The Shortlist That Beat a Library Half a Million Strong

The constraint on photosynthesis in most crops is not light, and it is not water. It is one enzyme. Rubisco pulls carbon dioxide out of the air and attaches it to a five-carbon sugar, slowly enough that plants compensate by building it in bulk: it accounts for 20 to 50% of the soluble protein in a leaf. It also grabs oxygen by mistake, at a cost in energy and carbon. The obstacle to fixing it is that the large subunit doing the catalysis is nearly identical across land plants, so random mutagenesis mostly yields broken enzyme, and the space of possible changes is far too large to search exhaustively.

A group at MIT tried shrinking the search rather than enlarging the library. They scored every possible single amino acid change in tobacco Rubisco's large subunit with ESM-IF1, a model that predicts which sequences fit a given protein backbone, feeding it crystal structures of the enzyme in its inactive open and substrate-bound closed states. It rated 202 substitutions at 49 positions more plausible than the real tobacco residue. The team built a saturated library of roughly 400 variants at those sites and selected it in a strain of E. coli engineered to die unless Rubisco is working. Sixty-four substitutions came out enriched. The best single change, threonine 391 to isoleucine, carboxylated 29% faster and needed 11% less carbon dioxide to reach half its maximum rate: a net 43% gain in carboxylation efficiency in air, with no significant loss of the enzyme's preference for carbon dioxide over oxygen.

Why it matters: Rubisco's sluggishness sets a ceiling on photosynthesis in soybean, rice and wheat, so a faster version is one of the few routes to higher yields that needs no extra land, water or fertiliser. Nothing here was grown as a crop, but it shows gains of this size are reachable.

The work appeared in Science Advances, from McDonald et al. in the chemistry department at the Massachusetts Institute of Technology, with collaborators at the Australian National University and Stanford University. Three authors are named on a provisional patent application filed by MIT relating to this work, and one declares cofounder interests in two biotechnology companies.

Why the sequence space is the real problem

The paper's most transferable result is about search efficiency. An earlier directed-evolution campaign on tobacco Rubisco used error-prone PCR and found 46 enriched variants in a library of about 500,000. Here the positive enrichment rate was 16%, which the authors put at 1700-fold higher, and the difference is in where the model looked. A sequence-only model the team tried first, ESM-1bv, mostly suggested amino acids already common in plants outside the nightshade family, at positions with an average conservation score of 28%. The structure-informed model went after more conserved positions, averaging 37 to 55% depending on the crystal structure, and 42% and 58% of its top 24 recommendations counted as rare against a survey of roughly 128,000 plant Rubisco sequences. The model was not reproducing evolution; it was proposing residues evolution has not used.

Growth in bacteria says nothing about whether a chloroplast would tolerate such residues. So the team built a transgenic line of a tobacco relative whose chloroplast genome carries a bacterial Rubisco in place of its own, then delivered the variant genes transiently into its leaves. Assembled tobacco holoenzyme was detectable for two of the rarest substitutions five days later.

What the winning variant says about the enzyme

Threonine 391 sits near the active site, on the side facing the small subunit, and its hydroxyl group hydrogen-bonds to the backbone at methionine 387; isoleucine or valine removes that bond. Two other substitutions there, to cysteine and to valine, improved carbon dioxide affinity without speeding carboxylation up, which makes position 391 look less like a lucky hit than a modulator of catalysis plants have left unexplored.

The benchmark the authors supply is worth stating plainly. Each naturally occurring substitution in this subunit is proposed to improve anaerobic carboxylation efficiency by only about 1%, while single changes here moved tobacco Rubisco across roughly 30% of the measured range for C3 plant enzymes. Specificity for carbon dioxide over oxygen was not significantly affected in any improved variant, which matters because the field long assumed speed had to be bought at specificity's expense. Stacking failed: the top five enriched substitutions combined grew slightly worse than threonine 391 to isoleucine alone, and the 10- and 20-substitution combinations did not grow at all. Credit also divides more finely than the result suggests. ESM-IF1 never recommended isoleucine at position 391; it recommended arginine, which was not enriched. The model found the site, the bacterial selection found the letter.

What the study can't say yet

Every kinetic number comes from enzyme purified out of bacteria, and the baseline is itself a variant, carrying a solubility-improving change the authors label wild type. The in-plant experiment measured assembly, not photosynthesis, and over days of transient expression rather than in a stably transformed line. The 15 to 25% faster carbon dioxide assimilation attributed to these variants at today's 420 parts per million is a projection from a standard photosynthesis model fed the measured kinetics at an assumed leaf Rubisco content, not a measurement. The authors also note that the same substitutions could reduce chloroplast Rubisco content or leave the enzyme less activated in a real plant, which may be why evolution never adopted them. Building stable transformed tobacco, they write, will be necessary to analyze and understand these potential limitations, as well as how the kinetic changes affect plant photosynthesis and productivity. One caution applies to the method rather than the enzyme: some recommendations landed on residues known to break the interaction with Rubisco activase, the chaperone that keeps Rubisco switched on, an interaction the authors say these models do not capture.

Quick questions

Does this mean faster-growing crops? Not yet. The enzyme is faster in a test tube and assembles inside a leaf, but no plant carrying it was grown and measured.

Why tobacco rather than wheat? Tobacco Rubisco can be made in bacteria alongside the chaperones it needs, which is what makes screening hundreds of variants practical.

What is the one-line takeaway? A structure-trained model cut the whole large subunit down to 49 sites worth saturating, and selection inside that shortlist produced a tobacco Rubisco 43% more efficient in air.