Single-cell perturbation prediction stops chasing the average: scDisentangle keeps each cell's identity in its counterfactuals
Synopsis
Single-cell perturbation-response prediction has struggled to beat simple mean baselines, and population-average metrics cannot show whether a prediction keeps the source cell's identity. scDisentangle splits each cell's expression into attribute embeddings and a sequence of residual factors extracted one after another, then generates counterfactuals at inference by replacing only the attribute embeddings while holding cell-specific factors fixed. On interferon stimulation, liver infection and unseen combinatorial genetic perturbations it beat the models and mean baselines compared in each benchmark, and in a myocarditis cohort it predicted donor-specific heart-infiltrating CD8+ T cell states from peripheral blood.
Interpretation
Single-cell perturbation-response prediction can now yield counterfactual expression profiles that retain the source cell's identity, rather than only a population-average perturbation shift. Earlier vector-arithmetic methods assumed perturbation effects are uniform across all cells or within predefined cell types, while disentanglement methods risked removing context-dependent biological variation. On the PBMC interferon-stimulation and mouse liver infection benchmarks, scDisentangle outperformed scGen, CPA, biolord, scDisInFact and mean baselines on correlation and distribution metrics; on the Norman combinatorial CRISPRa benchmark it was compared with CPA, GEARS and the match-mean and perturb-mean baselines.
The method encodes each cell's expression into an initial latent representation, then splits it into attribute-related embeddings and 16 scalar residual factors. Each successive factor fits only the residual left unexplained by the attributes and earlier factors, so shared information between factors is suppressed by construction without adversarial training or explicit independence penalties. Counterfactuals are generated by replacing the target attribute's embedding and class-mean baseline, holding cell-specific factors fixed and decoding; on the Kang data its factors show lower pairwise mutual information than CPA and scDisInFact.
Prediction accuracy no longer rises systematically as a perturbation response comes closer to the average perturbation effect. CPA and GEARS performance correlated positively with systematic variation (r = 0.40 and 0.21), meaning they do better when a response resembles the average effect. Across 128 Norman combinatorial perturbations, scDisentangle's perturbed-reference Pearson delta showed no significant association with systematic variation (r = 0.08, P = 0.38).
Perspective
The demonstrated scope is counterfactual prediction on single-cell expression data with annotated attributes: interferon-stimulated PBMCs, mouse liver infection annotated by spatial zone, Norman combinatorial CRISPRa perturbations, paired peripheral blood and cardiac data from immune checkpoint inhibitor-associated myocarditis, and a mouse prostate atlas spanning the castration-regeneration cycle. The intended reader is a single-cell researcher who needs cell-level and donor-level perturbation prediction; predictions are validated at the population level with correlation, distributional and RMSE-type metrics.
In the liver data, spatial zone and batch are not fully separable, so zone-specific effects should be read with that in mind. Prostate transition estimates rest on stages containing relatively few L2 cells and on RNA velocity inference, and are best read as a hypothesis about the timing of luminal progenitor expansion. Single-cell RNA sequencing is destructive, so per-cell counterfactuals have no truly matched ground truth and validation stays at the population level. The coarse-to-fine latent ordering is described as a property of the trained models; whether it extends to tissues with more complex hierarchies remains open.