FIG. 4.

An overview of DEEPsc training and inference. (a) Maseda et al. find the common genes in both spatial and scRNAseq data and perform dimensionality reduction on each data modality (with the final matrices having the same number of features). (b) During the training, DEEPsc uses spatial expression to “simulate” single-cell gene expression vectors. More specifically, every feature vector from the spatial expression is concatenated with all other vectors (labeled as “non-match”) and also itself (labeled as “match”) to form the input data to the neural network. (c) During inference, the scRNAseq feature vectors are concatenated with all spatial feature vectors, where the model should place a high probability for locations where the gene expression could have originated from. This figure was obtained from Maseda et al.⁶¹ [Front. Genet. 12, 348 (2021). Copyright 2021 Authors; licensed under a Creative Commons Attribution [CC-BY] license).