From new binders to new medicines

Since we released Chai-2 five months ago, the AI field has advanced rapidly, yet antibody design methods have remained limited to simplified fragments (single domains, scFvs) rather than the full monoclonal antibody (mAb) format commonly used in therapeutics. Furthermore, claims around binding affinity are necessary but insufficient: clinical candidates must meet stringent criteria for manufacturability, stability, safety, and biophysical behavior.

In the past few months, we discovered that Chai-2 has already crossed this gap. In new research, we find that the model can design full-length, drug-like mAbs, while maintaining the high hit rates we previously reported, testing at most dozens of designs. These designs show developability characteristics on par with well-behaved therapeutic antibodies. Cryo-EM validation confirms that Chai-2 reasons about structures at sub-angstrom scale, confirming precise epitope targeting and functional engineering. 

We also apply Chai-2 to traditionally “hard to drug” targets – six GPCRs and a peptide-MHC target – achieving similarly high success rates.

Developability: Antibodies With Drug-Like Properties

We tested 88 Chai-2-designed IgGs, spanning 28 target antigens, on four key developability criteria. We used thresholds defined in prior statistical analysis of therapeutic antibodies (Jain 2017, Jain 2023) to flag properties that may have to be further optimized. 

• 86% of designs showed zero or one flagged issue, meeting the standard for preclinical candidate selection

• 24 of 28 antigens yielded at least one design with a clean developability profile

Overall, these findings show that Chai-2 can produce drug-like antibody leads for the vast majority of targets with little or no developability optimization required.

Atomic Precision, Experimentally Verified

Accurate prediction of the bound complex between an antibody and its target is essential for programmable antibody design. Structural accuracy underpins Chai-2’s ability to design against a specified epitope, which can enable desired functions. 

We used Cryo-EM to examine five designed antibody-antigen complexes. Each antibody bound precisely where the model intended, and the predicted structures aligned closely with the experimental density. Even in the notoriously difficult CDR loops, we observed sub-angstrom level accuracy. This level of accuracy marks a meaningful step toward routine and computational, epitope-specific antibody engineering.

Expanding the scope of therapeutic applications

De novo antibody design should not only be fast and reliable. Chai is focused on validating applications on the hardest problems.

We challenged Chai-2 to design both mAbs and VHH antibodies binding to a panel of six GPCRs. This class of membrane-bound receptors accounts for one-third of all approved drugs (Lorente 2025), but only three of them are antibodies (Strohl 2025), given the difficulty of applying screening methods to membrane proteins. When testing just 10 to 73 designs for each target, we obtained at least one binder to each. Remarkably, we found agonists to two of the GPCRs straight out of the model, including in mAb format. To our knowledge, Chai-2 is first to generate mAb binders to GPCRs.

We turned to peptide-MHC targets to evaluate the model on a problem where atomic resolution is critical. Because MHC is highly conserved, success depends on the model’s ability to distinguish between one or two residues in the peptide. We tested between 27 and 50 designs against three therapeutically relevant peptide-MHC complexes, and obtained two hits on a MHC complex with KRAS G12V peptide. The binders are specific and don’t bind G12D or wild-type KRAS peptides, highlighting Chai-2’s performance on difficult specificity problems.

Looking Ahead

Chai Discovery’s mission is to transform biology from a science into engineering. We are continuously working toward the long-standing aspiration of rational drug design: generating drug candidates that are ready for IND-enabling studies in a single shot, entirely on the computer.

Read our technical report for additional information.