Protein design moved to the atom. The bottleneck moved to the bench.

The Institute for Protein Design at the University of Washington published a system this week, in Science, that designs biomolecules atom by atom.

Earlier models drew a protein backbone first and tried to fit the chemistry in afterward. RFdiffusion3 places the individual atoms around the reaction you want, then builds the protein to hold them.

The proof is in the enzymes. Of the first 96 phosphotriesterase designs, 4 were active against paraoxon. Of 96 protease candidates, 56 cut more than 10% of the target protein. It beat the previous model on 9 of 10 enzyme-scaffolding problems, at more than twice the speed of two rival methods.

These are early enzymes. The best protease worked at the low end of the range you see in natural ones, and the nucleic-acid designs were computational only, not yet made in a lab.

Here's what I think matters. Designing a candidate enzyme used to be the slow, expensive step. It's now fast and cheap. So the bottleneck moves to the wet lab: making the design, testing it, measuring whether it works. The scarce resource in this field is no longer the idea. It's the experiment that checks the idea.

For anyone funding AI in drug discovery, that changes the question you ask a founder. Not "how good is your model," but "how fast and honest is your loop from design to bench and back."

Read the primary source →