PREDICTIVE MODELS TO DISCOVER NOVEL RNA-BINDING PROTEINS

Abstract

RNA-binding proteins (RBPs) play significant roles in post-transcriptional gene regulation system. To gain a better understanding of this system, discovering new RBPs is needed. There have been several attempts to computationally predict new RBPs by using information derived from protein sequence or structure, where the performance is however not satisfying. In this dissertation, we introduced a new type of information derived from protein-protein interaction (PPI) data to RBP prediction and built an RBP classifier termed SONAR. We demonstrated the good predictive power of PPI information across species. Next, we proposed three well-performed RBP classifiers employing modified sequence and homology features and state-of-art machine learning algorithms, which also show predictive power on RNA-binding-related regions. Thereafter, we integrated above classifiers into a single ensemble classifier termed HYDRA, achieving the best performance among all available classifiers. Finally, we experimentally validated a couple of RBPs predicted by SONAR and HYDRA, expanding current RBP repertoire.