Xaira Therapeutics has assembled a large CRISPR perturbation dataset (X-Atlas) and trained a companion model (X-Cell) to predict how gene expression changes after interventions. Project leads Ci Chu and Bo Wang argue that the limiting factor for predictive performance is often the information content of the data rather than model size or compute.
Motivation and the information bottleneck
The teams observed a characteristic pattern: when test loss plateaus at a certain parameter count (for example around 1.5 billion parameters) while training loss keeps decreasing as scale increases, this indicates the model is limited by available information. Training on a single, relatively small dataset uncovered an information gap: a 3.1 billion-parameter model deviated from the expected scaling trend. In that situation, neither adding parameters nor more compute improves outcomes — what’s needed is richer experimental data.
The Xaira teams report that their new dataset provides roughly ~30× more useful information than prior datasets, and that this information gain permits further scaling of parameters and training compute.
From CELLxGENE and scGPT to perturbation data
CELLxGENE, created by the Chan Zuckerberg Institute, is a massive single-cell database that links hundreds of millions of cells to counts for ~20k–30k genes per cell plus detailed metadata. That resource has enabled many RNA-expression models and inspired models such as Bo Wang’s influential scGPT.
However, models trained on CELLxGENE-type observational data mainly capture relationships among cell types and states; they are not well suited to predict the outcome of active perturbations of gene expression. Gene expression values are highly correlated, and observational data alone often cannot resolve causal directions.
To get causal insight, Xaira ran CRISPR perturbations that effectively "turn down" individual genes and observe resulting expression changes. Such intervention data make it possible to infer upstream and downstream relationships (at least for first-order effects) and to train models that predict what happens when a gene’s expression is changed by a drug or an edit. The perturbation dataset collected by Xaira is called X-Atlas, and the model trained on it is X-Cell.
Methods and reported outcomes
X-Atlas was generated from millions of parallel CRISPR experiments that produced the raw measurements used to train X-Cell. The teams abandoned an autoregressive modeling approach in favor of diffusion-based methods and incorporated a broad set of prior knowledge alongside the data and architectural choices.
According to the team, X-Cell generalizes to real lab experiments on human cells and outperforms a linear baseline that had previously beaten earlier models. The researchers note that including many priors is justified, but priors alone cannot substitute for information-rich experimental data and appropriate model architecture.
Cost profile and organizational impact
Precise budgets are not disclosed. The reporting estimates that experiments and infrastructure for data collection likely cost a few tens of millions of dollars, while compute, headcount, and research were in the range of a few million dollars. The authors describe the budget profile as more akin to an RL rollout budget than to a large, data-dense pretraining program.
The project’s strategic importance at Xaira is underscored by later promotions: Ci Chu was promoted to Chief Discovery Officer and Bo Wang to Chief AI Scientist — promotions that occurred after the podcast interview referenced by the reporting.
Context and caveats
The teams stress that first-order perturbation effects are a tractable and informative starting point, though biological systems can include cycles and higher-order interactions (second, third order effects) that only appear when multiple genes change simultaneously. Bo Wang discussed advantages he has as an academic leader versus industry roles, and how his labs stay aligned with rapid AI advances.
Xaira’s work highlights that to progress predictive models of cellular response, collecting targeted, intervention-based, information-rich data can be more important than simply increasing model size or compute.
