Extended Data Fig. 2: In vivo programs of variability and comparison to in vitro cell cycle programs.

a, Heatmap depicts pairwise similarities between programs identified in tumor samples using NMF. Programs with limited similarity to all other programs were excluded. Top panel shows tumor type and correlations between program scores and cell complexity (that is number of genes detected per cell). Hierarchical clustering emphasizes multiple clusters (shown by squares), one of which is correlated with cell complexity and thus excluded as a potential technical artifact. b, NMF scores of G1/S genes (top panel) and G2/M genes (bottom panel) across all NMF programs associated with the corresponding cell cycle phase; each program is from a different cell line. Genes are ranked in each panel by average scores, and their assignment to in vitro and in vivo cell cycle programs is indicated in the right bar, demonstrating that G1/S programs differ both across cell lines and between cell lines and tumors, while G2/M programs are more consistent. Venn diagrams (right) illustrate the overlap of genes between in vivo and in vitro RHPs. c, Single-cell profiles (n = 264 cells from NCIH2126) showing G1/S and G2/M program score thresholds used to assign cells to different cell cycle phases. d, Examples of genes with distinct cell cycle upregulation in vitro and in vivo. Expression of HIST genes (preferentially induced in vitro) and MCM genes (preferentially induced in vivo) is shown along the cell cycle (relative to cells in G0) in cell lines (C, green lines) and tumors (D, yellow lines). e, Comparison of cell cycle phase distribution in vitro and in vivo. Scatterplot shows the percentage of cells in G0 (x-axis) and the ratio between the percentage of cells in G1/S and G2/M (y-axis) for each cell line (green, n = 198) and tumor (yellow, n = 25) analyzed. Cell lines display a significantly lower percentage of cells in G0 cells (P=2e⁻¹⁰, two-sided t-test).