Preprint: Why virtual-cell models struggle to generalize to unseen perturbations
Yongqi Huang’s first preprint from the lab is now on bioRxiv! Virtual-cell models compress unseen intervention geometry through a target-specific generalization bottleneck asks why “virtual cell” models that predict the effects of genetic and chemical perturbations often look accurate yet fail to tell unseen interventions apart.
Yongqi identifies a failure mode called Intervention Geometry Compression: predictions for unseen interventions collapse toward a few response directions learned from training data. The paper then shows that anchoring each target with a small amount of its own early empirical data transfers intervention identity across cellular contexts far better than training on more interventions. The work was done with Wilson lab PhD candidate Gary Wang.