AI designed antibodies from scratch and 4 of 5 came out the exact shape predicted
A University of Washington lab had AI design antibodies for a chosen spot on a target. Shapes matched, but they grip too weakly for drugs.
Antibodies are the basis of many modern medicines, and until now they have come from immunising animals, sifting huge random libraries, or taking them from patients. A paper in Nature on 5 November 2025, from David Baker’s lab at the University of Washington, instead designed them on a computer, aimed at a chosen spot on a chosen target protein.
The team adapted RFdiffusion, a protein-design model, so it could draw the flexible loops an antibody uses to grip. They made binders against four disease-related targets, then imaged five of them with an electron microscope: four matched the computer’s prediction down to the atom. The software, RFantibody, is free for academic and commercial use.
Strength is the weak point. The first designs gripped their targets only weakly, in the tens to hundreds of nanomolar range, and the team had to run a lab evolution step to reach single-digit nanomolar. Baker told GEN that the designs do not yet have what a drug needs, strong enough binding included.
Companies report better numbers that nobody outside has checked. Chai Discovery said in June 2025 that its Chai-2 model found at least one working binder for half of 52 target proteins that had no known antibody, testing up to 20 designs each.
What it means
Designing an antibody to hit a precise spot on a protein now works often enough to be checked under a microscope, which removes a slow step from early drug research. It does not yet produce medicines: the designs bind too weakly, and lab work is still needed to fix that. Treat company claims of higher success rates as unverified until someone independent repeats them.
Caveats
- These are proofs of concept, not drug candidates. Affinity, stability and manufacturability all need work, which the authors say themselves.
- The pipeline is not purely computational: it still needs yeast-display screening and, for the best binders, laboratory evolution.
- Success rates per target are not reported in the paper, so “it works” cannot be turned into “it works n% of the time”. The structures were confirmed for a handful of designs, not across the board.
- Chai Discovery’s hit rates come from the company that sells the model and were published as a preprint, not peer reviewed.