AI and protein design are everywhere today.
From conferences, to the news, to every other paper. I’m fully expecting my grandma to ask me about it one of these days! RFdiffusion is one of the main protagonists, and RFdiffusion2 promises things will get even better!
Can it tackle metallohydrolases, one of the toughest enzymes out there? Read to find out!
Don’t keep this newsletter a secret: Forward it to a friend today!
Was this email forwarded to you? Subscribe here!
AI Builds Enzymes
Researchers used the AI-based tool RFdiffusion2 to create completely new enzymes to catalyse metal-based reactions.
Enzymes: Catalyzing Life
Life runs on enzymes.
That’s not an exaggeration. Without these biological catalysts, reactions would take a lifetime. Instead, they happen in milliseconds! They’re behind everything, from digestion in your stomach to ATP synthesis and DNA repair.
They also power biotech: DNA polymerases, ligases, and CRISPR-Cas systems. Enzymes really come in all shapes and functions! But not all reactions are equal, and some enzymes have to work extra hard to do their job.
Hydrolases are a good example.
These enzymes use water to break chemical bonds. This is one of the most common reactions out there! Super important for the human body; for example, they power the digestion of proteins, fats, and sugars.
But not all hydrolases have such an easy life!
Metallohydrolases catalyse some of the hardest hydrolysis reactions in biology. They use bound metal ions to activate a water molecule next to the bond that needs to be broken, and zap! The bond is cut, and the molecule is broken.
This chemistry matters in many places:
Urea breakdown in the body.
Bone metabolism.
Pretty important!
And there’s more. If we could engineer new metallohydrolases, we could remove many of the pollutants humans have introduced into nature, like plastics or PFAS (the infamous forever chemicals!).
Natural enzymes can’t deal with these substances!
Engineering New Enzymes
Of course, scientists couldn’t resist such an interesting idea!
They have been trying to make artificial metallohydrolases for a long time, and they have had real success. Using good old protein engineering, they created enzymes that can act on new substrates, including nasty ones like chemical warfare agents.
The main strategies are:
Directed evolution: You introduce random mutations into a gene and screen the resulting proteins, then repeat the process.
Rational design: Uses structural and mechanistic insights to understand how a protein works → you make specific mutations to improve performance.
De novo design: Designs enzymes from scratch.
And it worked! But these new metallohydrolases have some big limitations:
Low activity and efficiency → need many rounds of refinement.
Promiscuous activity → makes optimization harder.
Can we do better?
Using AI To Design Better Enzymes
That’s exactly where today’s paper comes in!
The team used the new AI tool RFdiffusion2 to design active de novo zinc metallohydrolases. Directly from quantum-chemistry-derived active-site geometries! Without specifying where the catalytic residues should be in the sequence.
That’s the key idea.
They started with the fluorogenic substrate 4MU-PA and used quantum chemistry calculations to define the catalytic residues that make the reaction happen. Then, they let RFdiffusion2 generate protein backbones around those motifs!
The result?
In two rounds of 96 experimental designs each, they obtained enzymes with catalytic efficiencies up to 53,000 M⁻¹ s⁻¹, comparable to native enzymes! And all without any experimental optimization.
Let's look at their pipeline in more depth.
RfDiffusion2, Chai, and PLACER: A Winning Combo
I've actually written about Rfdiffusion2 before! But the short version is this.
Rfdiffusion2 is the successor to RFdiffusion, a generative model that creates proteins from scratch. Using a process called “diffusion”, it starts from a catalytic motif and builds a scaffold around it, progressively refining noise.
RFdiffusion is powerful, but it has a big limitation. It requires you to specify the exact sequence position of the catalytic residues. That limits the design space to enzymes for which you have this information, and it makes the process computationally harder.
RFdiffusion2 solves this in 2 ways:
It can scaffold amino acid motifs without knowing their sequence position in advance.
It can scaffold atom-level motifs, not just backbone motifs!
This results in much better enzyme predictions.
So, the pipeline from computer → lab is:
Quantum chemistry finds the active site.
RFdiffusion2 generates a protein scaffold around it.
ProteinMPNN designed the amino acid sequences.
AlphaFold2 and LigandMPNN were used to filter and optimize them.
PLACER and Chai-1 assess active-site geometry and organization.
The best designs go to the lab!
It’s a long pipeline, but that’s what you need to go from nothing to a working enzyme.
Bringing AI-Designed Enzymes into the Lab
Okay, how well did it actually work?
The team ran 2 design campaigns.
First design campaign: ZETA_1
The authors generated 5,120 (!) RFdiffusion2 designs and filtered them down to the 96 best candidates for experimental testing. Then they went to the lab:
86/96 designs are expressed and soluble.
5 show clear activity.
The best design, named ZETA_1, had a Kcat/KM of 16,000 M⁻¹ s⁻¹!
For comparison, previously designed metallohydrolases ranged from 3 to 60 M⁻¹ s⁻¹: an incredible improvement!
Second design campaign: ZETA_2, ZETA_3, ZETA_4
The second design campaign was even more successful! The team analysed ZETA_1 using computational and structural biology and used those learnings to improve their results.
This time:
85/96 designs are expressed and soluble.
11 have clear activity.
The best designs are ZETA_2, ZETA_3, and ZETA_4.
Their catalytic efficiencies were:
ZETA_2: 53,000 ± 5,000 M⁻¹ s⁻¹
ZETA_3: 19,000 ± 2,000 M⁻¹ s⁻¹
ZETA_4: 1,100 ± 200 M⁻¹ s⁻¹
Incredible! Especially that ZETA_3.
Is AI Coming for Enzymes?
This is such cool stuff!
It's incredible to see protein design moving so fast. I can’t keep up with it! But it's crazy what people are doing, and it's amazing to see the innovation! Outside and inside the lab.
This paper introduced a better AI design tool, RFdiffusion2:
It has a much higher hit ratio than before.
It's easier to start with -> you don't need the specific sequences, "only" the catalytic site.
It creates novel structures: the team checked and found no natural matches!
It still has limitations, of course:
Need for expert knowledge and intuition
Heavy computational pipeline
Enzymes are still not perfect: many designs aren't active, efficiency is still below that of natural enzymes, and the mechanism is still not dialed in
But it's an incredible advance! I can't wait to see what people do with it in biomedicine, bioremediation, and industrial biotech! Such an exciting time.
So, go here and read the paper; it's worth it! It's even a short one, this time.
If you made it this far, thank you! What do you think of RNA nanostructures? Do you think they have a place in biomedicine? Reply and let me know!
P.S: Know someone interested in RNA nanotech and SynBio? Share it with them!
What did you think of today's newsletter?
More Room:
3D Electronics for Everyone: The review I’ve been waiting for! Time for DNA origami to enter the clean room? I think so. This review summarizes recent advances in molecular electronics, focusing on the design and fabrication of reliable single-molecule devices such as switches and transistors. It highlights new strategies that combine atomic-scale manufacturing with 3D integration to improve device stability, scalability, and reproducibility, paving the way toward high-density molecular computing beyond the limits of conventional semiconductor technology. You know what I’m reading this week!
Little Critters, Cool Biology: I’m always amazed by biology: I guess it’s how I got into this to begin with, after all! In this paper, the authors use molecular and structural biology to identify two key proteins that control conoid assembly in Toxoplasma gondii. ASAF1 initiates formation of the conoid complex during cell division, while CGP stabilizes the mature structure by anchoring essential components. These findings reveal how this invasion-critical organelle is assembled and maintained.
Enzymes + DNA = Profits: Got inspired by today’s paper, but you work in DNA nanotech? Worry not, you can also use enzymes! This review examines how enzymes enhance nucleic acid nanotechnology, enabling dynamic, programmable DNA and RNA nanostructures with functions beyond enzyme-free systems. It highlights enzymatic strategies for assembly, reconfiguration, and signal processing, as well as applications in diagnostics, imaging, drug delivery, DNA computing, and data storage. Run and order some!