Extreme learning machine models for predicting the n-octanol/water partition coefficient (Kow) data of organic compounds

The partition coefficient of n-octanol/water (Kow) is generally considered a valuable characteristic in many natural sciences fields. Predictive quantitative structure-property and relationship (QSPR) models for Kow are encouraged since the experimental data are not always available. Extreme learning machine (ELM) method with the advantages of fast learning speed and good generalization ability has been applied to various domains. This work introduced COSMO descriptors (Sσ-profile) of molecules as descriptors and the ELM algorithm to construct estimation models for Kow of organics. Four ELM models were developed and their outcomes were compared with the MLR models produced by the same descriptors. Results revealed that the proposed ELM models were reliable and applicable to predicting the logKow values of compounds, especially the ELM-4 model with the best correlation coefficient R2 (0.949) and root-mean-square error RMSE (0.358). Therefore, the proposed approaches are effective and feasible, and can be extensively generalized to forecast other important properties of chemicals.