Predicting Passive Permeability of Drug-like Molecules from Chemical Structure: Where Are We?

Intestinal absorption in human is routinely predicted in drug discovery using in vitro assays such as permeability in the Madin-Darby canine kidney cell line. In silico models trained on these data are used in drug discovery efforts to prioritize novel chemical targets for synthesis; however, their proprietary nature and the limited validation available, which is usually restricted to predicting in vitro permeability, are barriers to widespread adoption. Because of the categorical nature of the in vitro permeability assay, intrinsic assay variability, and the challenges often encountered when translating in vitro data to an in vivo drug property, validation based solely on in vitro data might not be a good characterization of the usefulness of the in silico tool. In this work, we analyze the performance of three different in silico models in predicting the in vitro and in vivo permeability of 300 marketed drugs and 86 discovery compounds. The models differ in their approach (mechanistic vs quantitative structure-activity relationship) and the degree of complexity; one of them is a linear equation based on seven simple physicochemical descriptors and is presented for the first time in this work. Results show that in silico models can be successfully used to complement the discovery toolbox for characterizing in vivo intestinal permeability, defined using fraction of dose absorbed in human (Fa) and human jejunal permeability (P_eff). While the in vitro permeability models outperformed the in silico approach at predicting each of the in vivo end points explored, the gap in predictivity between the in vitro and the in vivo data was generally comparable to the gap between in silico and in vitro data. The in vitro and in silico approaches shared many of the same outliers, which can often be explained by the route of drug absorption (paracellular vs transcellular, active vs passive). Data suggest that the discovery process can greatly benefit from an early adoption of in silico models for predicting permeability as well as from a careful analysis of the in silico to in vivo disconnects.