Your Foundation Model Is Only as Causal as Your Data

Foundation models trained on observational single-cell data consistently fail at causal perturbation prediction, often losing to linear baselines. X-Cell proves the fix is not more parameters but causal training data: genome-wide Perturb-seq across 16 cell types, generating 25 million cells. The architecture shift from autoregressive to diffusion language models helps, but data quality matters most.

The implication for AI outside biology is stark. Any system requiring cause-effect reasoning cannot rely on passive data aggregation. Teams must budget for controlled interventional experiments, making data generation infrastructure as critical as model R&D. The organizational shift is from data lake accumulation to hypothesis-driven, high-throughput experimentation.

Tradeoffs are steep. The approach is capital-intensive, requiring specialized wet labs and months of quality control. The model still depends on curated priors (GenePT, PPI, etc.), some of which add marginal value. Generalization beyond measured cell types and into spatial or in-vivo context remains unproven.

Engineering leaders should note: the causal AI bottleneck is now data design, not compute. As causal models move into software, analytics, and product decisions, the ability to run large-scale interventions will determine who builds reliable AI and who gets stuck with expensive correlative tools.