Modelling of partition constants: linear solvation energy relationships or PLS regression?

The partitioning between octanol and water (K_ow) and the water solubility (S_W) are used to investigate similarities and differences between linear solvation energy relationships (LSER) and partial least squares regression (PLSR) models. The similarities in model structure are described, and shown to transform into a comparable prediction performance. Furthermore, the results demonstrate the opportunity for an analogous chemical interpretation and much of the alleged difference between the mechanistic or semi-empirical LSER and the statistical PLSR models then disappear.

Abstract

Estimation methods for partition constants are needed in many fields of engineering and science. The partitioning between phases is determined by the free energy of the transfer and all estimation methods must therefore describe the same entity. Linear solvation energy relationships (LSERs) try to split the contributions to van der Waals and polar interactions into directly interpretable solute descriptors, while projection-based regression methods can accomplish a similar dimensionality reduction from a set of theoretical descriptors. Here, we use the partitioning between octanol and water (K_ow) and water solubility (S_w) to investigate similarities and differences between LSER and partial least squares regression (PLSR) models. The similarities in model structure are described, and shown to transform into a comparable prediction performance. We also demonstrate the opportunity to accomplish an analogous chemical interpretation of a PLSR model—either directly or through a linear transformation of the PLS factors—as with an LSER model. Much of the alleged difference between the mechanistic or semi-empirical LSER and the statistical PLSR models will then disappear. The choice of a modelling approach should therefore primarily be driven by the availability of data and predictive performance. Copyright © 2009 John Wiley & Sons, Ltd.