Fig. 1

Schematic of unsupervised alignment and state variation mapping with scAlign. (a) The input to scAlign consists of cells sequenced across multiple scRNA-seq conditions. Expression can be represented as either gene-level expression or embedding coordinates from dimensionality reduction techniques such as PCA or CCA. (b) A deep encoding network learns a low-dimensional alignment space that simultaneously aligns cells from all conditions. (c) Paired decoders project cells from the alignment space back into the gene expression space of each condition and can be used to interpolate the expression profile of cells sequenced from any condition into any other condition. (d) For a single cell sequenced under any condition, we can calculate its interpolated expression profile in all conditions, then measure the predicted variance across all input conditions to calculate a state variation map for the same cell state under different conditions to identify cells whose expression profiles vary significantly across condition