A Duke University team has built an AI called Raygun that can make a natural protein smaller or larger while keeping its folded shape and its working behavior intact, a trick they say could hand biotech better-behaved versions of molecules already in use.
Raygun rests on ESM-2, a protein language model schooled on millions of amino-acid sequences. Where typical models dream up brand-new molecules, Raygun reshapes ones that already exist, borrowing tricks from evolution such as adding, deleting or swapping a single building block.
Earlier software that tried to resize real proteins typically ended up distorting them. Raygun avoids that by chopping a protein into segments, turning each segment into a numerical representation, then applying learned rules to reconstruct the protein at a new size while keeping its architecture sound, according to the team that includes computational biologist Rohit Singh.
The findings appeared in Nature on July 29. Fajie Yuan, a protein language model researcher at Westlake University who was not involved, welcomed the approach as a meaningful first step toward editing trusted proteins rather than starting from scratch.
The medical payoff could be significant. A smaller protein can fit into spaces a larger one cannot, and resized molecules may be easier to manufacture as biologics, antibodies and enzyme treatments. Refining existing proteins is also less risky than designing new ones from zero.
The bigger picture: protein AI is maturing from inventing molecules to fine-tuning familiar ones. Researchers expect that to matter for drug developers chasing biologics with better delivery, stability or tissue penetration.
