Flux Transformer drops expert cost-function tuning: it infers intracellular fluxes directly from isotope labeling and measured rates
Synopsis
Intracellular metabolic fluxes cannot be measured directly; the standard route feeds cells isotopically labeled nutrients, measures labeling patterns, then inverts them with nonlinear optimization, which is expensive and often needs expert adjustment of cost-function weights. Flux Transformer encodes labeling measurements and a few measured reaction rates as tokens carrying biological identity, integrates them with self-attention, and uses cross-attention to predict the few free fluxes that satisfy mass balance, yielding the full flux distribution. On published expert-curated data from four organisms it reproduced MFA fluxes (median R² = 0.90) and reported confidence intervals.
Interpretation
Metabolic flux analysis moves from a nonlinear fit that an expert tunes case by case to a single forward model inference: isotope labeling data and a few measured reaction rates go in, and a mass-balanced full flux distribution comes out. Traditional MFA guesses a flux distribution, computes the labeling it would produce, and iterates with nonlinear optimization toward the experimental data; the inputs mix labeling patterns with measured rates, the fit is sensitive to how each data type is weighted, expert adjustment of those weights is commonly needed for plausible fluxes, and the optimization is expensive and does not guarantee a global optimum. On simulated data for five organisms (E. coli, S. cerevisiae, I. orientalis, pennycress embryos, mouse T cells), Flux Transformer resolved more reactions (R² > 0.9) than traditional MFA, returned more accurate fluxes overall, and ran orders of magnitude faster than 13C-FLUX and INCA; machine-learning approaches using labeling patterns alone without measured fluxes, such as ML-Flux, recovered fluxes less reliably.
The model turns two heterogeneous measurement types into one representation, then lets each free flux to be predicted select which measurements inform it. Earlier machine-learning efforts either identified flux-related features from tracing data without determining fluxes, or trained surrogate models of the forward problem and fit their predictions, and a recent two-stage framework inferred relative fluxes from tracer data only, leaving unaddressed the common real-world case where labeling data and selected reaction rates are both measured and must be integrated. Each isotope labeling measurement becomes a token encoding metabolite identity, mass shift, tracer identity and labeling fraction; each measured rate becomes a token encoding reaction identity and value. Self-attention layers interpret each measurement in the context of the whole experiment, cross-attention with learned queries maps the contextualized embeddings to free fluxes, and a null-space basis reconstructs the full distribution; models are trained on noisy simulated data, with 10,000 steady-state flux vectors per organism and an ensemble of 20 models.
On real experimental data the model reproduces expert-curated fluxes, and where it cannot pin a flux down it reports a wider confidence interval. Training used simulated data only, while testing used published expert-curated MFA results and their underlying labeling and exchange-rate measurements, covering uniformly and positionally labeled glucose and uniformly labeled glutamine tracers. Across four organisms, Flux Transformer recovered the measured labeling patterns (median R² = 0.82) and the expert-curated MFA fluxes (median R² = 0.90); in batch cultures of E. coli, S. cerevisiae and I. orientalis, fluxes agreed with expert values over five orders of magnitude (R² > 0.9). The reactions with the largest disagreement were those the model reported with the widest confidence intervals, and for E. coli it reproduced Park et al. results despite training on a different network model.
Perspective
The demonstrated scope is flux inference at metabolic and isotopic steady state in cultured cells, spanning bacteria, fungi, plant embryos and mammalian cell lines, and tested on wild-type and 20 single-gene deletion E. coli strains; studies inferring fluxes from non-stationary labeling data are not yet covered, and the authors propose extending to them by generating suitable training data.
A model is specific to the organism, the analytical platform (GC-MS or LC-MS) and the tracers it was trained on, so a new tracer or organism requires retraining; quantitative predictions require measured uptake, secretion and growth rates because labeling alone constrains only relative fluxes. Some fluxes are not determined by the labeling data, particularly in larger networks with unidentifiable loops, a property of the experiment and network rather than the method, and the model reports them with wide confidence intervals. The models presented are restricted to cultured cells or plant embryos and do not yet cover organismal metabolism.
