State-of-the-art AI model accurately predicting unseen biological experiments

X-Cell is trained on X-Atlas/Pisces, the largest and most context-diverse genome-wide perturbation dataset ever reported

Powered by novel diffusion large language model architecture with more than 4 billion parameters, the largest virtual cell model to date

SOUTH SAN FRANCISCO, Calif.--(BUSINESS WIRE)--

Xaira Therapeutics, a company built to transform drug discovery through AI, today announced X-Cell, its first virtual cell model. Trained on causal, interventional data at unprecedented scale, X-Cell achieves state-of-the-art performance in accurately predicting the outcomes of genetic perturbations — including experiments the model has never seen. This marks a meaningful step toward transforming drug discovery from a trial-and-error process into a predictive engineering discipline.

X-Cell is trained on X-Atlas/Pisces, the largest genome-wide CRISPRi Perturb-seq dataset ever reported: 25.6 million perturbed single-cell transcriptomes across seven biologically diverse cellular contexts, more than three times the scale of Xaira's previous dataset, X-Atlas/Orion. The model reaches 4.9 billion parameters, the largest causal perturbation model built to date, and follows the same power-law scaling dynamics observed in large language models.

“In AI, strong performance comes from pairing the highest-quality data with the most capable models,” said Ci Chu, SVP of AI-Enabled Discovery at Xaira Therapeutics and co-senior author of the preprint. “X-Atlas/Pisces is the largest and most context-diverse perturb-seq dataset ever reported. X-Cell is a model architecture designed specifically to learn from data at that scale. Together, they allow the model to generalize and predict biology it has never seen before.”

X-Cell uses a diffusion language model architecture, a departure from the autoregressive single-cell models that have dominated the field. Rather than predicting a perturbed cell state in a single pass, X-Cell iteratively refines its predictions by progressively replacing control gene expression values with perturbed values, improving accuracy across each diffusion step. This iterative refinement process allows the model to calibrate its predictions in context.

The result of combining best-in-class data with this novel architecture is a model that generalizes to unseen experiments — unseen cell types, unseen exposures, and results from different laboratories using different technologies. This is a core milestone for the field of predictive biology.

“X-Cell provides an important step forward in enabling biologists to use computers to simulate how cells respond to different perturbations without running the actual experiments,” said Bo Wang, SVP and Head of Biomedical AI for Xaira, and also co-senior author. “As we further augment the data training sets, we expect that X-Cell should become increasingly useful for multiple purposes in drug discovery, including target identification, mechanism of action identification, matching targets to patients, and toxicity predictions.”

"The goal of building a virtual cell is to understand biology at a causal level, to be able to ask: if a cell is in a disease state, what does it take to bring it back to a healthy state?” said Xaira Therapeutics CEO Marc Tessier-Lavigne. “With X-Cell, we are already seeing the model make predictions that go beyond the data we provided. That's a big step toward a powerful model that will provide not just biological insights, but therapeutically actionable ones.”

Xaira is making a subset of the Pisces dataset and X-Cell model available to the scientific community. The company's roadmap calls for continued expansion of X-Atlas into primary cells, iPSC-derived cell types, organoids, and in vivo perturbations — building toward causal biology models capable of identifying the right targets, molecules, and patients across the full arc of drug development.

Read the preprint here.

Watch the video here.

About Xaira Therapeutics

Xaira Therapeutics is an AI life sciences company creating frontier machine learning models to discern the language of life and deliver transformative medicines to patients. Xaira seeks to recreate the drug discovery and development process from end-to-end by bringing together leading talent across three core areas: machine learning research to produce powerful, predictive models of disease, therapeutic design, and patient care; massive data generation at scale to train that artificial intelligence; and integrated drug discovery and development to create new medicines. Xaira is headquartered in South San Francisco with innovation centers in Seattle and London.​​

View source version on businesswire.com: https://www.businesswire.com/news/home/20260317710096/en/

Source: Xaira Therapeutics