Applications of Data-Driven Network Analysis in Metabolomics

Abstract

Metabolomics is a systems-wide study of small molecule metabolites. It provides a read-out of underlying cellular and biochemical events. Liquid Chromatography coupled with Mass Spectrometry (LC-MS) is one of the most common analytical platforms used to perform metabolomics studies. The analysis of LC-MS metabolomics data is a complex multi-step process. It involves data processing, normalization, followed by statistical analysis and functional interpretation. While several computational tools have been built to help perform these tasks, a major challenge remains linking alterations in metabolite levels to specific biological processes. In this dissertation, I develop and apply novel computational methods for the analysis and interpretation of metabolomics data, to help build testable hypotheses and derive novel biological insights.

Over the past decade, mapping and visualizing experimentally measured metabolites in the context of known biochemical pathways has become ubiquitous. However, pathway mapping is restricted to named metabolites from well-annotated biochemical pathways. Realizing the limitations of knowledge-based approaches, in Chapter Two, we developed a bioinformatics tool, Filigree, that provides a data-driven approach by inferring associations among metabolites directly from experimental measurements to construct metabolic networks. These associations can be quantified by ‘partial correlations' that measure the conditional dependence between metabolites, thus eliminating spurious and non-informative interactions. In a high-dimensional setting (n << p), the partial correlation network is computed using the l1-regularized graphical lasso method. The Differential Network Enrichment Analysis (DNEA) algorithm that Filigree implements computes the network using a joint estimation method (JEM) which allows the use of all samples in both experimental groups by modifying the graphical lasso penalty term. The network is then clustered using consensus clustering to identify highly interconnected subnetworks; and finally, the enrichment of these subnetworks is determined using the NetGSA algorithm. In addition, Filigree addresses common challenges that often arise in the analysis of “real world” metabolomics data like high dimensionality (n << p) and highly imbalanced experimental groups. To demonstrate Filigree’s applicability, I analyzed metabolomics datasets from type 1 and type 2 diabetes and lipidomics dataset from the Michigan Mother-Infant Pairs (MMIP) cohort and were able to identify previously known and some novel biochemical disruptions leading to an altered metabolic state.

In Chapter Three, I analyzed a COVID-19 metabolomics data to identify metabolic markers of disease severity. My analysis revealed that the plasma metabolome of COVID-19 patients and healthy controls is strongly influenced by clinical characteristics as well as anesthetic administration for intubation. There were distinct differences in the metabolic profiles of patients with mild and severe COVID-19. These differentiating metabolites included several acylcarnitines and acylglycerols and were better able to discriminate mild and severe COVID when compared to clinical risk factors.

In Chapter Four, I assessed the association of data-driven metabolic modules with the BMI trajectory of ALS (Amyotrophic Lateral Sclerosis) patients over 5- and 10-years preceding diagnosis. I showed that while individual metabolites do not show a significant association with BMI trajectory, metabolic modules obtained from partial correlation networks do, suggesting a nuanced relationship between BMI trajectory and the metabolome. Additionally, a subset of these metabolites was individually predictive of ALS survival as well, indicating a metabolic link between loss of BMI and ALS survival.