Deciphering the metabolic heterogeneity of cancer cell lines using metabolomics and computational analysis

It has been a century since Otto Warburg presented his observation of aberrant metabolic activity in cancer. He noted a propensity for tumors to metabolize glucose and ferment lactate, even when oxygen is available. This finding was the first indication of the central role of metabolism in cancer. However, later research has shown that this metabolic phenotype, coined Warburg effect, is merely one of many strategies used by cancer to grow rapidly. Why study metabolic reprograming in cancer? So far, the standard of care targeted mainly signaling pathways and cell interactions. These treatments resulted in some therapeutic success but also the emergence of resistance. Even though metabolism has played a crucial role in cancer diagnosis, i.e. through the uptake of radiolabeled glucose, the use of cancer metabolism as therapy has scarcely been explored. Targeting essential metabolic reactions could compromise its cancer's operation. Unraveling how cancer cells reprogram their metabolism, identifying which pathways are being used and at which rate could guide the development of novel therapies to improve patient outcome, overcome cancer resistance, or prevent relapse. Large-scale metabolomics studies, ranging from biotechnology to human diseases, have been booming in recent years due to technological advances in analytical methods. Despite the increased availability of large metabolomics datasets, most studies rely on simple statistics for data analysis, and the interpretation of such results remains manual. To date, one of the major challenges in the field of metabolomics is the lack of computational approaches that facilitate interpretation. More importantly, metabolomic measurements alone cannot reveal cells' metabolic operation, as one would need to measure intracellular metabolic fluxes. In mammalian cells, however, fluxes are hard to measure and their indirect inference using, for example, genomics or proteomics remains unsatisfactory. Thus, identifying the functional differences in cancer cells, i.e. how their metabolic fluxes differ across the metabolic network, has never been accomplished at large scale. In this thesis, we explore the use of metabolomics to infer metabolic operation in cancer cell lines. By comprehensively measuring metabolism's intermediates, metabolites, in a panel of 180 cancer cell lines, our goal is to unravel cancer's metabolic heterogeneity. In brief, we want to identify in each cell line which metabolic pathways are active, thus possibly important. To do so, we have applied and developed computational methods for omics integration. One of the methods enables the identification of metabolite changes related to metabolic activity, integrating metabolomics with fluxomics. Another deciphers possible molecular drivers of metabolic phenotype, hence linking metabolomics with all omics. To set the stage, we review in chapter 2 the current approaches for multi-omics integration and compare their contribution in generating novel biological insights. We find that most methods rely solely on data for integration, ignore the vast biochemical knowledge, e.g. cellular pathways, and rarely provide mechanistic insights. We argue that novel computational approaches, which would take advantage of our current biological knowledge and the recent advances in machine learning, could unfold the full potential of these multi-layer datasets. For a thorough investigation of cancer metabolism variability, we analyze in chapter 3 the intracellular metabolome of 180 pan-cancer cell lines using untargeted metabolomics. These cell lines are subject to in-depth normalization to ensure quality in our measurements. We apply metabotyping, a method based on pathway-centric clustering to identify pathway activity, which our lab previously developed to investigate breast cancer cell lines. We find that metabotyping does not scale well for large datasets. We devise an alternative approach that yields improved results. However, we argue that metabotyping using clustering renders artificial cluster separation and fails to identify functionally relevant metabolic types. In chapter 4, to tackle the goal of inferring pathway activity from untargeted metabolomics, we derive a new method which captures metabolite patterns that are functionally associated with fluxes. This method employs metabolites, metabolic pathways definition, and principal component analysis to infer pathway activity. We demonstrate that our method reveals differences in fluxes using published E. coli and S. cerevisiae datasets. Moreover, we show that metabolite levels can predict some fluxes. Lastly, in chapter 5, we apply this new method to the metabolomics data for 180 cancer cell lines of chapter 3. This procedure provides a global view of metabolic operation. We find that cancer cell lines fall into two main subtypes; one relies on mitochondrial pathways and lipid biosynthesis, and the other on some sugar metabolism pathways. In follow-up experiments, we demonstrate the validity of the obtained cancer pathway activity map using 13C labeling. Finally, to decipher the possible molecular drivers, we integrate all omics to our metabolic phenotypes and find many associations which include HIF1A, TGF-β, and EMT. In the concluding remarks, we summarize the key findings of this thesis and assess the significance of the rather not so heterogeneous cancer phenotypes. We discuss the main metabolic operation of cancer in vitro and its implications for therapeutics targeting metabolism. Overall, this thesis showcases the relevance of untargeted metabolomics for the discovery of metabolic alterations in cancer cells. Our computational method based on biochemical knowledge provides a powerful data-driven approach to investigate the metabolic operation of cancer