MLA-ac4C: A N4-Acetylcytidine Site Prediction Model Based on Multi-Layer Attention, Bidirectional GRU, and DenseNet

N4-acetylcytidine (ac4C) is a highly conserved post-transcriptional modification found in mRNA, known to play a critical role in the regulation of mRNA translation. Traditional experimental approaches for identifying ac4C sites are both time-consuming and costly, underscoring the need for more efficient computational prediction methods. Although various deep learning models have been proposed for ac4C site prediction, most rely on ordinary encoding strategies and simplistic network architectures, which limit their ability to capture high-dimensional features of mRNA sequences. In this study, we propose a novel framework that combines multiple encoding schemes with a hierarchical network structure to improve prediction accuracy.

We developed a new deep learning model, MLA-ac4C, based on a dual-pathway architecture. In the first pathway, one-hot encoding combined with positional encoding is used to represent the sequence, followed by a Bidirectional Gated Recurrent Unit (Bi-GRU) network to extract global features. In the second pathway, a hybrid encoding scheme integrating nucleotide chemical property (NCP), Electron-Ion Interaction Pseudopotential (EIIP), and Pseudo K-Tuple Nucleotide Composition (PseKNC) is used, and local features are extracted using Densely Connected Convolutional Networks (DenseNet). A multi-layer attention mechanism is then applied to integrate the features from both pathways and capture high-level representations for classification.

Rigorous evaluation on an independent test dataset shows that MLA-ac4C achieves superior performance, with Sensitivity (Sn), Specificity (Sp), Accuracy (Acc), Matthews correlation coefficient (MCC), and Area Under The ROC Curve (AUROC) reaching 83.88%, 85.14%, 84.51%, 69.03%, and 92.74%, respectively. These results outperform those of existing state-of-the-art models.

We present MLA-ac4C, a deep learning model based on Bi-GRU, DenseNet, and a multi-layer attention mechanism. Experimental results demonstrate that it provides a more accurate and efficient approach for predicting ac4C modification sites compared to existing methods.