Evolution is often described as the greatest engineer that ever existed. Four billion years of relentless optimization, the argument goes, have already found the best protein for every job. We think that framing quietly smuggles in a mistake — and that the mistake is the whole opportunity.
A greedy search over a vanishing fraction of space
Evolution is not a designer. It is a greedy, path-dependent search. Selection can only take a mutational step that is viable right now, in this organism, against this environment. It cannot pay a temporary fitness cost to reach a better solution on the far side of a valley. It cannot redesign a fold from scratch. It cannot anticipate.
The consequence is that natural enzymes are local optima shaped by constraints that have nothing to do with the chemistry they perform: codon usage, folding kinetics, the scaffolds an ancestor happened to carry, the neutral drift of populations. Whole classes of active-site geometry that would solve a reaction more cleanly were simply never reachable, because no continuous, viable path led there. The space of functional protein structures is astronomically larger than the space evolution actually explored.
That is what we mean by the slogan: function is conserved, structure is free. The chemistry we want — a transition state stabilized, a charge preorganized, a polymer threaded and turned over — does not care about lineage. It cares about geometry. And geometry is something we can design toward directly.
Designing the part evolution never reached
So we invert the question. Instead of mutating a natural enzyme and hoping, we encode the task itself as a scorable objective — substrate-binding geometry, transition-state stabilization, processivity, foldability — and let generative models search a learned manifold of designable backbones conditioned on that objective. The target is the reaction, not reproductive success. We treat the natural baseline as a number to beat, not a library to copy.
This is not a claim that we enumerate all structures or simulate every evolutionary path. We do neither. It is a claim that objective-driven generation can cross fitness valleys that selection cannot, and reach designs evolution had no route to.
The honest part: de novo enzymes start slow
Here is what the field does not always say out loud. De novo enzymes have historically been poor catalysts — often orders of magnitude short of their natural counterparts on rate enhancement. Computational scores are proxies. MD, FEP and QM-MM approximate catalysis; they are not turnover. A structure that predicts beautifully can still do nothing in a tube.
We do not claim that gap is solved. We claim something narrower and, we think, more defensible: a loop that climbs it.
The model proposes. The wet lab decides. The data closes the loop.
Every design that clears prediction is expressed and assayed, and the turnover we measure flows back into the objective layer and the generative models. Each turn sharpens what "catalytically competent" means in a way no public dataset can. Models commoditize. Functional data does not.
That is the bet. Not a bigger network — a tighter loop, pointed at the structures evolution forgot to invent.