From the source
NVIDIA and the University of Washington's Institute for Protein Design are showing how to run RFD4-Proteina, their BioNeMo open model for programmable protein design, from a browser.
- The guide walks through setting a design context, generating binders, and building a repeatable workflow in two
Hugging Face Spaces, one on dedicated GPUs and one on shared ZeroGPU capacity.
- RFD4-Proteina pairs a flow-matching model with an autoencoder that represents each amino acid, sugar, or nucleotide as an anchor-atom coordinate plus a learned latent vector, so the model generates sequences and all-atom structures together instead of handing off between separate backbone, sequence, and structure models.
- It can design proteins in contexts that include other proteins, small molecules, nucleic acids, glycans, metal ions, and modified amino acids.
- A thinking setting (low, medium, high) controls how much the model refines each design, trading more compute for designs more likely to pass in silico filters.
- NVIDIA says a benchmark across 24 protein-binding targets found high thinking produced 76.40 structurally distinct designs per GPU-hour that met the stated computational criteria, against 13.24 for Proteina-Complexa.
