Bayesian machine learning methods for predicting protein-peptide interactions and detecting mosaic structures in DNA sequences alignments

Short well-defined domains known as peptide recognition modules (PRMs) regulate many important protein-protein interactions involved in the formation of macromolecular complexes and biochemical pathways. High-throughput experiments like yeast two-hybrid and phage display are expensive and intrinsically noisy, therefore it would be desirable to target informative interactions and pursue in silico approaches. We propose a probabilistic discriminative approach for predicting PRM-mediated protein-protein interactions from sequence data. The model suffered from over-fitting, so Laplacian regularisation was found to be important in achieving a reasonable generalisation performance. A hybrid approach yielded the best performance, where the binding site motifs were initialised with the predictions of a generative model. We also propose another discriminative model which can be applied to all sequences present in the organism at a significantly lower computational cost. This is due to its additional assumption that the underlying binding sites tend to be similar. It is difficult to distinguish between the binding site motifs of the PRM due to the small number of instances of each binding site motif. However, closely related species are expected to share similar binding sites, which would be expected to be highly conserved. We investigated rate variation along DNA sequence alignments, modelling confounding effects such as recombination. Traditional approaches to phylogenetic inference assume that a single phylogenetic tree can represent the relationships and divergences between the taxa. However, taxa sequences exhibit varying levels of conservation, e.g. due to regulatory elements and active binding sites, and certain bacteria and viruses undergo interspecific recombination. We propose a phylogenetic factorial hidden Markov model to infer recombination and rate variation. We examined the performance of our model and inference scheme on various synthetic alignments, and compared it to state of the art breakpoint models. We investigated three DNA sequence alignments: one of maize actin genes, one bacterial (Neisseria), and the other of HIV-1. Inference is carried out in the Bayesian framework, using Reversible Jump Markov Chain Monte Carlo.