Development and application of computational methods to address current challenges in drug discovery

[eng] The recent advances in drug discovery have introduced many opportunities to solve previously unattainable medical needs, such as targeted protein degradation as a new drug modality. These opportunities challenge the status quo of the methodologies currently in use. Computational methodologies can help alleviate these challenges, and have been steadily growing in reliability and importance in drug discovery. In particular, I have performed a ligandability assessment of a set of 23 E3 ligases using mixed organic solvent molecular dynamics. E3 ligases are the main recognition component of the ubiquitin-mediated degradation machinery of the cell. Despite their crucial role for targeted protein degradation applications, they are widely considered undruggable and only a few have been employed. We aimed to expand the available E3 toolbox proposing a set of novel ligandable pockets, and validated them prospectively on FBW7, an E3 ligase of high pharmacological interest due to its implication in many cancer lines. We found a family of FBW7 allosteric modulators that reduces the levels of c-MYC in cells in a FBW7 and proteasome dependent manner. I evaluated its potential mechanism of action through a conformation analysis of the holo and apo forms of the SCFFBW7 complex. Finally, I performed a virtual screening to generate analogues of the FBW7 modulators to perform a SAR study. The use of virtual libraries to perform in silico hit identification has proven to be a successful alternative to high-throughput screening. The size of such libraries has grown exponentially the last years, especially since the appearance of ultra-large on-demand chemical libraries, reaching 1011 compounds. These ultra-large libraries provide a high chemical diversity and depth, which offer the opportunity to finding highly potent molecules in the early stages, alleviating the needs for hit optimization. However, current virtual screening technologies are not yet adapted to perform under such library sizes. I have developed a pipeline to efficiently explore the ultra-large on-demand chemical libraries from the bottom-up. Firstly, exhaustively screening the fragment space, identifying interesting scaffolds to grow. Then, performing a substructure search in the ultra-large spaces, recovering all the drug-like molecules with the selected scaffolds and docking them tethering the scaffold. Reducing the space to evaluate, allowed to allocate computational resources for more sophisticated post-docking methodologies such as Dynamic Undocking (DUck). We have validated the approach finding BRD4 inhibitors by growing scaffolds selected from three different scenarios: (i) computationally found scaffolds, (ii) crystallized scaffolds, and (iii) scaffolds derived from drug candidates. We obtained a high ratio of diverse and potent hits, 16 having IC50 comparable to the drug candidate (+)-JQ1. Coupled with a more efficient chemical space exploration algorithms, quick and accurate predictions of protein-ligand binding free energies are crucial to harness the opportunities presented by the ultra-large libraries. I have assessed the use of DUck as a fast tool for relative binding free energy calculations and activity cliff identification on different congeneric series of HSP90, CDK2 and BACE1 inhibitors. I defined and demonstrated the applicability domain of the method for compound series that follow a one-step dissociation barrier, and have a high and converged first dissociation barrier. Despite the new successful applications of the DUck framework, the current implementation, very rigid and dependent on licensed programs, hinders the widespread use of the methodology. I have developed and validated OpenDUck, an open-source python library to facilitate the creation of new applications within the DUck framework. The results obtained are highly reproducible between the different conditions and with the original DUck implementation, moreover, the usage of HMR coupled with a smaller box has led to a significant reduction in the simulation time. Finally, I have evaluated the use of the Crooks fluctuation theorem for free energy calculations from non-equilibrium simulations in CRBN-CK1a-lenalidomide ternary complexes and HSP90 protein-ligand complexes. I evaluated the performance and convergence at different speeds and sampling iterations, obtaining similar accuracy in 50 times less simulation time.