Atomic precision

Designing an enzyme means arranging specific atoms precisely enough to catalyse a reaction. “The catalytic and other sophisticated functions of biomolecules arise from the placement of many atoms in three-dimensional space,” the authors write. Most earlier design tools treat proteins as chains of whole amino acids, a level of detail that works for many proteins but is too coarse for building the active site of an enzyme.

RFD3 instead models every atom, for proteins and for nucleic acids such as RNA. The team trained it on the structures of biological molecules held in the Protein Data Bank, and users can tell it which atomic interactions a design must achieve, such as hydrogen bonding or contact with a target molecule. The model is also lighter than the AlphaFold3 system on which its architecture draws.

Beating earlier methods on screen

In computer-based tests, RFD3 outperformed previous approaches at designing proteins that bind a target, at producing RNA structures and at building enzyme active sites. It beat its predecessor, RFdiffusion2, on nine in ten enzyme-scaffolding problems drawn from a catalogue of natural enzyme mechanisms, and it was more than twice as fast as two rival methods for typical-length proteins. The RNA work was computational only.

Tested at the bench

The decisive evidence came from the laboratory. The team first designed enzymes to break down paraoxon, a toxic organophosphate. Native enzymes hydrolyse only a small subset of these compounds, so designed alternatives could prove valuable for environmental remediation and medical countermeasures, according to the paper. The target is difficult because the active site needs six catalytic residues coordinating two metal ions.

Of 96 designs tested, four were active. The best was confirmed by removing each catalytic residue in turn, which caused complete or substantial loss of activity, suggesting the designed site was doing the work. A second design round produced further active enzymes built on several different scaffolds.

The second target was transthyretin, a protein whose misfolding and aggregation are associated with systemic amyloidosis. More than half of 96 designed proteases showed activity against the chosen peptide, and the best cut most of it within an hour, as confirmed by gel electrophoresis and mass spectrometry. It also strongly preferred its intended target over six other sequences from proteins involved in amyloid diseases.

“These results are a considerable advance over protease design with RFD2,” the authors note. They described the work as “a major advance in enzyme design that will enable the computational design of more sophisticated biocatalysts”.

Limitations

Regarding limitations, the authors pointed out that the best protease worked at the “low end” of the range seen in natural proteases, and the designed phosphotriesterases appeared to be only partly carbamylated, a chemical modification of a lysine residue that they need for full activity. Predicted structures for the RNA aptamers also varied widely between samples.

Even so, the authors concluded: “We anticipate RFD3 will be broadly useful for a wide range of biomolecule design challenges and applications.”

Reference:
Butcher, J., Krishna, R., Mitra, R., et al. (2026). Atomic functional site scaffolding in biopolymers with RFdiffusion3. Science. https://doi.org/10.1126/science.aeh2982