Extended Data Fig. 3: Revisiting malignant state definition, and contextualization with state-of-the-art classification schemes.

a, Jaccard similarity indices of the robust NMF programs (n = 835) based on their top 50 genes. Programs are ordered by clustering and grouped into meta-programs (Methods). MP10 (Stress1) and MP15 (Stress2) were coalesced into a single Stress state. MP9 (ExN) and MP14 (NRGN neuron, NRGN) were coalesced into a single NEU-like state. MP1 (Ribosomal protein, RP), MP11 (Doublet), MP12 (Low quality, LQ) reflect low quality and are not considered a state. b, Correlation between cell state scores of meta-programs derived from this dataset (columns) and those derived by Neftel et al. (rows). Each cell was scored within tumor-of-origin for both sets of meta-program signatures. Pearson’s correlation coefficient was computed across all cells for each pair of meta-programs. c, Jaccard similarity between meta-programs derived from this dataset (columns) and those derived by Neftel et al. (rows). d, Distribution of scores in this dataset (upper panel) and Neftel et al. (lower panel) dataset of cells that scored maximally for both corresponding states (for example to both AC-like signatures derived from this dataset and from Neftel et al.) across all corresponding states. Dashed vertical lines mark the mean score of each distribution. e, Association between the malignant MPs from our datasets (rows) and malignant signatures from other studies (columns). Each cell in the heatmap represents the Pearson correlation coefficient between the gene expression scores of an MP and other signature scores across the malignant cells in the dataset.