Enrichment of virtual screening results using induced-fit techniques

Resumo

This thesis explains the design, development, test on the benchmarking dataset DUD-e and the application to an industrial virtual screening project of the PELE VS platform. The most common and quick tools used on virtual screening campaigns do not take into account the induced-fit effect, although there are some methodologies capable of reproducing this effect they are either time consuming or very limited on which transformations the protein may undergo. In this thesis with the development of the PELE VS platform we aim at using the simulation software PELE (Protein Energy Landscape Exploration) to account for the induced-fit effect. The PELE software uses a monte carlo algorithm coupled with an energy minimization step to explore the ligand conformations and model the protein. This approach allows the program to account for both big conformational changes and small local changes of the protein and to perform a good conformational search of the ligand, which coupled can account for the induced-fit effect with only a quick simulation. PELE has been traditionally, and successfully, used in the enzyme engineering field where only a few compounds per protein are usually tested and studied. In order to apply the program to the virtual screening field, where thousands of compounds are tested in silico, the first step was to automatize the whole procedure of preparing, launching and analyzing the simulations. Thus, during this thesis the PELE VS platform has been developed altogether with other auxiliary tools. Once the platform was developed, we wanted to test the behaviour of PELE on a well known benchmarking dataset, thus we tried our methodology on the DUD-e dataset. Since this dataset is formed by more than 100 proteins, we chose a few proteins for each of the families present in the dataset, reducing the number of proteins to 21 systems. Then, we tried to use a general protocol for all the chosen proteins in order to improve the results of currently used scoring functions in the field. After studying the simulations and trying several protocols on this subset we we observed that every protein (or at least family) that we want to study needs an specific simulation protocol in order to correctly reproduce the induced-fit effect and improve the results of the most used scoring function: glide from schrodinger. Finally we applied the platform and our previous hypothesis to an industrial virtual screening campaign, as part of the collaborative Retos project: Silicoderm. In this case we worked with only one protein and several compounds and we confirmed the need for a tailored simulation protocol for the receptor in order to improve results.