BoltzGen

An all-atom generative model for designing proteins and peptides that bind a chosen target, from MIT (first and corresponding author Hannes Stark, with senior authors Regina Barzilay and Tommi Jaakkola) and announced by MIT's Jameel Clinic in October 2025. Open Athena's Timothy O'Donnell is one of its equal core computational contributors; the paper's other affiliations are Boltz, CTU Prague, IOCB Prague and Boston, NVIDIA, UC Irvine, UCSF, MPI, HHMI and MIT's Jameel Clinic. Building on the structure-prediction line that runs through AlphaFold 3 and Boltz, BoltzGen unifies binder design with structure prediction in a single model that also reaches state-of-the-art folding performance, and a design-specification language steers generation through covalent bonds, structural constraints and binding sites. Targets can be proteins, peptides, small molecules or nucleic acids, and binders span nanobodies, minibinders, and linear and disulfide-bonded cyclic peptides.

The paper reports eight design campaigns with functional and affinity readouts across 26 targets, run by separate wet labs that each chose their own targets. The hardest test is 10 novel targets with no protein above 30% sequence identity in a bound context anywhere in the PDB: testing 15 or fewer nanobody designs per target produced hits for 60% of them, and the same experiment with protein binders hit 50%. Designing against three bioactive peptides gave nanomolar binders for two and micromolar for the third while testing only six designs each; peptides against two Ragulator proteins reached hit rates of 88% and 28%; and of 10 designs against the disordered region of NPM1, two bound and co-localized with the target in live cells. Weights, data, and inference and training code are released under the MIT license (about 1,100 GitHub stars).