SCALE Beats Every Virtual Cell Model on 100M Drug Response Predictions
A new virtual cell foundation model recasts perturbation prediction as latent transport, beating STATE by 12% on Tahoe-100M with 12.5x faster training.
ProTakeaways
- SCALE is a foundation model that reframes single-cell perturbation prediction as conditional transport between control and perturbed populations
- On Tahoe-100M, it beats STATE by 12.02% on PDCorr and 10.66% on DE Overlap
- Built on NVIDIA BioNeMo, it delivers 12.51x faster pretraining and 1.29x faster inference than STATE
- Uses a LLaMA-style set encoder plus DeepSets aggregation with JiT endpoint-prediction training
- Ablations show gains come from adaptive condition pooling, endpoint supervision, and mixed Gaussian priors
- Performance plateaus at 184M parameters, suggesting data and evaluation are the next bottlenecks
Predicting how a cell will respond to a drug, a gene knockout, or a cytokine without ever touching a pipette is one of the more ambitious promises in computational biology. A new foundation model called SCALE takes a serious swing at it, and the numbers suggest the field's current best approach has room to grow. SCALE (Scalable Conditional Atlas-Level Endpoint transport) is a specialized model built for virtual cell perturbation prediction, and its authors argue that progress here requires rethinking the training infrastructure, the modeling objective, and the evaluation metrics simultaneously.
The bottleneck no one wants to talk about
Virtual cell models aim to turn single-cell perturbation data into in silico systems for forecasting how cellular states respond to genetic, chemical or cytokine interventions. The core difficulty is fundamental: the same cell cannot be observed both before and after intervention. You are not learning a one-to-one mapping. You are learning how an entire population of untreated cells gets redistributed after a perturbation, in a gene expression space that is high-dimensional, sparse, and noisy.
The authors identify three coupled problems that have held the field back:
- Training and inference pipelines that do not scale to atlas-sized datasets
- Unstable modeling in the raw expression space
- Evaluation protocols that reward reconstructing average expression rather than recovering the biological signal that actually matters
Early approaches often relied on direct conditional mapping or randomly paired regression, which tend to emphasize average perturbation effects and can miss heterogeneous cellular responses. More recent methods shift the objective from cell-level matching to population-level prediction. STATE, the previous state of the art, moved to transformer-based set modeling. SCALE goes further by treating perturbation prediction as a transport problem between two observed endpoints.
Reframing the problem as conditional transport
Rather than reconstructing an unobserved trajectory from control to perturbed cells, SCALE learns a direct mapping between the two population endpoints you actually have. The architecture has two main pieces.