因果模型需要因果数据：Xaira 如何构建虚拟细胞

Cell-state atlases can describe associations, but they cannot by themselves predict what will happen after a targeted biological intervention.

Xaira uses pooled CRISPR perturbation experiments to create causal training data at a scale that observational datasets cannot provide.

When training loss keeps improving while test loss plateaus, additional parameters are no longer compensating for missing information in the data.

By increasing the information content of the training data, X-Atlas allows X-Cell to benefit again from larger models and more compute.

The team moved away from autoregression because diffusion better matched the continuous, high-dimensional structure of gene-expression responses.

A sophisticated biological model is not useful merely because it is nonlinear; it must beat the strong linear baseline on real perturbation outcomes.

The meaningful generalization test is whether predictions survive new perturbations, new contexts, and real human-cell experiments.

Some perturbation effects appear context-universal, while others depend strongly on the cellular setting in which they are measured.

Biological priors can improve inductive bias, but the episode treats them as complements to information-rich data rather than substitutes for it.

Xaira’s broader bet is to connect virtual-cell, protein-design, and patient-representation models across the full drug-discovery pipeline.