David Baker’s lab with Nvidia releases model that reasons about molecules
David Baker’s Institute for Protein Design has released RFD4-Proteina, an open generative model built with Nvidia that reasons about biomolecules rather than simply generating shapes that fit.
The University of Washington lab, whose director shared the 2024 Nobel Prize in Chemistry, says the model merges two lineages of drug-design AI. One is the institute’s RosettaFold Diffusion line. The other is Proteina-Complexa, contributed by Nvidia and trained on its BioNeMo platform for biology and chemistry.
The practical difference is joint refinement. RFD4-Proteina updates a protein’s sequence and structure together with the structures of whatever it is meant to bind, so a change in one propagates to the others instead of being fixed in place.
Nvidia reports fivefold to tenfold higher design success rates and comparable speedups in candidate throughput against previous models. Benchmarks covered four tasks: protein binding, small-molecule binding, enzyme scaffolding and DNA binding.
Wet-lab work backed the simulations. Baker’s team says bench results matched the computational gains, a claim worth watching because protein design papers have historically struggled to translate in vitro.
Like modern language models, RFD4-Proteina can spend more compute on harder problems, dialing effort up when a target resists easy solutions. That makes the model more useful for difficult interfaces than for routine scaffolds.
Releasing it openly is the strategic move. Baker argues the field advances faster when academic groups in low-resource settings can run the same tools as large pharma, and Nvidia gets its software embedded in the design stack that drug hunters reach for first.
