Modifications to the standard definition of polar surface area (PSA) are reported and tested against the octanol–water, chloroform–water and cyclohexane–water partition coefficients of 110 organic and drug-like molecules. It is shown that increasing the flexibility of PSA-based models can lead to some improvements in accuracy, but that these still fall well short of previously published methods. To compete with such methods, PSA-based descriptors must be scaled according to the known hydrogen bonding characteristics of common functional groups. Introducing this scaling markedly improves accuracy, with predictive errors typically around one-half of a log P unit, confirmed by splitting the data into training and test sets. All models developed follow known characteristics of the partition coefficients considered and are statistically and chemically valid.