RFD4-Proteina brings extra thinking time to protein design

The University of Washington and Nvidia introduced RFD4-Proteina on 21 October, bringing a familiar idea from reasoning models to the design of molecules: give a difficult problem more computation before settling on an answer. Their open generative model jointly develops a protein’s amino-acid sequence and three-dimensional structure, while allowing the molecules it must interact with to change shape too.[1]

The appeal is easy to see. A protein intended to recognise another molecule needs both the right chemical ingredients and the right arrangement of those ingredients. Optimising the two together creates room to revise a promising design instead of committing early to a shape that later proves difficult to realise.[2]

A design that can keep changing

RFD4-Proteina uses a shared internal representation for molecular identity and geometry. Its vocabulary extends beyond ordinary amino acids to modified residues, sugars, DNA and RNA. Recurrent refinement lets it revisit that internal representation during generation; a separate guided-search approach can use external scoring criteria to steer the search towards desired properties. These are ways to spend computation on improving a candidate, with low, medium and high refinement budgets available in the researchers’ comparisons.[2]

That matters when the specification describes behaviour as well as shape. A useful binder may need to recognise a molecule only in a particular state. An enzyme must hold its reacting ingredients in an arrangement that favours the intended chemical transformation. The interesting question is whether the resulting proteins perform those jobs when made in a laboratory.[2]

From a calculated structure to a working molecule

The accompanying preprint, which has not been peer reviewed, reports an enzyme designed for one bond-forming step in lenacapavir synthesis. Four of 67 tested designs produced the desired amide above background. The strongest converted 79% of the intended starting amine, compared with 4.3% in the no-enzyme control. This is a laboratory demonstration of one manufacturing reaction; it does not establish a complete production process or a reduction in the drug’s cost.[3]

A second experiment tested recognition of calmodulin, a protein that changes conformation when it binds calcium. Twelve of 18 designs bound in the calcium-containing condition with apparent dissociation constants below one micromolar, while showing no detectable binding without calcium. The distinction is functional: the designers asked for recognition of a particular molecular state, and measured whether the resulting proteins discriminated between conditions.[3]

Researchers can download the model’s code and weights under the BSD 3-Clause licence. Optional external evaluation software and weights retain their own terms, including restrictions in some upstream components. The model card also makes an unusually consequential version distinction: its computational benchmark table was frozen on 9 October using a previous checkpoint pair. Those figures have not been remeasured on the current 19 October checkpoints and should not be presented as measurements of that release.[4]

The strongest reason to watch this release is the richer specification it can attempt: a molecule that changes alongside its partner, recognises a chosen state or favours a particular reaction. Extra computation gives designers another control over that search. The laboratory remains where a convincing molecular proposal becomes a useful protein.

Report an error or suggest a correction