Charting the computational landscape of single-cell genetic perturbation

Single-cell genetic perturbation screens are increasingly used to study gene function at single-cell resolution. However, their analysis remains difficult because available computational methods differ in assumptions, input requirements, and output definitions, which can complicate cross-study comparison and consistent tool selection.

We summarize current methods by organizing single-cell genetic perturbation analysis into a five-step computational workflow: basic data preprocessing, data refinement and correction, perturbation-specific quality control and filtering, effect quantification and network inference, and functional annotation and visualization.

Within this structure, we group current algorithms by modeling principle, including regression and testing methods, latent-state and assignment-error models, matrix or manifold-based methods, and latent-variable or generative models. We compare their assumptions, limitations, and use cases to inform method selection. We also summarize strategies for multimodal perturbation analysis, in which transcriptomic, proteomic, epigenomic, spatial, and kinetic readouts affect multiple steps of the workflow. Finally, we review perturbation prediction models, from graph-based approaches to large-scale foundation models, with attention to their input requirements and practical limitations. By linking workflow steps with model assumptions and input requirements, this review provides a basis for method reporting and interpretation in single-cell perturbation studies.