Can AI create unbreakable proteins?
Protein design has evolved rapidly, with AI-based tools getting Nobel prizes and opening new research possibilities. So, how far can we push proteins? Can we create proteins that survive even at 150°C?!
Well, read on 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 Hardens Proteins
Scientists created ultrastable proteins using a combination of AI-based tools and molecular dynamics simulations. Image credits: Nature.
Proteins: Versatile Molecular Machines
Proteins do everything!
They catalyze reactions, start immune responses, transport nutrients, send chemical messages, and do a thousand more jobs! They truly are the workhorses of biology. And they also build all sorts of biomaterials!
Elastin gives tissues like skin and blood vessels flexibility. Adhesive proteins stick to surfaces with incredible strength. One example? Mussels! Their foot protein is even investigated as an industrial adhesive.
And proteins also create strong materials.
Collagen mixes with calcium to form our bones. Keratin builds hair and horns (!). Silk is woven into clothes. And spider silk? Weight for weight, it’s stronger than steel! Proteins create some of the strongest materials on Earth.
What’s the secret?
Hydrogen Bonds: The Secret to Strength
The answer hides at the molecular scale.
Okay, not really hiding. It’s hydrogen bonds.
Hydrogen bonds, combined with secondary structures, give proteins the ability to resist heat, force, and all kinds of stress. One of the strongest and most useful structures is the β-sheet, where protein strands are held together by hydrogen bonds into a twisted sheet-like architecture.
And the strongest proteins use a lot of hydrogen bond networks!
This β-sheet-based architecture gives incredible stability under stress. You find it in the toughest proteins:
Spider silk
Bacterial adhesion proteins
Titin, the giant protein in your muscles!
How does it work? In these proteins, hydrogen bonds are arranged so that when you apply force, you must break all of them simultaneously. This makes the structure much harder to unfold!
Think of it like this: one rubber band is easy to stretch or snap. But stack five or ten together, and now the whole thing is almost impossible to pull apart! It’s a similar idea!
This idea is not new, by the way. Scientists have known these principles for a long time. But if you want to use it to produce stronger proteins, you can only go so far with classical protein engineering.
AI to The Rescue
But protein design is now a different place, with AI-based tools.
And in today’s paper, the team used progress in AI-based protein design to generate stronger and more stable proteins! How? By combining AI-based rational design, molecular dynamics (MD) simulations, and experimental validation!
The core design principle is very straightforward: if you pack more hydrogen bonds along the force-bearing β strands, the protein should become much harder to unfold.
The results? Proteins with up to 6 times more hydrogen bonds, that were up to 4 times stronger than the starting template! And they could even survive autoclaving at over 120°C. Incredible!
The Design Pipeline
The starting point is the titin I27 domain, a classic mechanically stable protein. This domain has 4 hydrogen bonds, and it can resist an unfolding force of up to 200 pN! This is super strong for a protein.
They used a 3-step AI-based pipeline to improve its stability.
De novo structure and sequence generation
Starting from the I27 template, RFdiffusion generates many backbone candidates that preserve the motif but extend the hydrogen-bond-forming β strands. RFdiffusion is an AI-based model that denoises random noise into coherent protein structures around a structural motif. Rfdiffusion created over 1,000 different protein designs! About half of them are bad and end up in the bin. The good ones are passed to ProteinMPNN on the remaining designs. This is another AI-based model that can design amino acid sequences that will fold into specific geometries. This gives around 400 sequences per structure: a crazy total of ~200,000 sequences!
Structure validation
Validating 200,000 sequences is no easy task, and it can be slow. So, the team took a 2-step approach. First ESMFold (you guessed, another ML-based model) rapidly predicted the structure for the 200,000 sequences. After filtering, the top 1,000 structures were re-predicted using AlphaFold2 (more precise but slower). This dual-model balances speed and accuracy!
Screening physical properties with MD
The top 200 candidates are analysed for their mechanical properties using MD simulations. Two types of simulations:
steered molecular dynamics (SMD) for unfolding force.
annealing MD for thermal stability.
A super cool example of modern, in silico protein-design loop: AI for generation, structure predictors for filtering, and physics-based MD to screen properties.
What They Made: the SuperMyo Series
The proteins from the pipeline are called SuperMyo (because they are inspired by muscle proteins). The authors iteratively designed 6 series, A → F, each time with longer force-bearing β strands and more hydrogen bonds!
Across the series, the number of hydrogen bonds increases from 8 in the A series all the way to 33 in the F series! From a starting 4 in I27. While the production yield decreased, they could see a constant increase in mechanical strength!
The authors tested the designs with AFM-based single-molecule force spectroscopy, carefully stretching the proteins in a defined geometry. And the results are great:
I27: 4 hydrogen bonds, 246 ± 2 pN unfolding force.
A339: 8 H-bonds, 352 ± 1 pN. A 40% increase!
F553: 33 H-bonds, 1,050 ± 6 pN. The best performing one, an insane 4x improvement!
Mechanical stability is only half the story.
Do the added bonds affect thermal stability?
The team tested SuperMyo proteins in brutal ways! They run MD simulations from 0 K to 413 K to check whether the structures stay intact. For the best ones, they measured melting temperature, which turned out to be well over 90°C! While I27 unfolded >65°C.
Since the proteins survived, they went even harsher:
Repeated 121 °C autoclave-like heating/freezing cycles.
Direct heating to 150 °C for 1 hour!
While I27 precipitates, SuperMyo proteins stay soluble and clear, and they even maintain their AFM mechanical properties!
Hydrogels: From Nanoscale to Macroscale
Finally, the authors asked: Can we use these proteins?
So, they converted them into protein hydrogels using SpyTag–SpyCatcher crosslinking. They make gels from I27, A339, and B42 (the highest yielding proteins) and compared their behaviour under heat.
The results speak for themselves:
The I27 hydrogel precipitates and denatures around 80°C
A339 and B42 hydrogels remain transparent and gel-like even after heating to 121°C!
This means that the single-molecule stability goes from the nanoscale all the way up to the macroscale. They’re not just creating stronger proteins; they’re designing extremely tough biomaterials!
High-Throughput Protein Design
An interesting paper!
They created a versatile, high-throughput pipeline integrating many generative AI protein design tools (RFdiffusion, ProteinMPNN, and AF2/ESMFold). Combined with MD simulations, it creates a powerful screening tool!
It’s not limited to strong proteins. The generative AI part can be used for any protein campaign, and MD simulations are very versatile! For example, they could be used to screen for protein binders.
And these SuperMyo proteins work! They are stronger and more stable. This could be awesome for applications like wound healing or other applications in medicine. These often require sterilization, and autoclaving is the go-to method.
These proteins survive it, which is amazing!
Now, there are some limitations:
The pipeline is computationally intensive
SuperMyos are not completely de novo, because they’re based on I27
The yield of more complex SuperMyo was as low as 0.1 mg/L; very hard to produce at scale!
But still, a cool paper! Go read it here.
If you made it this far, thank you! What do you think of superstable proteins? And what about AI-based protein design? Reply and let me know!
P.S: Know someone interested in AI-based protein design and SynBio? Share it with them!
What did you think of today's newsletter?
More Room:
Optimizing DNA Origami Crystals: DNA origami crystals are cool to look at. They’re also hard to produce! DNA origami frames can self-assemble into ordered 3D nanostructures, but this requires slow thermal annealing over days. In this study, the authors identify key factors controlling assembly and show that optimized conditions can speed up crystal formation by nearly 100×. They demonstrate that assembly follows classical nucleation and growth behavior, allowing prediction and control of structure formation.
Switching Polymerases On and Off: DNA polymerases are at the base of biology and biotech. Controlling DNA polymerase activity improves amplification specificity, but precise regulation is hard. In this study, the authors develop a light-activated system using photocaged primers that block polymerase activity until UV exposure. This enables on-demand isothermal amplification with improved temporal control and specificity. Applied to diagnostics, the method achieves rapid and highly accurate detection of influenza A, demonstrating its potential for molecular testing and microfluidic applications.
DNA-Protein Hybrid Crystals: What’s better than DNA crystals? DNA-protein crystals! Protein crystallization is often unpredictable, limiting its use in materials design. In this study, the authors attach DNA shells to protein oligomers, enabling programmable assembly using DNA-guided interactions. By tuning DNA sequences, they create well-defined crystal lattices with improved crystallinity and controlled growth. This approach brings predictability and designability to protein crystals, opening new possibilities for programmable biomaterials.