Chai-2

Chai-2 is a multimodal, all-atom generative model for zero-shot de novo design of antibodies and protein binders, developed by the Chai Discovery Team and released in July 2025. It is the design-focused successor to Chai-1, the company's open-weights biomolecular structure predictor. Where Chai-1 predicts the structure of a given complex, Chai-2 generates new binder sequences and structures directly against a target epitope, addressing one of the hardest problems in therapeutic discovery: creating functional binders from scratch without large high-throughput screens or directed evolution.

The central result is that Chai-2 turns computational antibody design from a low-yield screening aid into a design-first workflow. Prompted with only a target structure and epitope, the model designs small batches of up to 20 candidates per target, which are then synthesized and tested in a single 24-well-plate experimental round. Across 52 novel antigens with no known binders in the Protein Data Bank, this procedure produced at least one validated binder for roughly half of the targets, at an overall antibody hit rate near 16% — an over 100-fold improvement over previously reported computational methods.

Chai-2 is described in a bioRxiv technical report, "Zero-shot antibody design in a 24-well plate," authored by the Chai Discovery Team (Jacques Boitreaud, Jack Dent, Danny Geisz, Matt McPartlon, Joshua Meier, Zhuoran Qiao, Alex Rogozhnikov, Nathan Rollins, Paul Wollenhaupt, and Kevin Wu). Unlike its open predecessor, Chai-2 is proprietary: it has no public weights, API, or web server, and access is granted selectively to early-access partners under Chai Discovery's Responsible Deployment policy.

#Key Features

• Zero-shot de novo antibody design: Chai-2 designs complementarity- determining regions from a target structure and epitope alone, without template binders or existing complexes, achieving a ~16% overall hit rate across 52 previously unaddressed antigens.
• Multimodal binder generation: A single model spans multiple binder formats — single-chain variable fragments (scFvs), VHH nanobodies, and miniproteins — and reports a 68% wet-lab hit rate for miniprotein binders.
• All-atom generative design plus folding: The system couples an all-atom generative design module with a folding model that roughly doubles the structural accuracy of its predecessor, reasoning explicitly over backbones, ligands, and post-translational modifications.
• Low-throughput, lab-in-the-loop workflow: Designs are validated at 20 candidates per target in a standard 24-well plate, compressing the full design-to-validation cycle to under two weeks and routinely yielding picomolar to low-nanomolar binders.
• Restricted, policy-gated access: Chai-2 is not publicly released; it is offered to selected academic and industry partners under a Responsible Deployment policy that prioritizes programs with clear health benefits.

#Technical Details

Chai-2 integrates all-atom structure prediction with generative sequence and structure design in a single multimodal framework. On antibody-antigen complex prediction, the folding component reaches DockQ scores above 0.8 for 34% of cases, a substantial gain over Chai-1. In the reported de novo campaign, the team prompted the model against 52 diverse antigens lacking known binders, designing at most 20 candidates per target and validating them by biolayer interferometry. Half of the targets yielded at least one binder in a single round, with an average antibody hit rate of about 15.5% (approximately 20% for VHH nanobodies and 13.7% for scFvs), while miniprotein design reached a 68% hit rate across the tested targets. Many binders showed picomolar to low-nanomolar dissociation constants, and the workflow supported cross-species-reactive and multi-specific designs.

#Applications

Chai-2 targets therapeutic discovery teams that need functional binders against novel or difficult antigens without building large screening libraries. Because it generates atomically resolved candidates across scFv, nanobody, and miniprotein formats, it fits antibody and biologics programs seeking rapid hit generation, epitope-directed design, and multi-specific constructs. The small batch sizes and short validation cycle make it practical for iterative, lab-in-the-loop campaigns where each round is characterized in days rather than months.

#Impact

Chai-2 demonstrates that generative design can deliver double-digit experimental hit rates for de novo antibodies, a regime long considered out of reach for computational methods and a step-change over screening-based pipelines. It extends Chai Discovery's move from structure prediction toward end-to-end molecular design and, together with wet-lab validation across dozens of targets, signals a shift toward design-first therapeutic discovery. Its significance is tempered by openness: unlike the Apache-2.0 Chai-1, Chai-2 is proprietary with no released weights, code, or public interface, so its results cannot be independently reproduced. The technical report is a preprint awaiting peer review, and the reported validation was conducted by the developer, leaving external, third-party benchmarking still to come.