A deep learning-based approach for predicting ubiquitination sites using word embedding techniques

Ubiquitination is a prevalent and reversible post-translational modification (PTM) that controls apoptosis and plays a crucial role in protein breakdown and cell disorders. Ubiquitination also involved in the control of a number of physiological mechanisms such as cellular proliferation, gene transcription, intracellular signaling, DNA correction, and synthesis. However, experimental identification of ubiquitination sites prediction is time-consuming and costly, so it is an urgent need to construct effective predictors. In this paper, a novel predictor called UbiSite-DeepWET has been developed to accurately predict ubiquitination sites from sequences using a convolutional neural network (CNN) model. First, we perform nine feature extraction method, namely Amino Acid Composition (AAC), Dipeptide Compostion (DPC), Composition of k-spaced Amino Acid Pairs (CKSAAP), Pseudo-Amino Acid Composition (PAAC), Amino Acid index (AAindex) database, Word2Vec, GloVe and fastText, to represent protein sequence patterns. Secondly, SHapley Additive exPlanations (SHAP) are used to remove unnecessary and irrelevant features for predicting ubiquitination sites. Finally, the best features are fed into the CNN classifier to build the UbiSite-DeepWET model for identifying the protein ubiquitination sites. Five-fold cross validation shows that the AUC values of Set1-Set6 datasets are 0.8846, 0.7764, 0.7881, 0.7740, 0.6870 and 0.7238, respectively. The Adaptive synthetic sampling approach (ADASYN) is applied in Set4 to Set6 unbalanced datasets, and the AUC values are 0.9929, 0.9975 and 0.9971, respectively. Additionally, we have created three independent test (IT) datasets which the AUC values are 0.8529, 0.6215 and 0.6001, respectively. The results demonstrate that the proposed technique UbiSite-DeepWET outperforms conventional ubiquitination prediction methods and offers new guidance for the identification of ubiquitination sites.