13CFLUX(v3): Next-Gen 13C Metabolic Flux Analysis

• 13CFLUX(v3) is a third-generation software platform for 13C MFA that simulates isotope labeling experiments and supports both steady and dynamic state analyses.
• It combines a high-performance C++ backend with a user-friendly Python interface, enabling multi-experiment integration, multi-tracer workflows, and advanced Bayesian inference.
• The modular architecture optimizes simulation speed, scalability, and maintainability while providing robust sensitivity analysis and uncertainty quantification.

13CFLUX(v3) is a third-generation software platform for $^{13}$C metabolic flux analysis ($^{13}$C-MFA) that simulates isotope labeling experiments and supports flux estimation in isotopically stationary and nonstationary workflows (Stratmann et al., 28 Sep 2025). It combines a high-performance C++ engine with a Python interface, is open-source, and is designed for modern fluxomics settings that include multiple experiments, multiple tracers, diverse analytical platforms, and advanced statistical inference such as Bayesian analysis. In the formulation presented for version 3, the platform generates labeling data from a FluxML model, fluxes and pool sizes, tracer or input labeling, and measurement configuration, with the stated ability to simulate “any desired labeling state of any metabolite” at any time point, including $t=\infty$.

1. Position within $^{13}$C metabolic flux analysis

$^{13}$C-MFA infers intracellular metabolic fluxes from isotope labeling data together with a metabolic network model. The paper characterizes it as the most informative fluxomics technique because it combines isotope labeling experiments, external rate measurements, metabolic network structure, and uncertainty estimation (Stratmann et al., 28 Sep 2025). The method is used across microbes, plants, and mammalian cells, and its computational demands have increased as experimental designs have expanded to include multiple experiments, multiple tracers, different analytical platforms, genome-scale models, and Bayesian inference.

Within that context, 13CFLUX(v3) is presented as a major update over 13CFLUX2. The update is described as a ground-up redesign that is faster, more flexible, more maintainable, open-source, and explicitly built to handle both classical and emerging workflows. The transition from a primarily stationary predecessor to a platform covering both stationary and nonstationary regimes is one of its central defining features.

The software is positioned not merely as a fitting tool but as a simulation and analysis engine for researchers who need to design tracer experiments, simulate isotopic labeling patterns, compute sensitivities and gradients, fit fluxes and pool sizes, assess uncertainty, run multiple experiments together, and perform Bayesian inference for complex INST data. This suggests that version 3 is intended to function as an infrastructural layer within broader computational fluxomics pipelines rather than as a single-purpose application.

2. Mathematical regimes: isotopically stationary and nonstationary analysis

The paper distinguishes two major regimes. In isotopically stationary (IST) analysis, isotope labeling has reached steady labeling equilibrium, so labeling states satisfy algebraic equations. In isotopically nonstationary (INST) analysis, labeling is tracked over time after tracer perturbation, so labeling states satisfy ordinary differential equations (Stratmann et al., 28 Sep 2025).

For INST, the supplement gives the governing equations in a cascaded form,

$$\mathrm{diag}({^k}\mathbf{X}) \cdot {^k}\dot{\mathbf{x}} = {^k}\mathbf{A}(\mathbf{v}) \cdot {^k}\mathbf{x} + {^k}\mathbf{b}(\mathbf{v}, {^0}\mathbf{x}, \dots, {^{k-1}}\mathbf{x}, {^\text{inp}}\mathbf{x}), \quad k=1,\dots,K.$$

Each cascade level $k$ is a lower-dimensional IVP whose solution depends on fluxes $\mathbf{v}$, pool sizes $\mathbf{X}$, earlier cascade levels, and input labeling. For IST, the time derivative vanishes at the asymptotic state,

$$\mathbf{0} = {^k}\mathbf{A}(\mathbf{v}) \cdot {^k}\mathbf{x} + {^k}\mathbf{b}(\mathbf{v}, {^0}\mathbf{x}, \dots, {^{k-1}}\mathbf{x}, {^\text{inp}}\mathbf{x}), \quad k=1,\dots,K.$$

The same framework also includes sensitivity systems for flux and pool-size estimation. These sensitivities are important because gradients and Jacobians are needed for parameter fitting, uncertainty quantification, and experimental design. The paper’s emphasis on both the forward labeling systems and the corresponding sensitivity systems indicates that 13CFLUX(v3) is designed around inference workloads, not only forward simulation.

A common simplification is to identify $^{13}$0C-MFA software with stationary workflows. The version-3 design explicitly rejects that reduction: INST support is one of the enumerated advances over 13CFLUX2, and the paper treats the asymptotic relation between INST and IST as a numerical consistency property rather than as a reason to collapse the two regimes into one.

3. Software architecture and implementation model

A principal contribution of 13CFLUX(v3) is its split architecture: a high-performance backend written in C++17 and a Python frontend built via pybind11 (Stratmann et al., 28 Sep 2025). The C++ engine is compiled into optimized machine code and uses Eigen for matrix and vector operations, SUNDIALS/CVODE for INST ODE integration, SparseLU for sparse linear solves, and OpenMP for multithreaded parallelism. The Python layer exposes the engine to Python 3.9–3.13.

The paper reports that the code base was refactored from over 130,000 LOC to under 15,000 LOC in the core C++ code, with the stated effect of improving maintainability and quality. This reduction is presented together with scalability and maintainability as explicit design goals, not as incidental implementation details.

The Python frontend is intended to make workflows easier to script, automate, and integrate with the scientific Python ecosystem. The paper specifically lists NumPy, SciPy, Matplotlib, Jupyter notebooks, third-party optimizers and samplers, and workflow tools like Airflow. The claimed practical effect is that users can write short, readable workflows rather than low-level simulation code; examples in the paper indicate that multi-start fitting or replacing the optimizer can be done with only a few lines of Python.

This division of labor between a compiled numerical core and a high-level scripting interface suggests a deliberate separation between throughput-critical simulation kernels and orchestration logic. In the paper’s presentation, performance and usability are therefore not treated as competing objectives but as properties attached to different layers of the same system.

4. State-space representations and numerical solvers

13CFLUX(v3) supports the two major state-space representations for isotope labeling, cumomers and EMUs (Stratmann et al., 28 Sep 2025). The software automatically performs graph-based topology analysis and dimension reduction to generate essential cumomer or EMU systems. The paper emphasizes that representation choice depends on the analytical setup: for some MS configurations, EMUs are faster, whereas for some MS/MS configurations, cumomers are faster. Accordingly, the software is described as automatically choosing the numerically most beneficial formulation.

That point is methodologically important because it avoids treating EMUs as universally superior. The software instead encodes representation selection as a problem-dependent numerical decision. A plausible implication is that version 3 aims to minimize avoidable formulation overhead before the actual linear algebra and ODE integration stages begin.

For INST, the platform solves stiff ODE systems using BDF methods via a customized CVODE implementation and also an SDIRK solver. The paper highlights a tailored linear solve at each solver step,

$^{13}$1

Instead of an iterative GMRES solve, the implementation exploits sparse structure and uses direct sparse LU factorization, which the paper reports as improving efficiency. Adaptive step-size control is built in, with default tolerances around

$^{13}$2

These defaults are presented as balancing accuracy and runtime.

The paper also states improved numerical reliability: adaptive solvers produce controlled accuracy, INST solutions converge consistently to IST behavior as $^{13}$3, and performance remains robust across parameter variations. These claims place numerical method selection at the center of the platform’s design, particularly for nonstationary labeling dynamics where stiffness and repeated solves dominate runtime.

5. Multi-experiment integration, multi-tracer workflows, and Bayesian inference

Support for multiple labeling datasets and multi-tracer studies is identified as a notable contribution of 13CFLUX(v3) (Stratmann et al., 28 Sep 2025). If there are $^{13}$4 labeling datasets, the naive approach would solve each separately. The paper instead reformulates the problem as an augmented system,

$^{13}$5

where

$^{13}$6

This lets the matrix inversion be done once rather than $^{13}$7 times, giving approximately an $^{13}$8-fold reduction in work. The same idea is reported to accelerate sensitivity calculations and experimental design.

The abstract and main text also emphasize support for diverse labeling strategies, including custom tracer designs, multiple isotope tracers, and different labeling states and time courses, as well as multiple analytical platforms such as MS, MS/MS, and other measurement configurations represented in FluxML. The combination of augmented-system simulation with broad input flexibility is central to the software’s role in integrated experimental analysis.

A major new use case highlighted in the paper is Bayesian INST. Bayesian MFA combines prior knowledge and data likelihood to produce a posterior distribution over fluxes and pool sizes rather than only point estimates. The stated advantages are the ability to reveal non-identifiable parameters, parameter correlations, and uncertainty structure. The workflow is implemented by coupling 13CFLUX(v3) to hopsy, described as a Python sampling framework for convex polytope sampling, via the translation

$^{13}$9

The paper states that this Bayesian workflow was only practical because 13CFLUX(v3) is fast enough to support the many millions of simulations MCMC may require.

6. Performance profile, interoperability, and research significance

The paper reports that 13CFLUX(v3) outperforms FreeFlux, INCA, and influx_si in the presented benchmarks and is about 40× faster than 13CFLUX2 for some multi-dataset simulations (Stratmann et al., 28 Sep 2025). Runtime is reported to scale well with system dimension and solver tolerance. These performance claims are tied directly to practical workflows: repeated simulation for parameter estimation, sensitivity analysis, multi-experiment fitting, and Bayesian posterior sampling.

Interoperability is treated as a first-class property. The software is described as open-source, distributed via source and containers, available as a Python package or wheel, containerized with Docker, and interoperable with Python tools and external libraries. The design is also said to support workflow orchestration with Apache Airflow and reproducible execution from laptops to HPC clusters. In practical terms, the platform is presented as suitable both for exploratory notebook-based analysis and for automated, high-throughput pipeline execution.

Several common misconceptions are implicitly corrected by the paper’s design choices. First, 13CFLUX(v3) is not limited to isotopically stationary MFA; INST support is an explicit architectural and numerical extension. Second, it does not commit users to a single labeling representation, since both cumomers and EMUs are supported and selected according to numerical benefit. Third, the Python layer is not a replacement for the compiled engine but an interface to it; the reported usability gains are therefore presented without abandoning a C++17 performance core.

The practical meaning of the platform for modern fluxomics research lies in the combination of speed, flexibility, reproducibility, interoperability, and extensibility. The paper’s bottom-line characterization is that 13CFLUX(v3) extends the classical stationary flux-analysis workflow into a flexible, scalable framework for isotopically nonstationary, multi-experiment, multi-tracer, and Bayesian fluxomics. That framing places the software within a broader shift in $t=\infty$0C-MFA from single-experiment point estimation toward integrated, uncertainty-aware computational analysis.