Machine learning and metabolic modeling-based identification of hypoxia-driven metabolic signatures in pediatric cancers

Metabolic reprogramming in pediatric cancers under hypoxia has been much less studied compared to adult-onset cancers; such studies are essential to identify relevant therapeutic targets at different hypoxic levels. Genome-scale metabolic modeling studies of cancer metabolism are used to elucidate reprogrammed metabolic pathways, analyze heterogeneity between cancer types and subtypes, identify synthetic lethality, perform in silico drug targeting, etc. We have used reactive-species-integrated genome-scale metabolic models of ten different pediatric cancer cell lines to understand metabolic rewiring under hypoxia. These pediatric cancer models exhibited a consistent metabolic signature under the given constraints, under each oxygen condition. We applied constraint-based modeling to simulate normoxic, hypoxic, and extreme hypoxic gradients and used machine learning to analyze the metabolic pathways that predominate across these gradients in pediatric cancers. Our machine learning classifiers showed improved accuracy in classifying the different gradients and generated metabolic features of importance. We performed a reaction-level flux analysis on the predicted discriminative metabolic features of importance. We observed that most metabolic pathways are not linearly rewired when hypoxic levels increase. Hypoxia-specific metabolic rewiring is enriched with reactive oxygen species, and extreme hypoxic cancer states exhibit a positive contribution from mitochondrial reactive sulfur species. Further, we propose that reactive sulfur species and reactive oxygen species detoxifying reactions contribute to redox balance in extreme hypoxia and hypoxia, respectively.