Artificial Intelligence for Long-Term Reproducible High-Throughput Untargeted Metabolomics

Abstract

Mass spectrometry (MS)-based assays suffer from the inherent variability of measurements across instruments and over time, caused by multiple sources of variation, such as differences in sample preparation and system setups, biological matrix effects, acquisition batch effects and so on. Across -omics, reproducibility of quantitative experiments is a well-known issue. Untargeted metabolomics is the method of choice for comprehensive characterization of all chemical compounds that occur in a cell of a biological sample. With growing demand for untargeted metabolomics in personalized health applications, it is crucial to achieve the level of reproducibility enabling robust sample quantification in longitudinal clinical studies. In this PhD thesis, we aim at improving reproducibility of untargeted metabolomics leveraging the most recent developments in AI. We build a platform for continuous system suitability testing (SST), develop a batch correction method and investigate calibration strategies for a high-throughput acquisition method. Complementary to these efforts, we investigate representation learning approaches across data modalities and develop an explainable deep learning application to demonstrate exciting opportunities for multi-modal biomedical research.