Constraint-based informed Machine Learning links non-canonical TCA cycle activity to Warburg metabolism and hallmarks of malignancy

Cancer cells undergo extensive metabolic rewiring to support proliferation, survival, and phenotypic plasticity. While a non-canonical variant of the tricarboxylic acid (TCA) cycle, characterized by mitochondrial citrate export and cytosolic metabolism, has been implicated in embryonic stem cell differentiation, its relevance in cancer metabolism remains poorly understood. Here, we introduce a hybrid computational framework that integrates constraint-based metabolic modeling and machine learning to systematically characterize non-canonical TCA cycle activity across more than 500 cancer cell lines and assess its association with metabolic and phenotypic hallmarks of malignancy. In the first step, we applied constraint-based modeling combined with flux sampling to infer pathway engagement, defining two complementary quantitative metrics: Cycle Propensity, quantifying the frequency of non-canonical TCA cycle activation across sampled feasible flux states in each cell line, and Cycle Flux Intensity, measuring the average flux through the predicted rate-limiting reaction. This analysis revealed tumor-type–specific patterns of pathway utilization. High Cycle Propensity was associated with preferential rerouting of cytosolic citrate through aconitase 1 (ACO1) and isocitrate dehydrogenase 1 (IDH1), promoting α-ketoglutarate and NADPH production. Importantly, elevated non-canonical TCA cycle activity correlated with Warburg-like metabolic features, including reduced oxygen consumption and increased lactate secretion. In the second step, we employed machine learning–based feature selection (ElasticNet and XGBoost) to identify transcriptional programs predictive of non-canonical TCA cycle activity. Over-representation analysis highlighted enrichment in pathways related to metastasis, angiogenesis, stemness, and oncogenic signaling. SHapley Additive exPlanations (SHAP) enabled prioritization of genes with the strongest predictive contributions, suggesting candidates for experimental validation. Integration with DepMap gene-dependency data further revealed distinct vulnerability profiles associated with high cycle activity. Overall, this study demonstrates how hybrid constraint-based modeling and AI approaches can uncover mechanistic links between metabolic pathway activity and malignant phenotypes, providing a scalable framework for the investigation of metabolic plasticity in cancer.