PREDICTION OF N4-ACETYLCYTIDINE IN mRNA USING A DEEP FOREST MODEL WITH HYBRID FEATURES

Abstract

Motivation: N4-acetylcytidine (ac4C) is the only experimentally confirmed acetylation modification in messenger RNA (mRNA) and has been associated with various human diseases. Despite its biological importance, accurate identification of ac4C sites remains challenging due to the time-consuming, costly, and often inaccurate nature of experimental detection techniques. Although several computational methods, including deep forest–based approaches, have been proposed for ac4C prediction, their performance is still limited and requires further improvement.

Results: We propose a novel computational framework, termed DeepAc4C, for the efficient and accurate identification of ac4C modification sites in mRNA sequences. The proposed model leverages convolutional neural networks to automatically learn discriminative sequence patterns associated with ac4C modifications. Extensive experimental evaluations demonstrate that DeepAc4C outperforms existing state-of-the-art methods, achieving an accuracy of 92%. Furthermore, an interpretability analysis was conducted to investigate the contribution of different sequence regions to the prediction process, revealing distinctive patterns that are characteristic of ac4C modifications.  The benchmark dataset used in this study was obtained from Zenodo (https://zenodo.org/records/5138047) and consists of DNA/RNA sequential data in FASTA format. The dataset comprises genomic sequences represented as alphabet-based nucleotide strings, suitable for deep learning–based sequence analysis.