Single-Cell Perturbation Prediction

Reducing a gene’s activity can disrupt an entire network of cellular processes. The same intervention may strongly affect one cell type and barely affect another. Can a model predict that response in a new cell context when it has seen only the unperturbed cells?

This zero-shot task tests two forms of transfer: familiar interventions in unfamiliar contexts, and targets whose perturbations were never observed during training. Memorizing an average response can help with the first, but cannot directly solve the second.

Training includes single-cell expression measurements from K562, Jurkat, HCT116 and HEK293T, before and after gene perturbations. Expression counts estimate how many RNA molecules from each gene were detected in a cell. Averaging many cells reveals a population response while individual measurements show variability.

Three anonymized test contexts provide only unperturbed cells and intervention requests. Eighty target genes are withheld from all training perturbations. The most distant context also uses different measurement chemistry, so biological and technical changes are entangled. The interactive cells above are synthetic.

The mean-effect baseline learns how each known target changes average expression across training contexts, then adds that change to cells from the new context. For an unseen target, it predicts no effect. A gated version suppresses responses when relevant genes are rarely detected.

The learned STATE model instead combines a population of basal cells with a protein-based representation of the target gene. A transformer predicts the changed population. Protein features give an unseen target a meaningful description even without an observed intervention, creating a route to novel-target transfer.

A model must improve on copying the unperturbed cells. Seven measures examine complementary properties: response magnitude, response direction, discrimination between perturbations and recovery of genes whose expression changes.

Each measure is converted to skill relative to that no-effect baseline. For example, reducing mean absolute error from 0.40 to 0.30 gives 25% skill. A zero component means no credited improvement over the baseline, not perfect prediction or necessarily no biological effect. Final scores average across the three contexts.

The mean-effect model scores about 0.103 overall, compared with 0.046 for the learned STATE model. Its advantage is strongest in the nearer contexts. But it receives zero normalized credit for one measure of changed-gene recovery in every context: a useful aggregate score can coexist with an important failure.

STATE shows a narrower strength on previously unseen targets in the far context, with a reported score of 0.12 versus zero for the mean-effect model. This supports a signal of novel-target transfer, not uniformly accurate expression prediction. Complete context-holdout tests and repeated training runs are not available for STATE.

An intervention has both a response pattern and a context-dependent strength. Separating these could help the model reuse conserved biology without inventing strong responses in unresponsive cells. Compare a learned gate with the same ungated response before changing the whole architecture.

Other useful controls replace protein features with target identifiers, or compare matched measurement chemistries where possible. These distinguish biological transfer from memorization and technical effects. The present far-context gap alone cannot tell us which explanation dominates.