Protein design methods are advancing rapidly, but designing high-affinity binders, scaffolding multiple functional motifs, and stabilizing large multi-domain proteins remains challenging.

In a new bioRxiv preprint, researchers introduce Genie 3, a diffusion model that uses atomistic SE(3)-equivariance and a branched polymer representation of protein structures. The approach aims to overcome the trade-off between the design capabilities of existing generative models and their computational cost.

Genie 3 demonstrates state-of-the-art performance across binder design, motif scaffolding, and unconditional protein generation, while achieving substantially faster design speeds than existing approaches. Furthermore, its capabilities are demonstrated by designing a nanomolar binder against Nipah virus Glycoprotein G, achieving a 12.5% success rate in the Adaptyv Bio Nipah Competition.