Prediction of Ubiquitination Sites with Feature-Weighting Scheme and Naive Bayes Vectorizer

Ubiquitination is one of the most common post-translational modifications (PTMs) and regulates the quantity and function of a variety of proteins. Many experiments show that the ubiquitin-proteasome system (UPS) plays a critical role in regulating a variety of biological processes, including the degradation of protein, gene transcription, DNA repair and replication, etc. Computational prediction of ubiquitination sites has become a popular research topic and enjoyed a huge burst of research activity. Since the composition of K-spaced amino acid pairs (CKSAAP) is quite helpful for predicting PTMs sites, we employ the scheme to encode samples in this study. Then we propose a category-based feature weighting scheme named the probability of positive sample (pps), which uses available labeling information to assign appropriate weights to amino acid pairs. The main idea of pps is that the more concentrated a high-frequency amino acid pair is in the positive category than in the negative category, the more contribution it makes in separating the positive samples from the negative samples. Then the Bayes formula is used to vectorize (as opposed to classify) the samples according to a probability distribution reflecting the probable categories that the sample may belong to. By replacing amino acid pair features with category-based features, the dimensionality of the sample feature space can be reduced from tens of thousands to a small number of categories. In the experiments, we investigate the effects of our method on the CKSAAP_UbSite benchmark dataset using the SVM as classifier. The results show that our proposed method has a consistently a better performance than other methods.