Antibodies against anything? AI tool adapted to make them

Antibodies are incredibly useful. Lots of recently developed drugs rely on antibodies that bind to and block the activity of specific proteins. They're also great research tools, allowing us to identify proteins within cells, purify both proteins and cells, and so on. Therapeutic antibodies have provided our first defenses against emerging viruses like Ebola and SARS-CoV-2.

 

But making antibodies can be a serious pain, because it involves getting animals to make antibodies for us. You need to purify the protein you want the antibodies to stick to, inject it into an animal, and get the animal to produce antibodies as part of an immune response. From there, you either purify the antibodies, or to purify the cells that produce them. It's time-consuming, doesn't always work, and sometimes produces antibodies with properties that you're not looking for.

 

But thanks to developments in AI-based protein predictions, all that hassle might become unnecessary. A recently developed diffusion model for protein structures has been adapted to antibody production and has successfully designed antibodies against flu virus proteins.

Making the antibody of your choice

Humans (and many other mammals) make antibodies that are four-protein complexes composed of two heavy and two light proteins. Both heavy and light proteins have constant regions, which are the same or similar among all antibodies produced. They also both have a variable region, which is unique to every antibody. It's the variable region that's responsible for recognizing proteins in viruses and other pathogens. Some other mammals, like camels, skip the light proteins and have antibodies that are simply a pair of heavy proteins (which still recognize pathogens via the variable regions of the heavy proteins).

 

The body doesn't know what proteins it will eventually need to recognize. So, it simply makes a lot of antibody-producing cells, each with a unique combination of heavy and light variable regions. When any of these cells run into the protein their antibodies recognize, they start dividing and produce a lot of the needed antibody. It takes time for these cells to mature and additional time to purify them. Plus, there's no guarantee that the specific combination of variable regions will be the optimal one for recognizing a protein.

 

The only way to avoid the hassle and uncertainty of getting an animal to generate antibodies for us is to figure out how to design antibodies that will recognize what we want. And that just hasn't been possible. We don't understand enough about how proteins fold up into a three-dimensional configuration to design one that will adopt a shape of our choice—one that wraps around a specific target.

 

We still don't really understand enough to do that intentionally. But in recent years, we've trained AI software to take a string of amino acids and accurately predict the three-dimensional structure that this protein would adopt. And, more recently, people have figured out how to merge these with diffusion models to create software that can design proteins that will adopt a specified configuration.

 

It turns out that this approach can be adopted for designing antibodies.

Training for antibodies

Antibodies present a distinctive challenge in that they don't simply have to adopt a specific configuration; they have to shape that configuration around another molecule and present the right sort of chemical environment to make any interactions sticky. However, the basic concept of training a diffusion model ended up working out well.

 

The basic concept behind diffusion models is that they're trained on both the end state (meaning a photo, or the structure of a protein), and noisy versions of that state. Once trained, if they're given something that's mostly noise, they'll readily hallucinate their way to something that resembles a reasonable end state.

 

So, the team behind the new work, led by the University of Washington's David Baker, adapted this approach to work with antibodies. There are many examples of detailed structures of an antibody bound to the protein it recognizes. So, the researchers started with these and added noise by shifting the position of the amino acids of the antibody. These target/noise combinations were fed to RFDiffusion, the previous software used for AI-based protein design.

 

The process worked in the sense that the resulting software could produce an amino acid sequence that it predicted would bind to a specific protein target. However, it produced a lot of predictions for each target, and the authors used other protein structure software to evaluate the strength of the interactions between the proposed antibody and its target.

 

In the end, they made antibodies the software predicted would bind to a number of disease-relevant proteins (they tested examples from the flu virus, the respiratory syncytial virus, and a toxin from Clostridium difficile). All of them stuck to their targets, although the strength of the interaction varied considerably. The team also purified the antibody-flu virus protein complex, and determined its structure, showing it largely reflected the software's prediction.

More to come?

The researchers acknowledge that they still have work to do, writing, "there is considerable room for improvements, as the binding affinities are modest." In animals, the genes for antibodies often undergo accelerated mutation, which produces additional variants that can fine-tune the interactions with pathogens. It might be possible to perform a similar process in software to improve affinities once a structure has been identified.

 

Still, even given the limited success, this could be a very useful tool. Antibodies are used for many things, and we could use them for many more if we could target them to additional proteins. And the researchers suggest that it would be relatively easy to modify the software to allow it to target non-protein chemicals.

 

The bioRxiv. Abstract number: 10.1101/2024.03.14.585103v1  (About the arXiv).