The Cancer Genome Atlas (TCGA), Chinese Glioma Genome Atlas (CGGA), and Gene Expression Omnibus (GEO) datasets were mainly used for the follow-up study. Ivy Glioblastoma Atlas Project (GAP) dataset (28) and Gill dataset (29) were included.

Copy number alternations (CNAs) associated with CD44 expression were analyzed using GISTIC 2.0.

The immune infiltrating cells were quantified by MCPcounter algorithm (30), ssGSEA algorithm (31), and TIMER algorithm (32). ESTIMATE algorithm (33) was used for calculating the microenvironment scores. The cancer immunity cycle was conducted using get set variation analysis (GSVA) (34, 35). 114 metabolic pathways from the previous literature were quantified using GSVA (36).

The scRNA-seq matrix of GSE138794 (37) was conducted using the R package Seurat. The detailed procedure was reported in our previous study (38). “FeaturePlot”, “Dimplot”, and “VlnPlot” were applied to visualize the expression of CD44. Single-cell pseudotime trajectories analysis was processed using the R package Monocle (39). Gene ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses were performed.

The western blotting assay evaluated the expression level of RUNX1, CD44, PD-L1, and β-actin. Anti-RUNX1 (ab240639, Rabbit, 1:1000, Abcam, UK), anti-CD44 (15675-1-AP, Rabbit, 1:2000, Proteintech, China), anti-PD-L1 (66248-1-Ig, Mouse, 1:2000, Proteintech, China), and anti-β-actin (66009-1-Ig, Mouse, 1:5000, Proteintech, China) were applied as the primary antibody. HRP goat anti-mouse IgG (SA00001-1, Mouse, 1:5000, Proteintech, China) and HRP goat anti-rabbit IgG (SA00001-2, Rabbit, 1:6000, Proteintech, China) were applied as the secondary antibody.

The primers of GADPH (F ACAGCCTCAAGATCATCAGC; R GGTCATGAGTCCTTCCACGAT) and CD44 (F CAGCTCATACCAGCCATCCA; R TGGGGTGTGAGATTGGGTTG) were designed by the primer 5.0. Total RNAs were extracted and reversely transcribed into cDNA by HiScript Q RT SuperMix for RT-qPCR. GADPH and CD44 expression levels were conducted using 2-ΔΔCT.

U251 cells were cultured in Dulbecco’s modified eagle medium (DMEM), with 10% fetal bovine serum (FBS) in the saturated humidity incubator (37°C, 5% CO2). The RUNX1 specific primers for overexpression (OV) plasmid were as follows: F CACACTGGACTAGTGGATCCCGCCACCATGGCTTCAGACAGCATATTTG; R AGTCACTTAAGCTTGGTACGTAGGGCCTCCACACGGCCTCCTC. The RUNX1-OV plasmid was transfected into U251 cells via Lipofectamine 2000. The short hairpin (sh)-RNA target sequences of CD44 (93684-1, GACCTCTGCAAGGCTTTCAAT; 93685-1, CTGCCGCTTTGCAGGTGTATT; 93686-11, GAGCATCGGATTTGAGACCTG) were used for constructing the lentivirus. RUNX1-OV and RUNX1-NC groups of U251 cells were transfected with a lentivirus vector and cultured in DMEM with puromycin for stable CD44-knockdown U251 cells.

The RUNX1-NC, RUNX1-OV, RUNX1-OV+CD44-shRNA, and RUNX1-NC+CD44-shRNA groups of U251 cells were seeded in 24-well plates, with a density of 1×10⁴ cells per well. 30 ul CCK8 reagent was diluted in 270 ul DMEM and subsequently added to each well. Later, U251 cells were cultured for 0h, 24h, 48h, and 72h. The absorbance was measured at 450 nm.

The RUNX1-NC, RUNX1-OV, RUNX1-OV+CD44-shRNA, and RUNX1-NC+CD44-shRNA groups of U251 cells were plated in 6-well plates for 2 weeks, with a density of 200 cells per well. The colonies were fixed with 4% methanol (1 ml per well) for 15 min and stained with 0.5% crystal violet. Representative images of each well were collected.

500 ul DMEM with 10% FBS was added to the lower chamber of a 6-well plate. The RUNX1-NC, RUNX1-OV, RUNX1-OV+CD44-shRNA, and RUNX1-NC+CD44-shRNA groups of U251 cells were digested and resuspended with a density of 2x10⁶ cells/ml using DMEM. 100 ul U251 cells were added to the upper chamber and cultured for 48h. The upper chamber was stained with 0.5% crystal violet for 5 min. Representative images of each well were collected.

The sections of different grades of gliomas were collected. Anti-CD44 (15675-1-AP, Rabbit, 1:500, Proteintech, China) was used as the primary antibody. The horseradish peroxidase-conjugated antibody (ZSGB-BIO, PV-9000, China) was used as the secondary antibody. The sections were stained with 3, 30-diaminobenzidine tetrahydrochloride (DAB) and counterstained with hematoxylin.

Spearman or Pearson correlation analysis was used to evaluate the correlations between continuous variables. For normally distributed variables, significant quantitative differences between and among groups were determined by a two‐tailed t-test or one‐way ANOVA, respectively. For nonnormally distributed variables, significant quantitative differences between and among groups were determined by a Wilcoxon test or a Kruskal–Wallis test, respectively. All statistical analyses were conducted by the R project (https://www.r-project.org/). It is considered to be statistically significant when P-values < 0.05. All tests were two-sided.

We delineated the pattern of CD44 mRNA expression levels regarding disparate WHO-grade gliomas by evaluating data from publicly available databases. Firstly, we checked the expression pattern of CD44 in WHO grades, and then subtypes were divided into LGG and GBM. Not surprisingly, the CD44 expression was observed to be significantly increased in higher WHO grades according to datasets from the CGGA and TCGA ( Figure S1A ). Also, compared to LGG patients, GBM groups showed higher levels of CD44 expression, which demonstrated an adverse role of CD44 in the proliferation and aggressive progress of gliomas. ( Figure S1B ). The representative images of CD44 in immunohistochemistry staining of different grades of glioma are shown in Figure S1C .

Moreover, we tested CD44 in two important gene mutation cohorts extensively used as clinical outcome prediction tools. Interestingly, the IDH-mutated gliomas, linked with better clinical outcomes (40), were associated with a lower CD44 level in pan-gliomas and GBM cases ( Figures S1D, E ). Furthermore, the ROC curve demonstrated that the value of CD44 is an effective predictor of IDH mutation in GBM and pan-gliomas cases (AUC value = 0.789; AUC value = 0.773, respectively Figure S1F ). Similarly, analysis in 1p/19q non-codeletion pan-glioma showed up-expressed levels of CD44, which can predict the poor survival of glioma. Still, analysis in 1p/19q codeletion pan-glioma showed down-expressed levels of CD44 ( Figure S1G ). Importantly, in LGG cases, 1p/19q codeletion and IDH mutation are correlated with lower expression levels of CD44 ( Figure S1H ). At the same time, we tested the sensitivity of CD44 as a prognostic biomarker in predicting the survival of 1 year, 3 years, and 5 years, respectively. ( Figure S1I ). We found the AUC of 1 year is 0.796, the AUC of 3 years is 0.767, and the AUC of 5 years is 0.766 according to the TCGA dataset, while the AUC of 1 year is 0.633, the AUC of 3 years is 0.687, and the AUC of 5 years is 0.688 according to the CGGA dataset. In addition, CD44 was found to be lower in expression levels in methylated glioma in the TCGA dataset ( Figure S2A ). The specific methylation probe significantly associated with gliomas regarding CD44 expression is displayed in Figure S3 . The levels of CD44 expression in the histology of glioma tissues and various cell lines are shown in Figure S2B .

There are three molecular categories with distinct sub-classes in human gliomas: classical (CL), proneural (PN), and mesenchymal (MS). Among these three sub-classes, subtypes of CL and MS show more aggressiveness in behavior (10, 41). To delineate the inter-tumor heterogeneity of CD44, we studied it among four molecular sub-classes according to the VERHAAK_2017 classification scheme (42). Based on the TCGA dataset, GBM and pan-gliomas of CL and MS subtypes had significantly higher expression levels of CD44 compared to subtypes of PN ( Figures S2C, E ). Furthermore, the ROC curve demonstrated that CD44 might play a role as a predictor of subtypes of CL and MS in GBM and pan-gliomas cases (AUC value = 0.833; AUC value = 0.782, Figures S2D, F ).

To detect the intra-tumor distribution of CD44 expression in glioma tissues, we examined CD44 expression levels in disparate structures. According to the Ivy Glioblastoma Atlas Project data, in peri-necrotic zones, pseudo palisading cells around necrosis, cellular tumors, and hyperplastic blood vessels, CD44 was highly expressed ( Figure S2G ). Interestingly, in radio graphical regions, different compositions were discovered in the T1 contrast-enhancing (CE) regions, compared with the non-contrast-enhancing (NCE) margins. The results of RNA sequencing analysis illustrated that CE regions have higher CD44 expression compared to NCE or normal tissue (NT) areas ( Figure S2H ).

Consequently, we evaluated prognostic value concerning the different CD44 levels in gliomas through Kaplan-Meier analysis. Compared to CD44^low patients, CD44^high patients showed significantly shorter OS in GBM, pan-gliomas, and LGG samples from TCGA and CGGA datasets ( Figures 1A–F ). Further, based on the univariate and multivariate Cox regression analysis, CD44 showed the potential as an independent prognostic factor with IDH, 1p19q, MGMT, and Subtype ( Figures 1G, H ).

After all our work, the prediction value of CD44 as a biomarker in clinical use still needs to be examined. We implied a prognostic nomogram model developed by CD44 expression and clinicopathological risk factors to test its prediction in the clinical prognosis of gliomas ( Figure S4A ). Nomograms, a widely used graphical statistical model in cancer, are formed by a statistical algorithm using a combination of all predictive factors that have been admitted to affect the probability of a clinical event. Calibration plots revealed nomograms’ potential to predict patients’ survival based on an ideal model ( Figure S4B ). Kaplan-Meier survival curve demonstrated the statistical significance of the survival difference between high and low-risk patients ( Figure S4C ). The AUC of 4 years is 0.865 in the TCGA dataset ( Figure S4D ).

PRISM and GDSC databases were used for drug prediction based on CD44 expression. The top correlated drugs in GDSC1 are shown in Figure S5A . The top correlated drugs in GDSC2 are demonstrated in Figure S5B . The top correlated drugs in PRISM were shown in Figure S5C . Notably, AZD6482, Bicalutamide, Bortezomib, Cyclopamine, Dasatinib, Lapatinib, Paclitaxel, Parthenolide, Pazopanib, Rapamycin, and Sorafenib were found with lower estimated IC50 in CD44^high glioma patients ( Figure S5D ).

Two central cells consist of solid tumors: stromal cells and cancer cells. At the same time, stromal cells and infiltrating immune cells build up most of the tumor microenvironment. First, the correlation between ESTIMATE scores and CD44 expression was explored. Data demonstrated a significantly positive association between CD44 expression level and stromal score, immune score, and ESTIMATE score according to the TCGA dataset ( Figure 2A ), indicating that changing stromal and infiltrating immune cells, CD44 can affect the glioma microenvironment. These results also inspired us to believe that there might be a specific role for CD44 in glioma development and migration. To further study the relationship between high CD44 levels in gliomas and tumor immunity function, we performed CIBERSORT, MCPcounter, and TIMER analyses to determine whether the typical immune cells responsible for the immune responses against tumors or participating in the inflammatory activity are related to high CD44 expression. We discovered the positive association between CD44 and multiple important infiltrating immune cell types like monocytes, macrophages, Myeloid-derived suppressor cells (MDSC), neutrophils, and natural killer (NK) cells in the TCGA dataset ( Figure 2A ). Altogether, our data revealed that CD44^high patients tended to increase immune cells infiltrating within their tumor microenvironment.

Then we turned to our most exciting part and evaluated the possible immune-related functions of CD44 in glioma cell growth in the TCGA dataset. Based on the cancer immunity cycle, CD44 was positively associated with the release of cell antigens, priming and activation, recruiting of CD4 T cell, CD8 T cell, Th1 cell, Th2 cell, Th17 cell, Th22 cell, Treg cell, B cell, eosinophil, neutrophil, dendritic cell, macrophage, NK cell, and the infiltration of multiple immune cells ( Figure 2B ).

Besides, CD44 was positively associated with several metabolic pathways, including glycogen biosynthesis, galactose metabolism, glycosaminoglycan biosynthesis, N-glycan biosynthesis, and starch and suctose metabolism ( Figure 2B ), indicating a highly activated tumor microenvironment.

To further evaluate the role of CD44 in immune infiltrating, the expression of CD44 in gliomas was also analyzed by scRNA-seq. The expression of CD44 in all eight identified cell types is visualized in Figure 3A . CD44 was richly expressed in neoplastic cells, macrophages, astrocytes, and T cells ( Figure 3B ). Additionally, the distribution of eight cell clusters was visualized by a three-dimension plot ( Figure 3C ). The correlation of CD44 and neoplastic cells, macrophages, astrocytes, and T cells is confirmed ( Figure 3D ).

The single-cell pseudotime trajectories and functional annotations of neoplastic cells and macrophages were explored. In neoplastic cells, a trajectory was reconstructed, containing three branch points and grouped cells into seven states ( Figure 3E ). In macrophages, a trajectory was reconstructed, containing five branch points and grouped cells into eleven states ( Figure 3F ). The relative expression level of CD44 in the cell states from neoplastic cells and macrophages is presented in Figures 4A , 5A , respectively. 100 genes were further identified with branch-dependent expression for branch point 3 of neoplastic cells. The expression of the different genes before and after branch point 3 and related clustering are presented in Figure 4B . Moreover, 100 genes that differentially expressed with branch-dependent expression for branch point 3 of macrophages were also confirmed ( Figure 5B ). Finally, in regards to CD44 in neoplastic cells and macrophages, the results of GO enrichment analysis and KEGG pathway analysis was presented in Figures 4C , 5C .

To investigate the potential mechanisms of the pathogenic role of CD44 in gliomas, in vitro validation was performed. A strong positive correlation between CD44 and RUNX1 was observed in the TCGA dataset ( Figure 6A ). Lentivirus targeting CD44 was transfected into U251 cells to select the shRNA (93684-1) with the highest knockout efficiency ( Figure 6B ). The plasmid of RUNX1 was further transfected into U251 cells for the overexpression of RUNX1. Four groups, RUNX1-NC, RUNX1-OV, RUNX1-OV+CD44-shRNA, and RUNX1-NC+CD44-shRNA, were used for the follow-up experiments. The western blotting assay revealed that the overexpression of RUNX1 promoted the expression of CD44 with statistical significance ( Figures 6D, E ). Notably, the silence of CD44 suppressed the expression of RUNX1 with statistical significance. CCK8 assay revealed that CD44 inhibition could decrease the U251 cell proliferation. At the same time, the up-regulated levels of RUNX1 alleviated the reduced accumulation of U251 cells ( Figure 6C ). Clone formation assay showed that the clone formation ability of U251 cells decreased under CD44 inhibition. In contrast, the overexpression of RUNX1 alleviated the decreased clone formation capacity of U251 cells ( Figure 6F ). Transwell assay revealed that CD44 inhibition could reduce the migration of U251 cells. In contrast, the up-regulated levels of RUNX1 alleviated the reduced migration of U251 cells ( Figure 6G ). A strong positive correlation between CD44 and PD-L1was observed in the TCGA dataset ( Figure 6A ). Furthermore, western blotting assay revealed that RUNX1/CD44 could promote the expression of PD-L1 ( Figures 6D, E ). These results suggested that RUNX1/CD44 axis could mediate the proliferation, migration, and immunotherapy of glioma.

CD44 was found with AUC values of larger than 0.6 in predicting the immunotherapy responses in 12 immunotherapy cohorts ( Figure 7A ). Compared to CD44^low patients, CD44^high patients showed shorter OS in the IMvigor210 cohort, Liu cohort, Braun cohort, Zhao cohort, and VanAllen cohort ( Figure 7B ). Compared to CD44^low patients, CD44^high patients showed longer OS in the Gide cohort, Lauss cohort, and Gide cohort ( Figure 7B ).