After a decade of false alarms, four labs from Chai Discovery to Xaira are designing real antibodies from scratch with AI, yet turning them into drugs remains a harder problem.
After ten years of watching AI-for-drug-discovery revolutions come and go, Carlos Outeiral has changed his mind, and the reason is not a press release. Outeiral, a former Meta protein-language-modeling researcher who worked on the company's evolutionary-scale modeling (ESM) effort, argues in a recent essay that designing brand-new antibodies from scratch with AI has finally crossed a real threshold. The harder question, of whether those antibodies become approved medicines, is one he leaves open.
In 2023, an AbSci-led team screened roughly 500,000 AI-designed variants of trastuzumab, the breast-cancer antibody approved as Herceptin in 1998, and recovered only three that actually bound HER2, the protein trastuzumab targets. That is a sub-0.001% hit rate on a target the field already understood, and it became a marker of what computable antibody design could not yet deliver.
Hit rates on novel antibody targets are now reported in the double digits across multiple labs with little to no optimization, according to Outeiral's survey of the field. Four startups anchor the wave: Chai Discovery, an AI antibody-design lab co-founded by Joshua Meier, a former Meta ESM researcher and a corresponding author on the 2023 trastuzumab paper; Nabla Bio, whose JAM-2 antibody-design preprint claims fully computational, drug-like antibody designs with high developability scores; Latent Labs, another AI antibody-design startup; and Xaira Therapeutics, a protein-design company that licensed technology from David Baker's lab. Parallel academic work has begun to back the trend: a July 2025 bioRxiv preprint describes zero-shot antibody design in a 24-well plate, a format where each design can be tested directly.
An independent stress-test of Nabla's JAM-2 preprint flags gaps in performance metrics and validation, and none of the JAM-2 or zero-shot results have cleared peer review yet. Outeiral's conversion is about the science, not the drug business.
Most drug candidates still fail in the clinic, and computational antibody design does not close that gap on its own. BIO's 2011–2020 industry-wide survey tracks that attrition across thousands of candidates: only a small minority survive from a working biological molecule to an approved medicine, and the expensive failures tend to live in clinical safety and efficacy, not in discovery. The pathway to a commercial antibody usually travels through a specific vehicle, an antibody-drug conjugate (ADC), a construct sometimes described as a biological missile that ferries a toxic payload to a tumor. A 2022 Nature review catalogs the clinical promise and limits of that lane, including the linker chemistry and payload release kinetics that govern whether the antibody actually delivers its toxic load in patients. ADCs dominate the commercial pathway for newly designed antibodies today, and that downstream chemistry is the part de novo design does not solve.
Three milestones to track: peer-reviewed publication of the JAM-2 and 24-well-plate results, which replaces preprint caveats with reviewer-imposed ones; independently reproduced hit-rate numbers from a lab with no commercial stake in the company whose designs it tested; and the first Phase 1 entry of a de novo-designed antibody from one of these four startups with public pharmacokinetic and immunogenicity data.