Fig. 2

Association of lymph node metastasis (LNM) signatures with epithelial dedifferentiation and dysplasia: A Heatmap showing linear regression z-scores for association of LNM gene signatures with tumor grade in 13 HNC bulk gene expression studies. Z-scores indicate the significance of associations between tumor grade and expression scores of LNM gene signatures within each study. Expression scores were calculated for each LNM gene signature (i.e., set of genes) as the mean of expression of genes within that signature. Row labels indicate meta-z-scores for the association of each signature with grade across studies, which were calculated by combining z-scores across studies using Liptak’s weighted meta-z test. LNM gene signatures include all genes that were negatively (anti-LNM) and positively (pro-LNM) associated with LNM, and genes within LNM gene clusters (L1-6). B Box plots showing mean expression of anti-LNM cluster 1 and pro-LNM cluster 4 genes within primary HNCs and tumor-adjacent normal tissue (normal) of the TCGA HNSC study, with HNCs stratified by tumor grade (G1-4). C Smoothed scatter plot showing the correlation between meta-z-scores for association of genes (points) with tumor grade (X-axis) and LNM status (Y-axis). Meta-z-scores for association of genes with grade and LNM were calculated based on separate meta-analyses. Dashed lines indicate meta-z-score significance thresholds (Absolute meta-z = 3.09). Regression lines (red dashed line) and Pearson correlation coefficient (r) are indicated. Text labels highlight genes that were among those with the 50 highest and lowest meta-z-scores for association with both LNM and grade (i.e., genes that were strongly association with both LNM and grade). D ScRNA-Seq analyses indicating the association of LNM signatures with epithelial differentiation and stemness within malignant cells. i Uniform Manifold Approximation and Projections (UMAPs) showing expression scores of LNM-associated gene signatures within the Puram scRNA-Seq dataset. Signatures shown include all genes that were negatively (anti-LNM) and positively (pro-LNM) associated with LNM, as well as genes within anti-LNM cluster L2 and pro-LNM cluster L4, the largest unsupervised clusters of LNM-associated genes. ii UMAPs corresponding to those shown in i, showing cell phenotypes including unsupervised cell cluster, cell type, and cell cycle phase, as well as CytoTRACE score, a measure of transcriptional diversity and stemness [70]. iii Heatmap showing correlations of LNM gene signatures with tumor plasticity signatures within the Puram and Stanford scRNA-Seq datasets. Points indicate correlations between expression scores within malignant cells of two scRNA-Seq datasets. Pearson correlation coefficients (r) is represented by the point color gradient, while point sizes represent negative log ten p-values (linear regression). Expression scores are calculated for each cell as the scaled mean expression (normalized counts) of all genes within a signature (i.e., set of LNM-associated genes). Tumor plasticity gene signatures shown include epithelial differentiation markers: Genes identified as part of an epithelial differentiation-related transcriptional program in HNC (referred to as “Epi dif. 1”) based on the original analysis of the Puram dataset [67], ESC markers: genes specifically expressed in embryonic stem cells (ESCs) [84], tumor grade-associated gene signatures: Genes positively (pro-grade) and negatively (anti-grade) associated with tumor grade in our meta-analysis, CytoTRACE [70] score (as described in ii), and EMT mesenchymal genes: mesenchymal genes used to calculated epithelial to mesenchymal (EMT) scores in this and previous studies [71, 72]. iv Scatter plots highlighting correlations between LNM signatures and a selection of the tumor plasticity signatures shown in in the heatmap in iii. These correlations are shown within malignant cells (points) of the Puram scRNA-Seq dataset. Point colors correspond to the unsupervised cell clusters shown in the UMAP in ii, illustrating the expression of gene signatures within specific malignant cell subpopulations. Pearson correlation coefficients (R) and regression lines (dashed lines) indicate the correlation between expression scores. Asterisks indicate linear regression p-values for the association of expression scores: ***p < 0.001. E Deregulation of LNM-associated genes associated with epithelial dysplasia in oral premalignant lesions (OPLs). Box plots showing mean expression of anti-LNM genes and pro-LNM gene cluster L4 in OPLs (n = 86), using a publicly available dataset [78]. OPLs are stratified based on their stage of premalignant disease, with deeper color indicating higher risk lesions with advanced dysplasia