At the intersection of computation and chemistry, a team spanning Washington, Oxford, and NVIDIA has crossed a threshold long imagined but rarely achieved: an artificial intelligence that does not merely predict the shape of life's molecules, but designs them — atom by atom — to perform specific work in the world. RFdiffusion3, led by Nobel laureate David Baker, has produced enzymes that exist, function, and validate a new chapter in humanity's relationship with the molecular machinery of living things. The question science has long carried — can we design biology rather than only observe it —
AI System RFdiffusion3 Designs Functional Proteins at Atomic Level
Designing proteins atom by atom, then watching them work
Mark
So this system designs proteins by moving atoms around. How is that different from what we could do before?
Mimi
Before, we could predict what a protein would look like if it folded a certain way. Now we can say: I want this enzyme to break down this chemical, and here are the constraints it needs to satisfy. The AI builds the protein from the atomic level up to meet those constraints.
Luke
But has anyone independently verified that these lab-tested enzymes actually work? The paper says they performed their intended reactions, but I want to know: who ran those experiments, and are the results published separately?
Mimi
The paper itself is in Science, so the experiments were peer-reviewed as part of that publication. The organophosphate-degrading enzyme and peptide-cleaving enzyme both worked in real chemical reactions.
Mark
Why does it matter that it's ten times faster than the previous version?
Mimi
Speed means you can iterate. You can design more variants, test more ideas, fail faster and learn faster. In protein engineering, that's the difference between a project taking months and taking weeks.
Luke
The comparison was against RFdiffusion2 on forty-one enzyme active sites. That's a specific benchmark. Do we know how it performs on novel protein designs that don't have existing enzymes to compare against?
Mimi
That's a fair question. The paper shows it works on known active sites and on two newly designed enzymes. Whether it generalizes to every possible protein design—that's what the open-source release will help answer.
Mark
Who can actually use this system?
Mimi
It's open-sourced, so any researcher with the computational resources can download it and start designing proteins. That's the whole point of releasing it publicly.
Luke
Computational resources being the operative phrase. This isn't something a small lab runs on a laptop. You need serious hardware.
Mimi
True. But compared to the cost of traditional protein engineering, it's still democratizing the field.
El Pulso
- The gap between predicting protein structures and engineering functional ones has been a stubborn frontier — RFdiffusion3 now crosses it, with lab-tested enzymes performing real chemical reactions as designed.
- The system's constraint-based architecture lets researchers dictate precise atomic positions, hydrogen bond formations, and spatial distributions simultaneously, compressing what once required enormous trial-and-error into a single generative pass.
- Running ten times faster than its predecessor and outperforming it in 37 of 41 benchmark comparisons, RFdiffusion3 doesn't just raise the ceiling — it lowers the cost of reaching it.
- Two engineered enzymes — one that degrades organophosphates, one that cleaves peptides — have already proven themselves in laboratory conditions, moving the achievement from algorithmic promise to physical reality.
- With the system already open-source since December 2025 and its full technical architecture now published in Science, the broader research community inherits both the tool and the blueprint to push further.
At the intersection of computation and chemistry, a team spanning Washington, Oxford, and NVIDIA has crossed a threshold long imagined but rarely achieved: an artificial intelligence that does not merely predict the shape of life's molecules, but designs them — atom by atom — to perform specific work in the world. RFdiffusion3, led by Nobel laureate David Baker, has produced enzymes that exist, function, and validate a new chapter in humanity's relationship with the molecular machinery of living things. The question science has long carried — can we design biology rather than only observe it — has begun to receive a practical answer.
A research team spanning the University of Washington, Oxford, and NVIDIA has published the full technical account of RFdiffusion3, an AI system that designs proteins and nucleic acids by directly manipulating the three-dimensional coordinates of individual atoms. The work carries significance beyond computation: proteins created by the system have already performed their intended chemical functions in real laboratory experiments.
David Baker, whose earlier work on protein structure prediction earned him the 2024 Nobel Prize in Chemistry, led the effort. His previous system, RosettaFold, underpinned two earlier versions of RFdiffusion. The new iteration makes a deeper architectural shift, adopting the framework of AlphaFold3 — a model that represents each amino acid through up to fourteen atoms and generates novel structures by working directly in three-dimensional space.
What distinguishes RFdiffusion3 is its approach to design with constraints. Researchers can specify not only where atoms should sit, but how hydrogen bonds should form and how atoms should distribute across the structure. The system then generates molecules that satisfy all conditions at once. Benchmarked against forty-one enzyme active sites, it outperformed its predecessor in thirty-seven — and does so roughly ten times faster.
Two enzymes have already moved from design to demonstration: an organophosphate-degrading enzyme and a peptide-cleaving enzyme, both performing the reactions they were built for. The system was made publicly available in December 2025; the Science paper published October 16, 2026 now explains the principles behind its performance. For other researchers, the open-source release means the foundation is already in hand — and the shift from predicting molecular structure to engineering molecular function has quietly, consequentially, begun.
A team of researchers spanning the University of Washington, Oxford, and NVIDIA has released the complete technical specifications for RFdiffusion3, an artificial intelligence system that designs proteins and nucleic acids by positioning individual atoms in three-dimensional space. The breakthrough carries weight beyond the laboratory: proteins created using this system have already performed their intended chemical functions in real experiments, marking a shift from computational prediction to functional molecular engineering.
David Baker, the University of Washington professor who won the 2024 Nobel Prize in Chemistry for his work on protein structure prediction, led the research. His earlier system, RosettaFold, became the foundation for two prior versions of RFdiffusion. The new iteration represents a significant architectural leap. Where its predecessors built on RosettaFold's framework, RFdiffusion3 adopts the network structure of AlphaFold3, a more sophisticated model that represents each amino acid using up to fourteen atoms and generates new protein structures by directly manipulating the three-dimensional coordinates of those atoms.
The system's power lies in what researchers call "design with constraints." A scientist can specify not just where particular atoms should be positioned, but also impose conditions on how hydrogen bonds form and how specific atoms distribute themselves within the protein's structure. The AI then generates designs that satisfy all of those requirements simultaneously. When tested against forty-one active sites from existing enzymes, RFdiffusion3 outperformed its predecessor in thirty-seven cases. It also runs roughly ten times faster than the previous generation, a practical advantage that accelerates the design cycle.
Two enzymes designed by the system have already demonstrated their functionality in laboratory conditions. An organophosphate-degrading enzyme and a peptide-cleaving enzyme both performed the chemical reactions they were engineered to perform. This is not a theoretical achievement. These are molecules that exist, that were designed by an algorithm, and that work as intended in the real world.
The research was published in Science on October 16, 2026. The team had already made RFdiffusion3 publicly available on December 3, 2025, but this paper provides the detailed explanation of how the system actually works—the principles underlying its design choices and its performance advantages. The open-source release means other researchers can now build on this foundation, testing the system's capabilities and pushing protein engineering toward even greater precision. The shift from predicting what proteins look like to designing them for specific functions represents a fundamental change in how molecular engineering can proceed.
Citas Notables
Researchers can specify not only where required atoms should be positioned but also set conditions for hydrogen bond formation and spatial distribution of specific atoms within the protein, and the system generates structures satisfying all constraints— Study findings on RFdiffusion3's constraint-based design capability
Quieres la nota completa?
Lee el original en KuCoin ↗
Technology