De novo Rubisco design with protein language models
Ribulose-1,5-bisphosphate carboxylase/oxygenase (Rubisco) fixes the majority of carbon dioxide globally but is challenged with low specificity for CO2 versus O2 and low catalytic efficiencies. Traditional engineering efforts have remained difficult because folding, assembly, specificity, and catalysis are tightly coupled, hampering efforts to explore sequence space. Therefore, we leveraged recent advances in protein large language models (PLMs) to generate sequences beyond those observed in nature, using both ProGen-2 that was fine-tuned on a limited dataset of non-Form I Rubiscos and an ESM-2 discriminator. With this approach, we generated 5.6 million novel Rubisco-like sequences and identified 21 highly diverse candidates predicted to be active that occupy regions of Rubisco phylogenetic space not previously observed in nature. Six designs were soluble in Escherichia coli, and five were shown to produce quantifiable 3PGA. One design produced an apparent CO2/O2 specificity estimate beyond the range of the natural representative Rubiscos assayed. We also solved the crystal structure of one de novo design that reproduced the predicted dimer and active-site geometry with sub-angstrom C agreement. Sequence-only generation followed by independent structural filtering therefore recovered soluble, active Rubiscos from regions of sequence space that are not represented in genomic databases. Together, these results establish a scalable strategy for accessing previously unexplored Rubisco sequence space, providing a broadly accessible path toward generating de novo Rubiscos that may have activity and specificity parameters needed to address longstanding limitations in biological carbon fixation.