Random Forest Algorithm-Based Prediction of Solvation Gibbs Energies

Solvation Gibbs energy of chemicals is a critical parameter in chemical industry and chemical reactivity. Predicting the solvation Gibbs energies for a large number of solvents and solutes through machine learning techniques is challenging area. In this work, the random forest (RF) algorithm, together with a combined descriptor set from solvents and solutes, was used for developing a quantitative structure-property relationship (QSPR) model for solvation Gibbs energies of 6238 solute/solvent pairs. The optimal RF (ntree = 25, mtry = 10 and nodesize = 5) model was obtained, whose training and test sets, respectively, have determination coefficients of 0.935 and 0.924, and root mean square errors of 2.477 and 2.464 kJ center dot mol(- 1). In predicting the solvation Gibbs energies for a large dataset, the optimal RF model is comparable to other QSPR models reported in the literature. (GRAPHICS)