RFdiffusion3 designs proteins and nucleic acids atom by atom with striking precision

A new generative artificial intelligence system can design proteins and nucleic acids atom by atom, giving scientists a level of control over molecular function that previous protein-design tools could not reach. The system, called RFdiffusion3, or RFD3, was described in a peer-reviewed study published in Science, and its developers say it can build complex functional sites, design RNA molecules, and create enzymes that coordinate multiple catalytic components from scratch.

Generative deep learning has transformed protein engineering over the past several years, making it increasingly routine to design proteins with specific functions from a blank slate. But most existing approaches operate at the level of amino acids, the building blocks that make up protein chains. That resolution is sufficient for shaping overall protein structures and engineering interactions between proteins, yet it falls short when researchers need to precisely arrange the individual atoms that form an enzyme’s active site or another functional region. The difference between a working catalyst and a failed design can come down to the exact positioning of a few atoms.

Newer methods have introduced atom-level modeling, but they have struggled with a different problem: designing detailed interactions between proteins and nonprotein molecules. Small-molecule ligands, nucleic acids, and catalytic groups such as metal ions all play critical roles in biology, and current atom-level tools have remained limited in their ability to create proteins, and particularly functional RNA and DNA molecules, with precisely engineered active sites that engage these partners.

Jasper Butcher and colleagues developed RFD3 to close that gap. The generative AI system designs proteins and nucleic acids by modeling individual atoms directly, which enables precise control over the molecular interactions that give biomolecules their functions. Rather than treating a protein as a chain of amino acid identities, the model reasons about the full three-dimensional arrangement of atoms, allowing it to scaffold the chemical groups that carry out catalysis and binding with atomic precision.

The system was trained on a broad collection of biological structures, giving it a rich statistical understanding of how atoms are organized in real biomolecules. According to the researchers, this training allows RFD3 to construct complex functional sites, design RNA molecules, and create enzymes that coordinate multiple catalytic components, tasks that combine geometry, chemistry, and sequence in ways that have challenged earlier generative models.

In computational tests, Butcher and colleagues showed that the model outperforms previous approaches in designing protein structures and organizing catalytic sites. The benchmarks evaluated how well the system could produce foldable protein backbones and correctly position the chemical groups that define a functional site, two capabilities that sit at the heart of molecular engineering. The results indicate that atom-level generative modeling can surpass amino acid-level design when the goal is precise functional chemistry.

The computational results were backed by laboratory experiments. The team demonstrated that RFD3 can produce functional enzymes from scratch, including an enzyme that breaks down paraoxon, a toxic organophosphate compound, and a protease engineered to cleave a peptide associated with transthyretin amyloidosis, a disease in which misfolded transthyretin protein accumulates as damaging deposits in tissues. Designing enzymes that perform specific, medically relevant chemistry is among the most demanding tests of any protein-design system, and the fact that these designs functioned in the lab suggests the model’s atomic-level reasoning translates into real molecular behavior.

The ability to design functional RNA molecules is a notable extension of the technology. RNA structures are notoriously difficult to model and engineer because their folding rules differ from those of proteins and their functional sites often depend on precise arrangements of nucleotide bases and metal ions. By modeling atoms directly, RFD3 can in principle address RNA design with the same framework it applies to proteins, opening a path toward custom ribozymes, RNA sensors, and other nucleic acid machines.

The work arrives at a moment when AI-driven biomolecular design is moving rapidly from demonstration to application. Earlier generative systems enabled the creation of novel protein folds and binders, but the frontier has shifted toward the fine chemical detail that determines function: the geometry of a catalytic triad, the coordination sphere of a metal cofactor, the contacts between a protein and a drug-like ligand. RFD3’s atom-level approach speaks directly to that frontier, and the published results suggest it can organize catalytic sites more reliably than its predecessors.

Potential applications span medicine, industry, and research. Enzymes that degrade toxic organophosphates could inform detoxification strategies, while proteases tailored to disease-associated peptides illustrate how designed catalysts might be tuned to specific biomedical targets. Because the same system handles proteins and nucleic acids, researchers anticipate uses ranging from engineered therapeutics and biosensors to basic investigations of how molecular function emerges from atomic structure. The study, published in Science, marks a step toward engineering biomolecules whose functions are specified not just at the level of sequence, but at the level of the atoms that do the work.

Subject of Research: Atom-level generative AI design of proteins and nucleic acids with engineered functional sites

Article Title: Introducing RFdiffusion3 – an AI tool for designing biomolecules with atomic precision

Article References: Introducing RFdiffusion3 – an AI tool for designing biomolecules with atomic precision. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: RFdiffusion3, protein design, generative AI, enzyme engineering, nucleic acids, RNA design, active sites, catalysis, deep learning, structural biology, Science journal, molecular engineering