GENESIS drops the need for control cells: it generates single-cell perturbation responses from experimental covariates alone and nears experimental reproducibility on Tahoe-100M
Synopsis
Single-cell perturbation atlases cover only a fraction of possible experimental combinations, and existing generative models map measured control cells to perturbed cells without a reference for how good a prediction can be. GENESIS instead generates single-cell expression distributions directly from covariates such as cell line or donor, drug or cytokine, dose and batch, training a lightweight multilayer perceptron with the energy score and needing no control cells at inference. On 530 repeated Tahoe-100M plate pairs its Pearson Δ was 0.430 versus 0.432 for experimental repeats and its energy distance 0.158 versus 0.154, neither significantly different; top-100 DEG overlap was 0.289 versus 0.299, 3.4% short.
Interpretation
Predicting single-cell perturbation responses no longer starts from measured control cells; the model generates whole populations of expression profiles directly from the variables describing an experimental condition. Earlier generative approaches framed the problem as a mapping from a measured population of control cells to a perturbed population, yet single-cell sequencing is destructive, so control and perturbed cells are different cells and that pairing is never directly observed. Across 530 Tahoe-100M plate pairs measuring the same condition, GENESIS reached Pearson Δ 0.430 (95% CI 0.381–0.482) versus 0.432 (0.378–0.491) for experimental repeats, p = 0.56; energy distance 0.158 (0.147–0.170) versus 0.154 (0.143–0.166), p = 0.36; top-100 DEG overlap 0.289 (0.263–0.317) versus 0.299 (0.271–0.330), p = 8×10⁻⁴, 3.4% short.
The model splits each condition into biological context, perturbation and experimental replication context, learns an embedding for each covariate value, concatenates them with noise drawn independently per cell and passes the result through a multilayer perceptron that outputs one cell's gene-expression profile. Training uses the energy score, a strictly proper scoring rule whose unique optimum over observations is the true distribution, so the conditional distribution is learned directly from individual observed cells without paired cells or fixed-size population targets. The loss adds a variogram term that reproduces the average magnitude of expression differences between gene pairs within cells, capturing gene–gene dependence; repeated noise draws generate populations of any size.
On all held-out conditions GENESIS performed best among the compared methods on both atlases, and still beat competing models trained on the full dataset when using only about 5% of perturbed training cells. On Tahoe-100M, top-100 DEG overlap rose from an average of 30.7 genes (95% CI 29.7–31.6) for the next-best method, STATE, to 51.3 (50.5–52.1) for GENESIS, a 67% increase; on the PBMC cytokine atlas it rose from 37.3 (34.5–40.2) to 41.7 (38.8–44.6), a 12% increase. With about 5% of perturbed training cells (4.6 million of 89.5 million), GENESIS raised top-100 DEG overlap by roughly 50% and Pearson Δ by 30% over the strongest full-data competitor (STATE), and cut energy distance by roughly 60% over the context-mean baseline.
Perspective
Evaluation covers the Tahoe-100M small-molecule atlas and the PBMC cytokine atlas; the former supplies an experimental-repeat reference from 530 repeated plate pairs, while the latter, with each donor on its own plate, supports only a split-half reference. Predictions are limited to combinations of covariate values seen during training.
GENESIS generalizes compositionally, so it cannot yet predict responses to a completely unseen compound, cell line or donor; treatment duration, culture medium and cellular environment shape responses but are often unrecorded in perturbation datasets and therefore cannot be modeled explicitly.