Virtual-cell models compress unseen intervention geometry through a target-specific generalization bottleneck

Predictive models of cellular perturbation are typically judged by their ability to accurately reconstruct molecular states after unseen interventions. However, research has revealed that even when these models exhibit high state-level similarity, they may fail to capture the relationships that distinguish perturbations—a phenomenon termed Intervention Geometry Compression (IGC).

Studies across various models and perturbation settings have consistently shown that unseen interventions display weakened global and local geometry, reduced variance between interventions, and spectral collapse.

The primary reason behind this loss of geometry is not due to limitations in the response space capacity. Instead, diagnostic projections indicate that much of the missing geometry is concentrated in a limited number of residual response directions. These directions not only outperform complexity-matched random subspaces but also appear consistently across a separate Jiang perturbation resource.

While polarity captures some aspects of this continuous orientation signal, it does not fully encapsulate the essence of the issue.

Time-resolved analyses further highlight that accurately entering the trajectory significantly enhances the downstream propagation of perturbations. Additionally, early responses specific to the held target can quickly reveal the endpoint orientation, even if the model's generalization capabilities are limited. Perhaps most importantly, anchoring the intervention identity using empirical data from the same target transfers intervention identity across different contexts far more effectively than simply increasing exposure to various interventions.

These findings underscore intervention-coordinate assignment as a critical information bottleneck in the generalization of virtual-cell models. They also support a design principle: by empirically anchoring the identity of interventions, models can more effectively generalize these anchored effects across diverse cellular contexts.