2D and 3D descriptors are typically used to characterize the structure of chemicals and then establish Quantitative Structure-Activity Relationships (QSAR). In a previous study, we showed that ranking a set of small molecule kinase inhibitors could be very challenging, even impossible for certain targets like the ERK2 kinase. In order to build more reliable models for this challenging datasets of relevance for lead optimization, we have developed a new workflow that encompasses (i) the structure-based docking of a series of known inhibitors in the binding site of a particular kinase, (ii) the independent molecular dynamics (MD) simulations of each protein-ligand complex (15 ns, NVT, 300K, TIP3P, 1fs), (iii) the computation of novel “MD descriptors”, which measure the variations of ligands’ 3D shape over the course of MD simulations as well as their dynamic protein-ligand (PL) interactions. We have constructed cross-validated QSAR models using several machine learning methods and various sets of MD-based descriptors. We discovered that a set of these MD-QSAR models have excellent prediction performance on held out data. Some of these models are highly interpretable, identifying features of ERK2 inhibitor MD conformation shape and PL interactions that are key determinants of activity. These features may help medicinal chemists design and identify new analogs likely to be more potent. These results demonstrate how MD simulations can provide an additional layer of information on ligands’ dynamic characteristics that are of importance for further focused structural optimization.

Characterizing the chemical space of kinase inhibitors using molecular descriptors computed from molecular dynamics trajectories

254th National Meeting (2017)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.