AI-Designed Antibodies: Machine Learning Redefines Molecular Medicine!

Lots of modern medicine is based on antibodies, from drugs to diagnostics. And yet, we still produce them in a slow (and expensive) way! Are there better methods out there?

Well, what about putting some machine learning in the mix?

Today we have an awesome paper!

Machine-Made Antibodies

Antibodies: The Workhorses of Biology

Antibodies are the backbone of biology.

With more than 160 FDA-approved drugs, they dominate biologics, even as peptides work hard to close the gap (GLP-1s and all that). They also fuel diagnostics and biomedical research, as you know if you ever stepped inside a lab.

But discovering or developing new antibodies? Not easy.

Traditional antibody discovery relies on immunizing animals (injecting them with antigens and collecting the antibodies) or selecting them from huge libraries. Powerful methods, but they are also expensive, time-consuming, and a lot of work.

Plus, they can’t guarantee that the antibody you get binds where you want it on the antigen (the epitope). And that can translate into lots of wasted money!

Computational Methods to the Rescue

Now, this sounds like the perfect problem for computational methods to solve.

And researchers have tried! They didn’t need me to tell them.

Over the past years, researchers have used computational tools to tweak antibodies. They have optimized CDRs (the hypervariable loops that bind the antigens) or improved affinity using deep learning trained on huge datasets.

But these approaches only refine existing antibodies.

There are no computational methods that can create antibodies from scratch to target specific epitopes. Such a structure-based method would let us block specific receptor-ligand contacts, target functional sites on viruses, or create new diagnostic tools!

We have seen something similar before. RFdiffusion has been used to create novel protein binders, but antibodies are tougher. The “vanilla” version doesn’t really work to capture the complexity of antibody and antigen interaction.

All Eyes on Antibodies

Here comes today’s paper.

The authors built exactly the tool they wished existed: a combined computational-experimental pipeline that designs antibodies from scratch. And not only do the designed antibodies target user-defined epitopes, but they also agree with the design models down to the atomic level!

The idea is :
decide on an epitope → design an antibody → test it experimentally → done.

Now, is it really this easy? Let’s see!

Laying the Groundwork

The pipeline is built on RFdiffusion, a generative model for protein design. This deep learning model is based on diffusion. RFdiffusion starts with a noisy “cloud” of amino acids and iteratively refines it into a 3D protein structure.

The model has been used to create binders before, but the “vanilla” model doesn’t work on antibodies. So, the researchers trained it on antibody complexes. They kept the non-variable regions of the antibody constant, with the model learning how to design the CDRs and to place the antibody on the target.

And just like that, RFdiffusion is ready to predict new antibodies!

The Pipeline Step by Step

Here’s how the magic happens:

1. Pick a framework and a target epitope
   Choose the epitope to hit and an antibody framework whose sequence and structure won’t change.

2. RFdiffusion protein design
   The fine-tuned RFdiffusion keeps the framework intact, while sampling diverse docking orientations and CDR structures.

3. Sequence design with ProteinMPNN
   The previous step creates the 3D structure of the antibody; now, it’s time for the sequence. That’s a job for ProteinMPNN, which designs the sequence from the 3D structure. So, you get a full set of antibodies and structures!

4. In silico filtering
   Many computational designs fail experimentally, so it’s important to filter using a structure-prediction “oracle”. They fine-tuned RoseTTAFold2 (RF2) on antibody-antigen data and then used it to re-predict the designed structures. If the prediction matches the design (self-consistency), it’s more likely to work experimentally.

5. Experimental screening and maturation
   The filtered designs are tested using yeast surface display (up to 9,000 designs per target!) and/or E.coli (lower throughput, up to 95 designs per target). The affinity of promising hits is improved using OrthoRep, a yeast-based hypermutation system.

The Results: Building Antibodies from Scratch

So, the pipeline was revving. Where did they unleash it?

VHHs: Single-domain binders

VHHs have only one heavy chain domain (so, no light chains, like other antibodies). They are smaller and easier to produce, with 3 CDRs instead of the canonical 6. A perfect starting point!

The researchers designed VHHs for 6 disease-relevant targets, ranging from bacterial and viral proteins to cell receptors. The pipeline worked! But the initial affinities were kind of modest, with K_ds in the tens to hundreds of nM: they didn’t bind well.

But after maturation with OrthoRep, the affinities improved 100x! To single-digit nM or sub-nM, comparable to normal antibodies. All while preserving the designed binding pose!

ScFvs: Two-chain designs

ScFvs (short for single-chain variable fragments, somehow) are a step up in complexity. They are fusion proteins, where a short peptide links the heavy chains and the light chains.

ScFvs have 6 CDRs, so they are also bigger than VHHs. So, more complicated to design! But they are still smaller than normal antibodies. And the genes for them are also hard to synthesize, apparently, because there is a lot of sequence homology.

Combining different heavy- and light-chain CDRs, the team designed scFvs for 2 targets, the bacterial protein TcdB and the PHOX2B peptide–MHC complex (relevant to neuroblastoma). And they even converted one of them into a full-length antibody!

How did they work? Well, not great, to be honest. They bound, but with affinities in the hundreds of nM. And when tested in functional assays in cells, they didn’t show effects.

Peering at the Structure: Cryo-EM validation

The team solved cryo-EM structures for several of the designed antibodies.

The results were great: the high-resolution maps confirmed that the VHHs bind exactly as designed! For scFvs, they produced cryo-EM maps showing the arrangement of both heavy and light chains, and, in one case, they resolved all six CDR loops!

These are the first atomically validated de-novo antibody designs ever reported. Incredible!

So… Does it Work?

Time for conclusions! A super cool study, with amazing figures.

It does work. They can design antibodies to specific epitopes. The structures match the designs atom-for-atom, and the affinity can be improved using affinity maturation. So that’s great!

But there are limitations:

• The success rate is not good.
  They produced plenty of designs, but most of them didn’t work. The final part of the paper is a sort of “retrospective”. They re-filtered their designs using the recently released AlphaFold3 and saw that most of them didn’t pass this stricter filtering. So, probably better filtering will help improve this!

• Limited affinity.
  Maybe related is the limited affinity of the initial computational hits, which still needs experimental validation to go from hundreds of nM to low single-digits.

• Design biases/immunogenicity risks.
  The designed CDR sequences are not “human-like”, and this could create problems with immune reactions.

But ehi, it’s a first step! And a great one.

I found it amazing how most of the designs targeted the epitope they aimed for. This opens up several targets that are unavailable today. Plus, this structure-based approach can also improve other pharmaceutical properties, such as aggregation, solubility, and stability.

But go here, read all for yourself, and let me know what you think! Just reply to this email.

More Room:

• Sensing Force with DNA Origami: During my PhD, I tried to use DNA origami to activate cellular receptors using force (I failed, btw). This study introduces single-molecule DNA origami tension sensors (smDOTS) to measure the tiny mechanical forces that drive T cell receptor (TCR) activation. These sensors can detect 8–19 piconewton forces at the single-molecule level—crucial for understanding how T cells recognize antigens. Built from DNA origami nanostructures with fluorescent barcodes and tunable membrane anchors, smDOTS allow high-resolution mapping of TCR–antigen forces and track ligand movement. By multiplexing sensors with different force thresholds, researchers captured diverse force dynamics, offering new insight into how mechanical forces shape immune signaling.

• Tricking DNA Origami with Electric Fields: When I think I’ve seen everything you can do with DNA origami, people go and do crazy things. This study investigates how electric fields affect DNA origami structures and the activity of DNA probes used in electrochemical sensing. Using single-molecule fluorescence imaging, the researchers tracked DNA hybridization events on individual DNA origami frameworks to quantify how applied voltage and scan duration influence probe behavior. They found that electric fields can alter DNA conformation and cause structural relaxation within the origami framework, impacting probe activity.

• DNA Copies Antibodies’ Homework: When you don’t like proteins, you make antibodies out of DNA. This study introduces a DNA framework-based strategy to precisely control the simultaneous recognition of different molecular targets by organizing bivalent aptamers on a programmable DNA scaffold. By tuning the distance and orientation between the aptamers targeting PDGF-BB, the researchers achieved optimized binding affinity, rapid kinetics, and ultrasensitive detection down to 0.5 pM. This programmable heteroligation approach offers a versatile platform for designing multivalent molecular recognition systems with potential applications in diagnostics and immunotherapy.

• 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.