ResUbiNet: A Novel Deep Learning Architecture for Ubiquitination Site Prediction

Abstract

Introduction: Ubiquitination, a unique post-translational modification, plays a cardinal role in diverse cellular functions such as protein degradation, signal transduction, DNA repair, and regulation of cell cycle. Method: Thus, accurate prediction of potential ubiquitination sites is an urgent requirement for exploring the ubiquitination mechanism as well as the disease pathogenesis associated with ubiquitination processes. Results: This study introduces a novel deep learning architecture, ResUbiNet, which utilized a protein language model (ProtTrans), amino acid properties, and BLOSUM62 matrix for sequence embedding and multiple state-of-the-art architectural components, i.e., transformer, multi-kernel convolution, residual connection, and squeeze-and-excitation for feature extractions. Conclusion: The results of cross-validation and external tests showed that the ResUbiNet model achieved better prediction performances in comparison with the available hCKSAAP_UbSite, RUBI, MDCapsUbi, and MusiteDeep models.