Addressing sources of variance in large-scale metabolomics

TY - GEN

T1 - Addressing sources of variance in large-scale metabolomics

AU - Lassen, Johan Kjeldbjerg

PY - 2025/4/8

Y1 - 2025/4/8

N2 - Metabolomics offers a direct insight into the biochemical processes of a biological sample, providing a snapshot of its metabolic state. By analyzing the complete set of metabolites – small molecules of the cellular processes – metabolomics allows researchers to investigate organisms and uncover biomarkers for disease diagnosis, progression, and treatment response. This powerful approach integrates complex data from various analytical techniques, such as liquid chromatography-mass spectrometry (LC-MS) and matrix-assisted laser desorption/ionization mass spectrometry (MALDI-MS), allowing screening of both known and unknown molecules for a holistic search of new markers. A significant limitation of metabolomics is the difficulties in separating variance that arise from technical and biological confounders from the outcome variance of interest. Biological variance can stem from genetic diversity, environmental factors, and general day-to-day fluctuations among samples. Technical variance, on the other hand, originates from factors such as inconsistencies in sample preparation, instrument performance, and computational data processing. In this thesis, I aim to explore and mitigate these sources of variance to enhance the reliability and reproducibility of metabolomics studies. By employing robust statistical models, machine learning techniques, and tailored experimental protocols, I optimize the extraction of true biological signals amidst the noise. I show that large-scale metabolomics studies are viable and have used them specifically to gain insights into aging processes, osteoporosis progression, and antimicrobial resistance predictions. These studies demonstrate the methodological perspectives of large-scale untargeted metabolomics, longitudinal sampling and end-to-end workflows. Specifically, I have modelled person age with root mean squared error (RMSE) of 5.77 years in ten thousand untargeted metabolomics samples, predicted osteoporosis one year ahead of the original diagnosis at an area under the receiver operating curve (ROC-AUC) of 0.72, and demonstrated that end-to-end modeling of MALDI-MS data may predict antibiotic resistance with near perfect accuracy. The proposed approaches not only improve the precision of metabolic profiling but also pave the way for more accurate biomarker discovery and better understanding of disease mechanisms. Finally, I highlight that the presented work needs methodological and biological validation in future studies to assess the overall generalizability.

AB - Metabolomics offers a direct insight into the biochemical processes of a biological sample, providing a snapshot of its metabolic state. By analyzing the complete set of metabolites – small molecules of the cellular processes – metabolomics allows researchers to investigate organisms and uncover biomarkers for disease diagnosis, progression, and treatment response. This powerful approach integrates complex data from various analytical techniques, such as liquid chromatography-mass spectrometry (LC-MS) and matrix-assisted laser desorption/ionization mass spectrometry (MALDI-MS), allowing screening of both known and unknown molecules for a holistic search of new markers. A significant limitation of metabolomics is the difficulties in separating variance that arise from technical and biological confounders from the outcome variance of interest. Biological variance can stem from genetic diversity, environmental factors, and general day-to-day fluctuations among samples. Technical variance, on the other hand, originates from factors such as inconsistencies in sample preparation, instrument performance, and computational data processing. In this thesis, I aim to explore and mitigate these sources of variance to enhance the reliability and reproducibility of metabolomics studies. By employing robust statistical models, machine learning techniques, and tailored experimental protocols, I optimize the extraction of true biological signals amidst the noise. I show that large-scale metabolomics studies are viable and have used them specifically to gain insights into aging processes, osteoporosis progression, and antimicrobial resistance predictions. These studies demonstrate the methodological perspectives of large-scale untargeted metabolomics, longitudinal sampling and end-to-end workflows. Specifically, I have modelled person age with root mean squared error (RMSE) of 5.77 years in ten thousand untargeted metabolomics samples, predicted osteoporosis one year ahead of the original diagnosis at an area under the receiver operating curve (ROC-AUC) of 0.72, and demonstrated that end-to-end modeling of MALDI-MS data may predict antibiotic resistance with near perfect accuracy. The proposed approaches not only improve the precision of metabolic profiling but also pave the way for more accurate biomarker discovery and better understanding of disease mechanisms. Finally, I highlight that the presented work needs methodological and biological validation in future studies to assess the overall generalizability.

M3 - PhD thesis

T3 - PhD Theses - Department of Molecular Biology and Genetics

PB - Aarhus Universitet, Institut for Molekylærbiologi og genetik

ER -