Knowledge-based libraries for predicting the geometric preferences of druglike molecules

Abstract

We describe the automated generation of libraries for predicting the geometric preferences of druglike molecules. The libraries contain distributions of molecular dimensions based on crystal structures in the Cambridge Structural Database (CSD). Searching of the libraries is performed in cascade fashion to identify the most relevant distributions in cases where precise structural features are poorly represented by existing crystal structures. The libraries are fully comprehensive for bond lengths, valence angles, and rotamers and produce templates for the large majority of unfused and fused rings. Geometry distributions for rotamers and rings take into account any atom chirality that may be present. Library validation has been performed on a set of druglike molecules whose structures were published after the latest CSD entry contributing to the libraries. Hence, the validation gives a true indication of prediction accuracy.