- Plenty of Room
- Posts
- AI Macrocycle Design: Revolutionizing Drug Discovery with Machine Learning!
AI Macrocycle Design: Revolutionizing Drug Discovery with Machine Learning!
How AI-designed peptides could reshape drug discovery.
Finding new drugs is always a challenge. And even more when it comes to macrocycles, the “Goldilocks” molecules that could be the new best thing in medicine! But what if it suddenly became easier? Yes yes, there is AI in there!
Such a cool paper today! And one that was suggested by a reader. So, shout out to Kyriaki for the suggestion! And if you also have a paper that could be worth sharing, don’t hesitate to send it to me!
Do you want to help us grow? Share this issue today!
Was this email forwarded to you? Subscribe here!
Let’s dive right in.
AI Builds Macrocycles from Scratch

Researchers used AI to create new macrocycle binders for proteins, included RbtA, a recently discovered protein. The binder can be seen in purple. Image credits: Nature.
Macrocycles: The “Goldilocks” Molecules
Most modern drugs fall into one of two camps:
Small molecules: around 90% of all drugs. They can enter cells to find proteins, but they don’t work well on targets with deep hydrophobic pockets
Biologics: mostly antibodies. They have amazing precision, but they are too big and polar to enter cells, so they are limited to extracellular targets
But there is a new, promising class of molecules sitting right in the middle: macrocyclic peptides, or just macrocycles. These are peptides whose ends are chemically linked, forming one or more rings. This circular structure enhances their stability, binding affinity, and cell permeability. And their medium size is the perfect sweet spot between small molecules and antibodies!
Macrocycles Play Hard to Get
So, macrocycles could have awesome applications in therapeutics and diagnostics. But finding them is hard. There are two main methods:
Natural product discovery: looking into plants, bacteria or fungi. It works, but it’s hard, slow and unpredictable!
High-throughput screening: screening trillions (!) of compounds. It’s powerful, but also resource-intensive ($$$).
Plus, these approaches don’t give you control over other important drug properties, such as target binding, selectivity, and membrane permeability. To optimize for these properties, one needs precise structural control (at the atomic level)!
One alternative to traditional methods is structure-guided design.
It uses computational methods to identify binders for target proteins, exploring diverse chemical and structural macrocycle designs. But this type of design is limited to well-studied structures and it lacks control at the atomic level!
So, computational methods are needed to design high-affinity macrocycle binders,which could create amazing innovation in diagnostics and therapeutics!
RFpeptides: AI-designed Macrocycle Binders
And this is where today’s paper comes in. The authors created RFpeptides, a deep-learning-based pipeline for designing protein-binding macrocycles from scratch.
They focused on diffusion models, which have already shown their strength in creating other proteins: small molecule binders, receptor ligands, and even antivenoms! For proteins, diffusion models start with a “cloud” of amino acids, turning it more and more into a real molecule with each iteration.
So, the team set out to expand two powerful tools to create macrocycle binders:
RF2 (RoseTTAFold2): A structure prediction model, modified to handle macrocycles, with their ends linked together.
RFdiffusion: The generative diffusion model in charge of creating peptide backbone structures.
Once they had their tools aligned, the new RFpeptides workflow looked like this:
Backbone Generation
RFdiffusion generates thousands (between 8,000 and 20,000!) diverse macrocycle backbones for a target protein.Sequence Design
ProteinMPNN designs sequences compatible with the generated backbones, optimized for solubility and structural compatibilityFiltering
The predicted models are first evaluated using a specific AlphaFold2-based tool for cyclic peptides, and then Rosetta evaluates additional physics-based metrics.Experimental Validation
The best candidates are synthesized and tested for binding to the target and structural agreement with the designed models.
And the pipeline worked great!
Working on Known Structures
They created binders for different therapeutic targets with experimentally validated structures:
Myeloid Cell Leukemia 1 (MCL1)
A promising target for anticancer therapeutics, with high-resolution structures available. Out of the 27 molecules synthetized, 3 showed micromolar binding affinity for MCL1! And the X-ray structures were nearly identical to the designed model.
MDM2
Another promising target for cancer therapy, MDM2 interacts with the infamous p53 (involved in everything cancer-related). Out of 11 top-ranked designs, 3 bound MDM2 with micromolar affinity! A very good hit rate.
GABARAP
This protein has one of the longest names I’ve ever seen: Gamma-aminobutyric acid receptor-associated protein.And apart from this, the protein is cool because it’s involved in autophagy, and modulators against it could be used to treat late-stage cancers. So, it’s an important target. And 2 of the macrocycles designed reached nanomolar affinity, making them the most potent macrocycle binders ever!
So, the team showed that their pipeline can create de-novo macrocycle binders with high affinity against important targets! But they were not done yet.
Designing Against Predicted Structures
All these targets had something in common: an experimentally validated structure.
But this is not the case for most proteins. It’s hard to know, but we probably have experimental structures for less than 0.1% of the proteins out there! But in the last few years, protein-prediction tools have gotten better, with help from public databases and machine learning. Just think about AlphaFold!
And the team used these new tools to push RFpeptides to its limits. They thought that the accuracy of RFpeptides could mitigate the risks of designing against a predicted, uncertain structure. And they were right!
They got to work on Rhombotarget A (RbtA), a recently discovered surface protein from an antibiotic-resistant bacterium, with no known structure. The team predicted its structure using 2 different tools, and then RFpeptides generated binders for the regions that were nearly identical in both predictions (smart).
The final result was that 3 binders were good enough to be characterized, and one of them was in the low nanomolar range! The team also generated the first structures for these proteins, with and without the binders, confirming the predicted structure!
Future Directions: Advantages And Final Thoughts
So, the authors created a powerful new way to design macrocycle binders. RFpeptides has several advantages over traditional methods:
It’s faster and more efficient: they only tested <20 designs per target, vs trillions in high-throughput approaches. They still found high-affinity binders, with a much higher success rate! And it could be combined with high-throughput approaches to get the best of both worlds, since the yield is still somewhat low
It’s atomically accurate, which enables further changes to the chemical structure, for example, to improve affinity, cell permeability or bioavailability
Lastly, and this is crazy, RFpeptides can start from just a structure or a sequence!
RFpeptides has, of course, great potential in therapeutics or diagnostics. But I can also see uses in synthetic biology, for example, to create new binders to activate proteins on demand!
This was a cool paper! Don’t hesitate and read it for yourself here!
If you made it this far, thank you! What do you think of this wave of AI-powered science tools? Are you excited or sceptical? Reply and let me know!
P.S: Know someone interested in AI in biology? Share this with them!
More Room:
Lighting Up DNA Origami: Fluorescent molecules on DNA origami are nothing new, but what about light-up RNA aptamers? This study shows that scaffolding RNA-based fluorescent light-up aptamers onto DNA nanostructures improves their fluorescence. Using the Broccoli aptamer as a model, the researchers demonstrated that DNA structures, from simple hairpins to DNA origami, increase fluorescence, with improvements linked to DNA stem length and stability. It looks useful!
DNA… NanoDonuts? If you’re hungry, look away. Or maybe not. This study shows that the self-assembly of DNA amphiphiles into complex structures can be precisely controlled by the DNA sequence, independent of base-pairing. Small sequence changes trigger the formation of unique non-equilibrium DNA nanotoroids (donuts for us mortals!) instead of typical structures like spheres or fibers. These nanotoroids can be stabilized using small molecule cross-linkers or co-assembled with other DNA amphiphiles. The work demonstrates protein-like sequence control in DNA-based materials, which is pretty cool!
DNA, Origami and Cell Receptors: DNA origami is very good at putting things precisely at the nanoscale. And the most common use is for surface proteins. This article highlights how DNA origami nanostructures are revolutionizing the study of cell signaling by enabling precise control of receptor-ligand interactions. While receptor clustering is key to efficient signal transduction, its regulation has been difficult to study. DONs now allow nanoscale manipulation of these interactions, providing new insights into signaling mechanisms and offering potential for developing targeted molecular therapies across various receptor classes.
Know someone who’d love this?
Pass it on! Sharing is the easiest way to support the newsletter and spark new ideas in your circle.Got a tip, paper, or topic you want me to cover?
I’d love to hear from you! Just reply to this email or reach out on social.