Deep Learning on Physicochemical Space of Metabolites in Mass Spectrometry
Open AccessMetabolites are small molecules found in a specific cell, organ, or organism and play various roles. As the endpoints of gene expression and cell activity, the metabolome is arguably the most sensitive measure of phenotype. Mass spectrometry (MS), coupled with high-resolution separation techniques, has rapidly become one of the most widely used methods in metabolomics because of the advantages of high sensitivity and selectivity and rich information. However, the metabolic signal detection and identification are still two main challenges in this field. In this dissertation, by exploring the MS (m/z, time, intensity) space, we study the MS signal detection and correspondence between the MS profile characterization and physicochemical properties of metabolites. We first focus on improving weak metabolic signal detection from single-cell MS studies by evaluating the local intensity profiles in MS (m/z, time) space. We developed Trace, a software framework that incorporates deep learning to automate MS signal image feature selection and optimization for the extraction of trace-level signals from high-resolution MS data. The method was validated using manually curated data sets from single-cell metabolomic studies of the South African clawed frog (Xenopus laevis) embryo using capillary electrophoresis electrospray ionization MS. We demonstrated that Trace combines sensitivity, accuracy, and robustness with high data processing throughput to recognize signals. Secondly, as a follow-up to our Trace protocol, we build a web server for upgraded Trace, which incorporates transfer learning from large open-source datasets and the pre- trained deep learning (DL) models, and ab initio scanning of MS datasets. The newly developed web server makes this protocol more user-friendly for the community. Users can process their MS data online, and a list of detected signals with their distribution will be generated for download. Users can select from multiple deep learning models for data analysis or even train the model chosen with their customized datasets for optimal accuracy and sensitivity. Then, given the high consistency of MS signal images among different experiments for individual metabolites, we aim for the real trace-level MS signal detection with the FaceNet model. FaceNet, a powerful tool for image space embedding and distance metric, is adopted to evaluate the signal similarity in multiple-sample MS studies in the embedding space. It provides a statistical strategy of trace-level MS signal detection by assessing the consistency of the potential signal across multiple samples. Higher consistency indicates a higher chance of the existence of a real chemical compound. Compared with the Trace protocol, this approach aims for higher sensitivity in real trace-level MS signal detection. Lastly, the correspondence between the physicochemical properties of metabolites and their intensity profiles in MS (m/z, time) space is explored. By deciphering the relation between the molecular characterization of metabolites and the first and higher moments (location and shape, respectively) of signal images expanded in (m/z, time) space of MS studies, the research aims to achieving direct detecting and identification/ classification of metabolites and other chemical compounds from MS datasets.
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.