Mixtures of QSAR models: Learning application domains of pK predicto rs

Abstract Quantitative structure‐activity relationship models (QSAR models) predict the physical properties or biological effects based on physicochemical properties or molecular descriptors of chemical structures. Our work focuses on the construction of optimal linear and nonlinear weighted mixes of individual QSAR models to more accurately predict their performance. How the splitting of the application domain by a nonlinear gating network in a “mixture of experts” model structure is suitable for the determination of the optimal domain‐specific QSAR model and how the optimal QSAR model for certain chemical groups can be determined is highlighted. The input of the gating network is arbitrarily formed by the various molecular structure descriptors and/or even the prediction of the individual QSAR models. The applicability of the method is demonstrated on the p K values of the OASIS database (1912 chemicals) by the combination of four acidic p K predictions of the OECD QSAR Toolbox. According to the results, the prediction performance was enhanced by more than 15% (root‐mean‐square error [RMSE] value) compared with the predictions of the best individual QSAR model.