Quantitative structure-retention relationships for rapid method development in hydrophilic interaction liquid chromatography of pharmaceutical compounds

The development of computer-assisted approaches capable of accurate prediction of the retention behaviour of analytes, leading to optimisation of chromatographic performance, is a major goal for method development in chromatography. Statistically-derived quantitative structure-retention relationships (QSRRs) represent a quite popular approach to retention prediction. Hydrophilic interaction chromatography (HILIC) is nowadays well known as a powerful technique for the separation of polar compounds. However, the detailed retention mechanism applicable in HILIC is still under some discussion and for this reason, method development in HILIC is difficult. The first part of this thesis concerns the application of QSRR methodology to predict the retention times of pharmaceutical test analytes on five HILIC stationary phases (bare silica, amine, amide, diol and zwitterionic), with a view to selecting the most suitable stationary phase(s) for the separation of these analytes. QSRR methodology seeks mathematical equations that correlate molecular features to chromatographic parameters. Genetic algorithm (GA) feature selection and partial least squares (PLS) regression were used to correlate experimental retention data to various densityfunctional-theory-computed molecular descriptors. The predictive power of the QSRR models was successfully evaluated performing an external validation to predict retention times of test compounds. The QSRR models developed were also utilised to provide some insight into the separation mechanisms operating in the HILIC mode. The second part of this thesis describes a Quality-by-Design workflow, which combines QSRR methodology with design of experiments (DoE) principles to successfully integrate predictive modelling into HILIC method development. DoE principles were first used to explore the chromatographic variables (percentage of acetonitrile, as well as pH and salt concentration) known to be effective in HILIC, followed by regression analysis to generate models capable of predicting retention parameters over a wide range of chromatographic conditions. The mathematical DoE model was shown to be highly predictive when applied to test conditions inside the design space. A QSRR model was then generated to predict retention times of test probes. A compound classification based on the concept of similarity was applied prior to QSRR modelling in order to enhance the predictive capability of QSRRs. Finally, the QSRR-DoE computed retention times of pharmaceutical test analytes and subsequently calculated separation selectivity factors were used to optimise the chromatographic conditions for efficient separation of targets. Quality assurance was achieved through the application of Monte Carlo simulation to propagate the prediction error. The desired separation for the target analytes was established experimentally, which confirmed the theoretical predictions. In the third and main part of the thesis, an in depth study on the strategies which enhance QSRRs prediction accuracy able to support HILIC method development was carried out. A similarity searching approach was applied in order to generate localised QSRR models, in which the retention of any given compound is predicted using only the most similar compounds in the available dataset. Two similarity measures were performed; retention factor ratio as a chromatographic similarity measure and Tanimoto index as the most popular similarity measure based on chemical structure. Prediction error was reduced when QSRR was based on similar compounds rather than using the entire dataset, with an excellent result for retention time (tR) similarity-based local models. However tR filtering is unable to be applied to a real-life situation, as the retention time of a new analyte is unknown. To tackle this challenge, a novel QSRR methodology was presented based on a dualfiltering strategy which combines Tanimoto similarity (TS) searching as the primary filter and tR similarity clustering as the secondary filter. To employ tR similarity filtering, correlation to a molecular descriptor was used as a measure of retention time. A comparison of diverse, global, TS-based and dual-filtering-based QSRR models over five different HILIC stationary phases showed that the proposed dual-filtering-based QSRR model was the most successful approach.