AI Protein Design: Binding Small Molecules and Creating Sensors!

Welcome to Plenty of Room!

Today we are dive deep into AI-based protein design to solve an important challenge: binders for small molecules! An amazing paper from last year, still worth highlighting!

Plenty of Room is your guide to the cutting-edge news in AI-driven protein design, DNA nanotech and more.

If you enjoy this issue, share it with a fellow science enthusiast: good ideas are meant to be shared!

Let’s get into it now.

AI, proteins and small molecules

This week we are jumping into AI-based protein design once again! And this time with an ambitious target: small molecules.

What are small molecules?

Small molecules are low molecular weight compounds that can easily diffuse across cell membranes, unlike proteins, DNA, RNA, etc. (which makes them large molecules, I guess?). Since they are so small, they often bind to receptors or enzymes, playing an enormous role in biology, medicine and more. There are just so many of them:

• Metabolites like glucose or cholesterol

• Hormones like thyroxine

• Neurotransmitters like dopamine

• Drugs: like, 90% of drugs are small molecules, for example aspirin

• Toxins and pollutants, like pesticides

Detecting small molecules is critical in biomedical research, diagnostics and environmental monitoring. But here is the problem: traditional methods like chromatography or mass spectrometry are expensive, slow and require specialized machines. So, there is a need for rapid and accessible alternatives: so why not turn to biology itself?

Small molecule binding proteins

Protein sensors are, on paper, great candidates for small molecule sensing, offering highly specific and real-time detection of these compounds. Unfortunately, creating small molecule binders is much more complicated than creating protein-protein binders:

• Natural proteins often are not optimized to bind to synthetic or non-biological molecules, which makes sense.

• Small molecules have fewer atoms, so the protein binders need high complementarity to create a strong bond, and this is especially hard for flexible compounds.

• Hydration challenge: Polar small molecules create strong bond with water, making it energetically difficult for a protein to break those bonds and bind the molecule instead.

• Finally, protein design for small molecules requires extensive experimental fine-tuning, which makes the whole process slow and expensive.

So, what to do?

Designed binding proteins: pseudocycles, AI and small molecules

In today’s paper, the authors (successfully!) tackle this problem, using AI-based protein design to create small molecule-binding proteins that have:

• High shape complementarity to the target molecules

• Modular and scalable structures

• Efficient sensing capabilities!

Their secret? The use of pseudocycles: synthetic protein structures with repeating structural units (the backbone) that surround a central binding pocket for the target small molecule.

The design of the pseudocycles consisted of AI tools, computational modeling and experimental steps:

1. Selecting the backbone for the pseudocycles: The team used a library of pseudocycles backbones previously designed with AlphaFold2, a computational tool for the prediction of 3D structures from amino acid sequences, and proteinMPNN, an AI-based tool for the creation of sequences that fold in a specific structure, optimized for expression.

2. Ligand docking and binding site optimization: Using the Rosetta software, the researchers simulate the small molecule binding, and optimized the interaction using LigandMPNN, a tool created excatly for this job!

3. Experimental screening of the designed binders: The most promising binders were synthetized and tested for their ability to bind target small molecules. Even with AI optimization, they still had to screen tens of thousands of variants!

The results are in: testing four small molecule binders

What are these “targets” I have been talking about? Well, the researchers designed and tested binders for 4 small molecules, and the results were pretty amazing.

1. Cholic acid: This is a bile acid, important for the monitoring of liver diseases. The binder achieved a great binding affinity (4.7 nM for those who like numbers) which was additionally increased 700x after a second round of design optimization!

2. Methotrexate: A chemotherapy drug, which requires regular blood monitoring to reduce adverse reactions in the patients. Here they achieved a good micromolar-range binding, with improvements after the second round.

3. Thyroxine: An important hormone, the binder designed for this molecule destroyed previous designs, achieving a nanomolar binding affinity.

4. AMA: the designed binders had a great affinity for this non-biological molecules, which shows that this approach can work also for these class of molecules.

Converting binders into sensors: nanopores and dimerization

The most exciting part? The researchers used the modular nature of the binders to create sensors. The team explored two main applications:

• Nanopore-based sensing: This is cool! They took the cholic acid binder and connected it to a protein nanopore. In this way, after binding to the ligand, the nanopores closes, blocking ion flow and allowing for real-time molecular detection.

• Dimerization systems: The researchers split the proteins into 2 independent domains, which can reassemble after binding a ligand. This can be useful to control gene expressions, for example!

What’s next then?

This work was definitely a great feat: creating a whole new pipeline and then 4 new protein binders, when others struggle to just create one, is definitely impressive!

It’s not easy though: even with AI, detailed ligand-protein knowledge is required for the design. In addition, experimental screening is still intensive: having to screen tens of thousands of proteins limits the accessibility of this approach.

But we do have the data needed for a lot of interesting biomarkers and metabolites, and both computational capacity and lab automation capability are improving at crazy speed. What I see today as a problem might not be one tomorrow! And this approach can create completely new classes of biosensors, synthetic enzymes and therapeutics. So, don’t take my word, just go here to read the paper!

And as always, thank you for reading! What do you think about synthetic proteins? Do you think AI tools are overrated? Do you think they are/will revolutionize science? Reply and let me know!

More room:

• DNA nanopores of all shapes: Did you miss DNA nanopores? Me too. Then, this review is for us: it explores key design principles, current challenges, and recent advances to enhance DNA nanopore stability and functionality. In particular, it focuses on how , ion leakage and structural instability limit their use in sensing and ion channel studies. There could be some inspiration in this review!

• Displacing strands in high resolution: Cryo-EM has been dominating structural biology, and it doesn't look like it’s going to stop anytime soon. For example, this study uses cryo-EM structures to reveal previously unknown elements that aid strand displacement in yeast Mip1 polymerase. This polymerase can unwind DNA independently, unlike other polymerases. Cool images!

• Not yet a subscriber to Plenty of Room? Sign up today — it’s free!

• You think a friend or a colleague might enjoy reading this? Don’t hesitate to share it with them!

• Have a tip or story idea you want to share? Email me — I’d love to hear from you!

• You have something you would love me to cover? Just reach out here or on my social!