Predicting Octanol-Water Partition Coefficients: Are Quantum Mechanical Implicit Solvent Models Better than Empirical Fragment-Based Methods?

In this work, we examined the performance of contemporary quantum mechanical implicit solvent models (SMD, SM8, SM12, and ADF-COSMO-RS) and empirical fragment-based methods for predicting octanol-water partition coefficients (log Pow). Two test sets were chosen: the first is composed of 34 organic molecules from a recent study by Mobley J. Chem. Theory Comput , 2016 , 12 , 4015 - 4024 , and the second set is based on a collection of 55 fluorinated alkanols and carbohydrates from Linclau Angew. Chem., Int. Ed. , 2016 , 55 , 674 - 678 . Our analysis indicates that the errors in the solvation free energies of implicit models are reasonably systematic in both solvents such that there is substantial cancellation of errors in the calculation of transfer free energies. Overall, implicit solvent models performed very well across the two test sets with mean absolute errors (MAEs) of about 0.6 log unit and are superior to explicit solvent simulations (GAFF and GAFF-DC). Interestingly, the best performers were empirical fragment-based methods, including ALOGP and miLOGP with significantly lower MAEs (0.2 to 0.4 log unit). The ALOGP method was further tested against the recent SAMPL6 log Pow challenge consisting of 11 drug-like molecules where it obtained an MAE of 0.32 log unit compared to the best-performing COSMOtherm model (0.31 log unit).