An enzyme-design model that generates more viable and more novel proteins, while training at a fraction of the cost.

‍

De novo enzyme design is a long-standing goal of protein engineering, with applications spanning industrial and medical biocatalysis. One of the key challenges is generating stable proteins that hold their catalytic machinery in exactly the right arrangement around a target molecule. Recent advances in generative protein models have made this problem more tractable, but current models remain expensive to train and produce limited structural diversity. Our new model, Emyx, is smaller and far cheaper to train, yet produces more successful and more novel designs than the leading alternatives.
‍

Doing more with less

Most AI protein-generation models were built by borrowing the heavy, complex machinery behind protein-structure prediction models. Motivated by the fact that generative protein models rely on much simpler inputs (the geometric position of a few amino acids) than structure-prediction models (which draw on rich, sequence-level information), we built Emyx to focus its capacity where it matters. The result is a 140M parameter highly efficient model that trains four times faster than the next-leading model and produces more successful protein samples that are also more novel.

Better designs, and more of them

We benchmarked Emyx on the standard AME benchmark for enzyme generation which includes 41 catalytic sites with increasing complexity. As shown in the figure below, Emyx produces successful designs about 1.5 times more than the next-best model. More importantly, the generated designs were not only more novel than ones generated by other models, they were also more varied resulting in a higher number of unique successful designs. Emyx achieves all of this despite being trained on less data than other models.

Why it matters

‍For our protein designers, Emyx's higher success rates and greater scaffold diversity mean starting each project from a richer pool of viable candidates, reaching into structural space beyond what natural enzymes have explored. And because it trains so cheaply, we can retrain and refine it far more often, iterating on the architecture and tuning it to our targets to push performance further with every cycle.