AI-Designed Protein Binders: BindCraft Revolutionizes Molecular Design!

In 2018, AlphaFold took the structural biology world by storm and won a Nobel Prize in 2024! Can it also design protein binders? Well, BindCraft puts that to the test!

Crafting Protein Binders

Protein-Protein Interactions: Don’t Leave Proteins Alone

Proteins like to work with each other. 

Protein-protein interactions are at the base of complex biological processes, from signal transmission and membrane transport to muscle contraction. They regulate all of them, and when things go bad, diseases follow (think Alzheimer's or prion diseases).

Because these interactions are so central, the ability to design protein binders that target and tune them would be super powerful for biotech and synbio!

How to Build New Binders

Traditional methods to generate protein binders are not the best, though.

Immunization, antibody library screenings, and directed evolution all work, but they are slow, require lots of work, and give you little control over the target site.

Computational design promises a better, more controlled route.

Physics-based models were the first to let you design binders, creating a scaffold and then optimizing the side-chains. Powerful, but with a low experimental success rate (often <0.1!). Plus, they are rigid because you must pick a scaffold and hold the target fixed.

Lately, we’ve seen incredible improvements in AI and machine learning methods! And they only seem to get better.

In the current state-of-the-art, RFdiffusion generates the protein backbone, and then ProteinMPNN optimizes the sequence. AlphaFold 2 (AF2) evaluates the plausibility of the designs and is used as a final filter before experiments.

This approach works well, but leaves a disconnect between the generation of the backbone (with RFdiffusion) and the design of the functional interfaces, filtered by AF2.

I would be happy with this, but some scientists are looking for every optimization opportunity!

BindCraft: Binder Design for Everyone

And here comes in today’s paper! The team built BindCraft, an automated, user-friendly, open-source pipeline that “hallucinates” de-novo protein binders, based only on AF2!

But what does AF2 actually do? Well, first of all, it’s the second version of AlphaFold.

AlphaFold is the deep learning model that rocked the structural biology world in 2018, thanks to its incredible capacity to predict protein structure from their sequences. An amazing achievement that even led to a Nobel prize!

And AF2 did even better! Plus, it introduced AlphaFold-Multimer, an additional model specialized in predicting protein-protein interactions.

BindCraft uses AF2 not only as a predictor, but as an active designer. And it works!

How BindCraft Works

The core idea is to leverage hallucinations from AF2.

Hallucinations are AI responses that contain false information presented as facts. Now, when ChatGPT does it, it can be annoying and sometimes even dangerous. When it’s a protein prediction model doing it, it generates new proteins!

BindCraft runs in 3 steps:

1. AF2 Multimer “Hallucination Loop”
   BindCraft starts from a random binder sequence plus the target structure. AF2 predicts the complex and calculates an error gradient (a measure of how far the model’s output is from the desired one). This error gradient is used to update the binder sequence to fit the target protein, and then it starts again. This loop designs the binder backbone, side chains, and protein interfaces at the same time!

2. Sequence Polishing
   After the loop, the sequences outside the interfaces are optimized using a ProteinMPNN variant (MPNNsol) to improve expression and solubility, while interface residues are left alone.

3. Re-prediction with AF2 Monomer
   Finally, AF2 monomer validates and filters the designs, to keep only the best ones!

And the results are high-quality binders, with just a few needing to be tested experimentally!

Real-World Testing: From The Cloud to the Lab

The authors applied BindCraft to 12 diverse, therapeutically relevant targets. Since there are quite a lot, I’ll focus on my favourites.

Cell-Surface Receptors

They started with clinically important cell-surface receptors.

The first one was PD-1, a key immune checkpoint receptor from T cells. The team screened 53 binder designs, and 13 of them showed binding! With the best one showing an apparent Kd of < 1 nM! Amazing!

Since they had such a high success rate, they tested fewer designs for the next binders:

• PD-L1: 7/9 designs showed binding.

• IFNAR2: 3/ 9 designs bound, and they also worked in cell assays.

• CD45: 4/16 designs worked, with the best one having a Kd of 14.7 nM!

This showed that the pipeline needs just a few attempts to find binders for complex receptors!

Toxin Blocking and Allergen Masking

CpE is a bacterial toxin that targets Claudin-1, a protein involved in the function of tight junctions. They designed a binder against Claudin-1 to outcompete CpE and stop it from killing cells.

Their binder12 showed nanomolar affinity for Claudin-1, and blocked the CpE cytotoxicity in cell assays!

They also tested a similar solution for allergens. In some countries, up to 50% of the population suffers from seasonal allergies! And current treatments are not great. Allergens are diverse and have highly charged surfaces that are hard to target with small molecules!

Could protein binders neutralize allergic reactions? To test BindCraft, they designed binders against 2 mite allergens and the major birch allergen.

The team successfully designed binders for all of them, and even checked their structures with cryo-EM! Plus, the binders blocked the allergens in in vitro assays.

Targeting Cas9 Nuclease

Protein-nucleic acid interfaces are usually undruggable by small molecules, since they are large and charged.

But nature can do it. Phages produce Acrs, small proteins that block the function of CRISPR-Cas by occluding the nucleic acid binding sites. So, the authors decided to mimic this function!

They designed binders against the REC1 domain that binds the guide RNA in SpCas9. 6/6 of the designs bound the protein! And they were also functionally active, reducing the gene editing of SpCas9 in cells.

AAV Retargeting: Optimizing Gene Delivery

Last but not least, they worked with adeno-associated viruses (AAVs). These are used for gene delivery, but it’s hard to send them to specific cells or tissues.

The team inserted miniprotein binders into an AAV capsid to precisely target the receptors HER2 and PD-L1. They showed that the targeted AAVs had better transduction specificity against receptor-expressing cells!

Strengths and Limitations

Cool work! BindCraft has lots going for it:

• High hit rate: The reported success rates vary between 10 and 100%, with an average of 46%! Far higher than other methods.

• Fast, user-friendly end-to-end pipeline: This will make it easier for people without protein design experience (like me) to enter the field.

• Functional outputs: The team not only showed binding but also practical utility.

Now, it’s still not perfect:

• High compute cost: The method is GPU-intensive, which is a problem.

• Filter sensitivity: The final filtering with AF2 monomer could toss genuine binders.

• Immunogenicity and delivery: The binders are synthetic and small, so there is a risk of immunogenicity, and there are delivery challenges.

But it’s a great work, and a pragmatic step towards making binder design more accessible! It’s an exciting time. Go and read all the details here!

If you made it this far, thank you! Do you think that AI-based protein design will be useful? Or is there too much hype? Reply and let me know!

More Room:

• What’s Inside a Lipid Nanoparticle? Lipid nanoparticles (LNPs) are used to deliver nucleic acids, but their internal structure and morphology remain poorly understood. Challenges such as characterization limits, particle dynamics, heterogeneity, and unclear formulation–structure relationships hinder progress. This review surveys current knowledge, emerging characterization methods, and efforts to link formulation parameters with LNP structure and function, aiming to improve therapeutic design.

• Mixing ML with T Cells: Can we use ML for even more? In this study, researchers developed deepAntigen, a graph convolutional network that identifies T cell antigens at the atomic level, improving accuracy over sequence-based methods. It predicts both antigen-HLA binding and antigen-TCR interactions, and its tumor neoantigen predictions in lung, breast, and pancreatic cancers were experimentally validated. This approach could accelerate the development of personalized mRNA cancer vaccines and immunotherapies. Looks cool, I’ll study and report back!

• Freezing DNA Origami: Studying ice with DNA origami? Yeah. This work shows how DNA origami can be used as a model system to study ice nucleation. Researchers found that well-folded tiles promoted a steady freezing rate, whereas unannealed DNA showed time-dependent decreases in nucleation efficiency. Interestingly, although annealed tiles generally enhanced freezing at low temperatures, unannealed DNA often froze at higher temperatures due to the formation of small numbers of highly efficient nucleators from molecular aggregation. This highlights how nanoscale structure and conformation strongly affect ice nucleation behavior.

• Share Plenty of Room with founders or builders

  I help biotech and deep tech companies transform complex technologies into engaging content that builds credibility with investors, partners, and potential hires. Let’s chat!

• 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.