Quantitative Structure–Retention Relationship Study of Benzodiazepines Using Adaptive Neuro Fuzzy Inference System as Feature Selection Method

Abstract

AbstractA Quantitative Structure–Retention Relationship (QSRR) study of 32 benzodiazepines is performed in this work. Two feature selection methods of Adaptive Neuro Fuzzy Inference System (ANFIS) and a stepwise regression approach adopted for the Multiple Linear Regressions (MLR) were used to predict the Liquid Chromatography‐Mass Spectrometry (LC‐MS) Retention Time (RT) of these compounds on a Xterra MS C‐18 stationary phase. ANFIS and MLR methods were used as variable selection tools and a neural network was employed for predicting the RTs. Three descriptors of 3D‐MoRSE‐signal 06/weighted by atomic polarizabilities (Mor06p), Radial Distribution Function‐1.0/weighted by atomic van der Waals volumes (RDF010v), and number of functional groups of RCoN<and >NCN (N‐072) reveal the importance of dispersion interactions between benzodiazepines and C‐18 stationary phases and also electrostatic and hydrogen bond interactions of these compounds with the polar mobile phase. The superiority of 3‐4‐1 ANFIS‐Artificial Neural Networks (ANN) model over MLR‐ANN with a linear feature selection method was illustrated by Leave‐Multiple‐Out (LMO) cross‐validation method. Values of R2L6O=0.771 and RMSEL6O=4.792 for ANFIS‐ANN model should be compared with the values of R2L6O=0.632 and RMSEL6O=6.481 for the MLR‐ANN model. LMO‐CV and Y‐randomization results indicate the robustness and predictive ability of the generated model of ANFIS‐ANN and show the ability of ANFIS in selecting the features for phenomena with nonlinear characteristics. Sequential Zeroing of Weights (SZW) as a sensitivity analysis method showed the importance of dispersion interaction in the retention mechanism of benzodiazepines in LC method.