Xaira представила X-Cell — «виртуальную клетку» на 4,9 млрд параметров для поиска лекарств

Xaira Therapeutics unveiled X-Cell on March 17, 2026 — the company's first "virtual cell": a 4.9-billion-parameter model trained on the largest genome-wide dataset of cellular perturbations in history, X-Atlas/Pisces.

What is X-Cell

X-Cell is a model that predicts how gene expression changes when individual genes are "switched off" one at a time. According to Xaira, it is the largest causal perturbation model to date: it scales according to a power law, the same way large language models improve in quality as parameter count and compute grow.

The training dataset was assembled through CRISPR experiments, running millions of tests in parallel.

• X-Cell launch — March 17, 2026, Xaira Therapeutics' first virtual cell model
• X-Atlas/Pisces dataset — 7 genome-wide Perturb-seq screens across 16 biological contexts, 25 million cells after quality filtering
• Model size — 4.9 billion parameters, the largest causal perturbation model
• Scaling follows a power law, as with large language models
• On July 6, 2026, Bo Wang was promoted to Chief AI Scientist and Si Chu to Chief Discovery Officer

Why causal models need causal data

Xaira argues that for a model to predict cause-and-effect relationships, it must be trained on causal data rather than descriptive genomics. Ordinary datasets capture correlations — which molecules co-occur in a cell; causal data show what happens when you deliberately intervene on a specific gene.

"Causal models need causal data," is how

Bo Wang (Chief AI Scientist) and Si Chu (Chief Discovery Officer) frame the central thesis of their approach on the Latent Space podcast.

The company's practical takeaway: a model trained on causal data from cancer cells can be transferred to healthy donor cells — the rules it learned still hold.

How the model broke through the "data wall"

The first version of X-Cell hit a ceiling: at 3.1 billion parameters, test error stopped decreasing — the model saturated at around 1.5 billion parameters and improved no further. The cause was a lack of informative data, not the architecture.

The X-Atlas/Pisces dataset delivered roughly a 30-fold increase in information density compared with the original approach. After that, the model began scaling again along with parameters and compute, and it overtook the linear baseline that had previously outperformed earlier models. It builds on scGPT — Bo Wang's earlier RNA model, published in Nature.

What it means

Xaira, an AI biotech with billions in initial funding, is betting not on the algorithm but on the data: its own factory of causal experiments turns out to matter more than someone else's ready-made model. This reverses the usual AI logic of "internet data plus a big model" — in biology, the necessary data simply isn't available in the open, so it has to be generated in-house.

FAQ

What is X-Cell in simple terms?

X-Cell is a "virtual cell," an AI model with 4.9 billion parameters that predicts how a cell will respond to the switching off of individual genes without running an actual experiment.

What was X-Cell trained on?

On the X-Atlas/Pisces dataset: 7 genome-wide Perturb-seq screens across 16 biological contexts, 25 million cells after filtering. According to Xaira, it is the largest such dataset ever disclosed.

Who developed the model?

The model was developed by Bo Wang (Chief AI Scientist, creator of scGPT) and Si Chu (Chief Discovery Officer); both were promoted on July 6, 2026.