Computational Methods for Learning and Predicting the DNA binding Specificities of Transcription Factors

The protein-DNA interactions between Transcription Factors (TFs) and Transcription Factor Binding Sites (TFBSs) are the key activities in gene regulation. Considerable efforts have been spent on elucidating their binding mechanisms in the past twenty years. In this thesis, I propose and describe three methods to learn and predict the DNA binding specificities of TFs.The first method named kmerHMM is a sequence pattern recognition method to discover the sequence patterns of TFBSs from protein binding microarray (PBM) data. The novelty of kmerHMM lies in two aspects. First, it has a competitive edge over the existing methods in using Hidden Markov Models (HMMs) to derive an HMM model to represent PBM data. Secondly, kmerHMM incorporates N-Max-Product algorithm and can derive multiple motif matrix models to represent PBM data. The second method named SNPdryad is a classification method to predict the deleterious effect of amino acid substitutions on proteins (including TFs). Comprehensive performance comparisons are conducted on benchmark datasets, reflecting that SNPdryad has the performance comparable to the best of the other methods. SNPdryad has been run on the complete human proteome, generating prediction scores for all the possible amino acid substitutions. In particular, case studies are conducted to show that SNPdryad can help identify the deleterious effect on TFs if a DNA-binding residue is mutated to others.The third method is a computational framework to learn specificity-determining pair-wise residue and nucleotide interactions between TFs and TFBSs. Its novelty lies in the fact that the existing methods usually focus on learning DNA-binding residues on the protein side (TF) only, ignoring the DNA side (TFBS). In contrast, our novel framework aims at learning the interactions between both sides. The proposed method is compared to state-of-the-art methods comprehensively, demonstrating its competitive performance. We also describe case studies showing how our method can be applied to reveal meaningful biological insights into protein-DNA interactions across different DNA-binding families.Ph.D