Comparison of Predictivities of Log P Calculation Models Based on Experimental Data for 134 Simple Organic Compounds

Predictivities of six log P calculation models (CLOGP, KOWWIN, ACD/LOGP, SLOGP, VLOGP and COSMO thermo) are compared using a common experimental dataset, which does not contain any of the compounds used in the training set for the six models. The former three models, which are based on fragment methods, showed better accuracy (r=0.87–0.91) than SLOGP, which is based on the atom contribution method and VLOGP and COSMO thermo, which are property‐based models. The trends of prediction by each model are described based on chemical structure and molecular weight. For all six models, several compounds had a significantly greater calculated log P value than the experimental log P value. It is interesting to note that most of these compounds had ionic moieties.