Virtual screening for novel mechanisms of action: applications and methodological developments

Abstract

The main motivation of this thesis has been to validate, improve and develop new methods with respect to the ones available nowadays in the area of drug discovery, in order to be able to study more challenging targets in the near future that currently are out of our reach. As the productivity of the pharmaceutical industry is decreasing year after year over the last decades, the improvement of such methods would be a step forward. As we are mostly a computational lab, this thesis has focused on different computational approaches such as docking, molecular dynamics or chemoinformatics.

On the first part of the thesis (first author on the publication in PLoS Computational Biology in 2014), I worked on Docking-based Virtual Screening (VS). Particularly, in validating rDock, a little-known but very powerful program that was published and released during this thesis as open source software. In order to validate it, we performed several benchmarking experiments with DUD and ASTEX sets to compare the performance of rDock against Glide and AutoDock Vina, two commonly used docking programs. The capabilities of rDock with respect to binding mode prediction (predict how a ligand structure will be upon binding to its receptor) and virtual screening (selecting the most likely active ligands amongst thousands or millions of drug-like molecules) were compared with Glide and Vina, and we demonstrated that rDock performed as well as them.

On the second project of the thesis (first author on the publication in Nature Chemistry in 2016), we wanted to develop a novel computational tool for drug discovery not only that was complementary to the existing ones, but also that improved them by adding new ways of interpreting the data. Taking advantage of the already known technique of Steered Molecular Dynamics (SMD), we proposed an approach consisting in reducing the size of the system, focusing around a key interaction point and running SMD to discriminate between active and inactive ligands. This approach, or as we call it: "Dynamic Undocking", is intended to foster drug design efforts in the lead optimization stage by improving the efficiency of the in silico assessment of protein-ligand binding affinity. After a positive retrospective assessment of the method using different systems of the DUD set, a prospective validation was required to evaluate its feasibility in a real drug discovery project. Hsp90 was selected as the test system: A fragment library was created and a subset of fragments was selected for a first stage of docking-based VS. About 300.000 ligands were docked with rDock and the top-scoring ones were subjected to Dynamic Undocking. In a collaboration with Vernalis, a pharmaceutical company in the UK, we tested tens of compounds selected with Dynamic Undocking and we were able not only to find positive and novel hits but also to improve hit-rate with respect to standard fragment screening by almost 10 fold.

Finally, we had the opportunity to participate in the D3R Grand Challenge 2015 where we could apply all the methods from this thesis (first author on the publication in Journal of Computer-Aided Molecular Design in 2016). This challenge was designed as a blind public test where different groups around the world tried to predict the binding mode and the affinity of a set of ligands for their respective protein target. Our approach consisted in a combination of docking and Dynamic Undocking and our results were placed amongst the best for the two systems of the challenge. We also discussed how the level of available data and previous knowledge on each of the systems impacted on the final results.

La motivación principal de esta tesis ha sido validar, mejorar y desarrollar nuevos métodos con relación a los disponibles hoy en día en el área del desarrollo de fármacos, para en un futuro poder estudiar dianas que actualmente están fuera de nuestro alcance. Debido a que la productividad de la industria farmacéutica está disminuyendo durante los últimos años, una mejora en los métodos disponibles sería un gran paso adelante. Esta tesis se ha centrado en diferentes métodos computacionales, como el docking o la dinámica molecular.

En la primera de las partes, trabajé en el cribado virtual (Virtual Screening) basado en docking. Concretamente, participé en la validación del programa de docking rDock mediante la comparación con dos programas muy usados hoy en día de su capacidad de predecir correctamente el modo de unión de un ligando con su proteína diana y de sus resultados en el cribado virtual de posibles fármacos.

En la segunda parte de la tesis, participé en el desarrollo de un método computacional novedoso en el diseño de fármacos que complementase y mejorase los métodos actualmente disponibles. Éste método, bautizado en inglés como “Dynamic Undocking”, consiste en una implementación específica de dinámica molecular mediante la cual somos capaces de detectar si un ligando puede ser activo o inactivo de manera rápida y eficiente. Se validó el método de manera retrospectiva y posteriormente se aplicó en otro proyecto con el objetivo de encontrar nuevos posibles fármacos para una proteína relacionada con cáncer. Gracias a una colaboración con una empresa del Reino Unido, encontramos nuevos ligandos de manera que aumentamos la tasa de éxito con relación a un método estándar en casi 10 veces.

Por último, participé en el “D3R Grand Challenge 2015”, un experimento a escala mundial donde los participantes aplicaron diferentes métodos y compararon sus resultados respecto a dos métricas distintas: la predicción del modo de unión y la capacidad de ordenar los ligandos proporcionados por la organización por su afinidad respecto a la proteína diana. En nuestro caso, aplicamos una combinación de docking y “Dynamic Undocking” con unos resultados excelentes.