Plasma protein binding affinity and its relationship to molecular structure: an in-silico analysis

In-silico plasma protein binding (PPB) models have been generated on human and rat in-house datasets, and on a human dataset sourced from the literature. From the results reported herein, it is apparent that models built on datasets relevant to the chemotypes under investigation in lead optimization programs will perform considerably better in this role than those generated on diverse compounds sourced from the literature. The in-house human and rat partial least-squares regression (PLS) models have cross-validated q2 values of 0.53 and 0.42 on the training sets, respectively. On the independent test and validation sets, they display similar predictive ability, with logK prediction errors of approximately 0.5 log units. This compares to approximately 0.25 log units variability expected for experiment. Given the considerable interspecies PPB differences, the prediction of PPB in one species using measurements in the other is no better than a prediction from an in-silico model generated on that species.