Nobel laureate David Baker’s lab at the University of Washington’s Institute for Protein Design has published a paper in Nature (volume 649, pages 183-193, 2026) demonstrating what the team calls “atomically accurate de novo design of antibodies” — a method for building antibodies entirely on a computer, targeting exact molecular surfaces, without ever exposing an immune system to an antigen.
How antibody discovery has always worked, until now
Traditionally, antibodies are discovered rather than designed: researchers immunize animals or screen vast natural antibody libraries, then hunt for a match that happens to bind the target of interest. It is an inherently probabilistic process — you search until you find something that works, then optimize it. Baker’s team inverted that process. Using a fine-tuned version of RFdiffusion, a generative AI model for protein structure, combined with yeast-display experimental screening to validate the computer’s designs, the researchers generated antibody variable heavy chains (VHHs), single-chain variable fragments (scFvs), and full antibodies that bind to user-specified epitopes with atomic-level precision.
Why atomic precision is the headline, not just a nice detail
The distinction between “roughly binds” and “binds with atomic-level precision to a specified epitope” is the crux of why this matters commercially and clinically. Antibody drugs that bind imprecisely, or bind unintended targets, are a major source of side effects and a major reason drug candidates fail in clinical trials. A design process that lets researchers specify exactly which atoms on a target protein an antibody should engage — rather than discovering a binder and then trying to characterize and refine it after the fact — could in principle produce cleaner, more specific drugs with fewer off-target effects, and do it faster than screening-based discovery allows.
A follow-on from a Nobel-winning body of work
Baker won the 2024 Nobel Prize in Chemistry for earlier AI-driven protein design work, and this latest paper is described as a direct follow-on breakthrough from the same lab. It extends the core idea behind Baker’s Nobel-recognized research — using generative AI models to design novel proteins with specific functions — into antibodies specifically, which represent a roughly $200 billion drug industry built almost entirely on discovery-based methods rather than from-scratch design.
The field is already inching toward the clinic
This is not purely theoretical. Generate:Biomedicines, a separate company working in AI-optimized antibody design, has reported promising Phase 1 data on an AI-designed antibody for asthma that lowers asthma-triggering protein levels with dosing as infrequent as every six months, without notable side effects. That result is a separate story in its own right, but it is worth noting here as evidence that AI-generated or AI-optimized antibodies are no longer confined to computational papers — they are beginning to produce real clinical data.
The skepticism: design success is not the same as a drug
Not every biologist is ready to declare victory. A recurring critique of computational protein design is the gap between a molecule that looks perfect on a screen — or even binds correctly in a yeast-display assay — and a molecule that can actually be manufactured at scale, avoids triggering unwanted immune responses (immunogenicity) in patients, and demonstrates real clinical efficacy and safety in human trials. Antibody drug development has a long history of promising early-stage candidates failing for reasons that have nothing to do with binding affinity, including manufacturability challenges and unexpected immune reactions. Baker’s team has demonstrated the binding problem can be solved computationally; whether de novo designed antibodies clear the rest of the drug-development gauntlet at meaningfully higher rates than discovery-based antibodies remains an open, and expensive, question.
What it means for the $200 billion antibody business
If de novo AI antibody design proves reliable and reproducible beyond Baker’s lab, it could restructure how the antibody drug industry approaches early-stage discovery — shifting resources away from massive screening campaigns and toward computational design pipelines that generate candidates against precisely chosen targets. That shift would not happen overnight, and manufacturability and immunogenicity testing will still gate how fast any of these designed antibodies reach patients. But combined with early clinical signals from AI-optimized antibodies elsewhere in the industry, Baker’s Nature paper is another marker of generative AI moving from a tool that helps optimize existing drug candidates to one capable of inventing entirely new ones from first principles.
