a Malignant cell clusters (CC0-CC5, top) and samples (bottom) are shown in t-SNE plots. Paired primary and metastatic tumors are indicated (dotted circle). b The hierarchical clustering of the malignant clusters shows a similar cell proportion of the paired primary and metastatic tumors (underlined). c The cumulative cell proportion of the malignant cell clusters is shown in TCGA-HNSCC (n = 500). d The HPV infection status is indicated in a t-SNE plot (top). Cell counts of the HPV-positive and HPV-negative cells are shown across the malignant cell clusters (bottom). e According to the tumor subsites for all the patients (ALL, n = 500), OP (n = 71), and OC (n = 308), the proportions of the CC0 and CC1 malignant cell clusters in the HPV-positive and HPV-negative tissues are shown. f A dot plot shows the average expression levels of the single-cell malignant programs of HNSCC, including hypoxia, stress, epithelial differentiation, p-EMT, and cell cycle in each malignant cluster. g Forest plots show the hazard ratios for OS and RFS between the high- and low-proportion groups (>70th percentile) for each malignant cell cluster in the deconvoluted TCGA-HNSCC data (top). CC2 is not shown because its 70th percentile of cell proportion is zero. Hazard ratios and 95% confidence intervals (CI) are indicated. *,P < 0.05 and **,P < 0.01, log-rank test. Kaplan–Meier’s plot analyses for OS (left) or RFS (right) between the CC1-high and CC1-low groups are shown in TCGA-HNSCC, GSE41613, and GSE42743, respectively (bottom). P = 0.02 (OS; TCGA-HNSCC), P = 0.006 (RFS; TCGA-HNSCC), P = 0.003 (OS; GSE41613), P = 0.016 (OS; GSE42743), log-rank test. Follow-up time for OS and RFS is truncated to 5 years. h Expression levels of LGALS7B in the malignant cell clusters are shown. i Representative images of immunohistochemical staining for LGALS7B expression are shown in the CC1-high and CC1-low samples. n = 1. Scale bar, 100 µm. In e, h, box plots show the median (center line), the upper and lower quantiles (box), and the range of the data (whiskers). Source data are provided in a Source Data file.