Unsupervised spatially embedded deep representation of spatial transcriptomics

Table 1 Summary of features of the methods for detecting spatial domains. Compared to other methods, SEDR allows the implementation of more types of data and provides more information for downstream analyses, including latent representation and de-noised feature values. In addition, it uses GPU to accelerate calculations

Methods	Model	Resolution	Latent representation	De-noising	Batch integration	Programming	GPU
Seurat	Principal component analysis	Spot or single cell	√	×	√	R	×
SpatialLDA	Latent Dirichlet allocation	Single cell	√	×	×	Python	×
Giotto (HMRF)	Hidden Markov random field	Spot or single cell	×	×	×	R	×
stLearn	Spaital morphological gene expression normalization	Spot or single cell	√	×	√	Python	√
SpaGene	Spatial network (KNN)	Single cell	√	×	×	R	×
SpaGCN	Graph convolutional network	Spot or single cell	×	×	√	Python	×
BayesSpace	Bayesian model with a Markov random field	Spot	×	×	×	R	×
DeepST	Variational graph autoencoder	Spot or single cell	√	×	√	Python	√
STAGATE	Graph attention autoencoder	Spot or single cell	√	×	√	Python	√
UTAG	Graph + clustering	Single cell	×	×	√	Python	×
SEDR	Variational graph autoencoder + masked self-supervised	Spot or single cell	√	√	√	Python	√

ISSN: 1756-994X