Prediction of gas chromatographic retention times of narcotic and hazardous drugs in blood using QSRR and machine learning models

Abstract

The reliable identification of narcotic and hazardous drugs in blood is of critical importance in forensic, clinical, and public health investigations. In this work, gas chromatography (GC) combined with quantitative structure–retention relationship (QSRR) modeling was employed to predict the retention times (RTs) of narcotic and hazardous drugs in blood samples. Experimental RTs of 75 drugs were determined using GC equipped with a non-polar HP-5 column, and a wide range of molecular descriptors was calculated from optimized molecular structures. Genetic algorithms were applied for descriptor selection, and linear and nonlinear predictive models, including GA-PLS, GA-KPLS, and artificial neural networks (ANN), were developed and evaluated using leave-group-out cross-validation and an external test set. The results demonstrated that nonlinear approaches provided superior predictive performance compared to linear models, with the ANN model showing the highest accuracy (R² = 0.969 for the training set and 0.932 for the test set) and the lowest prediction errors. Analysis of selected descriptors revealed that molecular hydrophobicity, structural complexity, hydrogen bonding capability, and three-dimensional molecular features play a significant role in chromatographic retention behavior. Overall, the proposed QSRR and machine learning framework enables accurate prediction of GC retention times, reduces the need for extensive experimental measurements, and offers an efficient tool for the screening and analysis of novel narcotic and hazardous drug derivatives.