AI-designed antibodies created from scratch

Research led by the University of Washington reports on an AI-guided method that designs epitope-specific antibodies and confirms atomically precise binding using high-resolution molecular imaging, then strengthens those designs so the antibodies latch on much more tightly.

Why epitope targeting matters

Antibodies dominate modern therapeutics, with more than 160 products on the market and a projected value of US$445 billion in 5 years. Antibodies protect the body by locking onto a precise spot—an epitope—on a virus or toxin.

That pinpoint connection determines whether an antibody blocks infection, marks a pathogen for removal, or neutralizes a harmful protein. When a drug antibody misses its intended epitope, treatment can lose power or trigger side effects by binding the wrong target.

In medical development, knowing exactly where an antibody lands on a molecule can decide whether it succeeds in patients or fails in trials. Researchers design epitope-specific antibodies to target disease-critical regions, such as the receptor-binding tip of a virus spike or the toxic domain of a bacterial protein. Reaching that level of accuracy normally requires a slow, iterative process with years of lab work, involving animal immunization, multiple rounds of screening, and structural studies to confirm the binding site.

A reliable way to plan those interactions on a computer could make antibody creation faster and more focused, aiming at the precise molecular surfaces that control infection, toxicity, or cell signaling.

Computational efforts have largely optimized existing antibodies, and some deep learning models have proposed variants once a binder already exists. Recent generative approaches have needed a starting binder, leaving de novo, epitope-specific antibody creation as an unmet goal.

In the study, “Atomically accurate de novo design of antibodies with RFdiffusion,” published in Nature, researchers trained an AI system to build antibodies that recognize user-specified molecular sites.

The model, known as RFdiffusion, used information about antibody frameworks and target surface “hotspots” to shape new binding loops. A second network, RoseTTAFold2, predicted whether each design would fold and bind as intended, filtering out unstable or misaligned candidates.

Llama assisted research

Design efforts focused first on single-domain antibodies known as VHHs. These miniature antibodies come from animals such as llamas and alpacas and are prized in research for being stable, compact, and easy to engineer. Their small size allows them to reach crevices on viral or bacterial proteins that full-size antibodies cannot.

Researchers used a humanized VHH framework as the scaffold and aimed designs at influenza hemagglutinin, Clostridium difficile toxin B, RSV sites I and III, and the SARS-CoV-2 receptor-binding domain.

Laboratory screens used yeast surface display to test 9,000 designs per target and E. coli expression with single-concentration surface plasmon resonance to evaluate 95 designs per target. Each target posed a distinct structural challenge, from the smooth surface of influenza to the complex folds of C. difficile toxin.

Hitting the targets

Influenza designs produced several lab-made antibodies that attached to the virus protein in test tubes. High-resolution imaging showed one of those matches lining up with the computer’s plan at near-atomic detail, including how a key loop on the antibody reached the target site. Microscopic sugar on the virus shifted aside when the antibody settled in, a movement seen in the images and consistent with the planned approach.

C. difficile toxin work yielded a compact antibody that grabbed the intended site and blocked a previously designed competitor from landing there. Lab tests on cells showed protection against the toxin’s damage.

Follow-up imaging captured the same docking behavior before and after lab evolution, indicating that improvements in grip did not change where or how the antibody latched on.

Missing a few

SARS-CoV-2 tests produced a compact antibody that attached only when the virus protein moved into the “up” position and blocked a known competitor at that spot. Imaging placed the connection on the right site while revealing a different angle of approach than planned, a result labeled a design failure by the authors.

Designs aimed at a cancer-related peptide on human immune proteins showed on-target recognition in two separate assays, yet engineered T cells built from those designs did not kill tumor cell lines in lab tests.

A good start

Reported success rates remain low at 0% to 2% across targets, and authors point to improved filtering with AlphaFold3 ipTM as a route to enrichment. Prospects include faster and potentially more targeted antibody discovery as models and filters improve, with particular value for applications needing precise epitope engagement such as receptor–ligand blockade, conformational modulation, and conserved viral sites.

Written for you by our author Justin Jackson, edited by Gaby Clark, and fact-checked and reviewed by Robert Egan—this article is the result of careful human work. We rely on readers like you to keep independent science journalism alive. If this reporting matters to you, please consider a donation (especially monthly). You’ll get an ad-free account as a thank-you.

© 2025 Science X Network