Predicting the Permeability of Macrocycles from Conformational Sampling – Limitations of Molecular Flexibility

Abstract

Macrocycles constitute superior ligands for targets that have flat binding sites but often require long synthetic routes, emphasizing the need for property prediction prior to synthesis. We have investigated the scope and limitations of machine learning classification models and of regression models for predicting the cell permeability of a set of denovo-designed, drug-like macrocycles. 2D-Based classification models, which are fast to calculate, discriminated between macrocycles that had low-medium and high permeability and may be used as virtual filters in early drug discovery projects. Importantly, stereo- and regioisomer were correctly classified. QSPR studies of two small sets of comparator drugs suggested that use of 3D descriptors, calculated from biologically relevant conformations, would allow development of more precise regression models for late phase drug projects. However, a 3D permeability model could only be developed for a rigid series of macrocycles. Comparison of NMR based conformational analysis with in silico conformational sampling indicated that this shortcoming originates from the inability of the molecular mechanics force field to identify the relevant conformations for flexible macrocycles. We speculate that a Kier flexibility index of ≤10 constitutes a current upper limit for reasonably accurate 3D prediction of macrocycle cell permeability.