Cancer cell state map built with XDec-SM deconvolution.

(A) The XDec-SM algorithm contains three stages (Stage 0, 1, 2). Stage 0 is generalized to utilize a set of reference RNA-seq or scRNA-seq profiles. Reference scRNA-seq profiles with user provided cell-type classifications are utilized to create pseudo bulk reference profiles by summing every five profiles ordered by total gene coverage. The reference RNA-seq profiles or pseudo bulk profiles are transformed to a 0–1 range to identify differentially expressed genes representing different cell types. The subset of informative genes with distinct expression levels across the groups of references are utilized in Stage 1. In Stage 1, using the gene expression profiles of complex samples transformed to 0–1, XDec-SM estimates the gene expression profiles (transformed counts) of constituent cell types and the proportions of constituent cell types in each sample. XDec-SM uses an iterative algorithm for constrained matrix factorization using quadratic programming. The estimated proportions of constituent cell types in each sample are used in Stage 2. In Stage 2, using the untransformed gene expression profiles of the same complex samples and the proportions estimated in Stage 1, XDec-SM estimates the gene expression profiles (counts) of constituent cell types. Stage 2 uses non-negative least squares optimization using quadratic programming. (B) Cancer cell state map of breast cancer. The cancer cell state map of breast cancer acts as a framework to determine a tumor’s cancer cell state (layer 1 bottom). The map is also indicative of a tumor’s microenvironment (layer 2), similarity to model organisms including murine models and cell lines (layer 3), and the activation of cancer associated pathways (layer 4 top). Figure components created with BioRender.com.