Chai-2: The "24-Well Plate" Revolution in Antibody Design

Original: Chai Discovery Team, bioRxiv 2025
DOI:10.1101/2025.07.05.663018

1. Background: The Antibody Discovery Dilemma

Antibody therapeutics have become a cornerstone of modern biomedicine. As of 2022, monoclonal antibodies accounted for more than half of all biopharmaceutical approvals in the US and Europe. However, traditional antibody discovery workflows face numerous challenges: animal immunization and library screening are resource-intensive and time-consuming; obtained leads typically require months of affinity or developability optimization; precise control over desired binding sites is difficult, and challenging targets often remain intractable.

Computational methods promise to transform this landscape, but existing approaches rarely exceed 0.1% experimental success rates, still requiring high-throughput experimental screening. The release of Chai-2 marks a substantive change to this situation.

2. Technical Approach: Zero-Shot Generation Architecture

2.1 Core Design Workflow

Chai-2 employs a "zero-shot" design strategy:

• Input: Target structure + epitope residue list (only a few residues needed)
• Generation: Design model generates sequences and full-atom structures
• Ranking: Computational models rank and filter designs
• Validation: Direct entry into small-scale biochemical experiments (e.g., 24-well plate format)
• Characterization: Determine quantitative binding affinity

The entire workflow from epitope definition to experimental validation requires only approximately two weeks.

2.2 Technical Architecture Improvements

Compared to its predecessor Chai-1, Chai-2's folding module achieves twice the experimental accuracy in antibody-antigen complex prediction.

• Multimodal generation: Supports scFv antibodies, VHH nanobodies, miniproteins, and other binder formats
• Multi-target prompting: Can design against multiple targets simultaneously, enabling customized cross-reactivity
• No per-target tuning required: All capabilities achieved through zero-shot learning without target-specific fine-tuning

3. Experimental Validation: From Miniproteins to Antibodies

3.1 Miniprotein Design: Establishing Benchmarks

The research team first used miniprotein design as a benchmark test, selecting five targets previously studied in the literature. Chai-2 achieved at least three times the success rate of previous best methods on each target.

Particularly noteworthy is TNFα—estimated to be in the top 1% of computationally difficult targets in the PDB, with no previous computational design successes reported. Chai-2 achieved a breakthrough.

Affinity: IL-7Rα, PD-L1, PDGFRβ, and InsulinR reached picomolar levels; TNFα achieved low nanomolar affinity.

3.2 Antibody Design: The Core Breakthrough

On the more challenging antibody design task, Chai-2 achieved a historic breakthrough:

• 52 novel targets with no known antibodies in PDB
• ≤20 antibodies or nanobodies designed per target
• Target success rate: 50% (26/52 targets with at least one successful binder)
• Overall hit rate: 16% (VHH 20.0%, scFv 13.7%)
• Comparison to previous methods: >100× improvement (previous <0.1%)

This result represents the first double-digit success rate for computational methods in de novo antibody design, a milestone achievement.

3.3 Design Quality Assessment

• Structural novelty: Majority of designs have RMSD > 10Å from known antibodies
• Sequence novelty: All designs have CDR edit distance > 10
• Diversity: Most targets contain multiple structural clusters
• Developability: Comparable to existing therapeutic monoclonal antibodies
• Affinity: EFNA 2.2 nM, VATF 80 nM, CSF1 17 nM, etc.

4. Application Showcase: Flexible Design Capabilities

Using CCL2 as an example, the research team demonstrated Chai-2's flexible prompting capabilities:

Same target, two different epitopes, generating both VHH and scFv formats for each epitope.

• VHH: 4/20 success (20% hit rate)
• scFv: 5/20 success (25% hit rate)

Chai-2 also supports cross-reactivity design, enabling simultaneous design against multiple proteins.

5. Implications and Outlook

5.1 Impact on Antibody Discovery Workflow

• Experimental scale: Can be completed in 24-well plates without high-throughput screening
• Timeline: From epitope to experimental validation in just two weeks
• Feedback loop: Tighter design-validation feedback loops enable rapid iterative optimization
• Target coverage: 50% success rate across tested targets indicates broad applicability

5.2 Technical Evolution Trajectory

Chai-2's success builds upon multiple technological advances: AlphaFold2 (structure prediction) → RoseTTAFold/RFdiffusion (diffusion models in protein design) → Chai-1 (multimodal structure prediction) → Chai-2 (generative modeling for antibody design). This evolution reflects the trend from "prediction" to "design," from "structure" to "function."

5.3 Limitations and Challenges

• High-affinity prediction: Predicting which designs will have high affinity remains challenging
• Epitope selection: Current methods require pre-specified epitopes
• In vivo performance: Stability, immunogenicity, and other in vivo properties of computationally designed antibodies require further validation

6. Conclusion

Chai-2 achieves a leap from <0.1% to 16% in de novo antibody design tasks, marking a substantive breakthrough for computational methods in biologics discovery. The "24-well plate" scale experimental workflow and two-week timeline open new possibilities for antibody discovery. This advance represents not only algorithmic progress but may reshape the standard strategy for biologics lead discovery. As computational design success rates continue to improve, traditional high-throughput screening methods may gradually give way to more precise and efficient computational-experimental closed-loop workflows.

References

Chai Discovery Team. (2025). Zero-shot antibody design in a 24-well plate. bioRxiv. https://doi.org/10.1101/2025.07.05.663018