Fig. 1.

The factorized embeddings (FE) model reconstructs data with high accuracy and preserves sample pair-wise distances. (A) Schema of the FE model. (B) Pair-wise Euclidean distance preservation between 1500 random pairs of samples in original feature space (gene expression) and reconstructed space. (C, D) The FE-trained representation preserves more accurately than t-SNE pair-wise distances between samples in the embedding space. (E) The FE model allows for precise imputation of transcriptomes on a patient-level