Protein-dependent prediction of messenger RNA binding using Support Vector Machines

RNA-binding proteins interact specifically with RNA strands to regulate important cellular processes. Knowing the binding partners of a protein is a crucial issue in biology and it is essential to understand the protein function and its involvement in diseases. The identification of the interactions is currently resolvable only through in vivo and in vitro experiments which may not detect all binding partners. Computational methods which capture the protein-dependent nature of the binding phenomena could help to predict, in silico, the binding and could be resistant against experimental biases. This thesis addresses the creation of models based on support vector machines and trained on experimental data. The goal is the identification of RNAs which bind specifically to a regulatory protein. Starting from a case study, done with protein CELF1, we extend our approach and propose three methods to predict whether an RNA strand can be bound by a particular RNA-binding protein. The methods use support vector machines and different features based on the sequence (method Oli), the motif score (method OliMo) and the secondary structure (method OliMoSS). We apply them to different experimentally-derived datasets and compare the predictions with two methods: RNAcontext and RPISeq. Oli outperforms OliMoSS and RPISeq affirming our protein specific prediction and suggesting that oligo frequencies are good discriminative features. Oli and RNAcontext are the most competitive methods in terms of AUC. A Precision-Recall analysis reveals a better performance for Oli. On a second experimental dataset, where negative binding information is available, Oli outperforms RNAcontext with a precision of 0.73 vs. 0.59. Our experiments show that features based on primary sequence information are highly discriminative to predict the binding between protein and RNA. Sequence motifs can improve the prediction only for some RNA-binding proteins. Finally, we can conclude that experimental data on RNA-binding can be effectively used to train protein-specific models for in silico predictions.