Table 1.

Summary of existing virtual KO methods and feature comparison with GenKI

Name	Input data required	Method	Supervised / unsupervised	Description	Reference
scGen	scRNA-seq (WT and KO samples)	Transfer learning	Supervised	Train a variational autoencoder that learns to generalize the response of the cells in the training set of perturbations	(5)
CPA	scRNA-seq (KO samples)	Generative modeling	Supervised	Train an autoencoder with adversarial that decomposes the data into a collection of embeddings associated with the cell type, perturbation, and other external covariates to study combinatorial genetic perturbation	(6)
CellOracle	scRNA-seq and scATAC-seq (WT sample)	Graph-based modeling	Unsupervised	Simulate gene expressions in response to transcription factor (TF) perturbation by signal propagation through an inferred gene regulatory network	(7)
scTenifoldKnk	scRNA-seq (WT sample)	Manifold alignment	Unsupervised	Simultaneously project inferred WT and virtual KO gene regulatory networks to a joint low dimensional space	(8)
GenKI	scRNA-seq (WT sample)	VGAE	Unsupervised	Train a VGAE model that learns the latent gene representations of WT sample and virtually construct a virtual KO counterpart to discern similarity	This study