QSRR models for predicting the retention indices of VOCs in different datasets using an efficient variable selection method coupled with artificial neural network modeling: ANN-based QSPR modeling

A combination of the smoothly clipped absolute deviation (SCAD) method and the artificial neural network (ANN) was utilized as a novel methodology (SCAD-ANN) in the quantitative structure-retention indices relationship (QSRR). The proposed SCAD method reduces the dimension of data before using the robust ANN modeling method. The efficiency of the SCAD-ANN methods was evaluated by the construction of a QSRR model between the most relevant molecular descriptors (MDs) and RIs for two sets of volatile organic compounds. The SCAD method was applied to training data, and effective MDs were selected in a lambda with the lowest cross-validation error (lambda(min)) and were defined as the inputs to the ANN modeling method. All ANN parameters were optimized simultaneously. Some statistical parameters were computed, and the obtained results indicate that the constructed QSRR models have acceptable values. Also, the applicability domain analysis reveals that more than 95% of the data are in the confidence range, indicating that the prediction results of the SCAD-ANN models are reliable.