Creating new proteins by AI diffusion

Welcome to Plenty of Room!

Today we have new and improved protein binders for some clinically relevant receptors! And of course they used AI. Will it be more interesting than others we have seen before? You just have to read to find out.

Plenty of Room is your guide to the cutting-edge news related to molecular machines.

Already subscribed? Share with a friend that might find this interesting! It really helps.

New here? Just go ahead and subscribe!

Let’s get into it now.

Creating new proteins by AI diffusion

Creating proteins with perfect affinity and specificity is what computational biologists dream of every night (okay, maybe every other night). This would create incredible opportunities for new therapeutics, better diagnostics and more robust bioengineering. Researchers have made huge progress in designing new proteins, and the machine learning/AI revolution definitely helped a lot. These computational tools have given us large sets of small, ideal scaffolds that fold reliably and bind to certain targets. But as always in biology, the devil is the detail. These small proteins have a limited capacity for shape matching, especially for targets with flat and featureless surfaces. Take the tumor necrosis factor receptor (TNFR) superfamily, for instance. These receptors are vital in inflammatory diseases and cancer, making them prime drug targets. But their flat, polar surfaces make traditional design methods stumble.

So, what to do? Well, today’s paper offers a possible solution. The authors employed RoseTTAFold diffusion (RFdiffusion) to create novel binders exactly for the TNFR superfamily. But what is RFdiffusion? Glad you asked. RFdiffusion is a generative AI-based protein design tool, that is based on diffusion, the same mechanism used in image generations (like DALL-E). In practice, there are a few key steps in this model:

1. Initialization: It starts with a random "cloud" of amino acid residues, essentially noise with no structure.

2. Conditioning: The algorithm is guided by the structure of the target protein, directing the design to fit the target's surface.

3. Progressive Denoising: Over many iterations, the noise resolves into a well-folded protein structure tailored to the target.

4. Output: The final product is a designed protein that matches the desired shape and properties.

Pretty cool, right? This method enabled researchers to design larger, more complex proteins that could tackle flat surfaces like TNFR1's. They even refined their results using partial diffusion, which reintroduces noise into initial designs and then optimizes them further.

And this approach worked very well! 90 out of 96 designs were well expressed in E. coli, even if they were much less regular and longer than previously designed binders. Six of these design were high affinity, with the best two designs having a K_D of around 25 nM. The researchers even improved on these two, with one of the designs reaching a K_D lower than 10 pM, making it the most potent monomeric binder for TNFR1 ever reported! And the designs were also very specific, with no binding to the closely related TNFR2.

Once they succeeded with TNFR1, the team turned their attention to other TNFR family members: TNFR2, OX40, and 4-1BB. Using partial diffusion, they created binders for these targets, too. The standout for TNFR2 achieved a K_D of 198 pM, confirmed by X-ray crystallography with near-atomic precision. Very exciting! The researchers then tested the functional applications of these new binders in cells. They showed two effects, depending on the target:

• Antagonist effect: this is when a molecule binding to the receptor blocks the biological response. The TNFR1 binders blocked TNF-α signaling, which drives inflammation. Unlike bivalent antibodies that might accidentally activate TNFR1, these monomeric antagonists safely inhibited the receptor.

• Agonist effect: an agonist activates a receptor to produce a biological response. For OX40 and 4-1BB, binders were fused to multivalent oligomers that mimic ligand clustering, activating signaling pathways crucial for T-cell expansion. These designs outperformed native ligands and antibody-based complexes, offering tunable control over therapeutic responses.

All together, a very cool paper! Some proteins are VERY hard to target (like p53 or the IL-6 receptor), so it’s great to see progress in this area. The team showed that RFdiffusion is a very powerful tool, with several advantages:

• Exploring New Design Spaces: Starting from random noise allows for entirely novel protein folds, not limited by known scaffolds.

• Target-Specific Designs: Conditioning ensures designs fit even flat, polar surfaces that usually resist traditional approaches.

• Efficiency: The method produces high-quality designs quickly and effectively.

• Versatility: Beyond binders, RFdiffusion can generate enzymes, structural proteins, or anything the application requires.

In summary, this was a great read that shows how AI-driven tools like RFdiffusion are reshaping the future of protein design! So, I highly recommend you read this one!

More room:

• Creating plant circuits: Plants don’t get a lot of love in biotech (and I am guilty of it here too). So, if you want to increase your intake of green papers, this is a cool one to read. This study expands the ABA sensing system in plants by engineering orthogonal chemical-induced dimerization (CID) modules. These modules were reprogrammed to detect substances with nanomolar sensitivity, with the plant changing colors as a readout! Pretty cool.

• DNA origami antennas: When you TV does not work properly, now you can try to improve it using DNA origami (who uses TV anymore?). This study uses DNA origami to fully control coupling between single-photon emitters and optical antennas. A fluorescent molecule was positioned in an antenna gap with precise alignment, showing up to 1400-fold fluorescence enhancement when aligned with the antenna’s axis.

• Machine-learning your way to silver nanoclusters: I have covered silver nanoclusters before: a handful of silver atoms are templated using ssDNA and they gain cool fluorescent properties. And what if you put machine learning in the mix? This work introduces generative model for designing DNA-stabilized silver nanoclusters (AgN-DNAs) with multiple fluorescence properties. Experimental validation showed the model's ability to design bright NIR-emitting AgN-DNAs with a 4-fold greater abundance.

• 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!