MetfluxLive · Docs
← App

MetfluxLive documentation

What the app computes, where its data come from, how to read the results, and a set of in-silico experiments to try.

1. What MetfluxLive is

MetfluxLive is a browser tool for exploring a whole-cell metabolic model. You choose a culture medium, the app predicts how fast every reaction of the cell runs (the fluxes), and draws them as an animated map. On top of that you can label atoms of a nutrient and watch where they end up, switch enzymes off, and follow the medium as it is consumed over time.

Everything runs inside the browser: there is no server doing calculations and nothing you enter leaves your computer. Saved media and preferences are kept in the browser's local storage.

The numbers are model predictions, not measurements. They show what the network can do under the assumptions explained below. Use them to form hypotheses and design experiments, not as a substitute for them.

2. Quick start

  1. Open the app. It loads Human-GEM, the standard medium, and runs a first simulation.
  2. The centre shows the cell: each bubble is a pathway inside a compartment; the bright moving lines are carbon flowing between them; the Lines selector below the map switches them to nitrogen, all redox carriers together, NADH, NADPH, ubiquinol, ATP or free energy (how they are computed). Double-click a bubble to open its pathway map.
  3. The left panel is the experiment: medium, knockouts and the objective. Change a concentration and the fluxes are recomputed by themselves while Auto, next to Simulate, is on (see below).
  4. The right panel explains whatever is selected: the summary, a metabolite, a reaction, a pathway or a gene.
  5. Use the search box (Ctrl+K) to jump to any metabolite, reaction, enzyme or pathway.

3. The interface

Simulate and Auto

The centre of the top bar holds the mode (Steady state / Dynamic, section 8), the ? with a summary of the methods, and Simulate with its Auto switch.

  • Simulate solves the model now with everything in the left panel (medium, tracers, measured fluxes, knockouts, objective, and in Dynamic the duration and events) and redraws the map with the new result. Nothing else computes fluxes: the map and the panels only display the last result.
  • Auto presses Simulate for you shortly after every change. It is remembered separately for each mode: on by default in Steady state, where a run takes a second or two, and off by default in Dynamic, where a run takes half a minute or more and you usually want to set several things (events, tracers) before starting. Turn it on in Dynamic too if you prefer; a new change stops the run in progress and starts again.
  • With Auto off, a change leaves the old result on screen, an amber dot on Simulate and the message "Parameters changed" in the status bar until you press it. Switching mode works the same way: it re-runs only if Auto is on for the new mode.
  • While a calculation runs, Simulate reads "Simulating…" and the status line at the bottom right turns amber; it turns green when the result is ready.

Light and dark mode

The moon/sun button at the right of the top bar switches between the dark theme (default) and a light theme with soft pastel surfaces. The data colours (label, change, free energy) are the same in both; only the background, panels and neutral marks change. The choice is remembered by the browser and also applies to this page.

Three ways to arrange the map

GroupWhat you seeGood for
CellReactions grouped by compartment and pathway, laid out as a cell: cytosol, mitochondrion, ER, Golgi, lysosome, peroxisome, nucleus and the medium outside. Reactions that cross a membrane sit on it.Seeing where things happen and what is transported.
PathwaysOne bubble per pathway, placed so that pathways sharing metabolites are close.An overview of which pathways are active.
Cascade (default)Pathways in rows, ordered by how many reaction steps separate them from the medium: nutrients at the top, distant products at the bottom. Inside a pathway the same idea gives a textbook-like top-to-bottom chain.Reading metabolism as a flow from nutrients to products.

Reading the map

  • Line width and particle speed show flux on a logarithmic scale, so small fluxes stay visible.
  • In a pathway map, circles are metabolites and diamonds are reactions. The ring colour of a metabolite is its compartment. Cofactors such as ATP or water are hidden to keep the map readable.
  • With a tracer, colour changes from grey (unlabelled) through amber to magenta (fully labelled).
  • With a knockout, choosing "Changed flux only" colours by increase (orange) or decrease (blue) against the unperturbed cell; a red ring marks the perturbed reactions.

What the lines between groups carry

In the Cell, Pathways and Cascade views a line joins two groups (bubbles) when one of them makes something that the other uses. The Lines selector below the map, on the left, chooses what that something is:

LinesWhat a line isUnit
Carbon (default)Carbon atoms that one group passes to another, not counting the atoms of carriers (CoA, nucleotides, thiamine, lipoamide...).mmol C/gDW/h
NitrogenThe same for nitrogen atoms. Ammonium counts as a carrier of nitrogen here.mmol N/gDW/h
All redox carriersThe sum of the lines of every redox carrier: NADH, NADPH, ubiquinol, FADH2, reduced vitamin K, and (counted half, as they carry one electron) cytochromes c and b5 and the ferredoxins. The overall map of reducing power.mmol e− pairs/gDW/h
NADH, NADPH, UbiquinolReduced cofactor made by one group and oxidised by another: where the reducing power comes from and where it goes.mmol/gDW/h
ATPATP, GTP, UTP and CTP made by one group and spent by another.mmol/gDW/h
Free energy (ΔG′m)No lines: each group is sized by the free energy its reactions dissipate and coloured by whether they run downhill or uphill (see below).J/gDW/h

The ? next to the selector opens this section. The choice is remembered by the browser.

How a line is computed

The lines are built from the flux solution, one metabolite pool at a time:

  1. For a pool (say pyruvate in the cytosol), every active reaction that touches it either produces it (flux × coefficient > 0) or consumes it. Productions are summed per group (PA), and so are consumptions (CB).
  2. The pool is assumed well mixed: a molecule made by group A is equally likely to be used by any consumer. Group A therefore sends B the share
    F(A→B) = w × PA × CB / max(P, C)
    where P and C are the pool's total production and consumption (equal at steady state) and w is what one molecule carries: its transferable carbon or nitrogen atoms, or 1 for a cofactor.
  3. The shares of all pools are added up for each pair of groups. Flow inside a group is not drawn.

So the atom count is a weight, not a normalisation: one mol/h of glucose (6 C) weighs twice as much as one mol/h of pyruvate (3 C). The direction comes from the flux solution itself, i.e. which group produces the pool and which consumes it. The well-mixed assumption is the same one the tracer uses. A line does not know which atom goes where; it only shares out each pool among its consumers.

Transferable atoms: leaving the carriers out

Many intermediates are bound to a carrier that is recycled and never passes from one pathway to another: the CoA of acetyl-CoA and malonyl-CoA, the UDP of UDP-glucose, the thiamine diphosphate of the pyruvate dehydrogenase adducts, the lipoamide of the 2-oxoacid dehydrogenases, the tRNA of charged amino acids, carnitine, tetrahydrofolate, glutathione. Counting their atoms would make, for example, malonyl-CoA weigh 24 C when only 3 belong to malonate.

To leave them out, the app runs the tracer on the current flux solution with a special label: every atom of a carrier is "labelled" (1), every nutrient atom is not (0). The label each atom ends up with is the share of that position that comes from a carrier. A metabolite's transferable atoms are then its atoms of the element times (1 − that share). In the default run this gives 1.9 transferable carbons for mitochondrial acetyl-CoA, 2.9 for malonyl-CoA, 6 for UDP-glucose and 2 for the hydroxyethyl-thiamine-diphosphate adduct. Because the tracer follows the atom maps, a molecule that is partly made from a carrier is counted in part: IMP has 9.6 transferable carbons of 10, because some of it comes from recycled adenine nucleotides.

The carriers are the organic cofactors hidden from the maps (nucleotides, NAD(P), FAD, CoA, ubiquinone, cytochrome c, thioredoxin) plus thiamine diphosphate, lipoamide and lipoyl-protein, ACP, tRNAs, carnitine, tetrahydrofolate and glutathione. Closed cycles that take in no atoms from outside (two reactions in different pathways turning the same molecule back and forth) count as carriers too, because their atoms only go round. Inorganic carriers of the element stay in: CO2 and bicarbonate for carbon, ammonium for nitrogen, so decarboxylations and deaminations do appear as lines.

In the default Human-GEM run (Pathways view) the thickest carbon lines are glucose from the medium into glycolysis (4.4 mmol C/gDW/h) and lactate from glycolysis to the medium (4.0), followed by amino acids going into protein through the charged tRNAs, and glutamine feeding the TCA cycle as 2-oxoglutarate. The thickest nitrogen lines are glutamine from the medium (0.34 mmol N/gDW/h), amino acids into protein, and ammonium released by glutaminase to the medium.

Redox cofactors and ATP

For NADH, NADPH, ubiquinol and ATP the pools followed are the reduced cofactor (or the triphosphates) and w = 1. A line from A to B says that A reduces (or phosphorylates) the cofactor and B oxidises (or hydrolyses) it. In the default run:

  • NADH flows from glycolysis to oxidative phosphorylation (0.37 mmol/gDW/h);
  • NADPH flows from the malic enzyme ("Pyruvate metabolism") and the pentose phosphate pathway to fatty acid and cholesterol synthesis;
  • ubiquinol flows from succinate dehydrogenase (TCA cycle) to complex III;
  • ATP flows from glycolysis (0.91) and oxidative phosphorylation (0.56) to maintenance and biomass ("Artificial reactions").

The four triphosphates are taken together because nucleoside-diphosphate kinase turns ATP into GTP, UTP and CTP. A line from the ATP-making groups to "Nucleotide metabolism", and from there to protein synthesis, shows that relay.

All redox carriers. This option adds up the reducing equivalents of every carrier, in electron pairs (one NADH, NADPH, ubiquinol or FADH2 = one pair = one hydride; a reduced cytochrome or ferredoxin = half a pair). Each carrier is still paired producer → consumer on its own, so NADH made in one pathway is never matched with NADPH used in another: the two pools are separate in the cell and only exchange through specific enzymes (transhydrogenase, malic enzyme), which appear as reactions of their own. The tooltip of a line names the carrier that contributes most. In the default run the total is 1.27 mmol pairs/gDW/h: NADH 0.94, ubiquinol 0.17, NADPH 0.15. The thickest lines are glycolysis → oxidative phosphorylation (0.37, NADH), serine synthesis → glycolysis (0.21, the NADH of phosphoglycerate dehydrogenase used by lactate dehydrogenase) and TCA cycle → oxidative phosphorylation (0.19, mostly ubiquinol). Glutathione is left out: GSH is also made and consumed as a molecule (synthesis, conjugation, biomass), which would be counted as redox traffic.

Free energy (ΔG′m)

With Free energy selected there are no lines, because free energy is not handed from one pathway to another like an atom. Instead, every reaction j has an estimated transformed Gibbs energy ΔG′m, and the product vj × ΔG′m,j (mmol/gDW/h × kJ/mol = J/gDW/h) is the free energy the reaction dissipates as it runs (when negative) or would need to take in (when positive).

  • Global views: a bubble's size is the free energy its reactions dissipate; its colour weighs that against the energy its uphill steps would need, from red (all downhill) through slate to violet (mostly uphill). Grey: active, but no reaction has an estimate.
  • Pathway map: each reaction is coloured by ΔG′m in the direction it runs, from red (−40 kJ/mol or less, far from equilibrium) through slate (near equilibrium) to violet (uphill). The reaction card gives the value, its uncertainty and the energy dissipated. The tooltip of a group gives the energy dissipated, the uphill part and the share of its flux with an estimate.

Where the values come from. ΔG′m is estimated once, offline, with eQuilibrator 3 and the component contribution method (Noor et al. 2013; Beber et al. 2022), by tools/build_thermo.py. Metabolites are matched by their MetaNetX, ChEBI, HMDB, BiGG or KEGG identifiers, else by the InChIKey of the model's structure. Conditions: 37 °C, ionic strength 0.25 M, pMg 3, and the pH of each compartment (cytosol, nucleus and ER 7.2, matrix 8.0, intermembrane space 6.9, peroxisome 7.0, Golgi 6.4, lysosome 4.7, medium 7.4; Casey et al. 2010, Porcelli et al. 2005). Every reactant is at 1 mM (water at 1). Reactions that cross a membrane use the multicompartment method of Haraldsdóttir et al. (2012), with a matrix potential of −160 mV against the cytosol (Zorova et al. 2018) and a plasma membrane potential of −60 mV (Yang & Brackenbury 2013 report −10 to −90 mV). Two redox couples take measured midpoint potentials instead of the estimate: cytochrome c, +0.25 V (Loveridge et al. 2026), placed in the intermembrane space where it really is; and membrane ubiquinone, +0.065 V at pH 7 (Urban & Klingenberg 1969), because the estimate for the long isoprenoid quinone is far off (it made succinate dehydrogenase +74 kJ/mol).

Coverage. In Human-GEM 2,295 of 4,102 metabolites are found and 6,476 of 12,748 reactions have an estimate. The rest involve proteins, R groups, lipids described as classes, or compounds outside eQuilibrator. In the default run, reactions with an estimate carry 94 % of the active flux.

What the default run shows. The cell dissipates about 330 J/gDW/h through the reactions with an estimate. Most of it is in glycolysis (126) and the transporters (112), then oxidative phosphorylation (42). Hexokinase (−20.8 kJ/mol), pyruvate kinase (−24.5), ATP synthase (−13.6) and cytochrome oxidase (−35.4) run downhill. Some steps that cells run in that direction come out uphill at 1 mM: glyceraldehyde-3-phosphate dehydrogenase (+18.8), complex I (+16.2) and complex III (+9.7). In cells, low 1,3-bisphosphoglycerate and NADH/NAD+ or QH2/Q ratios different from 1 make them run near equilibrium. A few uphill steps point to the model instead: the mitochondrial diphosphate transporter (+61.8 kJ/mol) moves PPi4− into the negatively charged matrix as a uniport, which the membrane potential forbids; in cells it would be an exchanger.

Limitations of ΔG′m. (1) The 1 mM standard is a convention, not a measurement: a reaction with |ΔG′m| below about 10 kJ/mol can run either way at real concentrations, and "uphill" means "needs concentrations far from 1 mM", not "impossible". (2) The estimates have uncertainties, shown on the reaction card; above 20 kJ/mol the sign is not reliable. (3) Fluxes are not constrained by thermodynamics: FBA may choose a direction that ΔG′m disfavours, and this view only shows it. (4) Compartment pH and membrane potentials are typical values, the same for every cell line; the potentials of the ER, Golgi, lysosome and peroxisome are set to 0. (5) Reactions added in the model editor have no estimate. (6) Enthalpy (ΔH) is not shown: estimates of it are far scarcer than of ΔG and, for the direction of a reaction, ΔG is what matters.

Display

  • Cell view: each compartment's pool is separate, and transporters are groups of their own sitting on the membranes, so a line runs from cytosolic glycolysis to the mitochondrial pyruvate carrier and on to the TCA cycle. Pathways and Cascade: the pools of all compartments are merged and transport reactions are ignored, so the same carbon goes straight from glycolysis to the TCA cycle.
  • Only the 220 largest lines (320 in the Cell view) are kept. Width and particle speed are logarithmic, relative to the largest line: lines below about 1/350 of it are hidden, and particles move only on lines above about 1/90 of it. Selecting a bubble keeps its own lines bright and dims the rest.
  • Colour: plain flux in cyan. With a tracer of the same element (a 13C tracer and the Carbon lines, a 15N tracer and the Nitrogen lines), the colour is the labelled fraction of what the line carries, weighted by production. In "Changed flux only", the colour is the mean change of the two groups it joins.
  • In a pathway map the lines are always reaction fluxes; there the selector only matters for free energy, which colours each reaction.
  • Download › pathway_flows.csv lists the lines of every quantity between pathways, and fluxes.csv gains ΔG′m columns.
Limitations. (1) The carrier share relies on the atom maps; where a reaction has none, the leftover atoms are mixed (as in the tracer), so the share of a molecule made by such a reaction is approximate. (2) Groups are the pathways (subsystems) assigned by the model: a reaction filed under an unexpected pathway moves its lines there. Open a bubble to see which reactions carry its lines. (3) pFBA picks one route among equally short ones. The app breaks one such tie on purpose: copies of a reaction that use a deoxynucleotide instead of the ribonucleotide (hexokinase with dATP, pyruvate kinase with dADP) cost 0.1 % more, since cells hold deoxynucleotides at about 1 % of the ribonucleotide levels or less (Traut 1994). Without it, a dATP/dADP cycle could stand in for ATP/ADP in glycolysis and send its lines through "Purine metabolism". Other ties remain: pyruvate kinase also makes GTP from GDP in the model (same enzyme), and pFBA uses it because it saves the nucleoside-diphosphate kinase step.

Other controls

  • Show: everything, only what carries label, only the flux that changed (after a knockout, or against a compared scenario), or only the labelling that changed against a compared scenario.
  • Panels: drag the border between a side panel and the map to resize it; double-click the border to restore it. The sections of the left panel fold when you click their title.
  • Language: English, Spanish, Portuguese, Italian, French, German or Greek (top right). Names of metabolites, reactions and pathways come from the model and stay in English. When the window is narrow or the labels are long, the top bar shrinks by itself: Edit, Import and Download become icons (their tooltip keeps the full name), then the brand name is hidden.
  • Download: exports tables, charts, the map and the session (see section 16). Import loads a session or model-edit file back.
  • Compare (left panel): saved scenarios to compare with (section 12). ✎ Edit model (top bar): add or remove reactions, enzymes and metabolites (section 13).

4. Data sources

WhatSourceUsed for
Human-GEM 2.1.1SysBioChalmers/Human-GEM; Robinson et al., Sci. Signal. 2020The human network: 12,748 reactions, 8,374 metabolites (one per compartment), 2,870 genes, 146 pathways, plus gene–reaction rules and metabolite structures.
Mouse-GEM 1.8.0SysBioChalmers/Mouse-GEM; Wang et al., PNAS 2021The mouse network. It shares identifiers with Human-GEM, so structures are reused.
Atom-to-atom mapsComputed for this app with RXNMapper (Schwaller et al., Sci. Adv. 2021) and RDKit, plus hand-curated corrections for central metabolismFollowing individual atoms through each reaction.
Plasmax and HPLMTable S1 of Vande Voorde et al., Sci. Adv. 2019; HPLM originally from Cantor et al., Cell 2017Physiological media.
Williams ESupplementary Table 2 of Grankvist et al., Nat. Metab. 2024The medium of the human liver slices (section 9).
Standard mediumLoosely follows DMEM / Ham's F-12Default medium.
Measured exchange fluxesJain et al., Science 2012 (NCI-60 cell lines), per-line values from S1 Table Q of Aurich et al., PLoS Comput. Biol. 2017; other sets: Dietmair et al., PLoS ONE 2012 (HEK293), Lemons et al., PLoS Biol. 2010 (fibroblasts), Grankvist et al., Nat. Metab. 2024 (human liver)The default uptake and secretion of glucose, lactate, glutamine and 15 more metabolites (section 5).
ATP maintenanceKilburn, Lilly & Webb, J. Cell Sci. 1969, value as used by Zielinski et al., Sci. Rep. 2017The ATP the cell spends whether it grows or not.
Uptake kinetics (Vmax, Km)Order-of-magnitude values for cultured mammalian cells, set by handMaximum uptake of the medium components whose exchange was not measured (vitamins, cystine, histidine…), and of all of them when measured fluxes are switched off. Not measured values.
Linear solverHiGHS, compiled to WebAssemblySolving the flux balance problem.
Gene expression (cell types)Human Protein Atlas 26.0 (cell lines, single cell types, tissues), Cellosaurus 57.0; Bgee 15.2 for mouseThe genes switched off in each cell line, cell type or tissue (section 6).

The models are converted once, offline, into the JSON files in data/; the app itself needs only those files.

How good are the atom maps?

Of the 11,094 reactions of Human-GEM that are not simple exchanges, 10,554 have an atom map:

MethodReactionsReliability
Identity (the same molecule on both sides: transport)4,172Exact.
RXNMapper5,196Usually right; a confidence score is shown in the reaction panel.
Largest common substructure1,138Lower: used for generic "R-group" reactions and very large molecules.
Partial48Only some atoms are mapped.

Transketolase, transaldolase, 6-phosphogluconate dehydrogenase and aconitase were checked and corrected by hand. Two systematic errors of the automatic mapper are corrected for every reaction they affect:

  • Transaminases (34 reactions): RXNMapper tends to give the carbon skeleton of one keto acid to the other amino acid, for example pyruvate carbons to the 2-oxoglutaramate made from glutamine. The skeleton of each amino acid ↔ keto acid pair is now kept by common substructure.
  • The CoA moiety (about 200 acyl transfers): in every acyl transfer CoA is carried over intact, so the CoA atoms of a product are taken from the CoA of a substrate. Where that is impossible (pseudo-metabolites such as "acyl-CoA pool"), unmapped CoA atoms are counted as unlabelled instead of receiving label from the acyl chain.

Before these corrections, labelled glucose produced spurious M+1 succinate and fumarate, and CoA itself appeared 40 % labelled. Where no map exists, label is spread evenly over the products and the metabolite is drawn with a dashed outline.

5. Measured values and their sources

Left to itself, FBA chooses the cheapest way to grow. For a cancer cell in culture that means taking up a fraction of the glucose real cells consume and secreting no lactate at all, so the label of a glucose tracer barely reaches the TCA cycle. By default the app therefore pins the main exchanges to values measured in human tumour cell lines. Choose "None" in "Measured exchange fluxes" to switch them off.

NCI-60 exchange fluxes

Jain et al. measured the consumption and release of more than 100 metabolites by the 60 cell lines of the NCI-60 panel (Science 2012). The per-line values used here are those published in S1 Table Q of Aurich et al. (PLoS Comput. Biol. 2017), who applied them as bounds of ±20 % around each measurement. The app takes the median over the 59 lines with data and keeps the 18 exchanges whose direction is the same across the panel. Aspartate, proline, glycine, ornithine and glycerol are taken up by some lines and released by others, so they are left out.

MetaboliteDirectionMedian (fmol/cell/h)Interquartile rangeUsed (mmol/gDW/h)
glucoseuptake278171–3990.928
L-lactatesecretion353241–5681.180
glutamineuptake7138–990.237
serineuptake8.906.48–140.0297
alaninesecretion134.39–230.0428
glutamatesecretion8.123.38–150.0271
leucineuptake5.663.91–8.590.0189
lysineuptake5.414.27–8.500.0180
isoleucineuptake4.743.05–6.570.0158
valineuptake4.043.09–6.210.0135
threonineuptake3.782.69–6.040.0126
arginineuptake3.682.09–5.830.0123
tyrosineuptake2.572.04–4.200.0086
phenylalanineuptake1.901.54–3.180.0063
methionineuptake1.701.30–2.530.0057
asparagineuptake1.670.88–2.130.0056
tryptophanuptake0.450.36–0.890.0015
cholineuptake0.660.40–1.020.0022
  • Unit conversion. Rates per cell are converted with a dry mass of 0.3 ng per cell, the approximate NCI-60 average used by Zielinski et al. (Sci. Rep. 2017): mmol/gDW/h = fmol/cell/h ÷ 300. It is an average: for any particular line the absolute fluxes may be off by the difference between its mass and this one.
  • Margin. Each flux may move within ±20 % of the value, as in Aurich et al. You can change the margin, edit, add or remove rows.
  • Culture conditions. The NCI-60 lines are grown in RPMI-1640 with 5 % fetal bovine serum and 2 mM glutamine (as described by Dolfi et al., Cancer Metab. 2013, for the same lines). When you use another medium, the cell keeps the phenotype it had in RPMI.
  • Unmeasured products. With a measured set on, organic metabolites that were not measured can only be secreted in traces (0.001 mmol/gDW/h). In the NCI-60 data the measured exchanges of glucose, lactate and amino acids account for over 99 % of the observed exchange flux (Zielinski et al. 2017). Without this limit FBA disposes of surplus carbon as whatever is cheapest, such as sorbitol. CO2, ammonia, urea, water and inorganic ions are not limited.
  • When they cannot be met. If a measured flux becomes impossible (you remove glucose from the medium, a nutrient runs out in dynamic mode, a knockout blocks lactate production…), the measured fluxes are relaxed: uptakes keep only their maximum and secretions become free. The status bar and the summary say so.
  • Mouse-GEM uses the same values. They were measured in human cells.

Other measured sets: cells that are not tumours

The NCI-60 lines are cancer cells. "Measured exchange fluxes" also offers four sets measured in human cells and tissue that are not tumours. They follow the same rules as the NCI-60 set (±20 % margin, unmeasured organic products only as traces, relaxed when they cannot be met); only the source unit and its conversion to mmol/gDW/h differ. The ranges in the tables are the variation reported by each source, given for reference: the app applies its own margin.

HEK293 (Dietmair et al. 2012)

Dietmair et al. (PLoS ONE 2012, Table 2) measured the exchanges of the parental HEK293 line (human embryonic kidney cells immortalised with adenovirus DNA, not taken from a tumour) during exponential growth in FreeStyle 293 medium, per gram of dry weight (514 pg of dry mass per cell), oxygen uptake included. Values are the mean ± standard error of 3 cultures. Alanine (11 ± 17 µmol/gDW/h) has no clear direction and is left out. ATP maintenance is not added: with oxygen uptake measured, respiration is already set by the data. Check: with this set, the Standard medium and growth as objective, Human-GEM grows at 0.0313 h⁻¹; the cells grew at 0.028 h⁻¹.

MetaboliteDirectionMean ± SE (µmol/gDW/h)Used (mmol/gDW/h)
glucoseuptake528 ± 630.528
L-lactatesecretion695 ± 920.695
O2uptake335 ± 290.335
NH4+secretion17 ± 30.0170
glutamineuptake29 ± 170.0290
serineuptake37 ± 60.0370
leucineuptake19 ± 40.0190
arginineuptake16 ± 30.0160
valineuptake14 ± 30.0140
isoleucineuptake13 ± 30.0130
lysineuptake12 ± 20.0120
threonineuptake8.4 ± 1.60.00840
aspartateuptake6.3 ± 1.60.00630
phenylalanineuptake5.8 ± 1.00.00580
methionineuptake4.9 ± 1.10.00490
tyrosineuptake4.6 ± 0.70.00460
asparagineuptake4.1 ± 1.00.00410
histidineuptake3.2 ± 0.50.00320
tryptophanuptake1.5 ± 0.30.00150
glutamatesecretion9.6 ± 1.40.00960
ornithinesecretion9.0 ± 1.50.00900
glycinesecretion6.2 ± 1.10.00620
prolinesecretion3.8 ± 0.40.00380

Human fibroblasts, proliferating and quiescent (Lemons et al. 2010)

Lemons et al. (PLoS Biol. 2010, Fig. S1A) compared primary dermal fibroblasts from newborn foreskin while proliferating and after 14 days of contact inhibition (quiescent), in DMEM with 25 mM glucose, 4 mM glutamine and 7.5 % dialysed serum. Only glucose, lactate, glutamine and glutamate were measured, so FBA still chooses the other exchanges. The values were read from the bar chart (reading uncertainty about ±0.05); the ± is the published standard error. They are per µg of protein and are converted with 0.59 g of protein per gDW, the protein content of the Human-GEM biomass reaction (5.34 mmol of amino acid residues per gDW × 111 g/mol per residue): mmol/gDW/h = nmol/µg protein/h × 0.59. ATP maintenance (1.07 mmol/gDW/h, below) is applied. Proliferating, with growth as objective and the Standard medium: 0.0370 h⁻¹. Quiescent cells do not grow: choose an objective such as "Maximum ATP production" (13.9 mmol/gDW/h).

MetaboliteDirectionProliferating (nmol/µg protein/h)QuiescentUsed, P / Q (mmol/gDW/h)
glucoseuptake2.92 ± 0.151.41 ± 0.061.72 / 0.832
L-lactatesecretion1.85 ± 0.041.40 ± 0.051.09 / 0.826
glutamineuptake0.59 ± 0.030.21 ± 0.020.348 / 0.124
glutamatesecretion0.08 ± 0.010.05 ± 0.010.0472 / 0.0295

Human liver tissue ex vivo (Grankvist et al. 2024)

Grankvist et al. (Nat. Metab. 2024) kept precision-cut liver slices from 5 donors for 24 h in serum-free Williams E with fasted-state nutrients and measured what the tissue took up and released (Supplementary Data 3 and the source data of Fig. 1i, 4a and Extended Data Fig. 4i). The app takes the median over the donors (urea: over all 15 slices) and keeps the 17 exchanges with the same direction in at least 4 of the 5 donors; alanine, glycine, serine, valine and histidine change direction and are left out. The range in the table is that of the donors, so it can cross zero (a negative value means the opposite direction). Rates per g of wet liver per day are converted with a wet/dry weight ratio of 3.3, measured in rat liver (Wimmer, Wilmering & Sasse 1985): mmol/gDW/h = µmol/g/day × 3.3 / 24 / 1000. ATP maintenance is not added.

MetaboliteDirectionMedian (µmol/g liver/day)Range of the donorsUsed (mmol/gDW/h)
glucosesecretion30.4−5.62 … 75.30.00417
L-lactatesecretion11661.9 … 1600.0159
ureasecretion87.540.9 … 2020.0120
glutamineuptake42.513.9 … 47.00.00584
glutamatesecretion22.618.0 … 31.60.00311
arginineuptake12.58.08 … 22.90.00172
threonineuptake5.810.863 … 16.00.000799
aspartateuptake5.131.94 … 9.010.000705
leucineuptake4.371.93 … 24.30.000601
lysineuptake4.19−7.67 … 15.00.000577
asparagineuptake2.691.03 … 4.900.000370
tyrosineuptake2.71−0.780 … 7.030.000372
phenylalanineuptake2.572.03 … 5.440.000353
methionineuptake2.171.34 … 3.340.000299
prolineuptake1.70−2.40 … 4.560.000233
isoleucineuptake1.45−0.311 … 10.90.000199
tryptophanuptake0.7750.531 … 1.540.000107

The tissue does not grow and its rates are one to two orders of magnitude lower than those of cultured cells. Use the set with the Williams E medium (section 9) and a liver objective: with Human-GEM, glucose release then reaches 0.0724 mmol/gDW/h and urea release 0.0676; growth is 0. When the objective is the release of glucose or urea, its own measured rate is set aside (section 7).

Considered and left out: CHO cells (Ahn & Antoniewicz, Metab. Eng. 2011: hamster cells, article not openly available), exchange across pig organs (Jang et al., Cell Metab. 2019: arteriovenous concentration differences, not rates per mass of tissue) and endothelial cells (De Bock et al., Cell 2013: no absolute rates in the published data).

ATP maintenance

Cells spend ATP whether they grow or not: ion pumps, protein turnover, and so on. Human-GEM leaves this at zero. With the NCI-60 set on, the cytosolic ATP hydrolysis reaction (MAR03964) must carry at least 1.07 mmol/gDW/h, the value Zielinski et al. took from measurements by Kilburn, Lilly & Webb (J. Cell Sci. 1969) in mouse LS cells. It forces the respiration and TCA cycle turnover that real cells have. Without it, a glucose tracer labels fumarate at 0 % instead of about 9 % M+2.

What is still assumed

  • Uptake kinetics (Vmax, Km) of the medium components that were not measured, and of all of them when the measured set is off.
  • The composition of the "Standard" medium, which only approximates DMEM/F-12.
  • In dynamic mode, the size of the intracellular pools.

6. Cell types: which enzymes the cell has

The generic model can use every metabolic enzyme in the genome, which no real cell does: in the cell lines of the Human Protein Atlas a median of 30 % of the model's genes are not detected. "Cell type (gene expression)" switches off the genes that are not detected in a chosen cell line, cell type or tissue, and with them the reactions that only they catalyse. By default nothing is switched off (the generic model), so the numbers quoted in the other sections stay valid.

What you can choose

GroupModelSourceNumberExpression measure
NCI-60 panel consensusHuman-GEMHPA cell line atlas1 (median of 50 lines)nTPM
Cancer cell linesHuman-GEMHPA cell line atlas; description from Cellosaurus1132nTPM (bulk RNA-seq)
Non-cancer cell linesHuman-GEMHPA cell line atlas; description from Cellosaurus74 (HEK293, BJ fibroblasts, hTERT-immortalised lines, HaCaT, MCF-10A…)nTPM (bulk RNA-seq)
Cell typesHuman-GEMHPA single cell type atlas155nCPM (single-cell RNA-seq pooled per cell type)
TissuesHuman-GEMHPA tissue consensus (HPA and GTEx)51nTPM
Mouse tissues and cell typesMouse-GEMBgee 15.2264 tissues, 12 cell typespresent / absent calls

Search by name, cancer type, organ of origin or Cellosaurus accession; with an empty search box the list starts with the contexts that have measured data. Lines that Cellosaurus flags as problematic (misidentified, contaminated or misclassified) are kept, and its warning is shown. Five contexts have measured exchange fluxes from the same kind of cells (section 5): the NCI-60 consensus, HEK293, BJ fibroblasts (newborn foreskin fibroblasts, like those of Lemons et al.), liver tissue and hepatocytes. The panel offers to load them. The full list is in the catalogue of cell types.

How the list of genes switched off is made

  1. Not detected. A gene is off when it is below 1 nTPM (cell lines, tissues) or 1 nCPM (cell types), the cut-off below which the HPA itself calls a gene "not detected". In mouse, when Bgee calls it absent. Mouse-GEM genes are matched to Bgee through their Ensembl identifiers (Mouse-GEM's MGI_gene_ID_mapping.tsv). A gene without data stays on. For the NCI-60 consensus the value of each gene is its median over the 50 NCI-60 lines in the HPA atlas, as the NCI-60 measured fluxes are a median over the panel.
  2. Blocked reactions. A reaction is blocked when its GPR rule is false with those genes off: isoenzymes (or) compensate one another, a complex (and) needs every subunit. Reactions without a GPR rule (most transports, spontaneous reactions) are never blocked, and neither are the transports of molecules that cross membranes by diffusion (water, O2, CO2, NH3, urea, NO, H2S). Human-GEM assigns these to aquaporins or channels, and blocking them would, for example, trap in the mitochondria the urea that arginase 2 makes there in cells without AQP9: with the HEK293 measured fluxes, that alone cut the growth of HEK293 to less than half.
  3. Metabolic tasks. Every context must still be able to do the 57 essential metabolic tasks of Human-GEM 2.1.1 (data/metabolicTasks/metabolicTasks_Essential.txt): make ATP from glucose, synthesise nucleotides, cholesterol, phospholipids, heme and cofactors, take up the essential amino acids, oxidise fatty acids and grow on Ham's medium. They are run as in the RAVEN toolbox, with all the model's exchanges closed and only the inputs and outputs of each task open; both full models pass all 57 (Mouse-GEM uses the same metabolite names). When a context cannot do a task, a linear program looks for the cheapest way to make it possible through blocked reactions (each costs 2 minus the expression of its best-expressed gene relative to the cut-off, so genes just below it are preferred) and the genes those reactions need are switched back on. This is how tINIT protects metabolic functions (Agren et al. 2014), applied here to a list of genes switched off instead of to a reduced model.
  4. Measured phenotype. The five contexts with measured fluxes must also reproduce them, in their medium and within ±20 %: one more task.

The panel lists the genes switched back on and the task that needed each one. Over the 1413 human contexts the median is 14 genes (range 5–152); over the 276 mouse contexts, 10 (0–72). Some are false negatives of the data: POLR2A, the largest subunit of RNA polymerase II and needed in every cell, is below 1 nTPM in 1157 of the 1206 HPA cell lines and in 50 of the 51 tissues, a problem of that gene in the data rather than a real absence; it is switched back on in 1228 of the 1413 contexts. Others say more about the tasks than about the cell: "Protein synthesis from AAs" asks for albumin, so ALB is switched back on in every context that does not make albumin (81 % of the human contexts).

Block, or only make costlier

  • Block (default): the reactions are closed in every run, including the reference of a knockout and every step of a dynamic run.
  • Only make them costlier: the reactions stay open, but in the second stage of pFBA their flux counts 11 times (1 + 10) as much as that of other reactions. The value of the objective does not change; among the solutions that reach it, the one that uses fewer non-detected enzymes is chosen. It only acts with pFBA on.

The details of a reaction say when it is blocked and which of its genes are not detected; the details of a gene say whether it is detected in the chosen context or was switched back on. The downloads record the context, the genes switched off and rescued and the reactions blocked (settings.json, section 16). To explore a context as a whole, open "What … switches off" under the knockout list, or ▸ in a row of the catalogue: a table of the blocked reactions (pathway and genes not detected) and of the genes switched off (reactions they block and their pathways), with a filter, sortable columns and CSV export.

What it changes: some numbers

Human-GEM, standard medium, pFBA, steady state, growth as objective unless stated, "Block":

ContextMeasured fluxesGeneric modelWith the contextReactions blocked
NCI-60NCI-60, median of 59 lines (Jain 2012)0.0341 h⁻¹0.0341 h⁻¹1811
HEK293HEK293, exponential growth (Dietmair 2012)0.0313 h⁻¹0.0313 h⁻¹1906
BJ [Human fibroblast]Human fibroblasts, proliferating (Lemons 2010)0.0370 h⁻¹0.0370 h⁻¹1791
A-549None (medium ceilings only)0.0370 h⁻¹0.0370 h⁻¹1563
HEK293None (medium ceilings only)0.0370 h⁻¹0.0370 h⁻¹1906
hepatocytes (Glucose output (gluconeogenesis), Williams E)Human liver tissue ex vivo (Grankvist 2024)0.0724 mmol/gDW/h0.0700 mmol/gDW/h855
liver (Urea output (hepatocyte, ammonia disposal), Williams E)None (medium ceilings only)0.281 mmol/gDW/h0.281 mmol/gDW/h408

In none of these cases does the context lower growth: the metabolic tasks keep biomass synthesis possible, and the network has alternative routes for most blocked steps. What changes is which routes carry the flux; compare the maps with and without the context (section 12) or look at "What limits the objective" in the summary, which names a blocked reaction when one becomes limiting.

What it does not tell you

  • mRNA is not enzyme activity. A gene below 1 nTPM may still give enough enzyme for a small flux, and a detected gene says nothing about how much flux its enzyme can carry: the context only switches enzymes off, it never caps a flux by expression level.
  • The cut-off is a convention: below it the HPA's sequencing does not detect the gene, which is not proof that it is absent.
  • The HPA cell lines were grown in their own media, and tissues were taken from donors: expression may differ in the medium you simulate.
  • Bgee's absent calls depend on how much data exists for a tissue: tissues studied in few experiments have more genes called absent (from 0.5 % to 42 % of the Mouse-GEM genes).
  • The tasks protect generic functions only. A function of a particular cell (bile acids in hepatocytes, neurotransmitters in neurons) is not protected: it is possible if its genes are detected, otherwise not.

Data and licences

Human: Human Protein Atlas version 26.0 (cell lines: Jin et al. 2023; single cell types: Karlsson et al. 2021; tissues: Uhlén et al. 2015), released under CC BY-SA 4.0. The derived file data/human-gem/contexts.json (genes switched off and rescued in each context) is therefore distributed under CC BY-SA 4.0, with attribution to the Human Protein Atlas. Cell line descriptions: Cellosaurus 57.0 (Bairoch 2018), CC BY 4.0. Mouse: Bgee 15.2 (Bastian et al. 2021), CC0; data/mouse-gem/contexts.json is CC0 too. The files are built by the scripts in tools/contexts/ (cellline_meta.py, build_expression.py, build_contexts.mjs, tasks.mjs), which are not part of the published app.

7. Flux balance analysis, in plain words

A cell runs thousands of reactions at once. We rarely know the speed of each one, but we do know two solid things: the recipe of every reaction (what it consumes and what it makes), and the fact that a cell cannot create or destroy matter.

Flux balance analysis (FBA) builds on that. Think of the cell as a city's water network:

  • Pipes are reactions; the flow through a pipe is the flux, in millimoles per gram of dry cell weight per hour (mmol/gDW/h).
  • Junctions are metabolites. At every junction, what flows in must equal what flows out: nothing piles up and nothing appears from nowhere. This is the steady-state assumption.
  • Some pipes are one-way, and each has a maximum capacity. The pipes that bring nutrients in from the medium are limited by what the medium offers.

Those rules leave many possible flow patterns. To choose one, FBA assumes the cell is trying to achieve something — by default, to grow as fast as possible — and finds the pattern that does it best. "Growing" is itself a reaction, the biomass reaction, which consumes amino acids, nucleotides, lipids and energy in the proportions a new cell needs.

maximise vbiomass
subject to S · v = 0 (every metabolite is balanced)
lowerj ≤ vj ≤ upperj (direction and capacity of every reaction)

This is a linear programme, which a computer solves exactly in a few seconds even for 12,748 reactions.

What FBA can and cannot tell you

  • It needs no enzyme kinetics, which is why it works for a whole cell. For the same reason it knows nothing about enzyme amounts, regulation or metabolite concentrations inside the cell.
  • The best growth rate is unique, but the fluxes that achieve it often are not: there may be several equally good routes. The app therefore applies parsimonious FBA (pFBA) by default: among all the best solutions it takes the one with the smallest total flux, on the idea that cells avoid making enzyme they do not need. This also removes pointless internal loops.
  • The objective is a hypothesis. Growth is reasonable for proliferating cells in culture; it is a poor description of a resting neuron or hepatocyte. You can choose another built-in objective or any reaction (see below).
  • What enters and leaves the cell is the part of metabolism that is easiest to measure. Pinning those exchanges to measurements (section 5) removes much of FBA's freedom and makes the internal fluxes far more realistic.

Flux range (FVA) and what limits the objective

FBA returns one solution, but usually many give the same objective value. Flux variability analysis (Mahadevan & Schilling 2003) asks, for one reaction, how low and how high its flux can go while the objective stays at its optimum (within 10−6). The button Flux range (FVA) in the details of a reaction solves those two linear programs, with the same medium, measured fluxes, cell type and knockouts as the simulation (steady state only, about 2 s). The result is classed as fixed (every optimal solution needs exactly that flux), required (always used, amount variable), optional (optimal solutions exist with and without it; the flux on the map is the one pFBA chose) or always zero. With the default settings, glucose uptake can range from 0.742 to 1.11 mmol/gDW/h and pFBA takes the lower end; pyruvate kinase ranges from 0 to 1000. An end at ±1000 is the model's default bound reached through a cycle of reactions, not a real capacity: FVA does not remove such loops.

The summary lists what limits the objective: the bounds whose relaxation would raise the objective most, from the reduced costs of the FBA solution (objective gained per mmol/gDW/h of relaxation). Only a closure that the published model does not have (cell type, knockout, edit) counts as a limit, and exchanges of metabolites absent from the medium are left out. In HPLM it names pantothenate; in the standard medium without measured fluxes, linoleate and linolenate. When FBA has alternative optima, another set of limits can be equally valid.

Choosing the objective

FBA needs an objective: what the cell is assumed to make as much of as possible. Growth is the usual hypothesis for proliferating cells, but most cells in a body do not grow: they make ATP, export a product, or detoxify. The Cellular objective section therefore lists, besides growth, a set of cell functions of the kind used as "metabolic tasks" in liver and human-cell models (Gille et al. 2010; Richelle et al. 2021), the objectives you have saved, and a search box to use any reaction. Every built-in objective was checked in Human-GEM and Mouse-GEM to give a finite, non-trivial optimum.

Most functions are written as the release of a product (its exchange reaction, maximised). When the objective is the release of a metabolite, the app closes the uptake of that same metabolite: otherwise a medium that contains it (glucose, urea, creatine or GSH in Plasmax...) would simply be passed through, and the objective would measure the medium instead of the cell. Its measured rate, if any, is also left aside. The value is then the net synthesis.

ObjectiveReaction (maximised)Cell and functionValue*What limits it*
Growth, energy and stress
Growth (generic biomass) — defaultbiomass (MAR13082; Mouse-GEM MAR00021)Proliferating cells0.037 h⁻¹O2, amino acids
Maximum ATP productionMAR03964, cytosolic ATP hydrolysisEnergy stress, non-growing cells; a classic alternative objective (Schuetz et al. 2007)13.2O2 (lower it to model hypoxia)
Maximum glutathione synthesisMAR09351, GSH releaseOxidative stress (Lu 2013)0.020cysteine, cystine, methionine
Liver and lipid metabolism
Urea outputMAR09438Hepatocyte: ammonia disposal0.30nitrogen supply (glutamine)
Glucose outputMAR09034Liver, kidney: gluconeogenesis in fasting0.27amino acid carbon
Ketone body outputMAR09134, β-hydroxybutyrateLiver: ketogenesis in fasting0.99glucose and O2
Bile acid outputMAR09281, glycocholateHepatocyte0.098carbon, O2
Cholesterol outputMAR09285Liver: cholesterol synthesis for lipoproteins0.096carbon
De novo lipogenesisMAR00611, palmitate releaseHepatocyte, adipocyte0.21glucose
Neurons and endocrine cells
Glutamate releaseMAR09071Glutamatergic neuron0.55nitrogen supply
GABA releaseMAR09091GABAergic neuron0.55nitrogen supply
Dopamine releaseMAR09092Dopaminergic neuron0.019tyrosine, phenylalanine
Cortisol outputMAR09293Adrenal cortex0.095carbon, O2
Other cell types
Nitric oxide outputMAR09381Endothelium, activated macrophage0.55nitrogen supply (arginine)
Heme synthesisMAR09107Erythroid precursor0.11carbon, O2
Creatine outputMAR09290Kidney, liver0.20nitrogen supply

* Human-GEM, standard medium, measured exchange fluxes set to "None", mmol/gDW/h unless stated (Mouse-GEM gives the same values within 5 %). With growth as objective, every other function is close to 0. In the standard medium many functions are limited by the glutamine and amino acids it offers; in a richer or physiological medium they change.

What each objective represents

The same explanations appear under the selector when an objective is chosen.

Growth, energy and stress
Growth (generic biomass)
Maximises the model's biomass reaction: the cell grows as fast as the medium and the measured rates allow. The usual objective for proliferating cells.
Maximum ATP production (energy stress)
Maximises cytosolic ATP hydrolysis (MAR03964, the ATP-maintenance reaction) instead of growth: the most ATP the cell can make from what it takes up. A classic alternative objective, for non-growing cells and energy stress; with less O₂ in the medium it shows how ATP supply shifts in hypoxia. Growth is 0.
Maximum glutathione synthesis (oxidative stress)
Maximises the net output of glutathione (its export, MAR09351), which the cell can only sustain by making GSH from glutamate, cysteine and glycine; cells raise GSH synthesis under oxidative stress. Maximising glutathione synthetase itself is not used: the model then makes GSH and cleaves it again at once, a futile cycle. In the standard medium the value (~0.02 mmol/gDW/h) is set by the supply of cysteine, cystine and methionine, as in real cells. Growth is 0.
Liver and lipid metabolism
Urea output (hepatocyte, ammonia disposal)
Maximises urea release (MAR09438): the urea cycle disposing of the nitrogen of amino acids and ammonium, the main nitrogen function of the hepatocyte.
Glucose output (gluconeogenesis)
Maximises net glucose release (MAR09034) with glucose uptake closed: gluconeogenesis from amino acids, lactate or glycerol, as in liver and kidney during fasting. Remove glucose from the medium or not: it cannot be passed through.
Ketone body output (β-hydroxybutyrate)
Maximises the release of (R)-3-hydroxybutanoate (MAR09134), the main ketone body made by the liver from fatty acids and ketogenic amino acids during fasting.
Bile acid output (glycocholate)
Maximises glycocholate release (MAR09281): bile acid synthesis from cholesterol and conjugation with glycine, a hepatocyte function.
Cholesterol output (synthesis)
Maximises cholesterol release (MAR09285): the capacity for de novo cholesterol synthesis, which the liver exports in lipoproteins.
De novo lipogenesis (palmitate)
Maximises palmitate release (MAR00611): the capacity for de novo fatty acid synthesis from glucose and amino acids, as in hepatocytes and adipocytes.
Neurons and endocrine cells
Glutamate release (glutamatergic neuron)
Maximises glutamate release (MAR09071), the transmitter of glutamatergic neurons.
GABA release (GABAergic neuron)
Maximises the release of 4-aminobutyrate (GABA, MAR09091), the transmitter of GABAergic neurons, made from glutamate.
Dopamine release (dopaminergic neuron)
Maximises dopamine release (MAR09092), made from tyrosine; its supply of aromatic amino acids sets the value.
Cortisol output (adrenal cortex)
Maximises cortisol release (MAR09293): steroidogenesis from cholesterol in the adrenal cortex.
Other cell types
Nitric oxide output (endothelium, macrophage)
Maximises nitric oxide release (MAR09381) by nitric oxide synthase from arginine, as in endothelial cells and activated macrophages.
Heme synthesis (erythroid precursor)
Maximises heme release (MAR09107): the heme synthesis pathway (glycine + succinyl-CoA, iron), the main metabolic task of erythroid precursors.
Creatine output
Maximises creatine release (MAR09290), made from arginine, glycine and methionine in kidney and liver.
Measured fluxes and cell functions. The measured rates loaded by default (NCI-60) belong to proliferating tumour cells, and pin glucose, lactate, glutamine and other exchanges to them. They do not describe a hepatocyte or a neuron, and they also limit unmeasured by-products to traces, which can cap a function artificially (cortisol falls from 0.095 to 0.001 because a side product of steroidogenesis may only leave in traces). When a cell-function objective is chosen with measured fluxes active, the objective note asks you to set them to "None". Growth, ATP and glutathione can be used with them. For the liver functions you can instead use the human liver set (section 5).

Objectives that were considered and left out, because FBA does not give a meaningful answer for them in these models:

  • Glutathione synthetase (MAR04326): 6.5 mmol/gDW/h of GSH that γ-glutamyl aminopeptidase (MAR04175) cleaves again at once, a futile cycle that only burns ATP. Net GSH release (above) is used instead.
  • Glutathione reductase (NADPH regeneration): unbounded, because the model can reoxidise GSH in a closed loop.
  • Lipid droplet formation (MAR00031): its product has no outlet in the model, so the optimum is 0. Lipogenesis is represented by palmitate release.
  • Lactate release: feasible, but it is a by-product, not something a cell is thought to maximise.

Every objective is applied in the direction in which the reaction is written in the model. To use your own, search any reaction in the box, choose Maximise or Minimise, and press Save: the objective is kept in the list under a name of your choice (by reaction ID, so it also works in the other model when the reaction exists there). Delete removes it. Saved objectives are kept in the browser; a downloaded session stores the objective in use.

Growth in Mouse-GEM

Why. Mouse-GEM 1.8.0 ships with MAR00021 as its objective. That reaction takes exactly one unit of each of 33 precursors (amino acids, cholesterol, lipid pools, DNA, RNA, cofactors…), with no weighting by how much of each a gram of cell contains, so its value is not a growth rate: about 0.00016 in the standard medium, against 0.0370 h⁻¹ for Human-GEM, and the fluxes it implies do not match the composition of a cell.

How it is solved. Mouse-GEM also contains MAR13082, the generic cell biomass reaction of Human-GEM, with exactly the same stoichiometry (5.34 mmol of protein pool, 0.22 of lipid pool, DNA, RNA, glycogen, metabolite and cofactor pools and 45 ATP per gram of dry weight), but closed (bounds 0/0). The built-in objective Growth (generic cell biomass, MAR13082), offered only for Mouse-GEM, uses it: while it is the objective the simulator opens its upper bound to 1000, and only then. Nothing in the published model is rewritten: MAR00021 remains the default objective, and choosing any other objective leaves MAR13082 closed again. With it, Mouse-GEM grows at 0.0370 h⁻¹ in the standard medium and 0.0341 h⁻¹ with the NCI-60 measured fluxes, the same as Human-GEM. The composition is that of a generic human cell; a mouse-specific composition is not available in the model.

8. Steady state versus Dynamic

Steady stateDynamic
Question"With this medium, what does the cell do?""What happens over the next hours as the medium is used up?"
MediumConstant: concentrations do not change.Finite: every uptake lowers the concentration, every secretion raises it.
FluxesOne FBA solution.FBA is solved again each time a nutrient runs low enough to change what the cell can take up.
BiomassNot tracked.Grows exponentially at the current growth rate.
LabelFinal (isotopic steady state): every pool has reached its ultimate enrichment.Fills the intracellular pools gradually, so you see label arrive first in glycolysis and later in distant products.
Time neededA few seconds.Around half a minute; twice that with a knockout, because the unperturbed cell is simulated too.

How the dynamic mode works

  1. Time runs from 10 seconds to the chosen duration in 30 steps, spaced logarithmically (short steps first, when label changes fastest). With scheduled events, the short steps start again after each event.
  2. At each step the maximum uptake of every nutrient is recomputed from its current concentration, and can never exceed what is physically left in the medium.
  3. If a limit that was actually constraining the cell has moved by more than about 8 %, FBA is solved again; otherwise the previous fluxes are kept.
  4. Concentrations are updated: ΔC = −uptake × biomass × Δt. Oxygen is held constant, as gas exchange replenishes it.
  5. Biomass grows: X ← X · exp(growth rate × Δt).
  6. Label is advanced by one implicit step in every metabolite pool.

Three numbers control it: the duration, the initial biomass (gDW per litre of medium: more cells deplete the medium sooner) and the intracellular pool size (how much of each metabolite the cell holds, which sets how fast label fills it).

Scheduled events (advanced)

In dynamic mode, the "Scheduled events" section of the left panel lets you change the medium while the simulation runs:

  • Add a metabolite at a given hour: choose it, set the concentration to add and, with ⚛, which atoms are labelled and in what fraction of the molecules. This is how you programme a labelling pulse.
  • Change the whole medium at a given hour: the medium is replaced by one of the built-in or saved media, unlabelled, and whatever the cells had secreted is washed away. After a labelled start this is a chase.

The medium then holds a mix of batches of the same metabolite, for example unlabelled glucose that was there plus labelled glucose added later. Cells take them up in proportion to what is left of each. FBA is solved again at every event. Events set after the end of the simulation are ignored.

A metabolite the cells secrete into the medium and later take up again (lactate, alanine …) is counted as unlabelled when it comes back.
The dynamic mode is an approximation ("dynamic FBA"). All intracellular pools are given the same size, which is not true in real cells: read labelling times as indicative, and the order in which things get labelled as the informative part.

9. The medium

A stoichiometric model has no concentrations, only limits on uptake. The app converts each concentration into a maximum uptake rate with a saturating rule:

maximum uptake = Vmax · C / (Km + C)

So doubling a low concentration roughly doubles what the cell may take, while doubling an already saturating one changes little. Hover a concentration box to see the resulting limit. The cell does not have to use the whole allowance: FBA takes what serves the objective.

With the measured exchange fluxes on (the default), a measured uptake replaces this rule: the cell takes up what was measured, as long as the metabolite is in the medium. Those components are marked ⊙ in the list. The rule still limits everything else.

  • Built-in media: Standard (rich, DMEM/F-12-like), Plasmax and HPLM (physiological concentrations), and Williams E (serum-free, the medium of the human liver slices of Grankvist et al. 2024, from their Supplementary Table 2; menadione, manganese and nitrate, absent from the models, are left out).
  • Your own: edit, add or remove components, then "save…" to keep the medium under a name in this browser.
  • Unlimited components: water, protons, phosphate and the bulk salts are never limiting.
  • Injections: you can add a metabolite that is not extracellular (for example cytosolic citrate). It is then supplied directly inside that compartment — an experiment that only a model allows.

For Plasmax and HPLM, components that the models do not contain (selenite, manganese, vanadate, nitrate, p-aminobenzoate, methyl acetoacetate, phenol red) are left out. Essential fatty acids and fat-soluble vitamins, which in the real media come from serum, are added at the standard medium's values.

10. Isotope tracing

Press ⚛ next to a medium component to open the molecule editor, click the atoms to label (13C, 15N, 18O or 34S) and choose which fraction of the molecules carries the label.

Positional enrichment

For each atom of each metabolite the app computes its enrichment: the fraction of molecules in which that atom is the heavy isotope. An atom inherits its enrichment from the atoms it was made from, weighted by flux:

enrichment of an atom = Σ (flux of each producing reaction × enrichment of the source atom) / Σ flux

This is solved exactly for the whole network, cycles included (the TCA cycle feeds label back into itself). Select a metabolite to see its structure coloured atom by atom.

Isotopologues (M+0, M+1, M+2 …)

A mass spectrometer does not see positions: it counts how many heavy atoms each molecule carries. M+0 is the unlabelled molecule, M+3 a molecule with three heavy atoms. This distribution cannot be deduced from positional enrichment alone. Fully labelled glucose gives pyruvate that is either M+3 or M+0, never a random mix of labelled positions.

The app therefore traces whole molecules backwards. To build one molecule of the selected metabolite it picks one of the reactions that make it (in proportion to flux), finds which substrate molecule each atom came from, and repeats for those substrates until it reaches the medium, where the label is known. Atoms that came from the same substrate molecule stay together, which is what keeps the pattern right. Doing this for 2,000 molecules gives the distribution shown in the metabolite panel.

  • Available with a tracer, in both modes. In steady state it is the final distribution. In dynamic mode it is the distribution at the time shown on the slider, and "Show over time" plots each isotopologue along the run.
  • In dynamic mode the walk also goes back in time. A molecule found in a pool entered it some time ago: how long is drawn at random from the pool size and its turnover. Fluxes and medium composition are taken as they were at that earlier moment, and a molecule older than the simulation is unlabelled. This is what makes a pulse added at hour 4 appear first in glycolysis and only later in distant products.
  • Right after the start or an event (the first seconds), the atom-by-atom enrichment and the isotopologue bars can disagree: the first is computed in coarse time steps, the second with exact waiting times. They agree again within minutes.
  • Without a tracer the panel shows the distribution expected from natural isotope abundance (13C 1.07 %, 15N 0.36 %, 18O 0.2 %, 34S 4.25 %, 2H 0.01 % …), computed exactly from the formula. This is the background any measurement starts from.
  • With a tracer, natural isotope abundance is not added on top: compare with measurements that have been corrected for it.
  • Being a random sample of 2,000 molecules, each bar has an uncertainty of roughly ±1 percentage point.
  • If part of the atoms went through reactions without an atom map, the panel says so and the distribution is approximate.
  • 18O and 34S add two mass units per atom, 13C and 15N one.

Where the label ends up

With a tracer, the summary shows what share of the labelled atoms leaves the cell in each product (CO2, lactate, biomass …).

11. Knockout and knockdown

Add a gene or a reaction in the "Knockout / knockdown" section and set its residual activity: 0 % is a knockout, anything in between a knockdown.

  • Genes act through gene–protein–reaction rules. If a reaction needs A and B (a complex), losing either stops it. If it needs A or B (isoenzymes), losing one changes nothing. The list shows, for example, "2/7": two of the seven reactions that mention the gene actually depend on it. A struck-through gene has no effect at all.
  • The reference. Every perturbed simulation is compared with the same simulation without the perturbation, computed automatically.
  • A knockdown limits the reaction to that fraction of its reference flux. A reaction that carried no flux in the reference is not constrained by a partial knockdown.
  • "Changed" means a flux that differs from the reference by more than 2 %.
  • Measured fluxes stay in force in the perturbed simulation, so that the comparison is between two cells with the same measured phenotype. If the knockout makes them impossible, they are relaxed and the summary says so.
FBA assumes the cell reroutes perfectly and instantly. A knockout that is harmless in the model means "an alternative route exists in the network", not "the cell will not notice". Conversely, a knockout that stops growth in the model points to a reaction with no bypass.

Screening essential genes

Screen essential genes, below the knockout list, knocks out every gene one by one and reports which ones lower the objective. Only genes that are detected in the chosen cell type and whose reactions carry flux in the current result are screened. Each knockout is evaluated through the GPR rules like a normal one, with the genes switched off by the cell type also off, so a non-detected isoenzyme does not compensate; genes that close the same set of reactions share one calculation, and genes covered by a detected isoenzyme need none. Every calculation is an FBA (no pFBA) with the same medium, measured fluxes and cell type; the knockouts already in the list are ignored, and measured fluxes that a knockout makes impossible are relaxed and marked with *. In HEK293 with its measured fluxes, 614 genes take 263 linear programs (about 2½ minutes): 168 are essential and 36 lower the objective, among them cholesterol (HMGCR, SQLE, LSS), coenzyme A (COASY), nucleotide (UMPS, RRM1) and fatty-acid synthesis (FASN).

The result has two lists: essential (objective below 1 % of the unperturbed one) and lower the objective (1–95 %). Each gene shows how many reactions it closes (their names on hover); click it to see its details, or ✂ to add the knockout. Genes marked ↺ are not detected in the cell type and were switched back on because a metabolic task needs them (section 6): their essentiality comes from that task, not from data on the cell. The genes screened without effect are listed at the end. FBA predicts what the network needs, not what a real knockout does: regulation, enzyme amounts and alternative nutrients the medium lacks are outside the model.

12. Comparing two situations

The comparison mode puts any two situations side by side: two media or initial concentrations, a pulse or a medium change, a knockout, a different tracer, or an edited model against the published one.

  1. Set up the first situation and simulate it. In Compare, press Save current as scenario… and give it a name (for example "Control").
  2. Change whatever you want to test and simulate again.
  3. Pick the scenario in the list (the radio button). A blue badge on the map says what you are comparing with.

Below the list, Show [X] compared with [Y] makes both sides explicit. "Show" is the situation drawn on the map and in the panels: the current one or any saved scenario, whose stored result is shown again without recomputing (if it was made with other model edits, it is loaded and recomputed). "Compared with" is the reference: every change is measured from it to the situation shown, so "Flux change from «B» to «A»" means A relative to B. ⇅ swaps the two. Changing anything afterwards turns the shown side back into the current, unsaved situation.

From then on, the scenario takes the place of the "reference" everywhere:

  • Map. "Show: Changed flux only" colours reactions by flux change (blue = less, orange = more). "Changed label" colours metabolites by the change of their mean enrichment and hides the rest. In the global and cell views each group is coloured by the change of its flux-weighted label.
  • Summary panel. The objective with its change, the most changed reactions, the exchanges that change, the pathways that change most (sum of absolute fluxes) and the metabolites whose labelling changes most. Every line reads "scenario → now"; click it to go there.
  • Metabolite panel. Turnover and mean enrichment in both situations, and the two isotopologue distributions as paired bars (grey: scenario, amber: now). A situation without a tracer shows as 100 % M+0.
  • Reaction panel and tooltips. The flux in the scenario next to the current one.

Notes:

  • The scenario keeps its result, so nothing is recomputed when you switch between scenarios. ↺ loads a scenario's settings back (medium, tracers, perturbations, events, model edits) and recomputes it.
  • Two dynamic runs are compared at the closest time point. A steady-state scenario is compared with every time point of a dynamic run.
  • If the two situations use different model edits, reactions and metabolites are matched by ID. Those that exist on one side only count as zero flux on the other.
  • While a scenario is selected it replaces the automatic unperturbed reference of a knockout. Stop comparing to get that reference back.
  • Scenarios live in memory. To keep them, download the session (section 16). Importing it recomputes every scenario.

13. Editing the model

Press ✎ Edit model in the top bar. The "Model edits" section lists every change, and the reaction and metabolite panels get edit buttons. You can:

  • add a reaction, with stoichiometry, bounds, enzymes (GPR rule) and an atom map;
  • add a metabolite: a new compound (from its SMILES), or an existing compound in another compartment;
  • change a published reaction: its bounds (for example make it irreversible), its GPR rule (add or remove enzymes) or its name;
  • remove a reaction or a metabolite. A metabolite can only go together with the reactions that use it, and the app offers to remove them;
  • copy a reaction as new: for example, put an enzyme in another compartment.

Edits are operations applied, in order, on top of the published model. The published model itself is never changed. Removed items stay in place, flagged, so results before and after an edit can be compared item by item. Edits are kept in this browser for each model. ✕ undoes one, and "discard all" returns to the published model. The edit list and the map badge flag rejected edits (✕) and edits with warnings (⚠).

What makes an edit valid

The editor checks every rule below while you type. An edit with an error (✕) cannot be saved, and a stored edit that has become invalid is skipped. Warnings (⚠) mean the edit is legal but will probably not do what you expect. The rules follow the usual quality checks for genome-scale models (Thiele & Palsson 2010; MEMOTE, Lieven et al. 2020).

CheckLevelWhy
Unique ID (letters, digits, _ . -)errorSessions, comparisons and exports refer to reactions and metabolites by ID.
Every metabolite exists (or is created) and appears once; coefficients are non-zero numberserrorThe stoichiometric matrix must be well defined.
Substrates and products, unless it has a single metabolite (exchange)errorA reaction with only one side destroys or creates mass. An exchange of an extracellular metabolite is the legal way to do that; a single metabolite inside the cell is flagged as a sink.
Lower bound ≤ upper bound, both numbers; ±1000 means unboundederror / warningBounds define direction and reversibility: 0 … 1000 is irreversible, −1000 … 1000 reversible.
Element balance (every element except H)error, unless you accept itOtherwise atoms are created or destroyed. Label could then appear from nowhere, and growth could be "free". Lumped or boundary reactions can be accepted explicitly.
Hydrogen and charge balancewarningUsually a missing H⁺. GEMs balance protons at pH 7.2.
Formula of every participantwarningWithout formulas the balance cannot be checked.
GPR syntax: gene symbols joined by and (complex) / or (isoenzymes), balanced parentheseserrorThe rule decides what a knockout does. New gene symbols are added to the model and can be knocked out. No rule means no enzyme, so no knockout reaches the reaction.
Not a duplicate of an existing reactionwarningTwo identical reactions split the flux arbitrarily.
Metabolites in at most two compartmentswarningA reaction normally takes place in one compartment or moves something across one membrane.
No dead ends or isolated metabolites created by the editwarningAt steady state every metabolite must be both produced and consumed. A new compound that nothing consumes blocks the reaction that makes it, so it carries zero flux. A new product therefore needs an exit: a transport to the medium plus an exchange, or a consuming reaction. Removing a reaction can leave other metabolites stranded; the confirmation lists them.
The objective (biomass) reaction and its metabolites cannot be removederrorChoose another objective first.
Atom map: indices inside the molecules, from substrate to product, same element, no product atom filled twiceerrorA wrong map moves label to impossible places.
Atom map covers all product carbonswarningUnmapped carbons get the average label of the unmapped substrate atoms and are flagged position-uncertain.

Which data does an edit need?

  • For fluxes (FBA): only the stoichiometry and the bounds, plus the GPR rule if knockouts should reach the reaction. A formula for every metabolite lets the balance be checked.
  • To trace a metabolite as a whole: nothing more. A metabolite without structure counts as one pseudo-atom weighted by its carbons. Its mean enrichment is right, but it has no positions or isotopologues.
  • For positional labelling and isotopologues: a SMILES of every new compound and an atom map of every new reaction. The SMILES alone is not enough: it says what the molecule is, not which substrate atom becomes which product atom. From the SMILES the app (RDKit, loaded on first use) draws the molecule, computes formula and charge, and finds its symmetric atoms the same way as the preprocessing. Write it in the protonation state of the model (pH 7.2, carboxylates as COO⁻); heavy atoms must match the formula.

The atom map can come from three places, in decreasing reliability:

  1. An atom-mapped reaction SMILES that you paste, with every stoichiometric copy written out. You can make one with RXNMapper, which is already installed for the preprocessing: tools/.venv/bin/python tools/map_reaction.py "substrates>>products". It runs locally. The app matches each molecule of the SMILES to a participant by its skeleton, so charges and bond orders may differ.
  2. Automatic rules: the same compound on both sides is carried over whole (transport, carriers). A published reaction made of the same compounds, for example one you copied to another compartment, lends its map.
  3. Automatic maximum common substructure, the same greedy method as the preprocessing fallback. It is fast but chemically naive. In the example below it pairs five atoms of one pyruvate with acetoin, where RXNMapper correctly takes two carbons from each pyruvate. Check it, or paste a mapped SMILES instead.
Exchanges you add are exempt from the "unmeasured products only as traces" rule of the measured fluxes (section 5): you added them to be used. If the measured phenotype should include the new product, add it to "Measured exchange fluxes" with its measured rate.

Edits are saved with the session or alone (model_edits.json), and the import replays them on the published model. model_edits_report.csv lists every edit with its checks, for a methods section.

14. The cascade order

Each metabolite gets a depth: the smallest number of reaction steps needed to reach it from the metabolites of the current medium. Medium components are at 0; glucose inside the cell is at 1; glucose 6-phosphate at 2; and so on. Reaction directions follow the model, not the simulated fluxes, so the order does not jump when you change a concentration — only when you add or remove a medium component.

  • Cofactors (ATP, NAD, water, CoA …) are ignored, otherwise everything would be one step from everything.
  • The metabolite panel shows the depth and its position on a 0–100 scale.
  • With a rich medium most of central metabolism is only two to four steps away, because amino acids enter it directly. Remove components to stretch the cascade.
  • The top of every cascade is the same: first the medium (extracellular metabolites), then "step 0", the transport reactions that carry them into the cell, and from step 1 on what happens inside.
  • Inside a pathway map the rows count steps within that pathway, starting from its metabolites closest to the medium. That is what gives glycolysis its familiar glucose-to-pyruvate ladder. The medium metabolites that feed the pathway and their transporters are drawn above it, so nothing appears out of nowhere.
  • On the global map, the medium row is the "Exchange/demand reactions" pathway and step 0 is "Transport reactions"; double-click either to list its reactions.
  • Cycles cannot be drawn without some arrows pointing up; that is expected.

15. In-silico experiments

Each example gives the steps and what we obtained with Human-GEM, the standard medium, the NCI-60 measured fluxes and pFBA on (all defaults). Your numbers should match closely. With these defaults the cell grows at 0.0341 h⁻¹, takes up 0.742 mmol/gDW/h of glucose and secretes 1.35 of lactate.

A. Knock out an enzyme

  1. In "Knockout / knockdown", type ASNS (asparagine synthetase) and pick the gene.
  2. Set "Show" to "Changed flux only".

Result. Growth falls only slightly, from 0.0341 to 0.0333 h⁻¹ (−2 %): the measured cells already take up asparagine from the medium, and that covers most of the need. Switch the measured fluxes to "None" and compare.

Compare with: PDHB (pyruvate dehydrogenase): growth unchanged, but about 220 reactions are rerouted. FASN, HMGCR or CAD: growth stops, there is no bypass. GAPDH: no effect, because the model has an isoenzyme (GAPDHS).

B. Knock out a transporter

  1. Knock out the gene SLC2A1 (GLUT1). The list shows 2/7: only two of its seven reactions depend on it alone.
  2. Now remove it and knock out the reaction MAR05029 (facilitated glucose transport, catalysed by all GLUT isoforms).

Result. In both cases growth is unchanged and glucose uptake continues at the measured 0.742 mmol/gDW/h: the model still has other glucose transport reactions, such as the sodium-coupled symporters. To really starve the cell of glucose, set its concentration to 0.

Lesson. A generic model contains every transporter of the organism. A real cell line expresses only a few.

C. Trace 13C from glucose and read the isotopologues

  1. Press ⚛ next to glucose, click "All C", apply.
  2. Search for pyruvate, then citrate, and look at "Isotopologues" in the right panel. Use the compartment tabs to compare cytosol and mitochondrion.

Result. Lactate is 96 % M+3. Mitochondrial pyruvate is about 52 % M+3 and 41 % M+0: whole labelled three-carbon units from glucose, diluted by unlabelled pyruvate made from amino acids. Mitochondrial citrate is about 49 % M+2 and 38 % M+0; M+2 is the signature of labelled acetyl-CoA entering the TCA cycle. Fumarate is about 8 % M+2. The summary shows that 87 % of the labelled carbon leaves as lactate.

Variation. Remove the glucose label and label glutamine instead ("All C"). Glutamate is 100 % M+5 and fumarate 86 % M+4 (oxidative TCA cycle). Mitochondrial citrate is about 51 % M+4 and 35 % M+6: the M+6 comes from reductive carboxylation of M+5 2-oxoglutarate with CO2 that is itself labelled from glutamine oxidation.

D. Trace nitrogen

  1. Press ⚛ next to glutamine, click "All N", apply.
  2. Set "Show" to "Labelled only" and switch to the cell view.

Result. Label appears in glutamate, aspartate, alanine and the nucleotides, showing which pathways take their nitrogen from glutamine. Cytosolic glutamate and aspartate are about 95 % M+1 and alanine 100 % M+1.

E. Watch the medium run out

  1. Switch to "Dynamic". Keep 24 h and 0.5 gDW/L.
  2. When it finishes, drag the time slider and look at "Medium over time" in the summary.

Result. Biomass doubles (0.50 → 1.00 gDW/L). At the measured uptake rate glucose runs out between 9 and 13 h, and glutamine and choline by about 18 h. From then on the measured fluxes can no longer be met and are relaxed (the status bar says so), and growth slows from 0.0341 to 0.018 h⁻¹ by 24 h. FBA was solved 8 times.

Variation. Switch to Plasmax: with its physiological glutamine (0.65 mM) and vitamin B12, growth stops at about 9 h. Or raise the initial biomass to 2 gDW/L to see depletion within a few hours.

F. Maximise an antioxidant: glutathione

  1. In "Cellular objective", choose "Maximum glutathione synthesis" (the same as searching MAR09351, "Exchange of GSH", and keeping "Maximise").
  2. Read the objective value and the uptake list.
  3. Raise cystine, then cysteine, to 1 mM.

Result. The cell can export at most 0.0201 mmol/gDW/h of glutathione: sulfur is limiting, as cysteine, methionine and cystine are taken up at their maximum. With 1 mM cystine the maximum rises to 0.0226, and with 1 mM cysteine as well to 0.0250.

Pitfall. If you maximise the synthesis reaction instead (MAR04326), you get 6.5 mmol/gDW/h: the model makes and degrades glutathione in a loop, burning ATP. To ask "how much can the cell produce", maximise the export of the product, not an internal reaction.

Note. Growth is no longer the objective, so the cell does not grow in this simulation. The same approach works for other products, for example urate or taurine: search for "Exchange of …".

G. Low oxygen

  1. Set O2 from 0.2 to 0.0003 mM.

Result. Growth falls from 0.0341 to 0.0209 h⁻¹. Glucose uptake rises to 0.95 mmol/gDW/h and lactate secretion to 1.42, the top of their measured ranges: the cell leans on glycolysis for ATP. At 0.0001 mM growth drops to 0.0098 h⁻¹. With no oxygen at all the model cannot grow, because some biomass components need O2 to be made.

Compare. With the measured fluxes off, the same low oxygen makes the model excrete succinate rather than lactate: both dispose of excess reducing power, and FBA alone has no reason to prefer the one real cells use.

H. Compare media

  1. Switch the medium selector between Standard, Plasmax and HPLM.

Result. Growth: 0.0341 h⁻¹ in Standard and Plasmax, 0.0152 h⁻¹ in HPLM. With the measured fluxes on, both media give the same measured exchanges. HPLM is limited by pantothenate (1.05 µM, against 4.2 µM in Plasmax), which the cell needs for CoA: with Plasmax's pantothenate, HPLM grows at 0.0341 too. This limit depends on the assumed uptake kinetics. With the measured fluxes off, growth is 0.0370, 0.0238 and 0.0152 h⁻¹.

I. Remove glutamine

  1. Set glutamine to 0.

Result. The measured glutamine uptake is struck through in "Measured exchange fluxes": it no longer applies. Growth falls from 0.0341 to 0.0295 h⁻¹, and serine and the essential amino acids are taken up faster. The network can make everything it needs without glutamine. Many cultured cells cannot — a reminder that FBA shows capability, not regulation.

J. A labelling pulse during the run

  1. Switch to "Dynamic". In "Scheduled events", search glucose: an event appears at 4 h adding 5.5 mM.
  2. Press ⚛ on the event, click "All C", apply. Press "Simulate".
  3. Search glucose-6-phosphate, move the time slider, and press "Show over time".

Result. Until hour 4 there is no label: the panel shows natural abundance. At hour 4 medium glucose jumps from 3.9 to 9.4 mM, of which 5.5 mM labelled. Glucose 6-phosphate then becomes about 60 % M+6 and 40 % M+0 — whole labelled or whole unlabelled molecules, never a mixture — and pyruvate about 55 % M+3. Further on the label is diluted: about 30 % M+2 in mitochondrial citrate and 5 % M+2 in fumarate two hours after the pulse.

K. Pulse and chase

  1. Keep the pulse of example J and press "+ change the whole medium". Set that event to 12 h and leave the standard medium.
  2. Simulate and follow glucose 6-phosphate with "Show over time".

Result. Label rises after hour 4 and is washed out after hour 12: two hours after the chase glucose 6-phosphate, pyruvate, citrate and fumarate are all back to M+0. Slower pools keep their label longer; comparing how fast different metabolites lose it is the point of a chase. The run takes about a minute.

L. Compare normoxia and hypoxia, isotopologue by isotopologue

  1. Label glucose ("All C"). In "Compare", save the result as Control.
  2. Set O2 to 0.0003 mM and pick "Control" in the scenario list.
  3. Read the summary, set "Show" to "Changed label", then open mitochondrial citrate and fumarate.

Result. Growth goes from 0.0341 to 0.0209 h⁻¹ (−39 %), glucose uptake from 0.742 to 0.948 and O2 uptake from 0.359 to 0.142 mmol/gDW/h. The paired bars show the TCA cycle becoming much more labelled. Mitochondrial citrate goes from 48 % M+2 / 5 % M+4 to 41 % M+2, 25 % M+4, 20 % M+5 and 13 % M+6. Fumarate goes from 88 % to 42 % M+0, with M+2, M+3 and M+4 appearing. Fumarate M+3 is the signature of pyruvate carboxylation. Citrate M+4 to M+6 forms when labelled oxaloacetate meets labelled acetyl-CoA.

M. Add a heterologous pathway: acetoin

  1. Press ✎ Edit model → + Metabolite → new compound acetoin, compartment c, SMILES CC(O)C(C)=O, Read structure (C4H8O2), Save.
  2. + Reaction: substrates 2 pyruvate [c] and 1 H+ [c]; products 1 acetoin [c] and 2 CO2 [c]; GPR alsS and alsD (bacterial acetolactate synthase and decarboxylase, lumped). For the atom map, choose "atom-mapped reaction SMILES" and paste the output of tools/map_reaction.py (shown in the test tools/test_edits.mjs).
  3. The model checks now say that acetoin is a dead end. Add acetoin in compartment e (+ In another compartment), a transport (acetoin [c] ⇌ acetoin [e], reversible: its atom map is found automatically) and an exchange (acetoin [e] alone, 0 … 1000).
  4. Label glucose. Maximise the new exchange as the objective.

Result. With growth as the objective the pathway carries no flux: it costs carbon and gives the cell nothing. Maximising acetoin export gives 0.98 mmol/gDW/h. Acetoin is 58 % M+4, 37 % M+2 and 5 % M+0, exactly what two independent pyruvates that are 76 % M+3 give (0.76² = 0.58; 2 × 0.76 × 0.23 = 0.35). The automatic MCS map would have mapped only 3 of the 4 acetoin carbons from one pyruvate: use a mapped SMILES when positions matter.

N. Remove an enzyme from the model

  1. Save the default result with a glucose tracer as a scenario.
  2. In edit mode, remove the four L-lactate dehydrogenase reactions MAR04280, MAR04281, MAR04388 and MAR06018 (search them in "Remove a reaction or metabolite…"). Compare with the scenario.

Result. Growth is unchanged, but lactate secretion falls from 1.35 to 0.94 mmol/gDW/h, the bottom of the measured range: L-lactate is still made by minor routes. The cell burns the rest of its pyruvate, so oxygen uptake rises from 0.36 to 0.82. Mitochondrial citrate gains M+4 to M+6 and fumarate falls from 88 % to 47 % M+0. A knockout of LDHA alone would leave these reactions to the isoenzymes in their GPR rules. Removing the reactions takes the step out whatever enzyme catalyses it.

More ideas

  • Titrate a knockdown: add FASN (its knockout stops growth) and move the slider from 100 % to 0 % to see where growth starts to fall.
  • Measured against free FBA: repeat any experiment with "Measured exchange fluxes" set to "None" and compare.
  • Find essential nutrients: remove medium components one at a time and see which ones stop growth.
  • Positional tracers: label only carbon 1 of glucose and see how much of it is lost as CO2 in the pentose phosphate pathway.
  • Feed an intermediate: inject cytosolic citrate or mitochondrial succinate and see what the cell does with it.
  • Combine: a tracer plus a knockout shows how label is rerouted.

16. Downloading results

"Download" opens a list of everything the current result can provide. Tick what you want; you get one .zip with a README.txt describing each file. Tables are CSV (open them in any spreadsheet), charts are SVG (vector, editable), and the map is a PNG of what is on screen.

FileContentAvailable
fluxes.csvEvery reaction: flux, bounds, pathway, equation, genes, EC numbers; reference flux and change when there is a perturbation.Always
exchange_fluxes.csvUptake and secretion of every exchanged metabolite, with its concentration and kinetic parameters.Always
pathways.csvPer pathway: reactions, active reactions, total flux, enrichment.Always
metabolites.csvPer metabolite: turnover, cascade depth, mean enrichment.Always
flux_changes_vs_reference.csvReactions that changed by more than 2 %.Knockout / knockdown
atom_enrichment.csvEnrichment of every labelled atom.Tracer
isotopologues.csvM+0, M+1 … of the 150 busiest labelled metabolites (in dynamic mode, at the selected time).Tracer
isotopologues_over_time.csv and chartIsotopologue fractions of the selected metabolite at up to 30 time points.Dynamic, tracer, a metabolite selected
timecourse.csv, timecourse_enrichment.csvBiomass, growth, medium concentrations, exchange fluxes and enrichment at every time point.Dynamic
chart_*.svgUptake, secretion, pathway activity, most changed reactions, label fate, isotopologues, medium depletion, biomass.As above
map.pngThe map as shown.Map views
settings.jsonModel, medium, objective, tracers and perturbations: what you need to repeat the run.Always
session.jsonThe whole session, ready to import: model edits, medium (with labelled atoms), measured fluxes, perturbations, events, objective, dynamic parameters and the saved scenarios.Always, even without a result
model_edits.json, model_edits_report.csvOnly the model edits, and a table of every edit with its checks.Model edits
comparison_*.csvExchanges, pathways, mean enrichment and isotopologue distributions in the compared scenario and now.A scenario compared

Sessions: repeat the calculations elsewhere

"Import" in the top bar reads a session.json, a model_edits.json, or directly the .zip of a download that contains one of them. A session replaces the current settings and model edits, switches model if needed and recomputes every saved scenario, so you see the same comparison as its author. Edits imported alone can replace the current ones or be added after them. Everything is stored by ID, so a session also works after the model has been updated. IDs that no longer exist are skipped, and the status bar says how many. A file with only session.json (or only the edits) is downloaded as plain JSON.

In dynamic mode the per-reaction tables refer to the time point selected on the slider.

17. Limitations

  • No kinetics or regulation. Concentrations only set uptake limits. Enzyme levels, allostery and signalling are absent.
  • A generic cell. Human-GEM contains every reaction known in any human tissue. A specific cell type expresses a subset; the app does not yet build tissue-specific models.
  • Alternative optima. Several flux patterns can give the same growth. pFBA picks one; small "changes" after a knockout can be an equally good alternative rather than a real effect.
  • Measured fluxes are an average phenotype. They are the median of 59 tumour cell lines in RPMI-1640, converted with an average cell mass. They are not the phenotype of any particular line, and they are kept when you change the medium.
  • Uptake kinetics are rough estimates. They still set the uptake of unmeasured medium components, and with the measured fluxes off they set everything.
  • Net fluxes only. Reversible reactions exchange label in both directions in reality; the model sees only the net flow, which can underestimate label scrambling.
  • Atom maps are computed, not all curated. Uncertain positions are flagged.
  • No hydrogen tracing (2H) and no natural-abundance correction.
  • Dynamic mode assumes one pool size for all intracellular metabolites. Labelling times, and therefore isotopologue curves over time, scale with that number: their order and shape are informative, their clock is not.
  • Model edits are checked, not curated. The checks catch impossible or unbalanced reactions and dead ends, not wrong biology. Automatic atom maps (maximum common substructure) can pair the wrong atoms; prefer a mapped SMILES.
  • Scenarios are kept in memory with their results. Download the session to keep them after closing the page.
  • Mouse-GEM 1.8.0 ships with an objective (MAR00021) whose value is not a growth rate; use the objective "Growth (generic cell biomass, MAR13082)" for growth in h⁻¹ (section 7).
  • Cell types switch enzymes off where their mRNA is not detected, but do not scale fluxes with expression (section 6).

18. Glossary

FluxRate of a reaction, in mmol per gram dry weight per hour.
Carbon / nitrogen flowTransferable carbon or nitrogen atoms per hour that one group of reactions produces and another consumes, through shared metabolites; atoms of recycled carriers such as CoA are not counted (section 3).
ΔG′mTransformed Gibbs energy of a reaction with every reactant at 1 mM, at the pH of its compartment. Negative: the reaction can run forward (section 3).
GEMGenome-scale metabolic model: all known metabolic reactions of an organism, linked to its genes.
FBA / pFBAFlux balance analysis / its parsimonious variant (least total flux).
ObjectiveThe reaction whose flux is maximised or minimised.
Biomass reactionA pseudo-reaction that consumes cell components in the proportions needed to make new cell mass.
Exchange reactionThe entry or exit of a metabolite between the medium and the outside of the model.
Measured exchange fluxAn uptake or secretion rate measured in cells, used to pin the corresponding exchange reaction.
ATP maintenanceATP spent on upkeep (ion pumps, turnover) regardless of growth.
GPR ruleGene–protein–reaction rule: which genes a reaction needs.
EnrichmentFraction of an atom position occupied by the heavy isotope.
IsotopologueMolecules differing only in their number of heavy atoms (M+0, M+1 …).
TurnoverTotal rate at which a metabolite is produced.
gDWGram of dry cell weight.
ScenarioA saved situation (settings, model edits and result) that the current one is compared with.
Atom mapWhich atom of each substrate becomes which atom of each product in a reaction.
Dead endA metabolite that no reaction can produce, or none can consume. The reactions that touch it are blocked at steady state.
SMILESA line notation of a molecule's structure, for example CC(O)C(C)=O for acetoin.

19. References

  • Orth JD, Thiele I, Palsson BØ. What is flux balance analysis? Nat. Biotechnol. 2010; 28:245–248. doi:10.1038/nbt.1614
  • Lewis NE et al. Omic data from evolved E. coli are consistent with computed optimal growth from genome-scale models. Mol. Syst. Biol. 2010; 6:390 (parsimonious FBA). doi:10.1038/msb.2010.47
  • Mahadevan R, Edwards JS, Doyle FJ. Dynamic flux balance analysis of diauxic growth in Escherichia coli. Biophys. J. 2002; 83:1331–1340. doi:10.1016/S0006-3495(02)73903-9
  • Robinson JL et al. An atlas of human metabolism. Sci. Signal. 2020; 13:eaaz1482 (Human-GEM). doi:10.1126/scisignal.aaz1482
  • Wang H et al. Genome-scale metabolic network reconstruction of model animals as a platform for translational research. PNAS 2021; 118:e2102344118 (Mouse-GEM). doi:10.1073/pnas.2102344118
  • Schwaller P et al. Extraction of organic chemistry grammar from unsupervised learning of chemical reactions. Sci. Adv. 2021; 7:eabe4166 (RXNMapper). doi:10.1126/sciadv.abe4166
  • Vande Voorde J et al. Improving the metabolic fidelity of cancer models with a physiological cell culture medium. Sci. Adv. 2019; 5:eaau7314 (Plasmax). doi:10.1126/sciadv.aau7314
  • Cantor JR et al. Physiologic medium rewires cellular metabolism and reveals uric acid as an endogenous inhibitor of UMP synthase. Cell 2017; 169:258–272 (HPLM). doi:10.1016/j.cell.2017.03.023
  • Jain M et al. Metabolite profiling identifies a key role for glycine in rapid cancer cell proliferation. Science 2012; 336:1040–1044 (NCI-60 exchange fluxes). doi:10.1126/science.1218595
  • Aurich MK, Fleming RMT, Thiele I. A systems approach reveals distinct metabolic strategies among the NCI-60 cancer cell lines. PLoS Comput. Biol. 2017; 13:e1005698 (per-line values, S1 Table Q). doi:10.1371/journal.pcbi.1005698
  • Zielinski DC et al. Systems biology analysis of drivers underlying hallmarks of cancer cell metabolism. Sci. Rep. 2017; 7:41241 (cell mass, ATP maintenance). doi:10.1038/srep41241
  • Kilburn DG, Lilly MD, Webb FC. The energetics of mammalian cell growth. J. Cell Sci. 1969; 4:645–654 (ATP maintenance). doi:10.1242/jcs.4.3.645
  • Dietmair S, Hodson MP, Quek LE, Timmins NE, Gray P, Nielsen LK. A multi-omics analysis of recombinant protein production in Hek293 cells. PLoS ONE 2012; 7:e43394 (Table 2). doi:10.1371/journal.pone.0043394
  • Grankvist N, Jönsson C, Hedin K, Sundqvist N, Sandström P, Björnsson B, Begzati A, Mickols E, Artursson P, Jain M, Cedersund G, Nilsson R. Global 13C tracing and metabolic flux analysis of intact human liver tissue ex vivo. Nat. Metab. 2024; 6:1963–1975. doi:10.1038/s42255-024-01119-3
  • Lemons JMS, Feng XJ, Bennett BD, Legesse-Miller A, Johnson EL, Raitman I, Pollina EA, Rabitz HA, Rabinowitz JD, Coller HA. Quiescent fibroblasts exhibit high metabolic activity. PLoS Biol. 2010; 8:e1000514 (Fig. S1A). doi:10.1371/journal.pbio.1000514
  • Wimmer M, Wilmering B, Sasse D. The relation of rat liver wet weight to dry weight. Histochemistry 1985; 83:571–572. doi:10.1007/BF00492462
  • Uhlén M, Fagerberg L, Hallström BM, et al. Tissue-based map of the human proteome. Science 2015; 347:1260419. doi:10.1126/science.1260419
  • Karlsson M, Zhang C, Méar L, et al. A single–cell type transcriptomics map of human tissues. Sci. Adv. 2021; 7:eabh2169. doi:10.1126/sciadv.abh2169
  • Jin H, Zhang C, Zwahlen M, et al. Systematic transcriptional analysis of human cell lines for gene expression landscape and tumor representation. Nat. Commun. 2023; 14:5417. doi:10.1038/s41467-023-41132-w
  • Bairoch A. The Cellosaurus, a cell-line knowledge resource. J. Biomol. Tech. 2018; 29:25–38. doi:10.7171/jbt.18-2902-002
  • Bastian FB, Roux J, Niknejad A, et al. The Bgee suite: integrated curated expression atlas and comparative transcriptomics in animals. Nucleic Acids Res. 2021; 49:D831–D847. doi:10.1093/nar/gkaa793
  • Agren R, Mardinoglu A, Asplund A, et al. Identification of anticancer drugs for hepatocellular carcinoma through personalized genome-scale metabolic modeling. Mol. Syst. Biol. 2014; 10:721. doi:10.1002/msb.145122
  • Mahadevan R, Schilling CH. The effects of alternate optimal solutions in constraint-based genome-scale metabolic models. Metab. Eng. 2003; 5:264–276. doi:10.1016/j.ymben.2003.09.002
  • Dolfi SC et al. The metabolic demands of cancer cells are coupled to their size and protein synthesis rates. Cancer Metab. 2013; 1:20 (NCI-60 culture conditions). doi:10.1186/2049-3002-1-20
  • Thiele I, Palsson BØ. A protocol for generating a high-quality genome-scale metabolic reconstruction. Nat. Protoc. 2010; 5:93–121 (quality checks for model edits). doi:10.1038/nprot.2009.203
  • Lieven C et al. MEMOTE for standardized genome-scale metabolic model testing. Nat. Biotechnol. 2020; 38:272–276. doi:10.1038/s41587-020-0446-y
  • RDKit: open-source cheminformatics, MinimalLib 2026.03.6 (SMILES, depiction, maximum common substructure in the model editor). doi:10.5281/zenodo.591637
  • Noor E, Haraldsdóttir HS, Milo R, Fleming RMT. Consistent estimation of Gibbs energy using component contributions. PLoS Comput. Biol. 2013; 9:e1003098. doi:10.1371/journal.pcbi.1003098
  • Beber ME et al. eQuilibrator 3.0: a database solution for thermodynamic constant estimation. Nucleic Acids Res. 2022; 50:D603–D609. doi:10.1093/nar/gkab1106
  • eQuilibrator compound database (compounds.sqlite), used offline by tools/build_thermo.py. doi:10.5281/zenodo.4128543
  • Haraldsdóttir HS, Thiele I, Fleming RMT. Quantitative assignment of reaction directionality in a multicompartmental human metabolic reconstruction. Biophys. J. 2012; 102:1703–1711. doi:10.1016/j.bpj.2012.02.032
  • Casey JR, Grinstein S, Orlowski J. Sensors and regulators of intracellular pH. Nat. Rev. Mol. Cell Biol. 2010; 11:50–61 (compartment pH). doi:10.1038/nrm2820
  • Porcelli AM et al. pH difference across the outer mitochondrial membrane measured with a green fluorescent protein mutant. Biochem. Biophys. Res. Commun. 2005; 326:799–804 (intermembrane space pH). doi:10.1016/j.bbrc.2004.11.105
  • Zorova LD et al. Mitochondrial membrane potential. Anal. Biochem. 2018; 552:50–59. doi:10.1016/j.ab.2017.07.009
  • Yang M, Brackenbury WJ. Membrane potential and cancer progression. Front. Physiol. 2013; 4:185 (plasma membrane potential). doi:10.3389/fphys.2013.00185
  • Urban PF, Klingenberg M. On the redox potentials of ubiquinone and cytochrome b in the respiratory chain. Eur. J. Biochem. 1969; 9:519–525. doi:10.1111/j.1432-1033.1969.tb00640.x
  • Loveridge KM et al. A divergent mitochondrial cytochrome c in malaria parasites with an anomalously low redox potential. J. Mol. Biol. 2026; 438:169996 (+250 mV for canonical mitochondrial cytochrome c). doi:10.1016/j.jmb.2026.169996
  • Traut TW. Physiological concentrations of purines and pyrimidines. Mol. Cell. Biochem. 1994; 140:1–22. doi:10.1007/BF00928361
  • Schuetz R, Kuepfer L, Sauer U. Systematic evaluation of objective functions for predicting intracellular fluxes in Escherichia coli. Mol. Syst. Biol. 2007; 3:119. doi:10.1038/msb4100162
  • Lu SC. Glutathione synthesis. Biochim. Biophys. Acta 2013; 1830:3143–3153. doi:10.1016/j.bbagen.2012.09.008
  • Gille C et al. HepatoNet1: a comprehensive metabolic reconstruction of the human hepatocyte for the analysis of liver physiology. Mol. Syst. Biol. 2010; 6:411. doi:10.1038/msb.2010.62
  • Richelle A et al. Model-based assessment of mammalian cell metabolic functionalities using omics data. Cell Rep. Methods 2021; 1:100040. doi:10.1016/j.crmeth.2021.100040
  • Huangfu Q, Hall JAJ. Parallelizing the dual revised simplex method. Math. Program. Comput. 2018; 10:119–142 (HiGHS). doi:10.1007/s12532-017-0130-5