Mechanistically informed machine learning links non-canonical TCA cycle activity to Warburg metabolism and hallmarks of malignancy

Cancer cells undergo extensive metabolic rewiring to support growth, survival, and phenotypic plasticity. A non-canonical variant of the tricarboxylic acid (TCA) cycle, characterized by mitochondrial-to-cytosolic citrate export, has emerged as critical for embryonic stem cell differentiation. However, its role in cancer remains poorly understood. Here, we present a two-step computational framework to systematically analyze the activity of this non-canonical TCA cycle across over 500 cancer cell lines and investigate its role in shaping hallmarks of malignancy. First, we applied constraint-based modeling to infer cycle activity, defining two complementary metrics: Cycle Propensity, measuring the likelihood of its engagement in each cell line, and Cycle Flux Intensity, quantifying average flux through the reaction identified as rate-limiting. We identified distinct tumor-specific patterns of pathway utilization. Notably, cells with high Cycle Propensity preferentially reroute cytosolic citrate via aconitase 1 (ACO1) and isocitrate dehydrogenase 1 (IDH1), promoting [Formula: see text]-ketoglutarate ([Formula: see text]KG) and NADPH production. Elevated engagement of this cycle strongly correlated with Warburg-like metabolic shifts, including decreased oxygen consumption and increased lactate secretion. In the second step, to uncover non-metabolic transcriptional signatures associated with non-canonical TCA cycle activity, we performed machine learning-based feature selection using ElasticNet and XGBoost, identifying robust gene signatures predictive of cycle activity. Over-representation analysis revealed enrichment in genes involved in metastatic behavior, angiogenesis, stemness, and key oncogenic signaling. SHapley Additive exPlanations (SHAP) further prioritized genes with the strongest predictive contributions, highlighting candidates for experimental validation. Correlation analysis of DepMap gene-dependency profiles revealed distinct vulnerability patterns associated with non-canonical TCA cycle activity, outlining a characteristic landscape of genetic dependencies. Together, our integrative framework uniting constraint-based metabolic modeling and machine learning systematically reveals how non-canonical TCA cycle dynamics underpin metabolic plasticity and promote malignant traits.