A RoseTTAFold capable of modelling proteins along with the non-protein components like nucleic acids, metals, small molecules and covalent modifications. They then fine-tune on diffusive denoising tasks and can generate binding pockets and proteins that bind to molecules of interest, and they experimentally (in vitro) validate the results of these designs.
This is nice because I remember reading with AlphaFold2 that you can see predicted structures that have side-chains in the right position to bind a metal group, for example, but AlphaFold doesn't explicitly predict this group (Figure 1c). The RoseTTAFold All-Atom model would be useful for modelling and designing more complex biomolecular systems.