RFdiffusion3 designs proteins atom by atom and works in lab experiments

A team from the University of Washington, Oxford and NVIDIA published RFdiffusion3 in Science, a generative AI system that designs proteins and nucleic acids by modeling individual atoms.

It represents each amino acid with up to 14 atoms and sets their 3D coordinates directly. On 41 enzyme active sites, it outperformed RFdiffusion2 in 37 tests and generates structures about ten times faster.

Two enzymes it designed, a organophosphate-degrading enzyme and a peptide cleavage enzyme, completed their intended chemical reactions in real experiments.