|
Abstract INTRODUCTION:
Metabolic dysregulation is a hallmark of diseases including cancer, cardiovascular disease, and diabetes. A quantitative understanding of how metabolism is altered in these conditions is critical for therapeutic development. Metabolic flux analysis, which examines the rate at which metabolites flow through biochemical pathways, offers deep insights into metabolic activity and regulation. Stable isotope tracing experiments remain the most optimal method for dynamic measurement of metabolic processes. When combined with holistic computational models, they also offer a powerful tool for the prediction and validation of clinical interventions. In order to guide co-clinical applications robust and easy to use solutions are urgently needed.
OBJECTIVES:
Here we aim to create an easy-to-use toolkit for quantitative 13C metabolic flux analysis in human cells, for use by clinicians and researchers investigating metabolic activity, drug action, and potential resistance mechanisms. By lowering the technical barrier to flux analysis, we aim to provide a practical tool for interpreting metabolic dynamics beyond metabolite abundance. To this end, we have developed a comprehensive and user-friendly workflow that integrates liquid chromatography-mass spectrometry and stable isotope tracer analysis with machine learning-driven flux inference using Bayesian neural networks. Currently the workflow supports the design of isotope tracer experiments, exploratory data analysis, and probabilistic flux inference, with a focus on key flux ratios in human central carbon metabolism.
METHODS:
Cell Culture and Isotope Labeling:
Human cancer cell lines including adult leukemia (HAP1), pediatric osteosarcoma (HOS), and neuroblastoma (SKNBE2), were cultured under standard conditions and treated with three targeted metabolic inhibitors: (1) 2-deoxyglucose (2DG): glycolysis inhibitor; (2) 6-aminonicotinamide (6AN): oxidative pentose phosphate pathway inhibitor (PPP); (3) CB-839 (CB): glutaminase inhibitor (glutaminolysis). Cells were incubated with the stable isotope-labeled tracer [1,2-13C]-glucose to focus on glycolysis and PPP, and prepared for LC-MS analysis using a simple solvent extraction and centrifugation to obtain a final extract of approximately 250k cells in 300 µL 1:4 (v/v) water/acetonitrile.
Metabolite Profiling:
We developed a semi-targeted high-resolution HILIC-LC-MS method specifically focused on central carbon metabolites, enabling sufficient separation within a timeframe appropriate for clinical translation. Calibration curves were obtained using non-enriched QC samples, enabling quantification in the presence of sample matrix in the approximate range of 5 nM - 0.1 mM of extracted sample for most compounds. As full scan acquisitions on Orbitrap instruments often have observably biased isotope ratios, semi-targeted MS1 acquisition windows were incorporated for more accurate quantification of the relevant compounds. Data were preprocessed and analyzed in R prior to downstream analysis and modelling.
Flux Inference Framework:
Fluxes were inferred using a simulation-based approach that couples 13C-metabolic flux analysis with Bayesian neural networks in Python. The inference workflow consists of: (1) atom-resolved metabolic network modeling (EMU-based), (2) simulation of expected isotopologue patterns for candidate flux distributions, (3) training of a Bayesian neural network to map observed data to posterior flux ratios, incorporating uncertainty.
This architecture enables rapid, interpretable, and probabilistically grounded flux inference across different flux ratios.
RESULTS:
To optimize tracer selection for flux inference, Bayesian neural networks were trained on synthetic datasets simulating isotope labeling patterns under different tracer conditions. This yielded design-of-experiment maps that guide users in selecting the most informative tracers for specific regions of metabolism. For example, [1,2-13C]-glucose was optimal for resolving fluxes between glycolysis and the oxidative pentose phosphate pathway, while [U-13C]-pyruvate and [3-13C]-glutamine provided complementary information on anaplerotic versus oxidative TCA cycle activity. The selection focused on five flux ratios covering upper glycolysis, PPP, pyruvate utilization regarding TCA cycle, serine biosynthesis, and glutaminolysis, each chosen to inform on flux partitioning in central carbon metabolism.
For biological validation, our workflow was applied to the three cancer cell lines described above using [1,2-13C]-glucose to specifically probe the flux split between glycolysis and PPP. Key findings include: (1) 6AN treatment significantly increased glycolytic flux relative to PPP in HAP1 and HOS, indicating compensatory rerouting under PPP inhibition. (2) SKNBE2 displayed a non-significant trend in the same direction. (3) 2DG, unexpectedly, increased glycolytic flux relative to PPP in HOS and SKNBE2. (4) CB-839 consistently increased relative PPP flux across all cell lines, potentially reflecting enhanced NADPH production under glutaminolysis-inhibition.
CONCLUSION:
Our workflow offers a streamlined approach for integrating isotope tracing, high-resolution metabolomics, and Bayesian neural networks to infer pathway activity. The initial application to three cancer cell lines illustrates the potential of this approach to detect drug-induced metabolic rewiring and to capture context-specific metabolic adaptations. While the method shows promise for accessible and interpretable flux analysis, further validation with additional tracers and experimental systems is needed. These early results support its future use in mechanistic studies, biomarker exploration, and therapeutic profiling. With continued development, this workflow may contribute to broader adoption of metabolic flux analysis in biomedical research.
|
= Discovery stage. (53.14%, 2025)
= Translation stage. (22.33%, 2025)
= Clinically available. (24.53%, 2025)
