Abstract
Accurate and efficient prediction of high energy ligand conformations is important in structure-based drug discovery for the exclusion of unrealistic structures in docking-based virtual screening and de novo design approaches. In this work, we constructed a database of 140 solution conformers from 20 druglike molecules of varying size and chemical complexity, with energetics evaluated at the DLPNO-CCSD(T)/complete basis set (CBS) level. We then assessed a selection of machine learning potentials and semiempirical quantum mechanical models for their ability to predict conformational energetics. The GFN2-xTB tight binding density functional method correlates with reference conformer energies, yielding a Kendall's τ of 0.63 and mean absolute error of 2.2 kcal/mol. As putative internal energy filters for screening, we find that the GFN2-xTB, ANI-2x and MACE-OFF23(L) models perform well in identifying low energy conformer geometries, with sensitivities of 95 %, 89 % and 95 % respectively, but display a reduced ability to exclude high energy conformers, with respective specificities of 80 %, 61 % and 63 %. The GFN2-xTB method therefore exhibited the best overall performance and appears currently the most suitable of the three methods to act as an internal energy filter for integration into drug discovery workflows. Enrichment of high energy conformers in the training of machine learning potentials could improve their performance as conformational filters.