Understanding and improving statistical models of protein sequences

In the last decades, progress in experimental techniques have given rise to a vast increase in the number of known DNA and protein sequences. This has prompted the development of various statistical methods in order to make sense of this massive amount of data. Among those are pairwise co-evolutionary methods, using ideas coming from statistical physics to construct a global model for protein sequence variability. These methods have proven to be very effective at extracting relevant information from sequences, such as structural contacts or effects of mutations. While co-evolutionary models are for the moment used as predictive tools, their success calls for a better understanding of they functioning. In this thesis, we propose developments on existing methods while also asking the question of how and why they work. We first focus on the ability of the so-called Direct Coupling Analysis (DCA) to reproduce statistical patterns found in sequences in a protein family. We then discuss the possibility to include other types of information such as mutational effects in this method, and then potential corrections for the phylogenetic biases present in available data. Finally, considerations about limitations of current co-evolutionary models are presented, along with suggestions on how to overcome them.