural features through a gated fusion network. The sequence branch employs a densely connected convolutional network integrated with the convolutional block attention module to capture local sequence patterns, whereas the structural branch utilizes the transformer-encoder to learn RNA secondary structure features. Comprehensive evaluations of independent datasets demonstrate that Fusion_f5C-Pred achieves superior prediction performance with an ACC of 0.7952 and an AUROC score of 0.8684, significantly outperforming existing methods. The t-SNE visualization analysis confirms that the fused features exhibit enhanced inter-class separation in the representation space. CONCLUSIONS: Notably, our model's learned sequence patterns strongly agree with known RNA regulatory motifs identified via MEME, indicating biological interpretability. The proposed framework not only offers a robust computational tool for f5C research but also establishes a transferable architecture for studying other RNA modifications. The source code and datasets are publicly available at: https://github.com/HuiCong123/Fusion_f5C-Pred .