Prediction of water solubility and Setschenow coefficients by tree-based regression strategies

The experimental determination of water solubility (log S 0 ) and Setschenow coefficient (k m ) of a compound is a time-consuming activity, which often needs large amounts of expensive substances. This work aims at establishing two “open-source” chemometric models based on a regression tree that is able to predict the two abovementioned quantities. The dataset used is the largest to appear up to now for the collection of k m values, containing information on 295 molecules and it is relevant also for the collection of logS 0 values (321 molecules); for each of them 32 descriptors were taken from freely available databases. Information about water solubility and Setschenow coefficients, necessary to train the models, were taken from available literature. Validation was performed on a separate test set of molecules. The precision reached in the prediction is fully satisfying, being RMSEP = 0.6086 and 0.0441 for logS 0 and k m , respectively.