Uncertainty-driven model-based search methods for method development in liquid chromatography.

Empirical retention modelling is a proven method development approach that uses a limited set of initial "scanning" experiments to build analyte retention models and predict optimal experimental conditions from a predicted chromatographic response function (CRF) surface. Once initial models have been established, literature shows how smart algorithms can be used to further update the surface during their search for the optimal separation conditions. In this study, several search strategies were compared in silico using a set of 122 randomized mixtures. Each strategy was evaluated under varying degrees of peak detectability. Three main strategies were assessed: (i) a classic CRF-driven approach, wherein new search runs were conducted at the working point with the highest estimated CRF-value (ii) a new global uncertainty-driven strategy aimed at first improving all initial models simultaneously before attempting to maximize the objective, and (iii) a per-analyte variant of strategy ii). It was found that, as soon as peak detection was hindered (by peak overlap), the uncertainty-driven strategies outperformed the conventional CRF-driven searches. For example, when a resolution of Rs = 0.5 was needed to correctly identify and track the peaks, the uncertainty-driven searches only failed for 1 out of 122 samples (0.8 %), while the CRF-driven search failed in 14 % of the cases. When Rs = 1.0 was needed, failure rates increased to 4 % for the best uncertainty-driven approach, while the failure rate for the conventional CRF-driven approach grew to a very problematic 42 %, despite the retention time parameter estimation being conducted with the same model as the one used to generate the retention time data in the simulated chromatograms. To conclude, our results show that in the current setting uncertainty-driven strategies achieve much greater accuracy in identifying the desired separation than the CRF-driven approach. This improvement comes at the cost of a higher total analysis time, as a consequence of the richer retention time information gathered by the algorithm.