AI-Designed Antibodies Cross an Experimental Threshold, but the Real Test Is Just Beginning

The starting point for antibody drugs is often a long search: researchers must find, within a vast molecular space, a molecule that can bind to a disease-related protein without binding indiscriminately to other targets. Chai-2, published in Nature by Chai Discovery and collaborators, attempts to move this step forward: rather than slowly modifying existing antibodies, it uses AI to directly design new antibodies that may bind to a specified epitope.

The core use case of this work is specific. Researchers provide a target region on a protein, namely the epitope the antibody is intended to recognize, and the model then generates candidate designs such as scFv antibodies, VHH nanobodies, or small protein binders. If this can work, it could shorten the early antibody discovery process, especially for targets that lack existing antibody templates and are more costly to screen using traditional methods.

The abstract of the Nature paper shows that, after designing against 100 targets, the team reported an approximately 16% single-round antibody hit rate, including experimental validation for 52 novel targets. Chai Discovery added in a subsequent announcement that these 52 diverse protein targets lacked existing antibody binders in the SAbDab database; for each target, up to 20 antibody or nanobody designs were tested, and about half of the targets yielded at least one binding molecule. These numbers are not the same as clinical success rates, but they mean that “from model output to wet-lab binding” is no longer merely a workflow shown in a schematic.

More importantly, binding alone is not enough for an antibody. Company materials said some hit molecules further showed characteristics of being stable, specific, and non-polyreactive. These properties matter for drug development because molecules that aggregate easily, bind indiscriminately, or express poorly are often eliminated before entering animal or human testing. However, these descriptions remain early property assessments, and there is still considerable distance between them and the efficacy, toxicology, immunogenicity, and manufacturability required of a drug candidate.

Chai-2 is not betting only on antibodies. The same company announcement said the model achieved a 68% wet-lab hit rate in the design of small protein binders; however, this figure comes from different molecule types and tasks and cannot be used directly to infer the success probability of antibody drugs. For biomedical AI, this kind of distinction is important: strong performance by a model in one class of protein design does not automatically mean it can handle all therapeutic targets, all binding modes, or all manufacturing constraints.

The weight of this paper lies in the fact that it has undergone peer review and is supported by experimental data; its limitations are equally clear. Public information is still insufficient to judge the difficulty distribution of each target, the structural reasons behind failure cases, and how hit molecules perform in affinity maturation, cellular functional testing, and in vivo models. If the technology is to move toward therapeutic antibodies in the future, regulators and industry will require more than “AI-designed binders”: they will require reproducible design rules, traceable data sources, rigorous quality control, and an evidence chain compatible with existing drug development standards.

At a time when the narrative around AI drug discovery is often told too quickly, Chai-2 provides a more testable milestone: molecules designed by a model have already encountered real proteins in the laboratory and produced hits. But the path to antibody drugs has never been wide, and early hits are only the entrance. What will truly change R&D practice is not one impressive hit rate, but whether this method can consistently deliver developable molecules across more targets, more laboratories, and conditions closer to the clinic.