Xaira Therapeutics launches X-Cell model for drug discovery using causal data from CRISPR experiments

Xaira Therapeutics argues that causal models require causal data and has built a system around that principle. The company’s new X-Cell model is trained on millions of parallel CRISPR perturbation experiments to capture cause-and-effect relationships in gene expression, rather than relying on observational datasets like CELLxGENE.

According to the podcast interview, earlier models trained on a single, smaller dataset plateaued after reaching 3.1B parameters, with test loss flatlining while training loss continued to drop. Xaira attributes this to an information gap in the training data and claims its dataset provides roughly 30 times the information density, allowing the model to scale with parameters and compute beyond the previous wall.

The company estimates that data-collection experiments and infrastructure cost “a few tens of millions,” while compute, headcount, and research added “a few million,” describing the budget as resembling a reinforcement-learning rollout rather than a data-rich pre-training effort.

Leadership changes at Xaira reflect the strategic importance of the initiative: Bo Wang was promoted to Chief AI Scientist and Ci Chu to Chief Discovery Officer after the work was completed.

Xaira’s dataset, X-Atlas, and model, X-Cell, were designed to address a limitation of prior Virtual Cell models: the inability to predict the effects of changing RNA expression. By perturbing genes one at a time at scale, the team aims to infer causal graphs and predict downstream effects, which could inform drug mechanisms or gene-editing strategies.

The team also discussed abandoning autoregression in favor of diffusion modeling and reported generalization to real lab experiments in human cells, outperforming a linear baseline that had previously led other models.