Chai-2 is Chai Discovery's generative AI system for designing new antibodies, nanobodies, and small protein binders against a chosen protein target. Introduced on June 30, 2025, it combines molecular structure prediction with sequence and structure generation to propose molecules for experimental testing.[1]
Its significance is the small experimental shortlist: the launch study found binders for half of 52 tested targets after assaying at most 20 designs per target. That is promising evidence for computational antibody design, but it does not mean every target will yield a binder or that a binding hit is ready to become a medicine.[1]
What does de novo antibody design mean?
In this context, de novo means designing a new target-binding antibody rather than starting with an antibody already known to recognize that target. It can still use an established antibody framework. The framework supports the variable regions and CDR loops that form the binding surface.
Zero-shot describes the absence of a target-specific experimental optimization loop before the first test. It does not mean the model has never seen related biology during training, generates only one candidate, or removes the need for experiments.
The molecular format matters:
- A single-chain variable fragment (scFv) joins the heavy- and light-chain variable domains into one polypeptide.
- A VHH, commonly called a nanobody, uses a single antibody variable domain.
- A miniprotein is a small designed protein binder, without the requirement to adopt an antibody fold.
These formats serve different experimental needs. An scFv or VHH is not a full-length monoclonal antibody; understanding the Fab and Fc regions helps explain what changes when a binding fragment becomes an IgG.
How does Chai-2 differ from Chai-1?
Chai-1 predicts biomolecular structures from supplied molecular inputs. Its code and weights are available under Apache 2.0. Chai-2 addresses the design problem: finding new molecules that could bind a target. Installing the Chai-1 package does not provide Chai-2's design capabilities.[2]
This distinction also separates antibody-antigen docking, where both partners are already specified, from generating a new antibody. A plausible predicted complex can help prioritize a candidate, but it is not a measured binding result.
How does Chai-2 work?
Chai Discovery describes Chai-2 as a multimodal generative system integrating all-atom structure prediction and molecular design. The public disclosures describe its capabilities, but do not provide enough implementation detail to reproduce the complete system.[1]
The practical starting point is a target and a design brief. Chai's product page accepts experimental or predicted target structures, including a sequence from which a structure can be predicted. Researchers can specify an epitope, antibody format or framework, and species-specificity or cross-reactivity requirements. The company also describes support for target conformations, bound ligands, and modifications such as glycans.[3]
That makes target preparation consequential. An epitope is the surface the binder should recognize. A site exposed on an isolated protein fragment may be inaccessible in the biological assembly or covered by a glycan. The design brief should therefore identify the intended target state and experimental format, not just a protein name.
What do the reported hit rates mean?
The original 2025 preprint reported a pooled 16% experimental hit rate for antibody and nanobody designs across 52 targets. None of those targets had a preexisting antibody or nanobody binder in the Protein Data Bank under the study's selection criteria.[4]
| Reported result | What was counted | How to read it |
|---|---|---|
| 16% antibody/nanobody hit rate | Tested designs that bound their targets | A pooled result, not a separate 16% estimate for both scFvs and VHHs |
| 50% target success | Targets with at least one hit | Half the targets yielded no hit in that round |
| At most 20 designs per target | Candidates taken into experimental testing | This is the wet-lab shortlist, not necessarily the number generated computationally |
| 68% miniprotein hit rate | Tested designs in a separate binder campaign | This result cannot be substituted for the antibody hit rate |
The same report describes picomolar miniprotein binders and a design-to-validation workflow of less than two weeks. The time claim concerns early hit discovery and characterization, not complete therapeutic development.[4]
The two success measures answer different planning questions. Design-level hit rate helps describe the yield of the tested candidates. Target-level success tells you how often a campaign produced anything useful. Applying the pooled 16% rate as an independent probability for every design on a new target would hide variation between targets and similarities among candidates.
Chai's launch announcement also claimed an improvement of roughly 100-fold over previous computational antibody methods.[1] Treat that as a historical comparison reported by the developer. A useful comparison between tools needs matched targets, binder formats, candidate selection, assay conditions, and hit definitions. Headline percentages from separate studies do not establish a universal ranking.
What did the later antibody experiments add?
Chai's November 2025 research update extended the evidence to full-length IgGs and more demanding target classes. It reported that 86% of 88 designed IgGs, spanning 28 antigens, had zero or one flagged issue across four developability assessments, which screen properties relevant to making stable, usable antibodies. Cryo-EM structures of five antibody-antigen complexes supported the intended binding sites and close agreement with the predicted structures.[5]
The G protein-coupled receptor (GPCR) campaign tested 10 to 73 designs per target and found binders to all six receptors, with agonists that activated two of them. These are separate experiments from the original 52-target study, so the earlier limit of 20 tested designs should not be applied to them.[5]
These results strengthen the case that generated binders can retain useful properties in therapeutic formats. They do not establish clinical safety or efficacy. Structural agreement, binding, functional activity, and developability each answer a different question. Broader antibody engineering still involves choosing and testing the properties required for the intended application.
Can you use Chai-2 online?
As checked on September 24, 2026, Chai Discovery's product page directs commercial organizations to an access form and invites academic researchers to register interest in limited non-commercial access. It does not present Chai-2 as an unrestricted, self-service download or publish a standard price list.[3]
Chai-2 is also no longer the company's newest disclosed model. Chai announced Chai-3 in its June 2026 Pfizer agreement.[6] When discussing access, confirm which model and capabilities are being offered. The results above belong to Chai-2's published experiments, not automatically to every service sold under the Chai platform name.
Which alternatives address the same task?
Choose an alternative by the molecule you need to make and the information you already have. Antibody design, general protein binder design, and complex prediction are related tasks with different outputs.
| Tool | Relevant task | Important distinction |
|---|---|---|
| RFantibody | Antibody and nanobody design from an antigen and framework | Uses antibody-adapted RFdiffusion, ProteinMPNN, and RoseTTAFold2 to design and filter candidates.[7] |
| BoltzGen | Protein, peptide, nanobody, and antibody binder design | Provides distinct protocols, including framework-based antibody and nanobody CDR design.[8] |
| BindCraft | De novo protein binder design | Uses AlphaFold2 backpropagation, sequence optimization, and filtering; it is not an antibody-specific design pipeline.[9] |
| RFdiffusion | Protein backbone generation, including binders | Backbone generation is followed by sequence design, commonly with ProteinMPNN, and structural evaluation.[10] |
The BoltzGen guide explains how to choose a molecular format and review ranked candidates. The BindCraft guide covers target preparation and interpretation for general protein binders.
If you already have candidate sequences, Chai-1 and Boltz-2 can help examine proposed complexes. The Boltz-2 guide explains its inputs and confidence outputs. Use these predictions to select hypotheses for testing; establish binding, specificity, and the desired biological effect experimentally before treating a design as a successful reagent or lead.
