The Slowest Engine in the World: RuBisCO, Photorespiration and the Limits of Optimization

The Hook: An enzyme that works thirty times a minute

Picture a machine sitting in a data center, one that everything depends on. Every transaction runs through it. There is no second path. Now you learn two things about this machine.

First: it is slow. Not a little slow, but slow enough that you have to look at the number twice. Second: it has an error rate no operator would accept — on roughly every fifth or sixth operation it confuses the input signal with another one and produces a toxic intermediate that then has to be laboriously disposed of.

The consequence of both is predictable: you build very, very many of these machines. That is exactly what the biosphere has done.

In 2019, Yinon Bar-On and Ron Milo worked out how much RuBisCO — ribulose-1,5-bisphosphate carboxylase/oxygenase, the enzyme that feeds carbon dioxide into biology — exists on this planet. Their result: about 0.7 gigatons, roughly 700 million tons, "more than an order of magnitude higher than previously thought" (PNAS 116(10):4738–4743, 2019). More than 90 percent of it sits in the roughly 2 · 10¹⁴ square meters of leaf area of land plants, where the enzyme makes up about 3 percent of leaf mass. Per human being on Earth, that is nearly a hundred kilograms of a single protein.

And how fast does this apparatus work? Bar-On and Milo back-calculate global carbon fixation against the available enzyme mass and arrive at an effective, time-averaged turnover rate of "approximately 0.03 s⁻¹ on land and approximately 0.6 s⁻¹ in the ocean." Zero point zero three reactions per second. A RuBisCO molecule in a meadow therefore binds about two carbon dioxide molecules per minute. Compared with the maximal rate measurable in vitro at 25 °C, that is roughly a hundredfold slower on land and sevenfold slower in the ocean.

The part of this story that is not trivial: the slowness is the smaller problem. The bigger one is the confusion. And the genuinely interesting part: after forty years of attempts to improve this enzyme, the first clear field success was not an improvement of the enzyme but a bypass of its error.

Part 1: The core concept – an enzyme that cannot tell two gases apart

What RuBisCO is supposed to do

The Calvin-Benson cycle needs an entry point: a reaction that attaches inorganic carbon to an organic scaffold. RuBisCO provides it. It takes the sugar phosphate ribulose-1,5-bisphosphate (RuBP), a five-carbon body, and appends a CO₂ molecule. The resulting six-carbon intermediate splits into two molecules of 3-phosphoglycerate, which enter the cycle and ultimately yield sugar.

That is carboxylation, and it is the bottleneck of the entire living world. Nearly all the carbon in every organism you know has at some point passed through the active site of a RuBisCO molecule.

What RuBisCO also does

In the active site, RuBP is first converted into a reactive enediol form. This form is open to electrophilic attack — and carbon dioxide is not the only small, apolar gas floating around in the cell. Oxygen fits too.

When O₂ attacks instead of CO₂, one molecule of 3-phosphoglycerate is produced alongside one molecule of 2-phosphoglycolate. That is the error. 2-phosphoglycolate is not merely useless but actively harmful: it inhibits enzymes of the Calvin cycle and has to be cleared away. The clearance service is called photorespiration.

In 2016, Walker, VanLoocke, Bernacchi and Ort compiled the balance sheet of this clearance service in the Annual Review of Plant Biology. Their wording for the process is precise and unforgiving: "Photorespiration is initiated when the initial enzyme of photosynthesis, ribulose-1,5-bisphosphate carboxylase/oxygenase (Rubisco), reacts with oxygen instead of carbon dioxide and produces a toxic compound that is then recycled by photorespiration." Photorespiration "detoxifies glycolate and recycles 75% of the carbon into phosphoglycerate, which can reenter the C3 cycle."

Seventy-five percent back. A quarter of the already invested carbon is lost again as CO₂. And the route there is expensive: "every Rubisco oxygenation consumes 3.5 ATP and 2 NADH equivalents."

How often does this happen?

Here is the number worth remembering. Walker and colleagues calculate: "at 25 °C and under the then-current ambient carbon dioxide concentration of 350 ppm, photorespiration loses carbon at ~26% the rate of net carbon dioxide assimilation. This corresponds to ~2 Rubisco oxygenations for every 5 carboxylations."

Two misfires for every five hits. Or, in the authors' own counting: "This stoichiometry dictates that there are 5 carboxylations for every 7 Rubisco reactions." Five of seven reactions are productive, two are not. That is not a footnote, that is nearly a third of operations.

The leaf's energy budget reflects it: photorespiration "consumes 32% of total ATP and 28% of total NADH equivalents in an illuminated C3 leaf under current atmospheric conditions at 25 °C." Nearly a third of the light energy of a C3 leaf goes into repairing an enzyme's mistake.

Why the error was never simply mutated away

The obvious question: three billion years of evolution, and nobody fixed this?

The answer begins with the atmosphere. When RuBisCO arose, there was essentially no free oxygen. The possibility of confusion existed in theory but was irrelevant in practice — there was no competitor in the active site. The oxygen RuBisCO struggles with today is a by-product of photosynthesis itself, which makes it a problem photosynthesis created for itself. (How water is split and oxygen released in the first place is covered in The Machine That Splits Water: Photosystem II and the Origin of Oxygen.)

The answer continues with a physical argument, and this is where it gets interesting.

Part 2: The optimality thesis – an enzyme at the edge of the possible

Tcherkez, Farquhar and Andrews, 2006

The paper's title is itself an argument: "Despite slow catalysis and confused substrate specificity, all ribulose bisphosphate carboxylases may be nearly perfectly optimized" (Tcherkez, Farquhar & Andrews, PNAS 103(19):7246–7251, 2006, DOI 10.1073/pnas.0600605103).

The thesis: the slowness and the unreliability are not two independent defects but two sides of the same compromise. The authors write that all Rubiscos may be "nearly perfectly adapted to the differing CO₂, O₂, and thermal conditions in their subcellular environments, optimizing this compromise between CO₂/O₂ specificity and the maximum rate of catalytic turnover."

The proposed mechanism is chemically elegant. CO₂ is a "featureless" molecule: linear, nonpolar, without handholds. An enzyme can barely bind it specifically in the ground state. Specificity must therefore arise in the transition state. Evolution, the hypothesis runs, selected a transition state in which the CO₂ moiety already closely resembles the later carboxylate product. That discriminates well against oxygen. But it comes at a price: a product-like transition state means the carboxyketone intermediate binds so tightly that its cleavage to products is slowed.

Specificity and speed are therefore not freely combinable. Whoever wants to discriminate better has to become slower. The hypothesis also makes a testable prediction: because a product-like transition state is more ordered, the strategy's effectiveness deteriorates with increasing temperature.

What is strong and what is weak about this thesis

It matters to read the strength of the claim correctly. The title says "may be nearly perfectly optimized" — hedged twice, by "may" and by "nearly." The abstract reads "We assert that all Rubiscos may be nearly perfectly adapted": a strong assertive gesture over a hypothetical core. The mechanism is explicitly introduced as a hypothesis, not as a finding.

This reading is not pedantry. The optimality thesis went on to become the standard argument for not even attempting to improve RuBisCO — and that argument rests on an evidence base thinner than it appeared.

Flamholz and colleagues, 2018: the correlations shrink

Avi Flamholz, Noam Prywes, Uri Moran, Dan Davidi, Yinon Bar-On, Luke Oltrogge, Rui Alves, David Savage and Ron Milo re-ran the analysis ("Revisiting tradeoffs in Rubisco kinetic parameters," bioRxiv 470021, posted 15 November 2018). Their starting point is a plain observation about the evidence: the reasoning that established the optimality view "was based on data from only ≈20 organisms."

They expanded the dataset to "data from >200 organisms" and found: "We find that most correlations are substantially attenuated, with the inverse relationship between carboxylation k_cat and specificity S_C/O being a key example."

That is the heart of it. Precisely the correlation meant to demonstrate the trade-off between speed and specificity becomes markedly weaker with more data. A sharp compromise, in which every second of speed must be paid for with discriminating power, can no longer be asserted on this basis.

One constraint does survive, however — and it is not the same one as before. One model holds up better than it did: "increased catalytic efficiency (k_cat/K_M) for carboxylation requires a similar increase in catalytic efficiency for the competing oxygenation reaction, evidenced here by a strong power-law correlation between those catalytic efficiencies." Whoever makes carboxylation more efficient inevitably makes oxygenation more efficient too. The authors draw a different conclusion from this than Tcherkez: "our results imply that Rubisco evolution is constrained mostly by the physicochemical limits of O₂/CO₂ recognition, which should reframe efforts to understand and engineer this very central enzyme."

The limit, then, lies not in a speed-versus-accuracy bargain but in the physical distinguishability of two small gases. That is a harder limit, but a different one — and it says something different about where engineering work would have to begin.

A note in my own voice: I am of the opinion that this is an instructive instance of a recurring pattern — a trade-off model built on few data points and cited as a law of physics, until someone multiplies the sample by ten.

Part 3: How slow is slow? Surveying the tree of life

de Pins and colleagues, 2025

The most thorough answer so far to the question of whether a faster RuBisCO exists anywhere in nature appeared in November 2025: Benoit de Pins, Cyril Malbranke, Jagoda Jabłońska, Assaf Shmuel, Itai Sharon, Anne-Florence Bitbol, Oliver Mueller-Cajar, Elad Noor and Ron Milo, "Rubisco is slow across the tree of life," PNAS 122(47):e2501433122, published 17 November 2025, DOI 10.1073/pnas.2501433122.

The motivation is methodologically clean. The picture of a kinetically constrained enzyme "is based on kinetic measurements heavily focused on eukaryotic rubiscos, a rather conserved group of low genetic diversity." The fastest known variants came from the sparsely sampled prokaryotes. Hence the question: "Could there be yet faster rubiscos among the uncharted regions of rubisco's phylogenetic diversity?"

The answer came from characterizing "more than 250 rubiscos from a wide range of bacteria and archaea, thereby doubling the coverage of the diversity of this key enzyme" — carboxylation rates of 221 representative variants were measured with a coupled assay, supplemented by literature values with active-site quantification.

The numbers

• The median is 7 s⁻¹. For comparison, the authors cite "≈20 to 80 s⁻¹" for other central metabolic enzymes.
• Most variants fall below 10 s⁻¹ at 30 °C, "and only a few of them show carboxylation rates in the range of 15 to 30 s⁻¹."
• Published eukaryotic form I rates fall between 4 and 6 s⁻¹ (0.25 to 0.75 quantiles, corrected to 30 °C assuming a Q₁₀ of 2.2).
• Form II has a median of 15 s⁻¹; the remaining forms have median rates between 1.9 and 4.4 s⁻¹.
• For all ≈68,000 RuBisCO variants sequenced to date, the rate was predicted with a simple machine learning model.

The authors' summary: RuBisCO is "a relatively slow enzyme across the tree of life, never exceeding ≈30 reactions per second at 30 °C."

A qualification is needed here, one the paper itself invites. The ≈30 s⁻¹ is an approximation, not a sharp bound: in the results section the authors report a reconstructed ancestor that reached 33.0 s⁻¹, "the fastest rate ever measured using our pipeline." In the second round of screening "we did not measure any rate exceeding 30 s⁻¹," and all predicted rates reach up to 26 s⁻¹. The defensible statement is therefore: in broadly sampled nature, RuBisCO practically never exceeds about 30 s⁻¹ — not: it cannot in principle.

What makes one fast

One finding is striking and links kinetics to ecology: "relatively faster subclades share similar evolutionary contexts, involving micro-oxygenic environments or a CO₂ concentrating mechanism."

Fast RuBisCOs are found, in other words, where oxygen is scarce or where the organism actively concentrates CO₂. That fits the model of Flamholz and colleagues: if distinguishing CO₂ from O₂ is the real limit, then one can become faster as soon as one depends less on that distinction — because the competitor has been physically removed.

The authors are careful in their conclusions. They consider it possible that "we might be reaching a plateau in discovering higher carboxylation rates," and speak of RuBisCO having "probably reached 'Pareto optimality'." The "probably" is not decorative. Elsewhere they qualify the trade-off itself: such a catalytic bargain "may stand only when there is an evolutionary pressure on rubisco," and certain slow anaerobic form I variants "could be seen as a contradiction with the hypothesis of a trade-off."

Part 4: What the error costs – in tons and in dollars

Up to this point the story is biochemistry. It becomes agricultural economics the moment the calculation is carried onto a field.

Walker, VanLoocke, Bernacchi and Ort did exactly that, with a regional model for the United States. Their result, verbatim from the abstract: "A regional-scale model reveals that photorespiration currently decreases US soybean and wheat yields by 36% and 20%, respectively, and a 5% decrease in the losses due to photorespiration would be worth approximately $500 million annually in the United States."

Thirty-six percent for soybean. Twenty percent for wheat. And even a five-percent reduction in losses would be worth roughly half a billion dollars a year — in the United States alone, for two crops alone.

The paper's second important finding cuts against a widespread hope. One might assume that as atmospheric CO₂ rises, the ratio in the active site shifts, oxygenation declines, and the problem solves itself. The authors counter: "photorespiration will continue to impact yield under future climates despite increases in carbon dioxide, with models suggesting a 12–55% improvement in gross photosynthesis in the absence of photorespiration, even under climate change scenarios predicting the largest increases in atmospheric carbon dioxide concentration."

The reason is the temperature response. The solubility of CO₂ falls with temperature more steeply than that of O₂, and the specificity of RuBisCO deteriorates with warmth — the latter being precisely what the Tcherkez hypothesis predicts. More CO₂ and more warmth work against each other. The 12-to-55-percent range is wide, but its lower bound lies well above zero.

For the theoretical order of magnitude, the authors cite an older estimate: photorespiration reduces the theoretically attainable efficiency of gross C3 photosynthesis by 48 percent, at 380 ppm and 30 °C.

Part 5: What nature itself does

Before turning to genetic engineering, it is worth looking at how evolution has solved the same problem several times over — and consistently not at the enzyme.

None of the natural solutions makes RuBisCO faster or more accurate. All of them instead change the enzyme's environment: they raise the local CO₂ concentration at the active site so that the competitor O₂ gets its turn relatively less often. C4 plants separate pre-fixation and the Calvin cycle spatially across two cell types; CAM plants separate them temporally across night and day; many cyanobacteria and algae pack their RuBisCO into protein compartments with actively enriched CO₂.

That this works can be read off the dataset of de Pins and colleagues: the faster subclades share an evolutionary context involving "micro-oxygenic environments or a CO₂ concentrating mechanism." Whoever shuts out the oxygen can afford speed.

From this follows a design lesson that reaches far beyond biochemistry: when a component cannot be improved, you change its operating conditions. Nature has done this at least three times independently. That the pattern holds in engineering as well is part of what makes this article transferable.

Part 6: Rebuild instead of repair – South and colleagues, 2019

The approach

In 2019, Paul F. South, Amanda P. Cavanagh, Helen W. Liu and Donald R. Ort took a third route: not improving the enzyme, not raising the CO₂ concentration, but making the disposal route for the faulty product cheaper ("Synthetic glycolate metabolism pathways stimulate crop growth and productivity in the field," Science 363(6422):eaat9077, 4 January 2019, DOI 10.1126/science.aat9077).

The logic: natural photorespiration is a long route through three compartments — chloroplast, peroxisome, mitochondrion — losing carbon and nitrogen along the way. If the glycolate is instead broken down directly inside the chloroplast, transport routes and energy are saved.

The authors installed three alternative pathways into tobacco chloroplasts (Nicotiana tabacum cv. Petite Havana):

• AP1: five genes of the E. coli glycolate oxidation pathway, including glycolate dehydrogenase, glyoxylate carboligase and tartronic semialdehyde reductase.
• AP2: glycolate oxidase from Arabidopsis, malate synthase from Cucurbita maxima and catalase from E. coli.
• AP3: glycolate dehydrogenase from Chlamydomonas reinhardtii plus malate synthase from C. maxima. This pathway replaces the plant glycolate oxidase and thereby avoids hydrogen peroxide formation, which makes additional catalase unnecessary.

The decisive second component was flux control. A new pathway in the chloroplast is of no use if the glycolate keeps leaving the chloroplast via the native transporter. The authors therefore suppressed the glycolate-glycerate transporter PLGG1 by RNAi by about 80 percent. In the abstract: "Flux through the synthetic pathways was maximized by inhibiting glycolate export from the chloroplast."

The result

Two numbers from the abstract. First: "The synthetic pathways tested improved photosynthetic quantum yield by 20%." Second: "Numerous homozygous transgenic lines increased biomass productivity by >40% in replicated field trials."

More than 40 percent higher biomass productivity, in the field, across two growing seasons (2016 and 2017). The best pathway was AP3 with RNAi; AP2 showed no benefit over wild type, and with AP1 the gain was lost as soon as the RNAi module was added.

What this result does not say

Care is required here, because this paper in particular is often over-cited.

It concerned biomass productivity in tobacco, not grain yield in a food crop. Tobacco was chosen because it is "an ideal model field crop." Transfer to cereals remains explicitly a hope in the paper, not a measurement: "We are optimistic that similar gains may be achieved and translated into increased yield in C3 grain crops." Vegetative biomass and seed yield are not the same thing — in a leaf crop the biomass is the product; in wheat it is not.

Furthermore: a published correction to this article exists (Science, 2 August 2019, volume 365, issue 6452). I was unable to access its content in this run, and therefore cannot say whether or which figures it affects. Anyone citing the >40 percent should work from the corrected version.

Why this does not contradict Parts 2 and 3

At first glance it looks like a contradiction: Parts 2 and 3 say RuBisCO can barely be improved, while Part 6 reports a gain of more than 40 percent. Both are true, because they concern different things.

South and colleagues did not touch RuBisCO. The oxygenation rate in their plants is unchanged; the enzyme makes just as many mistakes as before. What was changed is solely what happens after the mistake — the disposal route. The hard limit lies in distinguishing CO₂ from O₂ in the active site. The disposal costs lie outside that limit and were therefore open to attack.

Precisely this distinction between the bottleneck itself and the downstream costs of the bottleneck is the real content of this article.

Frameworks to take away: four points of attack, three outcomes

The right-hand column tells a clear story: the two approaches that target the enzyme itself deliver nothing. The two that target its environment or its downstream costs deliver.

The Central Takeaway

Three points that carry beyond biochemistry.

First: distinguish the bottleneck from its downstream costs. RuBisCO is slow and inaccurate, and neither could be fixed so far. What could be fixed was the effort the system expends after every mistake. Anyone optimizing a system should ask these two questions separately: can I improve the component? And, independently of that: what does dealing with its failure cost me? The second question is often the more rewarding one, because it is not attached to the hard limit.

Second: check how many data points a cited law rests on. The trade-off between speed and specificity was treated as a quasi-physical bound. It rested on about 20 organisms. With more than 200 organisms, most of the load-bearing correlations were "substantially attenuated." What remained was a limit — but a different one, with different implications for engineering. I am of the opinion that "that is physically impossible" is, in technical discussions, similarly often an extrapolation from a handful of measurements.

Third: massive redundancy is a legitimate answer to an unimprovable bottleneck. The biosphere has an enzyme that binds two CO₂ per minute, and with it achieves the entire primary production of the Earth — by building 0.7 gigatons of it and devoting 3 percent of every leaf's mass to it. That is not elegant. It has worked for billions of years.

The practical prompt, if you are working on a system: take the bottleneck everyone considers unimprovable and write two lists. On the left, what would make the component itself better. On the right, what it costs that the component is the way it is — retries, error handling, transport paths, waiting times. The right-hand list is usually longer and almost always cheaper to shorten.

Cross-References in the Vault

• The Machine That Splits Water: Photosystem II and the Origin of Oxygen – the source of the oxygen RuBisCO has struggled with ever since
• The Strongest Bond in Nature: Nitrogenase and How Bacteria Split Nitrogen at Room Temperature – the second great slow and expensive key enzyme of the biosphere
• The Molecular Turbine: ATP Synthase and the Engine of Life – the counterexample: a biological catalyst that is fast at the physical limit
• Four That Act in Concert: Allostery, Hemoglobin, and the Logic of Cooperative Binding – how proteins bind and distinguish gases
• The Shape of Life: Levinthal's Paradox, Chaperones, and How Proteins Find Their Form – why assembling large enzyme complexes is a problem in its own right
• The Enemy That Became a Power Plant: Endosymbiosis and the Bacterial Origin of Complex Life – how the chloroplast got into the plant cell in the first place

Something to Think About

In your own domain: which component does everyone treat as given and unimprovable — and how much of your total effort actually goes not into that component, but into cleaning up behind it?

Fact Check

41 claims were checked (14 hard facts, 11 quotations, 9 source attributions, 7 synthesis claims) against the primary sources, each of which was opened individually during this run. 9 claims were corrected or weakened.

The three most important corrections:

1. "RuBisCO is the most abundant protein on Earth" was weakened. Bar-On and Milo themselves introduce this only as a claim made by others ("has been claimed to be the most abundant protein on Earth") and do not confirm it; what is documented is the mass of ≈0.7 Gt, not the ranking.
2. ">40% higher yield" was made precise as ">40% higher biomass productivity in tobacco in the field." South et al. measure vegetative biomass in Nicotiana tabacum; transfer to grain yield in C3 cereals is explicitly a hope in the paper ("We are optimistic that similar gains may be achieved"), not a measurement. In addition, a correction to this article exists (Science 365(6452), 2 Aug 2019) whose content was not accessible during this run.
3. "RuBisCO never exceeds 30 reactions per second" was weakened to "practically never." The abstract of de Pins et al. states "never exceeding ≈30," yet in the results section of the same paper a reconstructed ancestor reaches 33.0 s⁻¹.

Further weakenings concerned the optimality thesis (Tcherkez et al. phrase it as "may be nearly perfectly optimized," a hypothesis rather than a finding), the trade-off between speed and specificity (per Flamholz et al. "substantially attenuated"; what remains is a limit on CO₂/O₂ recognition), Pareto optimality (de Pins et al. write "probably"), the dollar figure (the abstract of Walker et al. gives ≈$500 million while the body text gives a different number — the abstract figure was used), and the attribution of the work by Flamholz et al., cited here as a bioRxiv preprint of 15 Nov 2018 because a peer-reviewed version could not be verified in this run. The transfer claims in "The Central Takeaway" are my own conclusions and are marked accordingly with "I am of the opinion that ..."

Note: Automated AI verification against primary sources; no human expert review.

Sources and Further Reading

1. Yinon M. Bar-On, Ron Milo: The global mass and average rate of rubisco. PNAS 116(10):4738–4743, 2019. DOI 10.1073/pnas.1816654116 — https://doi.org/10.1073/pnas.1816654116
2. Berkley J. Walker, Andy VanLoocke, Carl J. Bernacchi, Donald R. Ort: The Costs of Photorespiration to Food Production Now and in the Future. Annual Review of Plant Biology 67:107–129, 2016. DOI 10.1146/annurev-arplant-043015-111709 — https://doi.org/10.1146/annurev-arplant-043015-111709
3. Guillaume G. B. Tcherkez, Graham D. Farquhar, T. John Andrews: Despite slow catalysis and confused substrate specificity, all ribulose bisphosphate carboxylases may be nearly perfectly optimized. PNAS 103(19):7246–7251, 2006. DOI 10.1073/pnas.0600605103 — https://doi.org/10.1073/pnas.0600605103
4. Avi Flamholz, Noam Prywes, Uri Moran, Dan Davidi, Yinon Bar-On, Luke Oltrogge, Rui Alves, David Savage, Ron Milo: Revisiting tradeoffs in Rubisco kinetic parameters. bioRxiv 470021, 15 November 2018. DOI 10.1101/470021 — https://doi.org/10.1101/470021
5. Benoit de Pins, Cyril Malbranke, Jagoda Jabłońska, Assaf Shmuel, Itai Sharon, Anne-Florence Bitbol, Oliver Mueller-Cajar, Elad Noor, Ron Milo: Rubisco is slow across the tree of life. PNAS 122(47):e2501433122, 17 November 2025. DOI 10.1073/pnas.2501433122 — https://doi.org/10.1073/pnas.2501433122
6. Paul F. South, Amanda P. Cavanagh, Helen W. Liu, Donald R. Ort: Synthetic glycolate metabolism pathways stimulate crop growth and productivity in the field. Science 363(6422):eaat9077, 4 January 2019. DOI 10.1126/science.aat9077 — https://doi.org/10.1126/science.aat9077 (Correction: Science 365(6452), 2 August 2019)

← All articles