TABLE 1.

Some examples of representation learning methods reviewed in the text

Method	References	Loss type	Function f	Batch treatment	Pretraining	Counts
PCA	Pearson 1901	Reconstruction	Linear	None	None	Log-norm.
GLM-PCA	Townes et al. 2019	Reconstruction	Implicit	None	None	Raw
LEMUR	Ahlmann-Eltze and Huber 2025	Reconstruction	Linear	Conditional	None	Log-norm.
scVI	Lopez et al. 2018	Reconstruction	Fully connected	Conditional	None	Raw
scArches	Lotfollahi et al. 2022	Reconstruction	Fully connected	Conditional	Pretrained	Raw
scMAE	Fang et al. 2024	Imputation	Fully connected	None	None	Log-norm.
CLEAR	Han et al. 2022	Contrastive loss	Fully connected	None	None	Log-norm.
CLAIR	Yan et al. 2023	Contrastive loss	Fully connected	Overlap	None	Log-norm.
scConcept	Bahrami et al. 2025	Contrastive loss	Transformer	None	Pretrained	Ranked
Geneformer	Theodoris et al. 2023	Imputation	Transformer	None	Pretrained	Ranked
scGPT	Cui et al. 2024	Imputation	Transformer	None	Pretrained	Binned
STATE	Adduri et al. 2025	Reconstruction	Transformer	Conditional	Pretrained	Log-norm.

For an explanation of the columns, see “A Comparative Taxonomy.” Examples were chosen to include well-known methods and also to cover a broad range of modeling choices.

Representation learning of single-cell RNA-seq data