Xaira

Xaira Therapeutics is an AI drug discovery and development company whose work includes X-Cell, its first virtual cell model, and an active pipeline of de novo biologics. It announced X-Cell in March 2026, trained on the perturbation dataset X-Atlas/Pisces. Its therapeutic-development approach requires computationally designed molecules to pass experimental binding, developability and functional checks before advancing.[1][6][3][7]

These research and therapeutic programs followed Xaira’s April 2024 launch, which established its organizational origins and committed financing.[4][3]

Xaira launched on April 23, 2024, after being founded and jointly incubated by ARCH Venture Partners and Foresite Labs. David Baker of the UW Medicine Institute for Protein Design was a co-founder. At launch, the company was headquartered in the San Francisco Bay Area and had integrated a proteomics group from Interline Therapeutics.[4][3]

The launch announcement disclosed more than $1 billion in committed capital, led by ARCH Venture Partners and Foresite Capital. Other participating investors included F-Prime, NEA, Sequoia Capital, Lux Capital, Lightspeed Venture Partners, Menlo Ventures and Two Sigma Ventures. The announced amount was committed launch capital.[4]

Xaira announced X-Atlas/Orion on June 17, 2025, describing it as the largest publicly available genome-wide Perturb-seq dataset at that time. It also reported a method for identifying dose-dependent genetic effects in such data. On March 17, 2026, it announced X-Cell, trained on X-Atlas/Pisces.[3][5][6]

The research addresses a distinction between describing cell states and predicting responses to intervention. In his July 21, 2026 Latent Space introduction, RJ Honicky explained that observational RNA-expression data contain correlated changes that make causal relationships difficult to identify. Turning down individual genes and measuring the resulting expression changes supplies intervention data for predicting what happens when a gene is altered. He credited Ci Chu’s and Bo Wang’s teams with the X-Atlas dataset and X-Cell model, and identified Wang’s scGPT as a starting point.[1]

Honicky described an information bottleneck: on a smaller dataset, test loss stopped improving beyond roughly 1.5 billion parameters even as training loss continued to fall, and a 3.1-billion-parameter model departed from the scaling trend. His account reported renewed scaling with richer data and identified the team’s abandonment of autoregression for diffusion as a consequential modeling decision. Xaira announced Chu’s promotion to Chief Discovery Officer and Wang’s promotion to Chief AI Scientist on July 6, 2026.[1][3]

In their June 29, 2026 article, “Progressable Binders: The Binders that Matter”, Chief Scientific Officer Debbie Law and Chief Technology Officer Hetu Kamichetty described Xaira’s active de novo biologics pipeline across multiple indications. Their framework distinguishes an initial screening hit from a validated binder and a progressable binder: a molecule that jointly satisfies reliable binding, developability, sequence, immunogenicity and biological requirements.[7]

The authors illustrated the importance of experimental validation with COVID-RBD binding assays. They reported validated zero-shot scFv designs with affinities of 180 pM and 26 nM, measured using monovalent antigen. Alongside these, they showed unreliable software-generated affinity fits caused by insufficient saturation or mass-transport limitation, and heterogeneous interactions associated with biphasic dissociation and aggregation. These examples explain why an apparently strong affinity value alone does not establish a usable binder.[7]

Xaira’s stated early-stage criteria include affinity of approximately 100 nM or tighter, first unfolding temperatures of at least 65°C for Fab formats and 60°C for VHH formats, and at least 90% monomer content. The company also checks sequence liabilities, predicted immunogenicity, specificity and predominantly CDR-mediated antigen binding. Functional requirements depend on the target and modality.[7]

Internal programs enter lead optimization only after obtaining a structurally and functionally diverse panel of progressable binders. Diversity spans targeted epitopes, binding geometry, CDR sequence and structure, reducing dependence on a single early candidate. Law and Kamichetty emphasized the limits of this policy: “These criteria are enrichment filters, not a clinical oracle.”[7]

The authors credited the work to the broader Xaira team and thanked Andy Deng, Anna Lauko, Do Soon Kim, Tilelii Amimeur and Zimple Matharu for input on the article.[7]