Application of machine learning in predicting potentiometric selectivity (Mg 2+ /Ca 2+ ) of some amide ligands

In this study, in order to model the prediction of potentiometric selectivity (Mg ²⁺ /Ca ²⁺ ) of several amide ligands, the quantitative structure-property relationship (QSPR) approach was used along with two stepwise (SW) and genetic algorithm (GA) methods as variable selection techniques. Experimental data and molecular structures were entered into the models after calculating and screening the descriptors. The SW method identified five descriptors and GA five descriptors (with two common items) as key parameters. Using these variable selection methods and machine learning modeling methods of multiple linear regrassions (MLR) and support vector machine (SVM), various models including SW-MLR, GA-MLR, GA-SVM were created, and also, by identifying an outlier and removing it, GA-MLR and GA-SVM models were re-developed. The model obtained by GA-SVM (with one outlier removed) had R ² _train =0.893, R ² _test =0.757, RMSE _train =0.322, and RMSE _test =0.705, indicating the high predictive power and fit of the models. These values were a significant improvement compared to the reference paper (R ²  = 0.66 and RMSE = 0.53 even after removing 11 outliers). In addition, findings provide important mechanistic insights into the role of molecular features in potentiometric selectivity (Mg ²⁺ /Ca ²⁺ ) of amide ligands.