Bayesian optimization of separation gradients to maximize the performance of untargeted LC-MS

Abstract

Abstract Liquid chromatography (LC) with gradient elution is a routine practice for separating complex chemical mixtures in mass spectrometry (MS)-based untargeted analysis. Despite its prevalence, systematic optimization of LC gradients has remained challenging. Here we develop a Bayesian optimization method, BAGO, for autonomous and efficient LC gradient optimization. BAGO is an active learning strategy that discovers the optimal gradient using limited experimental data. From over 100,000 plausible gradients, BAGO locates the optimal LC gradient within ten sample analyses. We validated BAGO on six biological studies of different sample matrices and LC columns, showing that BAGO can significantly improve quantitative performance, tandem MS spectral coverage, and spectral purity. For instance, the optimized gradient increases the count of annotated compounds meeting quantification criteria by up to 48.5%. Furthermore, applying BAGO in a Drosophila metabolomics study, an additional 57 metabolites and 126 lipids were annotated. The BAGO algorithms were implemented into user-friendly software for everyday laboratory practice and a Python package for its flexible extension.