2026-07-29
Structure-aware protein model ESM-IF1 guides directed evolution of plant Rubisco; top single mutation T391I lifts carboxylation efficiency 43% and assembles in chloroplasts.
Most of the biomass on Earth is built by a single enzyme. Rubisco fixes atmospheric CO2 into sugars, sits at the heart of photosynthesis, and makes up 20 to 50 percent of the soluble protein in a leaf. It is also slow, error-prone (it often grabs oxygen instead of CO2), and notoriously hard to engineer. Assembling it needs several dedicated chaperones, and land-plant Rubisco is so conserved that random mutagenesis mostly hits deleterious changes. The question is whether machine learning can find the handful of useful sites that natural evolution never tried.
The team uses ESM-IF1, a structure-aware protein language model, as the filter. Unlike sequence-only models, ESM-IF1 is an inverse-folding model trained on roughly 12 million AlphaFold2 structures; it reads the biophysics of a 3D fold rather than what nature happened to sample.
The pipeline runs in steps:
The choice of a structure model over a sequence model matters because plant Rubisco is so conserved that sequence-only models (ESM-1bv) merely recommend what already exists. Site saturation instead of random mutagenesis both shrinks the library and recovers substitutions that need two nucleotide changes in one codon, which single-round error-prone PCR under-samples (A438R is the example).
Selection efficiency is the headline number. The library's positive enrichment rate is 16 percent, about 1,700 times higher than prior error-prone PCR work, which pulled 46 enriched variants out of roughly 500,000. The gain is not a higher hit rate per se; it is a library three orders of magnitude smaller.
On kinetics, the best single variant, T391I, raises the carboxylation rate (kcat) by 29 percent and aerobic carboxylation efficiency (kcat/KC) by 43 percent, with no loss in CO2/O2 specificity.
| Variant | Key change | Efficiency gain |
| T391I | kcat +29% | +43% |
| I465V | Better CO2 affinity | +32% |
| A438R | Better CO2 affinity | +15% |
Of 19 enriched variants, 7 genuinely grew better from improved catalysis rather than higher solubility. Stacking does not help: the 5-mutation variant is slightly worse than T391I alone, and the 10- and 20-mutation variants fail to grow because folding breaks.
The conceptual punch: several improved variants, T391I included, carry substitutions that are rare or absent across the plant lineage, yet they still assemble correctly in real chloroplasts (verified in transgenic N. benthamiana). Plant Rubisco has not reached its ceiling; there is real signal in sequence space that nature never sampled.
This is the strongest result yet for structure-aware ML on Rubisco, the most stubborn conserved enzyme in photosynthesis, and it breaks past the ceiling set by natural diversity. If the kinetic gains translate to whole plants, it points to a real route to higher crop photosynthesis. More broadly, it validates the recipe of ML site selection plus saturation plus in-vivo selection for conserved enzymes that random mutagenesis cannot crack.
The authors are candid. ESM-IF1 does not always pick the best substitution at a given site: it predicted T391R, which was not enriched, while the winning T391I came out of saturation. ML points to the right site; the final call still needs experimental screening. Unsupervised models also miss protein-protein interactions: recommendations such as R89P and K94E disrupt the Rubisco-activase interaction and could be harmful in a whole plant, so they need supervised models or manual curation.
The biggest open question is translation. The gains were measured in E. coli and in vitro. In real chloroplasts, lower Rubisco content, more side-product formation, or reduced activation state could erode them. Stable transgenic tobacco lines still have to be built to say whether whole-plant photosynthesis actually rises.