Quantitative structure-retention relationship model for the determination of naratriptan hydrochloride and its impurities based on artificial neural networks coupled with genetic algorithm

Abstract

Mathematical modeling of Quantitative Structure – Property Relationships met great interest in fields of in silico drug design and more recently, pharmaceutical analysis. In our approach we proposed automated method of creation Quantitative Structure-Retention Relationship (QSRR) for analysis of triptans, selective serotonin 5-HT1 receptor agonists used for the treatment of acute headache. The method was created using hybrid machine learning approach, namely Genetic algorithm (GA) coupled with artificial neutral networks (ANN). Performance of proposed hybrid GA-ANN model was evaluated with predicting relative retention times of naratriptan hydrochloride impurities. Several ANN types were coupled with GA and tested: single-layer ANN (SL-ANN), double-layer ANN (D-ANN) and higher order architectures: pi-sigma ANN (PS-ANN) and sigma-pi-sigma ANN (SPS-ANN). Partial Least Squares (PLS) method was used as a reference. The separation of naratriptan hydrochloride and its related products (impurities and degradation products) was obtained by developing a gradient high-performance liquid chromatography method with diode-array detector (HPLC-DAD). Degradation products during acid-basic hydrolysis were identified with an electrospray ionization tandem mass spectrometry (Q-TOF-MS/MS) detector. Independent data for outer validation of QSRR model was obtained from the determination of related products of sumatriptan succinate via an HPLC-DAD method.Accuracy of QSRR was measured by inner-validation on naratriptan data and outer validation on sumatriptan succinate samples. The best performing model were PS-ANN and SPS-ANN with mean errors of 8% (Q2=0.87) and 15% (Q2=0.77) on an inner-validation data set, respectively. Validation on similar samples from an outer validation data set of sumatriptan succinate impurities gave mean errors of 18% (R2pred=0.64) and 17% (R2pred=0.63) for the PS-ANN and SPS-ANN models, respectively.