New strategies based on machine learning and cheminformatics for untargeted metabolomics data annotation

Identifier:  TDX:4569

Handle: https://hdl.handle.net/20.500.11797/TDX4569

Authors:  Martinez De Cripan Vazquez, Sara

Abstract:
Untargeted metabolomics experiments generate hundreds to thousands ot ion signals trom metabolites, and reterence library searching is the primary method for compound identification. However, only a small portion of the detected signals are annotated as known compounds. Although the availability of reference libraries and reference standards has increased in recent years, they are still insufficient. This thesis aims to develop new computational strategies based on machine learning (ML) and chemoinformatics methods as reliable alternatives to libraries in untargeted metabolomics annotation. Also, the thesis aims to apply and validate these strategies in untargeted metabolomics experiments. The thesis is divided into two main parts. First, an algorithm called Structure-based fragment annotation (SBFA) was developed to guide the identification of metabolites based on the annotation of in-source fragments al the MS1 leve! without using reference MSIMS librarles. Second, the use of ML models to predict molecular properties, such as retention index (RI) and Collision Cross Section (CCS) values, was assessed. A framework for applying these predicted values for metabolite annotation was provided, which is of special utility in the absence of reference data.