RFD4-Proteina is a biomolecular design model from the Institute for Protein Design at the University of Washington and NVIDIA. It co-designs sequence and structure for proteins and for their complexes with other proteins, small molecules, DNA and RNA.

An autoencoder places amino acids, sugars and nucleotides in a shared 32-dimensional chemical and conformational latent space, supervised with force-field energies so that it favours physically plausible conformations, and the model generates designs by flow matching in that space. Recurrent refinement lets it spend more computation on harder designs, with low, medium (the default) and high "thinking" settings, and generation can also be steered by external scores such as energies or geometric criteria.

The released checkpoint supports unconditional generation, motif scaffolding, enzyme design, and protein, small-molecule and DNA binder design.

Corley N, Didi K, Kazman MW, Brent RI, et al. Biomolecular design through latent reasoning with RFD4-Proteina. Preprint, 2026. The code is released under the BSD 3-Clause license and the weights on Hugging Face (nvidia/RFD4-Proteina).