Predicting the binding modes of protein complexes: new strategies for molecular docking

Docking approaches and homology modelling procedures have recently become protagonists in the structural prediction of protein complexes. Therefore, in the last years it raised the need of developing more and more reliable tools and defining some guidelines to obtain accurate prediction of those structures. These goals are particularly relevant when only poorly accurate information about the proteins or the interaction (e.g. no interface indication for protein-protein docking, or protein models instead of experimental structures) is available. In this thesis several strategies based on combining different computational techniques are proposed to overcome the limitations of the sampling stage of ligand- and protein-protein docking approaches and to broaden the possibility of predicting the structure of protein complexes. In particular, in the field of protein-protein docking algorithms, the combination of two existing docking methods (HADDOCK and ZDOCK) was proposed, allowing to obtain a method (called ZADDOCK) able to overcome the limitations of the single strategies used and to sum their strengths. Moreover, both for ligand- and protein-protein docking, an analysis of the relationships between docking results and the quality of homology models was performed to assess the possibility of an a priori prediction of the accuracy of docking results on the basis of the evaluation of model quality indices. The results obtained for ZADDOCK show on average a very good performance of the method, that produced reliable predictions without the need of any interface data to guide the sampling step. This allows its employment in the study of complexes for which no experimental information is available and bioinformatics interface prediction fails. Moreover, an accurate description of the intermolecular interactions occurring in protein complexes was obtained, which is a key information to drive subsequent experimental work. The quality of ZADDOCK results indicates that the strategy of combining different computational techniques is a very promising avenue for the development of new docking approaches. As for the use of homology models in docking calculations, in this work it is demonstrated the possibility of a priori predicting the accuracy of ligand-protein docking results on the basis of model quality indices, with the development of a strategy based on the comparison with homologous proteins. Moreover, the results found for protein-protein docking give the bases for a future development of a general prediction strategy of docking accuracy on protein models.