Abstract

BoltzGen is a new generative model for designing protein and peptides of any modality to bind a wide range of biomolecular targets. It unifies design and structure prediction, resulting in a single model that also achieves state-of-the-art folding performance. BoltzGen’s generation process can be controlled with a flexible design specification language over covalent bonds, structure constraints, binding sites, and more. BoltzGen was developed at MIT and experimentally validated in a large-scale distributed effort involving multiple academic and industry labs. These groups independently validated designed nanobodies, minibinders, peptides, and cyclic peptides against diverse and novel targets such as small molecules, peptides, and proteins with disordered regions, with robust experimental validation including functional readouts in live cells. We explicitly focus our experimental validation on targets that are highly dissimilar to any proteins for which bound structures exist – more faithfully mirroring a real discovery campaign. BoltzGen, like Boltz-1 and Boltz-2, is open source under MIT license and freely available for unrestricted academic and commercial use, including data, model weights, training, and inference code.

https://github.com/HannesStark/boltzgen