AI-Designed Antibodies Enter the Lab: Xaira Reports a Preclinical Lead Obtained in Seven Weeks

Designing an antibody that can bind a disease target is only the starting point for drug development; molecular stability, binding specificity, and suitability for further development also determine success. On October 2, Xaira Therapeutics released experimental results for its generative AI model X-Design, including an oncology program that yielded a preclinical lead in seven weeks, adding another example of the path from computational design to practical drug development.

The model version announced this time, named Vega, was built using public and proprietary company data. Its design objectives include not only binding ability but also similarity to human antibody sequences and developability. Xaira’s platform page describes X-Design as a tool for generating therapeutic molecules, currently focused on antibodies and protein therapies; in its development pipeline, XA-1 is in oncology, while XA-4 is listed under immunology and inflammation.

In the XA-1 program, the team designed 182 single-domain antibody sequences for a preselected epitope—the specific site recognized by an antibody—and confirmed 18 binders after quality screening. The company reported that one molecule had a dissociation constant of 17 nanomolar for the human target and passed the developability tests used; the dissociation constant measures binding strength, with lower values generally indicating stronger binding.

The seven-week timeline also has a defined starting point. The company said that the antigen, relevant reagents, and target epitope had been prepared in advance, after which the process from DNA synthesis to obtaining a lead took seven weeks, including one round of optimization to improve cross-reactivity with a surrogate experimental species. This figure therefore reflects an experimental workflow under specific conditions and cannot be taken directly as the total timeline from identifying a disease target to completing drug discovery.

The other program, XA-4, addresses a more challenging G protein-coupled receptor (GPCR), for which the relationship between antibody recognition sites and function remains unclear. According to the company’s announcement, previous antibody library screening and immunization approaches had not yielded functional binders. This time, the team designed 60,000 sequences and, through screening and cell-based testing, identified candidate antibodies with antagonistic activity that could undergo further optimization, attempting to overcome obstacles encountered by traditional methods.

The two cases provide different levels of evidence: XA-1 focuses on binding and developability, while XA-4 also has support from cell-based functional testing. These results were all released by the company itself, and the platform page lists the relevant programs at stages ranging from early discovery to lead optimization. The announcement provided no evidence of efficacy in humans, and clinical safety cannot be inferred from these results.

The significance of this advance is that AI-generated sequences are beginning to undergo the multiple experimental tests required for drug development. However, a systematic assessment of whether these two cases represent the model’s performance across a broader range of targets is still lacking; Xaira said it would publish the relevant analyses separately. Whether the candidate antibodies can ultimately become therapies still depends on subsequent validation and clinical trials.