Bioactive Conformational Biasing: A New Method for Focusing Conformational Ensembles on Bioactive-Like Conformers

Computational approaches that rely on ligand-based information for lead discovery and optimization are often required to spend considerable resources analyzing compounds with large conformational ensembles. In order to reduce such efforts, we have developed a new filtration tool which reduces the total number of ligand conformations while retaining in the final set a reasonable number of conformations that are similar (rmsd < or = 1 A) to those observed in ligand-protein cocrystals (bioactive-like conformations). Our tool consists of the following steps: (1) Prefiltration aimed at removing ligands for which the probability of finding bioactive-like conformations is low. (2) Filtration based on a unique combination of two-/three-dimensional ligand properties. Within this paradigm, a filtration model is defined by its workflow and by the identity of the specific descriptors used for filtration. Thus, we developed multiple models based on a training set of 47 drug compounds and tested their performance on an independent test set of 24 drug compounds. For test set compounds after prefiltration, our best models have a success rate of approximately 80% and were able to reduce the total number of conformations by 36% while maintaining a sufficiently large number of bioactive-like conformations and slightly increasing their proportion in the filtered ensemble. We were also able to reduce by 39% the number of conformations that are remote (rmsd > 2.5 A) from the bioactive conformer (nonbioactive conformations). In accord with previous reports, prefiltration is shown to have a major effect on model performance. The role and performance of specific descriptors as filters is discussed in some detail, and future directions are proposed.