Quantitative structure – property relationship studies of liquid chromatography – mass spectrometry responsiveness of aripiprazole and its impurities using artificial neural networks

Abstract

The importance of studying electrospray ionization (ESI) technique stems from its widespread liquid chromatography – mass spectrometry (LC-MS) character, particularly emphasized in scientific areas with most of non-volatile and highly- or semi- polar analytes. Although it is known that ESI operating gas-phase ions are formed under the influence of electrical – thermal – pneumatic energy, applied to analyte-containing liquid after its removal from capillary tube, insufficiently explained complexity of process decelerates finding optimal instrumental settings and, hence, providing high ionization efficiency. Special consequence represents obtaining varying responsiveness among group of analytes of identical concentration levels, measured under the same experimental conditions. In this regard, developing quantitative analytical methods could be compromised. Taking into account previously associated findings, the purpose of this study was to examine LC-MS response behavior of test substances - atypical antipsychotic aripiprazole and its impurities in dependence with instrumental, solvent and analyte-related properties by employing Quantitative Structure – Property Relationship (QSPR) approach. QSPR methodology primarily quantifies correlation between property of interest, determined for a series of analytes in a chosen analytical system, and molecular descriptors - in silico calculated or experimentally obtained numerical values attributed to certain chemical information. Assuming nonlinear relationship between observed variables, artificial neural network (ANN) was used for establishing a comprehensive QSPR model with good response prediction ability. Data table for model building was composed of molecular descriptors, as well as, LC and ESI source parameters varied according to Box-Behnken design of experiments. Set of molecular descriptors included polar surface area, pKa, logP, logD, molecular volume, surface tension, vapor pressure, number of proton acceptors of analytes, as well as, viscosity, conductivity and surface tension of utilized mobile phases, due to encoding ESI-relevant chemical information. Additionally, the amount of methanol in mobile phase, pH of aqueous portion, flow rate of mobile phase, sheath gas flow, auxiliary gas flow, spray voltage and capillary temperature were considered as model’s input variables, based on significant impact on peak areas, shown within previously performed Placket-Burman design. Within it, unexpectedly occurred confounding patterns, guided by the alias matrix methodology, directed settings of involved, but statistically insignificant screened factors. Predictive power of finally obtained model was estimated using internal, 10-fold cross-validation procedure. Developed QSPR model showed satisfactory performance in terms of low root mean square errors (RMSE) and relatively high values of cross-validated coefficient of determination (Q2). In particular, study has demonstrated suitability of utilized modeling approach in finding the optimal experimental parameters and, hence, usefulness for efficient LC-MS method development. However, generalization of manifested underlying physicochemical mechanisms must be precluded and limited only to the examined system, regarding relatively small number of available structures.