Editorial: Molecular Dynamics and Machine Learning in Drug Discovery

Sergio Decherchi, Francesca Grisoni, Pratyush Tiwary, Andrea Cavalli
2021 Frontiers in Molecular Biosciences  
The drug discovery process is very long and expensive, and many factors hamper its final success. In the attempt to accelerate a drug candidate's progress along the discovery pipeline, computational modeling represents a key tool to address the design and optimization of lead compounds. While physics-based white-box modeling (e.g., docking and molecular dynamics), has represented the standard de facto for many years in the computational chemistry community, nowadays machine learning
more » ... s represent a powerful modeling alternative. The deep learning paradigm in particular can be considered a black box methodology as it can be difficult to extract rules or laws from the trained model. This Research Topic collects selected contributions that deal with both types of modeling approaches, some of which lie at the interface between the two. This "gray box" hybrid approach should not surprise as machine learning and statistical mechanics share several theoretical principles (Ferrarotti et al., 2019; Noé et al., 2019; Agliari et al., 2020; Decherchi and Cavalli, 2020; Ferraro et al., 2020; Tsai et al., 2020) as they both deal with distributions, manifolds, and hence free energies. Molecular dynamics (MD) is based on statistical mechanics. Setting up a MD run for complex systems can be still a not trivial task, requiring continuous automation tools to allow for a wider exploitation in academic and industrial settings. In this regard, the contribution from Schneider et al. discusses the implementation of a webserver for the setup of hybrid molecular mechanics and coarse-grained simulations for Human G-Protein Coupled Receptors (GPCRs) and ligands complexes. GPCRs represent the most important class of druggable targets, hence the importance of having handy tools to setup their systematic simulations. Analyzing and understanding MD outcomes can then be rather complicated, mainly because of the large amount of raw data. Bunker and Rog present a review on the mechanistic understanding of MD generated data for drug delivery in pharmaceutical research. Mechanistic interpretations can be supported by proper machine learning tools: it is often convenient to devise/use clustering, projections or feature extraction algorithms to extract actionable knowledge. This greatly facilitates the interpretation of results and can also allow to define order parameters in some cases, often dubbed collective variables in the MD realm. In the contribution from Arthur et al. Authors devise a combination of MD simulations of proteins and hierarchical pharmacophore features extraction. This strategy represents a smart and widely applicable paradigm (Spyrakis et al., 2015) which combines MD sampling (to recover some of the target flexibility) with a non-dynamical tool (e.g., virtual screening, static docking etc.). While this paper derives features directly for drug discovery, Spiwok and Kriz propose a more
doi:10.3389/fmolb.2021.673773 pmid:33928128 pmcid:PMC8076858 fatcat:l3366fjrkjaehbqz2nffe6swii