Proceedings · Session S-138 · filed November 19, 2026
AI & Emerging Tech in R&DSession paper
Baker Lab and NVIDIA Model Designs Paraaxon-Degrading Enzymes at 89/288 Hit Rate
89 of 288 designs from RFD4-Proteina degraded paraoxon; 192 RFD3 designs showed zero activity. Six new enzymes hit catalytic efficiencies 10-1,000x higher, after four days of compute.
By Tom Whitfield4 min read732 words
Summary
- 89 of 288 RFD4-Proteina enzyme designs degraded paraoxon in lab tests; 0 of 192 RFD3 designs did
- Six new enzymes showed catalytic efficiencies 10 to 1,000 times higher than earlier six-residue designs
- The computational design work took four days on the merged RFdiffusion-Proteina model
- Lead author Kieran Didi worked on-site with NVIDIA's Seattle team for more than a year
- The enzymes remain lab-validated only — no field deployment yet, per NVIDIA's Anthony Costa

Eighty-nine of 288 enzyme designs from a new AI model broke down the pesticide paraoxon in laboratory tests — while none of the 192 designs from the previous-generation model showed any activity on the same task. That result, from David Baker's lab at the University of Washington and NVIDIA, marks a step change in computational enzyme design, and it took four days of compute.
The numbers get better under closer inspection. Researchers examined six of the new enzymes in detail and measured catalytic efficiencies 10 to 1,000 times higher than the earlier six-residue designs. Kieran Didi, the study's lead author and a researcher at UW's Institute for Protein Design, described the target in a university video: enzymes for "things that nature never had the time or chance to develop against, such as degrading man-made pesticides."
What changed between RFD3 and RFD4-Proteina?
The problem was packing. Baker's team had already used a model called RFD3 to design paraoxon-degrading enzymes with six amino acids at the reaction site. They hypothesized that adding two more residues around a metal ion would speed up the reaction. Positioning all eight correctly was the hard part.
"That's a packing problem. It's a sterics problem. It's an electrostatics problem," said Anthony Costa, NVIDIA's director of digital biology, in an interview with R&D World.
The collaboration merged two model lines: Baker's RFdiffusion and NVIDIA's Proteina. Costa said NVIDIA contributed "a lot of the details about how the model reasons" from an architecture standpoint. The resulting RFD4-Proteina model generated multiple physically plausible ways to arrange residues around the metal ion, which the team then assayed in the lab.
How does the workflow differ from the old approach?
Before this, the process was slow and manual. "You simply hypothesize some structures, you write them down, you give them to a molecular dynamics calculation or to a quantum chemistry calculation," Costa said. Those methods are computationally intense and operator-heavy. In his words: "You're guessing with a lot of knowledge, but you're guessing."
The new workflow splits labor between human and model. The chemist supplies mechanistic insight — for example, that more close contacts with the metal ion should increase reaction rate. "That's insight that we ask the human being to actually reason about," Costa said. The model then works out the geometry.
The model can't take plain English directly. Instead, an agent translates natural-language specifications into model input and configuration. "You can literally specify these things in plain English. Have an agent construct the input and the configuration of the model, even download the model, set it up, put it on a GPU somewhere," Costa said.
The reasoning loop is familiar to anyone using commercial AI tools. "You go to ChatGPT and you pick low reasoning or medium reasoning or high reasoning, and we literally have those buttons in these models," Costa said. The model checks its own output, decides "Is this right?" and iterates with slight modifications.
The collaboration itself ran deep on personnel. Didi, the lead author, relocated to Seattle and worked on-site with the NVIDIA team for more than a year.
What are the caveats?
Sample size matters here. The headline 89/288 figure comes from one assay on one target, paraoxon. The detailed kinetic characterization covered only six enzymes. And the lab result is not a field result.
"We make these things in the lab and we test how they actually work. The lab is an isolated environment in which we get to see if we're right about the science, but we haven't put this thing in the field yet," Costa said.
For R&D managers, the practical takeaway is the compression of design cycles: a four-day computational run replaced a hypothesis-and-simulate loop that previously depended on slow molecular dynamics and quantum chemistry calculations. For portfolio decisions, the 10x-to-1,000x catalytic efficiency gain over prior designs suggests AI-designed enzymes are crossing from proof-of-concept toward measurable performance.
As agents take on more of the workflow, Costa argues the scientist's judgment shifts upstream: knowing which scientific question to ask, and standing up the right workflow to answer it. Whether RFD4-Proteina's lab success survives field deployment is the next test for the Baker-NVIDIA partnership.
via youtube.com (Original)
Filed under
- ai-protein-design
- enzyme-design
- rfd4-proteina
- paraoxon-degradation
- digital-biology
More from Tom Whitfield
Show full bio
Senior reporter covering technology and product news in research and development at Hypothesis Wire.
272 articles
References
- Nvidia and UW Unveil Open-Source Protein Design Model With 10x Throughput
- Quantum Simulation Passes 12,000-Atom Mark; Lab Filters Flagged
- Peptide Screens Should Build Knowledge, Not Just Hit Lists
- Deep Eutectic Solvents Boost Enzyme Half-Lives Ninefold in Biocatalysis
- Anthropic's Claimed AI Biology 'Discovery' Draws Sharp Pushback