Can you create biosensors using small and stable ML-based proteins?
You would say no. You need some big conformational change to make them work! Well, today’s paper challenges that. Welcome to a world of small, easy-to-design and produce biosensors!
An exciting new paper today!
Don’t keep this newsletter a secret: Forward it to a friend today!
Was this email forwarded to you? Subscribe here!
AI Builds Protein Switches
Researchers created allosteric biosensors using machine learning and AI-designed proteins, opening the way for new diagnostic modalities. Image credits: Nature.
Inside cells, proteins do almost everything.
Most cellular processes are controlled by protein networks composed of protein switches. They act as molecular “on/off” toggles, controlling signal transduction, gene expression, energy processing, and other essential biological tasks!
What makes them tick? Allostery.
Protein allostery is a classic concept. In its simplest form, it works like this:
A ligand binds to a protein at a site distant from the active site
The protein changes shape
These changes alter activity at the active site
It’s super cool, and everywhere! From hemoglobin to GPCRs and the Lac repressor! It’s simple in principle, but it can be very powerful in practice.
Artificial Protein Biosensors
Since allostery is so powerful, scientists naturally wanted to use it.
And they did! Artificial protein switches are widely used as biosensors in cell biology, neuroscience, and molecular biology. They’re used to measure all kinds of things inside cells: ions, small molecules, even other proteins!
Most of these biosensors have two parts:
A ligand-binding domain, providing specificity and affinity.
A reporter, such as fluorescent proteins or enzymes.
The dream? Build sensors for lots of biomarkers, make them specific and easy to design, and use them for diagnostics and biological control!
But there aren’t a lot of natural ligand-binding domains that show clear allosteric regulation, with large changes in the protein conformation. Not every ligand-binding protein does that.
Machine learning or AI-designed artificial binders are an alternative.
These proteins bind to small molecules, peptides, and other proteins. They’re small, simple to design, and single-component: perfect for sensing! But they’re also structurally stable: they bind to a ligand with no conformational change.
Can we still turn them into biosensors?
Turning Binders into Biosensors
Here is where today’s paper comes in!
The team shows that you can create allosteric protein switches, even without large conformational changes. How? Using machine-learning-designed ligand-binding proteins.
They created lots of chimeric biosensors by inserting artificial or natural binders into reporter enzymes such as
β-lactamase,
PQQ-glucose dehydrogenase,
NanoLuc,
LuxSit Pro.
The result?
47 functional protein switches that respond to different ligand classes: small molecules, peptides, and proteins. Not satisfied, they turned these switches into single-component logic gates, modulated antibiotic resistance in E. coli, and created electrochemical sensors!
But let’s take this one step at a time.
How To Build Biosensing Switches
The central idea of the paper is that allostery can emerge without large conformational changes. Instead, the switch can emerge from shifts in conformational entropy, the disorder intrinsic in biomolecules.
But how do you design and select for that?
The team built a 3-step method:
Start with a receptor and a reporter
The main reporter is β-lactamase. The team inserted a ligand-binding domain into the enzyme at selected positions, using circular permutation, so that the binder fits the chimera design. Circular permutation “rewires” the order of amino acids in a protein, without altering the 3D structure. This can give the protein improved catalytic activity, ligand binding, or thermal stability.
Choose binders with minimal shape change
The team first worked with natural ligand-binding domains that show no large global shape change. Then, they moved to machine-learned binders of cortisol and 17α-hydroxyprogesterone (17-OHP), and later to peptide and protein binders.
Screen circularly permuted libraries
For each binder, they built libraries of circularly permuted variants to insert into β-lactamase. The point is not to just identify binders, but sensors: the binding (input) activates enzymatic activity (output).
In this way, they identified 47 working switches!
What Did They Build?
The team demonstrated a wide variety of switches.
Natural binder example
As a proof of concept, they used a natural binder. The anticalin–colchicine domain only has modest local rearrangement after binding, but it can still control β-lactamase activity.
ML-designed steroid binders
They used 2 ML-designed binders:
HCY129.1 for cortisol
OHPFA1952 for 17-OHP
The best biosensor gives a 400-fold dynamic range, with an apparent Kd of ~0.9 μM. The coolest part? The sensors are functional, with only minimal conformational changes!
Peptide and protein binders
The team also tested ML-designed peptide binders, obtaining chimeras with low-micromolar affinity, and for viral and cancer proteins, showing activity here too. The method seems to generalize pretty well!
Alternative reporters
They showed that the synthetic receptors work with other reporters, not only β-lactamase (colorimetric readout). They created steroid-responsive versions of:
PQQ-glucose dehydrogenase (PQQ-GDH) → electrochemical readout
NanoLuc → luminescence
LuxSit Pro → luminescence
This shows that you can get different readouts!
What in the Biophysics is Happening?
The team also explored how these sensors worked.
Sure, they knew there were no big conformational changes, but what made them go from inactive to active after binding their ligands? Their hypothesis revolved around ligand-induced refolding or allosteric entropy changes.
The circular dichroism data showed no large refolding: one option gone!
NMR and mass spectrometry showed lots of small changes. This suggests that the ligand binding stiffens or orders the receptor structure, reducing conformational entropy and changing the activity of the reporter domain.
In other words, the sensors are not completely structurally stable, and the binding of the ligand stabilizes them. At that point, the reporter domain starts being enzymatically active, and you get an output!
Practical Applications of Biosensors
The team showed that these sensors aren’t just curiosities, but have practical applications.
Logic gates: YES and AND
The receptors can be combined into logic gates.
By duplicating the same receptor in one chimera, you make an intramolecular YES gate (the simplest logic gate that confirms the presence of input). Placing 2 17-OHP receptors in one β-lactamase chimera boosts the dynamic range by more than 20-fold!
The team also built an intramolecular AND gate (active only when 2 different inputs are present). They combined 2 different receptors in one chimera, a 17-OHP receptor and a C-peptide receptor. The sensor shows weak activation with either ligand alone, but 5x activity when both ligands are present!
The system is not just a sensor; it can be used for logic processing.
Ligand-dependent antibiotic resistance in E. coli
The researchers expressed the β-lactamase switches in E. coli so that the cells grow on ampicillin only when the right ligand is present → The cells become ligand-dependent in their antibiotic resistance.
So now you have steroids-addicted bacteria! And proof that the sensors work in cells.
Steroid-responsive bioelectrodes
Finally, the team attached a 17-OHP-responsive PQQ-GDH switch to electrodes. The PQQ-GDH reporter has electrons as a side product, so you can create an electrochemical readout!
The electrode current rises in response to 17-OHP concentration, and the system can be reset by simply washing the ligand away. The system even has a sub-nanomolar detection limit!
This is great for point-of-care diagnostics → single-component and electrochemically readable.
Expanding the Biosensing Landscape
Cool work!
The central idea is that you don’t need a big conformational change in a binder to make allosteric sensors. This is important because it increases the number of ligand-binding domains we can use.
Especially now that ML/AI-based designs are getting really good at binding! This will allow the creation of all kinds of sensors. ML binders are useful because they’re small, stable, and easy to produce.
Perfect for diagnostics! But not only: synthetic biology could also use more orthogonal switches. And this could also be the start of easier protein-based computation… Who knows!
The main problem here is that the switches have reduced affinity and sometimes slow kinetics. This could be solved via optimization, but the authors didn’t really try. Fair! But something to keep in mind.
But a super cool work, go here to read all of it!
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 ML-driven protein design and SynBio? Share it with them!
What did you think of today's newsletter?
More Room:
Responsive DNA Nanotech: Dynamic DNA nanotech is super cool, combining the precision and programmability of DNA self-assembly with stimulus-responsive control mechanisms. This review summarizes recent advances in DNA nanostructures that respond to light, pH, temperature, ions, and enzymes, enabling reversible and controllable nanoscale behavior. It also highlights applications in biomedicine and molecular computing, while discussing current challenges and future directions for precision medicine.
Engineering DNA Crystals: DNA crystals could be used to organize biomolecules in precise 3D positions, but we’re not there yet. This paper explores the use of DNA T-junctions as building blocks for such crystals, providing an early proof of concept for using designed DNA frameworks to crystallize difficult biomolecules without relying on their natural interactions. The approach could help study challenging targets such as aptamers, ribozymes, and ribonucleoproteins.
Triangles, Aptamers, and Cancer: Fighting cancer is one of the most common uses for DNA nanotech. In this paper, the authors develop aptamer-guided DNA nanostructures (Apt-tFNAs) for targeted cancer delivery and investigate how their size and aptamer valence affect performance. They show that smaller structures improve tumor targeting while reducing accumulation in the liver and kidneys. An optimized single-aptamer design achieved the best delivery and therapeutic efficacy, highlighting the importance of tuning DNA nanocarrier architecture for precision medicine.