Flux Studio

Constraint-based metabolic modelling, in your browser. No install, no server, no queue.

Pick a strain, pick a growth medium, and solve. Thirteen analyses, from flux balance analysis to synthetic lethality to strain design, over 4,659 genome-scale metabolic models (2,313 Escherichia coli and 2,346 Lactobacillaceae) and 12,340 semi-curated growth media. The linear program is solved in your browser tab by a WebAssembly build of GLPK: your models and your media never leave your machine.

Every result is checked against COBRApy 0.27.

---

Thirteen analyses	Solve, and see the flux map
Knockout: what changed, and is it real	12,340 media, in every analysis

---

What it does

Simulate — FBA · parsimonious FBA · linear MOMA · loopless flux (CycleFreeFlux) · dynamic FBA

Explore the solution space — flux sampling (ACHR) · flux variability · production envelopes · phenotype phase planes

Screen and design — reaction essentiality · gene essentiality (through the GPR rules, so isozymes survive and complexes do not) · synthetic lethality (double deletions) · FSEOF strain design

Compare — two conditions side by side · multi-model analytics · group comparison

Model quality — blocked reactions, mass and charge imbalance, dead ends, orphan genes

The knockout study

Knock out any set of reactions and see the wild type and the mutant solved on the same medium: growth before and after, two Escher maps, and five figures that each answer a different question. How much of the reaction did the cell actually need (a titration, because a knockout is one point on a curve). Where in metabolism the damage landed. The whole flux redistribution on one signed-log scatter, so reversals and on/off are both visible. Which reactions moved, by enzyme name. What the cell now eats and excretes.

And one thing most tools skip: a flux-difference plot is a lie if the fluxes came from plain FBA. Many flux vectors give exactly the same growth rate and the solver returns an arbitrary one, so diffing two arbitrary choices draws re-routing that is solver noise. Every change named here is checked by running FVA on both states. If a reaction's feasible range at the wild-type optimum is disjoint from its range at the knockout optimum, the knockout forces the change and it is real. If the ranges overlap, the plot says so and draws the point hollow. On a PGI knockout in E. coli DH1, only 21 of the 40 largest changes survive that test. Nineteen were the solver choosing.

What it cannot do, and will not pretend to

glpk.js exposes a linear program only: no integer variables, no quadratic objective. So ROOM, exact add_loopless, gap-filling and OptKnock (which need MILP) and quadratic MOMA (which needs QP) are absent rather than faked. Linear MOMA is the form COBRApy uses for large models anyway, and CycleFreeFlux gives the same loop-free flux distribution as add_loopless at the same growth rate.

Validated against COBRApy 0.27

Same model, same medium, run in the browser and diffed against a Python reference:

	COBRApy 0.27	Flux Studio
Growth (FBA)	1.120796	1.120796
Essential genes	5	5 (identical set, max diff 4.6e-11)
Linear MOMA distance	62.1449	62.1451
Loopless growth	1.120796	1.120796
Blocked reactions	14	14
Knockdown titration	21 points	max diff 4.8e-07
Alternate-optima verdict	21 forced / 19 ambiguous	21 / 19

Link straight into an analysis

Any page can hand a specific strain to a specific analysis, which is how the model browsers do it:

https://omidard.github.io/FluxStudio/?model=<gem_file>
                                     &tab=<analysis>     explore | fva | dfba | sampling |
                                                         envelope | phaseplane | genes |
                                                         synlethal | design | essential |
                                                         multi | cohort | qc
                                     &medium=<preset>    M9_glucose_aerobic | M9_glucose_anaerobic |
                                                         MRS | BHI | CDM | Feces | Urine | Serum
                                     &ko=<rxn,rxn>       preload a knockout study

?model=536056.3.json.json&ko=ATPS4rpp            knock out ATP synthase in E. coli DH1
?model=GCF_020539925.1.json&tab=genes&medium=CDM gene essentiality in Pediococcus on CDM
?models=a.json,b.json,c.json                     compare three strains

gem_file is the key used by gems_metadata.json.

Where the data comes from

Flux Studio ships no data. It reads, at run time:

• panGEMs — the 4,659 strain models, from EcopanGEM and LactoPanGEM
• Media — 12,340 curated growth media, keyed to BiGG exchange reactions, from DSMZ MediaDive, HMDB, USDA, FooDB and the primary literature

The two pangenomes were built a decade apart and use different BiGG naming generations: EcopanGEM writes EX_glc__D_e, LactoPanGEM writes EX_glc_D_e. A medium that names an exchange the model does not have is a closed exchange, so binding an E. coli medium to a Lactobacillus model would silently delete glucose and every amino acid, and the strain would read as dead. Flux Studio resolves across both spellings and always shows you the coverage, because an unbound compound is not a cosmetic miss: it is a removed nutrient.

Notes on the biology

Lactobacillaceae are fastidious. None of the LactoPanGEM strains grow on M9 + glucose, and only about half on MRS or BHI. They need amino acids, nucleotides and vitamins. The default medium for that collection is therefore CDM, the chemically defined medium from the LactoPanGEM paper (Table 2), on which 16 of 16 strains tested grow.

Architecture

docs/
  index.html            the app
  fba/
    fba_engine.js       LP construction and every analysis; no DOM
    fba_ui.js           Explore, Compare, Knockout study
    ko_plots.js         the five knockout figures + the alternate-optima check
    studio.js           the other ten analyses
    media.js            Media database client
    media_ui.js         the shared media picker, mounted on every analysis
    media_presets.json  eight quick presets
    subsystems.json     BiGG reaction to pathway
    core_map.json       Escher map
  vendor/               glpk.js (WASM), Escher, Plotly

Everything is a plain ES module. There is no build step.

Cite

If Flux Studio is useful in your work, please cite the models it runs on:

• Ardalani O., Phaneuf P.V., Mohite O.S., Nielsen L.K. Pangenome reconstruction of Lactobacillaceae metabolism predicts species-specific metabolic traits. mSystems 9(7):e00156-24 (2024). doi:10.1128/msystems.00156-24
• Ardalani O. et al. Annotating the pangenome reveals the diversity in the genetic basis for metabolic enzymes. Science Advances 12(27) (2026). doi:10.1126/sciadv.aeb3363

Licence

MIT. Escher, Plotly and glpk.js are vendored under their own licences.