New AI Models Are Turbo-Charging Protein Design

Getting a new biologic to market is laborious, expensive, and time-consuming, and one of the biggest drains on resources is watching promising candidates fail in preclinical or clinical testing. AI models that can design new proteins may be the ticket to optimizing drug development and providing companies better shots on goal. Companies are taking different approaches to the de novo design, but all share a common need: copious amounts of training data to level up their models. Will it eventually lead to fully autonomous de novo drug design?

Optimizing multispecifics development

Sanofi is moving AI-based protein design from proof-of-concept toward real-world drug discovery. Norbert Furtmann, PhD, head of biologics AI and design at Sanofi, says his team’s program is three-pronged: foundational models; property prediction and virtual screening; and de novo design.

The foundational models team develops sequence- and structure based models, analogous to large-language models but trained to “speak the language of proteins,” Furtmann says. These models have been built using a combination of public and in-house data. The property prediction and virtual screening team builds models to predict properties essential for function and developability, including stability, solubility, affinity, and so on. They run large in silico screens on hundreds of thousands of sequences, identifying which designs will move on to lab testing.

This is particularly valuable for multispecifics, where the enormous combinatorial design space makes comprehensive lab testing impossible. “If we have optimized monospecific antibody or nanobody building blocks, let’s say against target A, against target B, and against target C, and we want to plug them together, then we are facing enormous design spaces,” says Furtmann. “That could easily go into the hundreds of thousands of molecules, and you could not test all of them in a lab.”

The de novo design team starts with a target antigen structure and then has the AI model help design antibodies or nanobodies against it. Furtmann says the company’s first proof-of-concept studies on therapeutically relevant targets have been so successful that they are now looking to scale up the program. “We had a target where traditional discovery approaches faced challenges,” he says. Using their de novo design platform, the scientists successfully created functional binders, which Furtmann says are now moving forward into multispecifics.

When designing the binders, Furtmann says, it’s important to look beyond basic characteristics. “We try to think end-to-end and to understand what the monovalent building blocks need to look like to become successful, developable molecules in a multispecific format,” Furtmann says. “If you have a monospecific molecule, and then you plug five of them together for a multispecific, and they don’t behave well anymore, you need to start optimization again.”

The “lab-in-the-loop” workflow is key to this strategy, Furtmann says. This means taking the AI-generated designs, synthesizing them in the lab, and then feeding that data back into the model to fine-tune its designs. To achieve this, the company invests heavily in automation to achieve rapid, reproducible measurements and reliable data. “We aim for integration of the AI with the wet lab to get to fast feedback loops,” he says. “Yes, for monospecific optimization, but also specifically for the more complex multispecific molecules.”

A whole new type of antibody

Biolojic Design, based in Ness Ziona, Israel and Washington, DC, is using AI to build “multibodies,” standard human antibodies engineered to bind two or more targets. Unlike bispecific antibodies, which are built from fragments of two different antibodies, multibodies retain the familiar structure of a conventional human IgG antibody. This approach benefits from the established manufacturing, regulatory, pharmacokinetic, and safety advantages of ordinary antibodies, while enabling multiple therapeutic effects from a single drug.

Antibody structure typically includes quite a bit of “unused” space. Antibodies recognize their targets through six complementarity-determining regions (CDRs), but only a small fraction of that surface area is needed for antigen binding. “If you look at known antibody-antigen pairs, most of the CDRs are not involved in binding,” says Yanay Ofran, PhD, founder and CEO of Biolojic. “Each antibody has a lot of ‘free real estate.’ Our models allow us to precisely identify the free real estate and ‘develop’ it to accommodate additional binding.”

The AI model redesigns this part of the molecule, allowing it to recognize additional epitopes without compromising specificity. Each arm of a multibody conditionally binds only one target at a time, but the drug binds differently depending on the tumor, tissue, or patient. This adaptability could help address tumor heterogeneity while limiting off-target interactions.

To build their multibodies, Biolojic fine-tunes temporary, highly specialized models for each project, Ofran says. “This is not a general solution intended to solve all the problems,” he says. “It’s a concentrated effort of a couple of months to build a dedicated model that will make one drug that no other technology can make and then will be retired.” First, existing models propose roughly a billion antibodies, each predicted to possess some desired property, such as binding a particular target or having a good developability profile. Researchers synthesize and test the variants and feed measurements of binding affinity, expression, specificity, and stability back into the model to refine it. That model proposes candidates combining all the required characteristics, which are subsequently evaluated in the lab.

One Biolojic product, a conditional agonist/antagonist, has already entered Phase II trials. The antibody binds the cytokine IL-2 in a context-dependent fashion, enabling it to stimulate T cells that attack the tumor while blocking T cells that protect it. This asset is being developed by Biolojic’s spinout company, Aulos Bioscience. So far, safety and efficacy data are encouraging, Ofran says.

Ofran points out that Biolojic’s approach goes beyond simply speeding up drug discovery. “We want to be fast, we want to be inexpensive, but what we want to do most is to come up with molecules that we are unlikely to find in traditional discovery,” he says. “Not to accelerate the time for getting the same type of molecules that you will get otherwise, but to get molecules that do things that traditionally discovered molecules cannot do.”

Evolving methods need new benchmarks

It can be difficult to objectively compare models when proof-of-concept results rely on different targets, assays and metrics, and that’s where Seattle-based A-Alpha Bio comes in. “We are particularly interested in answering this question earnestly: what is the kind of data that helps build better models?” says Natasha Murakowska, PhD, director of applied machine learning at A-Alpha. “What will it require to do this sort of de novo design perfectly every time?”

To that end, A-Alpha is running a large VHH de novo design benchmark. They have invited about a dozen groups to each submit 10,000 designed antibody sequences, which A-Alpha will synthesize and test against a panel of 10 targets, yielding some 110,000 VHHs. Their aim is to build the largest apples-to-apples benchmark of de novo-designed antibodies, evaluating not just whether antibodies bind, but how strongly and how specifically, and what sequence or structural features distinguish the winners.

Central to this effort is AlphaSeq, A-Alpha’s library-on-library yeast display assay. One library encodes the antibodies, the other encodes the targets; when combined, AlphaSeq effectively measures all pairwise interactions. This generates, on the scale of one million quantitative protein-protein interaction data points per experiment, Murakowska explains, revealing information about not only target binding affinity but also off-target effects.

The benchmark’s target panel is stratified into three tiers of difficulty. In Tier 1, a known antibody–antigen crystal structure is provided, and the models must design novel variants that improve binding. Tier 2 provides a crystal structure of the antigen alone, challenging models to choose a bindable epitope and design a complementary interface.

Tier 3 is what Murakowska calls the “holy grail”: targets where no structural information is provided, requiring the model to infer the antigen’s structure, pick a druggable region, and then design an antibody to bind it. “It stacks complexity, and it tasks model makers in multiple ways,” she says. “If you want to earnestly solve this problem, teams need to think about the holistic problem of representing the right protein, representing the right epitope, which is that surface that the antibody binds, and then also designing for that surface.”

Leveling up AI-assisted drug design requires vast amounts of high-quality data that doesn’t currently exist. A-Alpha recently launched the Atlas Consortium, a collaborative model where companies of very different sizes jointly generate new, shared AlphaSeq datasets under a common experimental framework. One of the key selling points of the Atlas Consortium is that every member, large or small, increases the value of participation for the other members by contributing datasets.

Human out of the loop?

Anthropic has published results showing that its Claude models can design miniprotein binders with essentially no human involvement. The models designed binders against 15 protein targets and sent the designs to Adaptyv Bio and Twist Bioscience for synthesis and evaluation. Testing confirmed that the Claude-designed minibinders were successful for 14 of the 15 targets.

To achieve this, Claude drew on existing open-source software, including backbone generators like RFdiffusion3; sequence-design tools such as ProteinMPNN; and structure predictors used to screen candidates before synthesis. Within that framework, Claude selected which surface of each target to bind, chose and combined design tools, filtered candidates, and delivered 30 ranked designs per target after 24- to 48-hour autonomous runs. Human involvement was limited to selecting targets, granting permissions, and ensuring the infrastructure was running properly.

Among 1,320 designs, 354 bound their targets, giving an overall hit rate of 27%, although results varied widely depending on the target. Claude was not able to come up with a binder for maltose-binding protein, for instance, although it did produce several binders against TNFα, a therapeutically relevant target that has stumped other AI design efforts.

The designs were evaluated solely for binding affinity, not for activity, and none of the designs were solved structurally. Anthropic describes the result as foundational, a step toward the goal of AI systems autonomously handling drug-discovery from end to end.

“Minibinders are not meant to be drugs,” says Emily Leproust, PhD, CEO of Twist Bioscience. Designing therapeutics will require the model to advance to larger, more complex models, which will require more data points to train them. “That’s why we are adding more and more capacity, so that people can build more data points with us at a lower cost and at speed.”

“What we like about AI is you can do co-optimization of binding and developability at the same time,” Leproust adds. Traditional drug screening involves first optimizing for binding and then engineering in developability after the fact, which can be frustrating and inefficient. Using AI to optimize for both in parallel can boost success rates. “That can significantly, I think, increase the probability that an antibody is going to behave well in the clinic, because you don’t engineer developability after the fact,” she says. “It’s a subtle thing, but it really makes a big impact.”

The post New AI Models Are Turbo-Charging Protein Design appeared first on GEN - Genetic Engineering and Biotechnology News.