Wiki · Bio × AI Methods · concept
Perturbation prediction
Predicting how a cell's gene expression changes in response to a perturbation — without running the experiment.
Perturbation prediction is guessing how a cell will react to being poked — before you poke it. Knock out a gene, add a drug, and a cell’s gene-expression profile shifts. The goal is to predict that shifted profile computationally, so you can ask “what would this cell do if…?” without doing the wet-lab experiment. It is one of the sharpest tests of whether a model has learned real biology rather than surface patterns.
How it works
The training fuel is Perturb-seq and related atlases: experiments that apply thousands of individual perturbations to cells and read out single-cell gene expression for each one. A model learns the mapping from (starting cell, perturbation) to (resulting expression). Approaches span a wide range — from simple linear baselines that just add an average perturbation effect, up to single-cell foundation models fine-tuned for the task. The hard part is generalization: predicting the effect of a perturbation the model never saw during training, or the same perturbation in a new cell type. That makes it fundamentally an intervention problem — closer to asking about cause and effect than to fitting a curve to observed data.
Why it matters (for bio × AI)
This is the virtual experiment. If prediction works, you can screen millions of gene knockouts or drug candidates in silico and take only the promising few to the bench — which puts perturbation prediction at the center of disease-mechanism and drug-discovery work. But the honest caveat matters: generalizing to truly unseen perturbations remains hard, and careful benchmarks keep finding that simple baselines are surprisingly competitive with elaborate models. The promise is real; so is the gap between demo and dependable tool.
Related
See single-cell foundation model for the cell representations these predictors build on, and foundation model for the pretrain-then-adapt recipe behind them.