Whole tissue transcriptomic analyses have been helpful to characterize molecular subtypes of hepatocellular carcinoma (HCC). Metabolic subtypes of human HCC have been defined, yet whether these different metabolic classes are clinically relevant or derive in actionable cancer vulnerabilities is still an unanswered question. Publicly available gene sets or gene signatures have been used to infer functional changes through gene set enrichment methods. However, me-tabolism-related gene signatures are poorly coexpressed when applied to a biological context. Here, we apply a simple method to infer highly consistent signatures using graph models. Using The Cancer Genome Atlas Liver Hepatocellular cohort (LIHC), we describe the main metabolic clusters and their relationship with commonly used molecular classes, and with the presence of TP53 or CTNNB1 driver mutations. We find similar results in our validation cohort, the LIRI-JP cohort. We describe how previously described metabolic subtypes could not have therapeutic rel-evance due to their overall downregulation when compared to non-tumoral liver, and identify N-Glycan, Mevalonate and Sphingolipid biosynthetic pathways as the hallmark of the oncogenic shift of the use of Acetyl-coenzyme A in HCC metabolism. Finally, using DepMap data, we demonstrate metabolic vulnerabilities in HCC cell lines.