Open Collections

Wang, Yukai

University of British Columbia

Metabolomics, the study of small molecules within biological systems, offers valuable insights into biochemical processes. This thesis addresses two challenges in mass spectrometry-based metabolomics: processing complex breathomics data and enhancing the structural annotation of metabolites using deep learning. In the first study, I introduce BreathXplorer, an open-source Python package designed to process real-time exhaled breath data from secondary electrospray ionization high-resolution mass spectrometry (SESI-HRMS). BreathXplorer tackles the challenge of non-Gaussian metabolic signal shapes by employing topological algorithms or Gaussian mixture models (GMM) to identify exhalation intervals and density-based spatial clustering (DBSCAN) to cluster m/z values. It accurately determines the start and end points of exhalation, ensuring precise quantitative measurements. In a proof-of-concept study on exercise breathomics, BreathXplorer identified exercise-responsive metabolites, showing its potential in real-time metabolomics research. In the second study, I explore deep learning as a solution for compound annotation in tandem mass spectrometry (MS/MS). This offers a predictive strategy where traditional library search tools are limited due to small spectral libraries. Generative models have become a fundamental approach for various tasks; however, their application to MS/MS structural annotation is less developed and often underperforms compared to traditional fingerprint-based counterparts. In this work, I investigate the potential of generative models by building and comparing transformer-based fingerprint models and generative models. This comparison helps to understand their strengths and limitations in annotating chemical structures. Training and testing on 616,594 unique structures, I identified three key limitations of direct structure generation with generative models: (1) error accumulation, (2) generation of invalid compounds, and (3) generation of unrecorded compounds. I then propose a solution using the generative model as a ranker. My results demonstrate that generative-based ranking outperforms fingerprint-based systems. Furthermore, I observed a positive correlation between high generation scores and the correctness of predicted structures. My analysis shows that generative models can leverage richer structural information during training, leading to improved accuracy in end-to-end chemical structure identification.

Text

2024-10-18

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/89449

Bioinformatics

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nc-nd/4.0/