Aberrations in Notch-Hedgehog Signalling Reveal Cancer Stem Cells Harbouring Conserved Oncogenic Properties Associated with Hypoxia and Immunoevasion

www.nature.com/bjc

ARTICLE Molecular Diagnostics Aberrations in Notch-Hedgehog signalling reveal cancer stem cells harbouring conserved oncogenic properties associated with hypoxia and immunoevasion

Wai Hoong Chang1 and Alvina G. Lai1

BACKGROUND: Cancer stem cells (CSCs) have innate abilities to resist even the harshest of therapies. To eradicate CSCs, parallels can be drawn from signalling modules that orchestrate pluripotency. Notch-Hedgehog hyperactivation are seen in CSCs, yet, not much is known about their conserved roles in tumour progression across cancers. METHODS: Employing a comparative approach involving 21 cancers, we uncovered clinically-relevant, pan-cancer drivers of Notch and Hedgehog. GISTIC datasets were used to evaluate copy number alterations. Receiver operating characteristic and Cox regression were employed for survival analyses. RESULTS: We identified a Notch-Hedgehog signature of 13 genes exhibiting high frequencies of somatic amplifications leading to transcript overexpression. The signature successfully predicted patients at risk of death in five cancers (n = 2278): glioma (P < 0.0001), clear cell renal cell (P = 0.0022), papillary renal cell (P = 0.00099), liver (P = 0.014) and stomach (P = 0.011). The signature was independent of other clinicopathological parameters and offered an additional resolution to stratify similarly-staged tumours. High-risk patients exhibited features of stemness and had more hypoxic tumours, suggesting that hypoxia may influence CSC behaviour. Notch-Hedgehog+ CSCs had an immune privileged phenotype associated with increased regulatory T cell function. CONCLUSION: This study will set the stage for exploring adjuvant therapy targeting the Notch-Hedgehog axis to help optimise therapeutic regimes leading to successful CSC elimination.

British Journal of Cancer (2019) 121:666–678; https://doi.org/10.1038/s41416-019-0572-9

BACKGROUND their downstream targets could influence tumour progression and Tumours are far from homogeneous masses, yet many contem- prognosis. The role of Notch signalling in oncogenesis was initially porary therapies continue to treat them as such. It has become discovered in T cell acute lymphoblastic leukaemia.6 Since then, increasingly clear that a minor population of tumour cells known multiple studies on Notch signalling have demonstrated both as cancer stem cells (CSCs) contribute to treatment resistance as oncogenic and tumour suppressive functions in haematological they have the propensity to tolerate DNA damage1,2 and evade and solid malignancies, implying its pleiotropic nature that is very immune detection3 to give rise to new tumours post therapy. much dependent on cellular types.7 Hedgehog is a morphogen Identification of CSCs has remained a challenging endeavour since that regulates a signalling cascade involving the Smoothened they only make up a small proportion of the tumour and are protein to influence morphogenetic processes such as prolifera- histologically similar to non-stem cancer cells. Moreover, mole- tion and differentiation.8 Interactions between Notch and Hedge- cular markers that identify CSCs are often cancer-type dependent, hog signalling have been demonstrated in multiple cancers. Hes1, which limit their broad scale applications.4 CSCs share many a Notch effector, is targeted by sonic hedgehog in neural cells.9 qualities with embryonic or adult stem cells. For example, When Patched, a negative regulator of Hedgehog, is abrogated in activation of signalling pathways involved in coordinating cellular mice, this gives rise to medulloblastoma with enhanced Notch homoeostasis, morphogenesis and cell fate determination (TGF-β, signalling.10 Hedgehog signalling promotes the expression of Wnt, Notch and Hedgehog) are often seen in CSCs. These Jagged2 (a Notch ligand)11 and in ovarian cancer mice models, pathways rarely act in isolation and significant crosstalk between inhibition of Jagged1 would sensitise tumours to docetaxel them have been reported.5 treatment by affecting GLI2 function.12 Concurrent activation of In order to fully exploit these pathways for CSC therapy, pan- Hedgehog and Notch signalling was observed in prostate cancer cancer explorations are warranted to reveal conserved compo- cells that were resistant to docetaxel.13 Glioblastoma treated with nents that can be prioritised as therapeutic targets. Concentrating a Notch inhibitor was subsequently desensitised to further Notch on Notch and Hedgehog signalling pathways, we seek to attain a suppression as they upregulate Hedgehog signalling.14 comprehensive understanding of how somatic copy number These studies highlight the importance of Notch-Hedgehog alterations and expression profiles of pathway genes along with interactions in cancer, which calls for a better understanding of

1Institute of Health Informatics, University College London, 222 Euston Road, London NW1 2DA, UK Correspondence: Alvina G. Lai ([email protected]) Received: 21 January 2019 Revised: 13 August 2019 Accepted: 20 August 2019 Published online: 16 September 2019

© The Author(s), under exclusive licence to Cancer Research UK 2019 Aberrations in Notch-Hedgehog signalling reveal cancer stem cells. . . WH Chang and AG Lai 667 their relationship and also to reveal crosstalk with other pathways visualise samples’ distance (tumour and non-tumour) in reduced involved in regulating CSC function. Harnessing genomic and 2-dimensional space. The R vegan package was employed for MDS transcriptomic sequences of 21 cancer types, we performed a ordination using Euclidean distances. Permutational multivariate comprehensive investigation linking genomic alterations to analysis of variance (PERMANOVA) was used to investigate transcriptional dysregulation of Notch-Hedgehog pathway genes. statistical differences between tumour and non-tumour samples. We discovered conserved patterns of Notch-Hedgehog hyper- The linear model and Bayes method were employed for activation across cancers and revealed putative driver genes that differential expression analyses, followed by the Benjamini- were associated with CSC phenotypes underpinning poor clinical Hochberg false discovery rate method. Kaplan-Meier, Cox propor- outcomes. We also examined the relationship between the tional hazards and receiver operating characteristic survival tumour microenvironment (hypoxia and immune suppression) analyses were performed using R survminer, survival and and Notch-Hedgehog+ CSCs. In-depth knowledge of the Notch- survcomp packages. Hedgehog signalling axis afforded by this study will set the stage for exploring combinatorial chemotherapy targeting both path- Functional enrichment and transcription factor (TF) analyses ways simultaneously to potentially eradicate CSCs. Differential expression analyses as mentioned previously were performed on patients separated into quartiles 4 and 1 based on their 13-gene scores. Differentially expressed genes were mapped MATERIALS AND METHODS against KEGG and Gene Ontology (GO) databases using GeneCo- A total of 72 genes associated with Notch and Hedgehog dis25 to determine pathways that were enriched. The Enrichr tool signalling were retrieved from the Kyoto Encyclopedia of Genes was used to determine whether differentially expressed genes and Genomes (KEGG) database listed in Table S1. were enriched for stem cell TFs binding targets by comparing chromatin immunoprecipitation sequencing proﬁles from ChEA Study cohorts and ENCODE databases.26 We retrieved transcriptomic and genomic proﬁles of 21 cancer The R ggplot2 and pheatmap packages were used to generate types (n = 18,484) including their non-tumour counterparts from all plots. The Cancer Genome Atlas and Broad Institute GDAC Firehose15 (Table S2). RESULTS

1234567890();,: Somatic copy number alterations analyses Recurrently amplified driver genes associated with Notch and We retrieved Firehose Level 4 copy number variation datasets in Hedgehog activation in 21 diverse cancer types the form of GISTIC gene-level tables, which provided discrete To characterise the extent of Notch and Hedgehog signalling and amplification and deletion indicators.16 A sample was defined as identify common molecular subtypes, we examined somatic copy ‘deep amplification’ for values that were higher than the number alterations (SCNAs) and differential expression (tumour maximum median copy-ratio for each chromosome arm (+2). versus non-tumour) patterns of 72 genes in 18,484 cases of Samples with values less than the minimum median copy-ratio for clinically annotated stage I to IV samples representing 21 cancer each chromosome arm were called ‘deep deletions’ (−2). GISTIC types (Fig. 1a; Tables S1 and S2). We found that 70 out of 72 genes indicators of +1 and −1 represented shallow amplifications and were recurrently amplified in at least 20% of samples per cancer deletions respectively. type in at least one cancer type (Fig. 1a). Lung squamous cell carcinoma (LUSC) had the highest fraction of samples harbouring Calculating Notch-Hedgehog 13-gene scores, hypoxia scores and amplified Hedgehog genes, while endometrial cancer (UCEC) had regulatory T cell (Treg) scores the fewest somatic gains (Fig. 1b). When considering Notch gene The Notch-Hedgehog 13-gene signature was employed to calculate amplifications, LUSC also emerged as the top candidate while a score for each patient. It comprised of the following genes: JAG1, clear cell renal cell carcinoma (KIRC) had the fewest number of LFNG, DTX2, DLL3, GPR161, PSENEN, GLI1, HES1, PTCRA, DTX3L, Notch gene amplifications (Fig. 1b). In terms of focal deletions, this ADAM17, KIF7 and NOTCH1. Hypoxia scores were calculated from 52- was also the highest in LUSC for Hedgehog genes and renal hypoxia signature genes.17 Treg scores were calculated based on the chromophobe carcinoma (KICH) for Notch genes (Fig. 1b). overlap between four Treg signatures,18–21 consisting of 31 genes: Focusing on recurrently amplified genes, we identified 35 genes FOXP3, TNFRSF18, TNFRSF9, TIGIT, IKZF2, CTLA4, CCR8, TNFRSF4, IL2RA, (Hedgehog pathway: 13 genes; Notch pathway: 22 genes) that BATF,IL2RB,CTSC,CD27,PTTG1,ICOS,CD7,TFRC,ERI1,GLRX,NCF4, were gained in >20% of samples and in at least one-third of PARK7, HTATIP2, FCRL3, CALM3, DPYSL2, CSF2RB, CSF1, IL1R2, VDR, cancer types (>7 cancers) (Fig. 1c). GLI3, SMURF1, RBPJL, JAG1, ACP5 and MAGEH1. Scores were calculated from the average log2 LFNG and DTX2 were some of the most amplified genes present in expression values of 13, 52 or 31 genes representing Notch- at least 18 cancers (Fig. 1c). In contrast, KIF7, NOTCH1, MAML and Hedgehog, hypoxia and Tregs respectively. Kaplan–Meier analyses ADAM17 were the least amplified genes (Fig. 1c). LUSC had the of the Notch-Hedgehog signature were performed on patients highest number of amplified genes (34 genes) followed by 33 separated into quartiles based on their 13-gene scores. For analyses genes in oesophageal carcinoma (ESCA) and stomach and in Figs. 4–6, patients were separated into four groups using median oesophageal carcinoma (STES) and 32 genes in stomach 13-gene scores and median CSC transcription factor expression adenocarcinoma (STAD) and bladder urothelial carcinoma (BLCA) levels (EZH2, REST and SUZ12), hypoxia scores or Treg scores as (Fig. 1c). In contrast, only 8 genes were amplified in UCEC (Fig. 1c). thresholds for Kaplan-Meier and Cox regression analyses. Nonpara- SCNA events associated with overexpression could represent metric Spearman’s rank-order correlation tests were used to candidate driver genes since positive correlations between gene investigate the relationship between 13-gene scores and TF amplification and overexpression are indicative of a gain-of- expression levels, hypoxia scores or Treg scores. function.27 Differential expression analyses between tumour and adjacent non-tumour samples revealed that 13 of the amplified Multidimensional scaling, differential expression and survival genes were also significantly upregulated (>1.5-fold-change, analyses P < 0.05) in tumours of at least 7 cancer types (Fig. 1c). These As per the journal’s guidelines, we have not repeated methods genes were prioritised as a Notch-Hedgehog signature potentially here as we have previously published detail methods for representative of multiple cancers: JAG1, LFNG, DTX2, DLL3, multidimensional scaling (MDS), differential expression and GPR161, PSENEN, GLI1, HES1, PTCRA, DTX3L, ADAM17, KIF7 and survival analyses.22–24 Briefly, MDS analysis was employed to NOTCH1 (Fig. 1c). Aberrations in Notch-Hedgehog signalling reveal cancer stem cells. . . WH Chang and AG Lai 668 a 34 Hedgehog and 38 Notch pathway genes Prognostic cohorts

21 cancer types (n = 18,484) Glioma 415 Astrocytoma Renal clear cell 371 194 Oligoastrocytoma Papillary renal cell 669 270 290 Glioblastoma Liver 130 533 75 Oligodendroglioma Somatic copy number alteration (SCNA) Stomach

Tumour vs. non-tumour differential expression Total = 2278 patients Glioma

13 driver gene signature 70 genes amplified in >20% samples per cancer type Association of driver genes with: 35 genes amplified in >20% samples in at least 7 cancers Survival analysis: Kaplan–Meier, Cox regression Tumour hypoxia 13 genes upregulated >1.5 fold (Driver genes) and receiver operating characteristic Cancer stem cells

b Hedgehog Notch 1.00

0.75

0.50

0.25

0.00 Fraction of samples with SCNA LIHC LIHC GBM GBM KIRP KIRP KICH KIRC KICH KIRC STES STES BLCA BLCA LUAD LUSC STAD LUAD LUSC STAD PAAD PAAD ESCA ESCA SARC SARC BRCA CESC CHOL BRCA CESC CHOL HNSC UCEC HNSC UCEC COAD COAD KIPAN KIPAN GBMLGG GBMLGG

Deep amplification Shallow amplification No alteration Shallow deletion

GLI3 SMURF1 CUL1 SHH SMO SMURF2 GPR161 SPOP GLI1 Hedgehog BOC ARRB1 DHH KIF7

RBPJL JAG1 LFNG DTX2 DLL3 NCSTN PSENEN RFNG APH1A DTX3 PSEN2

Notch NOTCH2 HES1 NUMBL NOTCH4 PTCRA DTX1 NCOR2 DTX3L ADAM17 MAML1 NOTCH1 KIRP GBM GBMLGG KICH KIPAN KIRC COAD STAD STES SARC PAAD UCEC LUAD CHOL BRCA LIHC LUSC BLCA ESCA CESC HNSC CESC KICH GBM GBMLGG CHOL SARC LUAD LUSC UCEC BRCA KIRP BLCA ESCA LIHC KIPAN KIRC COAD PAAD HNSC STAD STES 0 5 10 15 20 Number of cancers

Somatic amplification Differential expression

0.2 0.4 0.6 0.8 1.0 –4 –2 0 2 4 Fraction amplified Log2 fold change

A 13-gene Notch-Hedgehog signature predicts survival outcomes proportional hazards regression analyses to test the prognostic in five cancers roles of individual Notch-Hedgehog signature genes across 20 Tumours displayed various degrees of somatic gains and over- cancer types where survival data is available. Prognosis appeared expression of Notch-Hedgehog pathway genes (Fig. 1), suggesting to tissue type-dependent (Fig. S1). All 13 genes were prognostic in that aberrant activation of these pathways may influence disease the glioma dataset (GBMLGG), consisting of samples from patients progression and survival outcomes. We employed univariate Cox with astrocytoma, oligoastrocytoma, oligodendroglioma and Aberrations in Notch-Hedgehog signalling reveal cancer stem cells. . . WH Chang and AG Lai 669 Fig. 1 Pan-cancer drivers of Notch-Hedgehog signalling. a Schematic diagram illustrating the study design and the identification of Notch- Hedgehog driver genes, which represent the 13-gene signature. SCNA and expression profiles of 72 Notch-Hedgehog pathway genes were interrogated in 21 cancer types involving 18,484 patients. We identified 70 genes as amplified in at least 20% of samples and 35 genes that were amplified in at least 20% of samples in at least 7 cancer types. Differential expression analyses between tumour and non-tumour samples revealed that the 13 recurrently amplified genes were also upregulated, potentially indicating a gain-of-function. These 13 genes were prioritised as a Notch-Hedgehog signature, which was prognostic in five cancer types involving 2278 patients. Associations of the signature with tumour microenvironmental features of hypoxia and immunity were also investigated. Pie slices indicate the number of samples within each cancer type. b Stacked bar graphs represent the proportion of samples in each cancer type with SCNA of Hedgehog and Notch pathway genes. Width of the bars reflect the number of samples within each cancer. c Somatic gains and differential expression profiles of 35 Notch- Hedgehog genes that were recurrently amplified in at least 7 cancer types (one-third of all cancers). Cumulative bar charts on the left represent the number of cancers with at least 20% of samples with somatic amplification. Heatmap on the left represents the extent of somatic gains for each of the 35 genes separated into Hedgehog and Notch signalling pathways across 21 cancers. Heatmap intensities depict the fraction of the cohort in which a given gene is amplified. Columns (cancer types) were ordered using Euclidean distance metric and hierarchical clustering to reveal cancers that were similar. Heatmap on the right represents tumour and non-tumour differential expression values (log2) for the 35 genes. Genes highlighted in red represent the 13 Notch-Hedgehog signature genes. Cancer abbreviations were listed in Table S2

glioblastoma multiforme (Fig. S1). A majority of the genes (9 out of prognostic in astrocytoma (P = 0.038), oligoastrocytoma (P = 13) were associated with poor prognosis (hazard ratio [HR] > 1, 0.0018) and glioblastoma multiforme (P = 0.045) (Fig. 3b). Patients P < 0.05) (Fig. S1). However, despite showing high frequencies of with low-grade gliomas stratified by the signature into the 4th SCNAs (Fig. 1c), none of the 13 genes harboured prognostic quartile had significantly higher death risks compared to those information in patients with LUSC, cholangiocarcinoma (CHOL) or within the 1st quartile: astrocytoma (HR = 2.535, P = 0.021), oesophageal carcinoma (ESCA) (Fig. S1). oligoastrocytoma (HR = 4.169, P = 0.014) and glioblastoma multi- We next considered all 13 genes as a group in assessing forme (HR = 2.163, P = 0.042) (Table S3). prognosis. For each patient, we calculated their 13-gene scores by To evaluate the predictive performance of the signature, we taking the average expression of all genes. Patients were employed receiver operating characteristic (ROC) analyses and separated into survival quartiles based on their 13-gene scores. compared area under the curves (AUCs) derived from the Remarkably, Kaplan–Meier estimates and log-rank tests revealed signature versus those derived from TNM staging. The signature that the 13-gene signature accurately predicted patients at higher had greater sensitivity and specificity in predicting 5-year overall risk of death in five cancer types (n = 2278): glioma (P < 0.0001), survival compared to TNM staging: papillary renal cell (AUC = clear cell renal cell (P = 0.0022), papillary renal cell (P = 0.00099), 0.796 vs. AUC 0.640) and stomach (AUC = 0.710 vs. AUC = 0.561) liver (P = 0.014) and stomach (P = 0.011) (Fig. 2a). Patients within (Fig. 3c). Importantly, when used as a combined model with TNM the 4th quartile had significantly poorer survival rates compared staging, it outperformed either the signature or TNM when to those within the 1st quartile: glioma (HR = 3.386, P < 0.0001), considered alone, suggesting that incorporating molecular sub- clear cell renal cell (HR = 2.177, P = 0.00048), papillary renal cell type information on Notch-Hedgehog signalling allowed more (HR = 4.881, P = 0.0053), liver (HR = 2.627, P = 0.0039) and sto- precise stratification: clear cell renal cell (AUC = 0.802), papillary mach (HR = 2.217, P = 0.014) (Table S3). When comparing tumour renal cell (AUC = 0.812), liver (AUC = 0.720) and stomach (AUC: and non-tumour expression patterns, Mann–Whitney–Wilcoxon 0.728) (Fig. 3c). In terms of predicting prognosis in glioma tests revealed that a vast majority of the 13 genes were subtypes, performance of the signature was the best in significantly upregulated in tumours of these cancers (Fig. S2) oligoastrocytoma (AUC = 0.823), followed by glioblastoma multi- where hyperactivation of Notch-Hedgehog signalling was asso- forme. (AUC = 0.761) and astrocytoma (AUC = 0.743) (Fig. 3c). The ciated with adverse survival outcomes (Fig. 2a). Multidimensional signature also performed well when all glioma subtypes were scaling analyses revealed that the 13 genes could accurately considered as a group (AUC = 0.815) (Fig. 3c). distinguish tumour from non-tumour samples in these cancers (Fig. 2b), suggesting that Notch-Hedgehog transcriptional states The Notch-Hedgehog signature identifies molecular subtypes with could be used to identify cells with oncogenic properties. stem cell-like features Multivariate Cox proportional hazards regression was used to Notch-Hedgehog hyperactivation is associated with increased determine whether the signature was confounded by other mortality rates (Figs. 2 and 3). To further investigate the underlying clinicopathological features. Tumour, node, metastasis (TNM) biological consequences of augmented Notch-Hedgehog signal- staging is frequently used for patient stratification. Even after ling and how they lead to adverse outcomes, we performed accounting for TNM staging, the signature remained an indepen- differential expression analyses on all transcripts comparing high- dent predictor of survival: clear cell renal cell (HR = 1.731, P = and low-risk patients as predicted by the 13-gene signature. The 0.014), papillary renal cell (HR = 2.297, P = 0.042), liver (HR = liver cancer cohort had the highest number of differentially 2.146; P = 0.024) and stomach (HR = 2.161, P = 0.017) (Table S3). expressed genes (DEGs): 3,015 genes (−1.5 > log2 fold change Given that both the signature and tumour stage were indepen- >1.5; P < 0.01) (Table S4). This was followed by glioma (1,407 dent of each other, we reason that the signature could be used to genes), stomach (906 genes), papillary renal cell (817 genes) and improve TNM staging. We observed that Notch-Hedgehog driver clear cell renal cell (545 genes) carcinoma (Table S4). Despite genes offered an additional resolution in tumour classification for having very different pathologies, there was a great deal of DEG further stratification of similarly-staged tumours in these cancers: overlap between these cancers. 14 DEGs were found in all five clear cell renal cell (P < 0.0001), papillary renal cell (P < 0.0001), cancers, 164 DEGs were observed in at least four cancers and 470 liver (P < 0.0001) and stomach (P = 0.0068) (Fig. 3a). DEGs in at least three cancers (Fig. S3a), implying conserved Glioma samples are classified into four histological categories biological roles of Notch-Hedgehog signalling in driving disease with varying severity: low-grade astrocytoma, low-grade oligo- progression. dendroglioma, low-grade oligoastrocytoma (consisting of both KEGG pathway analyses on DEGs demonstrated enrichments of abnormal astrocytoma and oligodendroglioma cells), and grade IV pathways involved in regulating self-renewal and pluripotency, i.e. glioblastoma multiforme. Kaplan-Meier analyses of glioma sam- Wnt, TGF-β, MAPK, JAK-STAT and PPAR signalling (Fig. 3d; Fig. S3b), ples grouped by histology revealed that the signature remained suggesting that tumours with hyperactive Notch-Hedgehog Aberrations in Notch-Hedgehog signalling reveal cancer stem cells. . . WH Chang and AG Lai 670

a Glioma Renal clear cell 1.00 1.00

0.75 0.75

0.50 0.50

0.25 0.25 Survival probability p < 0.0001 p = 0.0022

0.00 0.00 012345 012345 Years Years Number at risk Number at risk Q1 147 104 72 47 29 17 Q1 131 109 87 71 51 43 Q2 148 119 74 51 28 24 Q2 131 115 99 83 74 50 Q3 147 110 59 37 20 15 Q3 130 109 88 71 48 28 Q4 148 101 46 25 14 9 Q4 132 102 81 64 48 31

Papillary renal cell Liver Stomach 1.00 1.00 1.00

0.75 0.75 0.75

0.50 0.50 0.50

0.25 0.25 0.25 Survival probability p = 0.00099 p = 0.014 p = 0.011

0.00 0.00 0.00 012345 012345 012345 Years Years Years Number at risk Number at risk Number at risk Q1 66 60 38 23 19 13 Q1 76 63 36 23 16 10 Q1 75 48 25 13 7 5 Q2 62 56 41 24 17 11 Q2 79 60 29 20 16 13 Q2 80 62 30 12 7 6 Q3 63 53 32 24 17 14 Q3 80 61 34 22 15 8 Q3 84 53 24 11 5 3 Q4 65 47 27 16 12 8 Q4 78 50 32 19 12 8 Q4 82 63 20 7 2 1

b Glioma Renal clear cell Papillary renal cell Liver Stomach Non-tumour Tumour Dimension 2

P < 0.001 P < 0.001 P < 0.001 P < 0.001 P < 0.001 Dimension 1 Fig. 2 The Notch-Hedgehog 13-gene signature predicts patient survival in five cancers. a Kaplan–Meier estimates for overall survival using the signature. Patients were ranked and quartile stratified into Q1 (<25%), Q2 (25–50%), Q3 (50–75%) and Q4 (>75%) based on their 13-gene scores. P-values were determined using the log-rank test. b Separation of tumour from non-tumour samples using the signature. Ordination plots of MDS analysis of the signature using Euclidean distances to represent tumour and non-tumour samples in 2-dimensional space. PERMANOVA test confirmed statistically significant differences between tumour and non-tumour samples

signalling were characterised by molecular footprints of stemness while targets of REST and SMAD4 were enriched in all cancers and that there was significant crosstalk between Notch-Hedgehog except for clear cell renal cell carcinoma (Fig. 3d). These TFs were and other pathways involved in controlling tumour initiation.28,29 thought to induce epithelial-mesenchymal transition and promote Additionally, Gene Ontology analyses revealed significant enrich- invasion and metastasis consistent with their roles in tumour ments of processes related to cell differentiation, cell proliferation, initiation and maintenance.30–32 embryo development and morphogenesis (Fig. 3d), supporting To independently confirm that the 13-gene signature is a the hypothesis that tumour aggression and elevated mortality potential pan-cancer marker of CSCs, we performed Spearman’s could be caused by the presence of CSCs that are likely to be correlation analyses to compare 13-gene scores with expression refractory to therapy. Consistent with these results, Enrichr profiles of other CSC markers where we would expect to see transcription factor (TF) analyses revealed that TFs associated positive correlations. We examined expression profiles of nine with stem cell function appeared amongst top enriched genes implicated in CSC regulation: CD105, CD133, CD200, CD24, candidates (Fig. 3d). DEGs were enriched as binding targets of CD29, CD44, CD73, CD90 and NESTIN. Putative neural CSC markers SUZ12, REST, EZH2, SMAD4 and FOXM1 as supported by both are CD133, NESTIN, CD105 and CD44.33 We observed significant ChEA and ENCODE databases (Fig. 3d). Binding targets of SUZ12 positive correlations between 13-gene scores and all four markers and EZH2 were consistently enriched across all five cancer types, in glioma samples (Fig. S4). CD105, CD29, CD44, CD73, CD90 and Aberrations in Notch-Hedgehog signalling reveal cancer stem cells. . . WH Chang and AG Lai 671 a c Renal clear cell: TNM staging Papillary renal cell: TNM staging Glioma 1.00 1.00 1.00

0.75 0.75 0.75

0.50 0.50 0.50 Classifier AUC 0.25 All glioma 0.815 Astrocytoma 0.743

Survival probability Oligoastrocytoma 0.823 0.25 0.25 0.00 Glioblastoma 0.761 p < 0.0001 p < 0.0001 0.00 0.25 0.50 0.75 1.00

0.00 0.00 Renal clear cell 012345 012345 1.00 Years Years Number at risk Number at risk 0.75 S1 Low 130 117 98 81 67 46 S1 Low 81 74 51 28 21 11 S1 High 131 115 93 78 54 36 S1 High 80 67 38 23 15 11 S2 Low 28 27 25 23 18 16 S2 Low 10 10 8 7 7 6 S2 High 29 25 18 15 12 8 S2 High 97 4 2 1 1 0.50 S3 Low 60 48 39 32 23 17 S3 Low 21 18 12 7 4 3 S3 High 61 49 43 30 23 12 S3 High 23 18 10 8 7 4 S4 Low 42 31 24 18 14 11 S4 Low 52 0 0 0 0 S4 High 41 22 15 12 10 6 S4 High 72 2 2 2 2 0.25 Classifier AUC TNM 0.717 Signature 0.718 Liver: TNM staging Stomach: TNM staging 0.00 Signature + TNM 0.802 1.00 1.00 0.00 0.25 0.50 0.75 1.00 Papillary renal cell 1.00 0.75 0.75

0.75

0.50 0.50 0.50

Survival probability 0.25 0.25 0.25 Classifier AUC p < 0.0001 p = 0.0068 TNM 0.640 Signature 0.796 0.00 Signature + TNM 0.812 0.00 0.00 0.00 0.25 0.50 0.75 1.00 012345 012345 Years Years Liver Number at risk Number at risk 1.00 S1 Low 71 56 31 23 19 13 S1 Low 21 17 9 6 5 4 S1 High 73 62 33 24 15 9 25 20 7 3 3 1 S1 High S2 Low 57 39 18 6 3 2 S2 Low 34 28 13 9 7 5 S2 High 57 40 12 5 0 0 0.75 S2 High 35 24 17 9 6 4 S3 Low 65 48 26 13 5 4 S3 High S3 Low 38 26 13 8 5 5 66 45 20 5 2 1 S4 Low 963 211 S3 High 37 20 11 6 6 2 S4 High 10 6 3 3 2 2 0.50

b 0.25 Classifier AUC Astrocytoma Oligoastrocytoma Glioblastoma multiforme TNM 0.697 1.00 1.00 1.00 Signature 0.624 0.00 Signature + TNM 0.720 0.00 0.25 0.50 0.75 1.00 0.75 0.75 0.75 Stomach 1.00 0.50 0.50 0.50

0.75 0.25 0.25 0.25 Survival probability p = 0.038 p = 0.0018 p = 0.045 0.50 0.00 0.00 0.00 012345 012345 012345 0.25 Classifier AUC Years Years Years TNM 0.561 Number at risk Number at risk Number at risk Signature 0.710 Q1 49 31 21 16 8 4 Q1 33 27 18 13 5 3 Q1 18 10 3 0 0 0 0.00 Signature + TNM 0.728 Q2 48 36 22 17 9 8 Q2 32 26 14 12 7 7 Q2 19 10 3 1 0 0 Q3 48 38 26 17 9 7 Q3 32 27 10 7 2 1 Q3 19 8 1 0 0 0 0.00 0.25 0.50 0.75 1.00 Q4 49 36 17 7 5 3 Q4 33 25 14 8 3 2 Q4 19 10 1 0 0 0

d KEGG Gene Ontology ENCODE & ChEA

Cytokine−cytokine receptor interaction Cell adhesion SUZ12_CHEA Focal adhesion Signal transduction REST_CHEA Pathways in cancer Transmembrane transport REST_ENCODE ECM−receptor interaction Transport Cell differentiation Jak−STAT signaling pathway FOXM1_ENCODE Cell−cell signaling Neuroactive ligand−receptor interaction IRF8_CHEA Cell proliferation Calcium signaling pathway SMAD4_CHEA Response to stress MAPK signaling pathway Extracellular matrix organization TP53_CHEA TGF−beta signaling pathway Biosynthetic process EZH2_CHEA Wnt signaling pathway Lipid metabolic process Cell adhesion molecules (CAMs) EZH2_ENCODE Embryo development Complement and coagulation cascades Cytoskeleton organization SUZ12_ENCODE Hedgehog signaling pathway Amino acid metabolic process TRIM28_CHEA PPAR signaling pathway Morphogenesis NFE2L2_CHEA Arachidonic acid metabolism Growth Tyrosine metabolism Neurological system process AR_CHEA

Liver Liver Liver Glioma Glioma Glioma Stomach Stomach Stomach

Renal clear cell Renal clear cell Renal clear cell Papillary renal cell Papillary renal cell Papillary renal cell 20 40 60 80 100 200 300 50 100 150 200 250

Gene number Gene number Gene number

NESTIN were positively correlated with 13-gene scores in renal nature of these genes, we would not expect to see positive cancers (Fig. S4); an observation which is consistent with these correlations in all cases. Nonetheless, our results overall suggest genes being markers of renal CSCs.34 Seven and four CSC markers that hyperactive Notch-Hedgehog signalling is associated with were positively correlated with 13-gene scores in liver and CSC phenotypes, contributing to tumour aggression and poor stomach cancers respectively (Fig. S4). Given the tissue-specific survival outcomes. Aberrations in Notch-Hedgehog signalling reveal cancer stem cells. . . WH Chang and AG Lai 672 Fig. 3 The Notch-Hedgehog 13-gene signature is independent of TNM stage and predicts overall survival in glioma histological subtypes. a Kaplan–Meier analyses were performed on patients stratified according to TNM stages and 13-gene scores. Patients were first separated into TNM stage and then median-stratified into low- and high-score groups based on their 13-gene scores. P-values were determined using the log-rank test. b Kaplan–Meier estimates for overall survival using the signature on glioma subtypes ranging from low-grade (astrocytoma, oligoastrocytoma) to high-grade gliomas (glioblastoma multiforme). Patients were first stratified by histological subtypes followed by quartile stratification into Q1 (<25%), Q2 (25–50%), Q3 (50–75%) and Q4 (>75%) based on their 13-gene scores. P-values were determined using the log-rank test. c Predictive performance of the signature. The receiver operating characteristic (ROC) analysis was used to assess specificity and sensitivity of the signature in predicting 5-year overall survival. ROC curves generated from the signature were compared to those generated from TNM staging and a combined model uniting TNM stage and the signature. AUCs for TNM stage were in accordance with previous publications employing TCGA datasets.2,22,51 AUC: area under the curve. TNM: tumour, node and metastasis. d Enriched biological pathways and transcription factor binding associated with DEGs. Differential expression analyses were performed between Q4 and Q1 patients followed by mapping of DEGs against KEGG, Gene Ontology, ChEA and ENCODE databases

Transcription factors involved in self-renewal processes influence significant positive correlations between Notch-Hedgehog+ CSCs survival outcomes in patients with hyperactive Notch-Hedgehog and hypoxia scores in glioma (rho = 0.33, P < 0.0001) and clear cell signalling renal cell carcinoma (rho = 0.16, P = 0.00031) (Fig. 5a). By group- Previously, we observed that binding targets of TFs associated ing patients based on their 13-gene and hypoxia scores, this joint with stem cell function were enriched amongst DEGs (Fig. 3d). model allowed the identification of patients with potentially more Polycomb proteins, EZH2 and SUZ12 have been implicated in CSC hypoxic tumours harbouring Notch-Hedgehog+ CSCs, which formation and maintenance.35,36 REST is a transcriptional repressor influenced overall survival rates: glioma (P < 0.0001) and clear cell involved in maintaining embryonic and neural stem cell pheno- renal cell carcinoma (P = 0.00013) (Fig. 5b). Indeed, patients with types.37 Given their roles in CSC maintenance, we would expect to high CSC and hypoxia scores had significantly poorer survival see the elevated expression of these TFs in tumours with outcomes: glioma (HR = 6.008; P < 0.0001) and clear cell renal cell hyperactive Notch-Hedgehog signalling. Indeed, we observed carcinoma (HR = 2.389, P < 0.0001) (Fig. 5c). The CSC-hypoxia significant positive correlations between 13-gene scores and EZH2 model is also prognostic in glioma subtypes: astrocytoma (HR = levels in glioma (rho = 0.45; P < 0.0001), clear cell renal cell (rho = 5.052, P < 0.0001), oligoastrocytoma (HR = 16.717, P = 0.0066) and 0.22; P < 0.0001), papillary renal cell (rho = 0.33; P < 0.0001) and glioblastoma (HR = 2.686, P = 0.022) (Fig. 5b, c). Our results liver (rho = 0.26; P < 0.0001) cancers (Fig. 4a). Additionally, in the suggest that hypoxic zones within tumours could very well glioma cohort, positive associations between 13-gene scores and represent CSC niches. REST (rho = 0.39; P < 0.0001) or SUZ12 (rho = 0.17; P < 0.0001) profiles were observed (Fig. 4d). Putative Notch-Hedgehog+ CSCs are potentially immune To determine whether these associations harboured prog- privileged nostic information, patients were categorised by their 13-gene Cancer progression is negatively correlated with immunocompe- scores and expression profiles of individual TFs into four tence of the host and evidence points to the role of CSCs in categories: (1) high 13-gene score and high TF expression, (2) immunomodulation.3,42 CSCs reside within niches that are often high 13-gene score and low TF expression, (3) low 13-gene score protected from environmental insults as well as attacks by the and high TF expression and (4) low 13-gene score and low TF immune system. Hypoxic zones not only serve as CSC niches expression (Fig. 4a, d). Interestingly, combined relationship of (Fig. 5),43 but also attract immunosuppressive cells such as the signature and TF expression profiles allowed further regulatory T cells (Tregs),22,44 tumour-associated macrophages45 delineation of patients into additional risk groups: glioma and myeloid-derived suppressor cells.46 Given that positive associa- (EZH2: P < 0.0001; REST: P < 0.0001 and SUZ12: P < 0.0001), clear tions between Notch-Hedgehog+ CSCs and hypoxia were linked to cellrenalcell(EZH2: P < 0.0001), papillary renal cell (EZH2: P = poor progression in glioma and clear cell renal cell carcinoma, we 0.029) and liver (EZH2: P < 0.00057) cancers (Fig. 4b, e). Patients hypothesise that tumours characterised by these features would be with high 13-gene scores that concurrently harboured high TF immune privileged or hypoimmunogenic. expression had the poorest survival outcomes: glioma (EZH2:HR To test this hypothesis, we retrieved a list of 31 genes that = 5.141, P < 0.0001; REST:HR= 3.646, P < 0.0001; SUZ12:HR= represent tumour-infiltrating Tregs. This gene list was identified 3.596, P < 0.0001), clear cell renal cell (EZH2:HR= 2.854, P < from the overlap of four Treg signatures18–21 to yield a more 0.0001), papillary renal cell (EZH2:HR= 4.391, P = 0.0099) and representative profile of tumour-infiltrating Tregs that is not liver (EZH2:HR= 2.685, P = 0.0005) cancers (Fig. 4c, f). Taken specific to a single cancer type. A Treg score for each patient together, our results suggest that coregulation by Notch- within the glioma and clear cell renal cell carcinoma cohorts was Hedgehog signalling and CSC TFs could synergistically con- calculated as the mean expression of the 31 genes. We observed tribute to more advanced disease states. significant positive correlations between Treg scores and the Notch-Hedgehog 13-gene scores in both cohorts, supporting the Tumour hypoxia exacerbates disease phenotypes in Notch- hypothesis that CSCs are potentially hypoimmunogenic: glioma Hedgehog+ CSCs (rho = 0.43; P < 0.0001) and clear cell renal cell carcinoma (rho = Hypoxia is intricately linked to pluripotency as it promotes stem 0.31; P < 0.0001) (Fig. 6a). As performed previously, patients were cell maintenance and self-renewal in both embryonic stem cells separated into four groups based on their 13-gene and Treg and CSCs,38 in part, through modulating hypoxia-inducible factor scores. When used in combination with the Notch-Hedgehog (HIF) function.39 For example, glioma stem cells are typically found signature, Treg expression profiles allowed further separation of in the vicinity of necrotic regions that are hypoxic.40 Glioma stem patients into additional risk groups that influenced overall survival: cells have increased ability to stimulate angiogenesis through glioma (P < 0.0001) and clear cell renal cell carcinoma (P < 0.0001) VEGF upregulation41 and inhibition of HIFs could reduce CSC (Fig. 6b). Intriguingly, patients characterised by high 13-gene and survival, self-renewal and proliferation.40 We reason that hypoxia Treg scores had significantly higher mortality rates compared to functions to maintain CSC niches. To assess the levels of tumour those with low 13-gene and Treg scores: glioma (HR = 4.921, P < hypoxia, we employed a 52-hypoxia gene signature17 for 0.0001) and clear cell renal cell carcinoma (HR = 2.968, P < 0.0001) calculating hypoxia scores in each patient by taking the average (Fig. 6c). This was also true for other histological subtypes of expression of hypoxia signature genes.17 Indeed, we observed glioma: astrocytoma (HR = 2.721; P = 0.0032), oligoastrocytoma Aberrations in Notch-Hedgehog signalling reveal cancer stem cells. . . WH Chang and AG Lai 673

a High signature score and high EZH2 expression High signature score and low EZH2 expression Low signature score and high EZH2 expression Low signature score and low EZH2 expression

Glioma: EZH2 Renal clear cell: EZH2 Papillary renal cell: EZH2 Liver: EZH2 = 0.45 = 0.22 = 0.33 = 0.26 p < 2.2e−16 p = 3.1e−07 p = 3.7e−08 8 p = 1.7e−06 9 8 8

7 8 7 7

Signature score 6 7 6

6 46810 468 56789 6810 EZH2 expression

b High signature score and high EZH2 expression High signature score and low EZH2 expression Low signature score and high EZH2 expression Low signature score and low EZH2 expression

Glioma: EZH2 Renal clear cell: EZH2 Papillary renal cell: EZH2 1.00 1.00 1.00

0.75 0.75 0.75

0.50 0.50 0.50

Survival probability 0.25 0.25 0.25 p < 0.0001 p < 0.0001 p = 0.029

0.00 0.00 0.00 012345 012345 012345 Years Years Years Number at risk Number at risk Number at risk 182 139 96 65 37 27 146 127 106 88 74 57 75 67 46 29 20 14 185 130 58 27 15 9 148 115 92 70 46 24 76 61 32 21 13 7 111 82 48 36 19 15 115 97 77 65 50 35 52 39 27 19 16 15 112 83 49 32 20 14 115 96 80 66 51 36 53 49 33 18 16 10

Liver: EZH2 c 1.00

Hazard Ratio (95% CI) P-value 0.75 Glioma High signature score and high EZH2 vs. low signature score and low EZH2 5.141 (3.374–7.831) 2.48E–14 High signature score and low EZH2 vs. low signature score and low EZH2 1.475 (0.863–2.521) 0.15 Low signature score and high EZH2 vs. low signature score and low EZH2 1.936 (1.179–3.180) 0.0091 0.50 Renal clear cell High signature score and high EZH2 vs. low signature score and low EZH2 2.854 (1.884–4.353) 8.46E–07 High signature score and low EZH2 vs. low signature score and low EZH2 1.360 (0.835–2.215) 0.22 0.25 Low signature score and high EZH2 vs. low signature score and low EZH2 1.759 (1.116–2.774) 0.015 p = 0.00057 Papillary renal cell High signature score and high EZH2 vs. low signature score and low EZH2 4.391 (1.427–13.513) 9.90E–03 0.00 High signature score and low EZH2 vs. low signature score and low EZH2 1.856 (0.498–6.924) 0.35 012345 Low signature score and high EZH2 vs. low signature score and low EZH2 2.008 (0.566 - 7.119) 0.28 Years Number at risk Liver High signature score and high EZH2 vs. low signature score and low EZH2 2.685 (1.540–4.683) 5.00E–04 95 76 35 24 20 16 High signature score and low EZH2 vs. low signature score and low EZH2 1.089 (0.548–2.163) 0.81 99 59 32 20 12 6 59 52 34 21 15 10 Low signature score and high EZH2 vs. low signature score and low EZH2 1.886 (1.009–3.524) 0.046 60 47 30 19 12 7

d e Low signature score and low REST/SUZ12 expression Glioma: REST High signature score and high REST/SUZ12 expression = 0.39 High signature score and low REST/SUZ12 expression p < 2.2e−16 Low signature score and high REST/SUZ12 expression 9

Glioma: REST Glioma: SUZ12 8 1.00 1.00 Signature score 7 0.75 0.75

0.0 2.5 5.0 7.5 0.50 0.50 REST expression

Survival probability 0.25 0.25 p < 0.0001 p < 0.0001

Glioma: SUZ12 0.00 0.00 = 0.17 012345 012345 p = 1.9e−05 9 Years Years Number at risk Number at risk 188 143 96 63 37 28 160 117 79 52 31 22 190 132 71 45 25 17 161 112 57 30 14 7 8 106 80 35 18 9 7 135 100 49 33 20 17 106 79 49 34 20 13 134 105 66 45 26 19 Signature score 7 f

Hazard Ratio (95% CI) P-value 8910 Glioma SUZ12 expression High signature score and high REST vs. low signature score and low REST 3.646 (2.461–5.402) 1.12E–10 High signature score and low REST vs. low signature score and low REST 1.811 (1.097–2.990) 0.02 Low signature score and high REST vs. low signature score and low REST 1.626 (0.994–2.658) 0.053 High signature score and high REST/SUZ12 expression

Low signature score and high REST/SUZ12 expression Glioma High signature score and high SUZ12 vs. low signature score and low SUZ12 3.596 (2.332–5.545) 7.03E–09 High signature score and low REST/SUZ12 expression High signature score and low SUZ12 vs. low signature score and low SUZ12 1.956 (1.227–3.118) 0.0048 Low signature score and high SUZ12 vs. low signature score and low SUZ12 1.210 (0.737–1.987) 0.45 Low signature score and low REST/SUZ12 expression

(HR = 5.431; P = 0.0091) and glioblastoma (HR = 3.065; P = DISCUSSION AND CONCLUSION 0.0068) (Fig. 6c). Taken together, our results suggest that CSCs Aberrations in the Notch-Hedgehog signalling axis are frequently found within immunosuppressed environments are likely to be implicated in malignant progression. Hedgehog genes, Shh, PTCH1 more aggressive. and GLI1, were detected in over 50% of liver cancer tumours and Aberrations in Notch-Hedgehog signalling reveal cancer stem cells. . . WH Chang and AG Lai 674 Fig. 4 Prognostic significance of a combined model of the Notch-Hedgehog signature and transcription factors (EZH2, SUZ12 and REST) involved in pluripotency maintenance. a Scatter plots demonstrate significant positive correlations between 13-gene scores and EZH2 expression profiles in four cancers. Patients were stratified into four categories based on median 13-gene scores and EZH2 expression. Density plots depict the distribution of 13-gene scores and EZH2 expression at the y- and x-axes, respectively. b Kaplan–Meier analyses were performed on the four patient categories to ascertain the combined relationship of the signature and EZH2 expression on overall survival. c Table inset depicts univariate Cox proportional hazards analyses of the relationship between EZH2 and the signature in four cancer types. d Scatter plots demonstrate significant positive correlations between 13-gene scores and REST or SUZ12 expression levels in glioma. Patients were stratified into four categories based on median 13-gene score and REST or SUZ12 expression. Density plots depict the distribution of 13- gene scores and REST or SUZ12 expression at the y- and x-axes, respectively. e Kaplan–Meier analyses were performed on the four patient categories to ascertain the combined relationship between the signature and REST or SUZ12 expression on overall survival in glioma. f Table inset depicts univariate Cox proportional hazards analyses of the relationship between REST or SUZ12 and the signature in glioma. CI: confidence interval. Significant P-values are highlighted in bold

inhibition of Hedgehog signalling by cyclopamine, Smoothened revealed associations of CSCs with hypoxia and immunosuppres- antagonist or anti-SHH resulted in decreased cell growth and sion. We observed that CSCs characterised by hyperactive Notch- increased apoptosis.47 Notch signalling is also activated in liver Hedgehog signalling exhibited immune privileged features cancer and this leads to the formation of liver tumours in mice.48 associated with the attenuation effects of Tregs (Fig. 6). Effective- Notch blockade using γ-secretase inhibitors reduced cell viability ness of immunotherapy is biased towards differentiated cells that in hepatoma cell lines.48 In clear cell renal cell carcinoma, make up the tumour bulk due to distinct antigen presentation in inhibition of Notch signalling reduced anchorage-independent CSCs.59 CD133+ glioma CSCs fail to express NK cell ligands or growth and mice treated with Notch inhibitors had impaired MHCI, which prevents immune detection.60 Stimulatory NK cell growth of transplanted cancer cells.49 Elevated expression of ligands are also downregulated in breast CSCs, contributing to Notch ligands correlated with aggressiveness and poor survival evasion from NK cell killing.61 Pan et al. elegantly reviewed recent rates in stomach cancer.50 initiatives focusing on immunotherapeutic agents against CSC These studies have paved the road for understanding the role antigens employing dendritic cell vaccines, myeloid-derived of Notch-Hedgehog signalling in carcinogenesis. However, large- suppressor cell-based approaches and the use of immune scale comparative studies investigating the similarities and checkpoint blockades recognising PD-1 or CTLA4.59 The Notch- differences in Notch-Hedgehog signalling across multiple cancer Hedgehog signature may be used to stratify patients prior to types have remained limited. We interrogated expression and immunotherapy. mutational profiles of 72 genes from Notch and Hedgehog Immunoevasion can be exacerbated by tumour hypoxia as the pathways in 21 diverse cancer types involving 18,484 patients. Our latter not only promotes CSC survival, but also creates an integrated analysis of genomic, transcriptomic and clinical data environment that facilitates further immune suppression.22 It revealed molecular distinct tumour subtypes that were charac- may be possible that Notch-Hedgehog+ CSCs in glioma and clear terised by Notch-Hedgehog hyperactivation. Concentrating on 13 cell renal cell carcinomas are more frequently found within Notch-Hedgehog driver genes that were recurrently amplified and immunosuppressed hypoxic zones (Fig. 5). Indeed, hypoxia could overexpressed, we found that these genes were associated with stimulate self-renewal of CD133+ glioma stem cells and this is clinically relevant molecular features of stemness. The biological abrogated by HIF-1α knockdown.62 Hypoxia promotes the consequences of elevated expression of driver genes were maintenance of undifferentiated states through the activation of manifold. High-risk patients showed overexpression of genes Notch-responsive genes in neuronal progenitors.39 Hypoxia also associated with other stem cell-related pathways such as Wnt, activates cellular reprogramming of non-stem cancer cells into JAK-STAT and TGF-β signalling (Fig. 3d. and S3b).51 Simultaneous CSCs in glioblastoma by inducing the expression of Oct4, Nanog inhibition of Notch and JAK-STAT pathways by combined AG-490 and c-Myc.63 Glioma stem cells are pro-angiogenic due to and GSI IX therapy impaired pancreatic cancer progression.52 GLI2 promiscuous secretions of VEGF that is further induced by is regulated by both Hedgehog and TGF-β pathways and others hypoxia.41 Bevacizumab, which targets VEGF, could suppress have surmised that TGF-β may potentiate Hedgehog signalling xenographs derived from glioma stem cells but not those derived cascade by increasing GL12 availability, contributing to metasta- from non-stem glioma cells.41 In renal cancer, we observed that sis.53 Hence, our study reveals molecular targets with overlapping Notch-Hedgehog+ CSCs are likely to be enriched in hypoxic functions that can be prioritised to improve therapeutic outcomes. tumours and the combined effects of hypoxia and augmented Furthermore, binding targets of stem cell-related TFs (EZH2, Notch-Hedgehog signalling resulted in further elevation of death SUZ12 and REST) were enriched amongst genes upregulated in risks (Fig. 5). However, Sjölund et al. observed that Notch high-risk patients (Fig. 4). EZH2 synergises with Notch-Hedgehog+ signalling is not enhanced by hypoxia in renal cancer.49 Another CSCs to worsen survival outcomes in patients with glioma, renal study on renal CSCs revealed that hypoxia did not affect the and liver cancers (Fig. 4). Pharmacological inhibition of EZH2 differentiation potential of CD105+ CSCs.64 Nonetheless, hypoxia impaired glioblastoma CSC tumour-initiating capacity and survi- was found to induce the expression of stem cell markers, Oct4, val.35 EZH2-mediated transcriptional silencing leads to the Nanog, c-Myc and Klf4 in renal cancer cell lines, supporting our maintenance of undifferentiated states in glioblastoma through observation, and in another ten cancers including cervix, lung, STAT3 activation.54 In liver cancer, EZH2 overexpression is colon, liver and prostate.65 associated with vascular invasion, malignant progression55 and Although prospective validation is warranted, the results activation of β-catenin/Wnt signalling.56 Inhibition of EZH2 in renal presented in this work support a model where Notch-Hedgehog cancer cell lines led to increased apoptosis.57 Additionally, hyperactivation is linked to stemness and that hypoxia contributes enrichments of SUZ12 and REST targets in glioma patients with to the maintenance of undifferentiated phenotypes and the hyperactive Notch-Hedgehog signalling were linked to signifi- reduction of anti-tumour immunity. The use of immune check- cantly poorer prognosis (Fig. 4d, e). REST is implicated in point blockade has been increasingly tried in malignancy.66 transcriptional regulation of neuronal stem cells,37 while the Hence, molecular signatures capable of discerning responders overexpression of SUZ12 is linked to tumour progression.58 from non-responders will be valuable prior to the administration An exploration of the relationship between Notch-Hedgehog of these expensive drugs. As an independent prognostic indicator hyperactivation and tumour microenvironmental qualities in five cancer types involving 2278 patients, the Notch-Hedgehog Aberrations in Notch-Hedgehog signalling reveal cancer stem cells. . . WH Chang and AG Lai 675 a High signature and high hypoxia scores High signature and low hypoxia scores Glioma Low signature and high hypoxia scores Low signature and low hypoxia scores = 0.33 9 p < 2.2e−16

8 c

Hazard Ratio (95% CI) P-value

Signature score 7 Glioma High signature and high hypoxia scores vs. low signature and low hypoxia scores 6.008 (3.939–9.161) 2.00E–16 High signature and low hypoxia scores vs. low signature and low hypoxia scores 0.996 (0.545–1.821) 0.98 Low signature and high hypoxia scores vs. low signature and low hypoxia scores 2.048 (1.245–3.367) 0.0047 91011 Hypoxia score Astrocytoma High signature and high hypoxia scores vs. low signature and low hypoxia scores 5.052 (2.292–11.134) 5.89E–05 High signature and low hypoxia scores vs. low signature and low hypoxia scores 1.578 (0.589–4.223) 0.36 Low signature and high hypoxia scores vs. low signature and low hypoxia scores 3.827 (1.571–9.322) 0.0031

Renal clear cell Oligoastrocytoma = 0.16 High signature and high hypoxia scores vs. low signature and low hypoxia scores 16.717 (2.189–127.670) 6.60E–03 High signature and low hypoxia scores vs. low signature and low hypoxia scores 4.310 (0.447–41.540) 0.21 p = 0.00031 8 Low signature and high hypoxia scores vs. low signature and low hypoxia scores 5.383 (0.646–44.830) 0.12

Glioblastoma High signature and high hypoxia scores vs. low signature and low hypoxia scores 2.686 (1.151–6.270) 2.20E–02 7 High signature and low hypoxia scores vs. low signature and low hypoxia scores 1.635 (0.771–3.473) 0.2 Low signature and high hypoxia scores vs. low signature and low hypoxia scores 0.799 (0.363–1.759) 0.58

Signature score 6 Renal clear cell High signature and high hypoxia scores vs. low signature and low hypoxia scores 2.389 (1.575–3.625) 4.23E–05 High signature and low hypoxia scores vs. low signature and low hypoxia scores 1.330 (0.829–2.135) 0.24 91011 Low signature and high hypoxia scores vs. low signature and low hypoxia scores 1.350 (0.857–2.126) 0.19 Hypoxia score

b Glioma Renal clear cell 1.00 1.00

0.75 0.75

Low signature and low hypoxia scores 0.50 0.50 High signature and high hypoxia scores High signature and low hypoxia scores 0.25 0.25 Low signature and high hypoxia scores

Survival probability p < 0.0001 p = 0.00013

0.00 0.00 012345 012345 Years Years Number at risk Number at risk 181 141 97 66 37 26 140 125 100 82 66 49 183 120 53 30 19 11 141 113 85 70 51 30 113 92 53 33 15 13 122 99 84 65 45 29 113 81 48 31 20 15 121 98 86 72 59 44

Astrocytoma Oligoastrocytoma Glioblastoma 1.00 1.00 1.00

0.75 0.75 0.75

0.50 0.50 0.50 p = 0.018

0.25 0.25 0.25

Survival probability p < 0.0001 p < 0.0001

0.00 0.00 0.00 012345 012345 012345 Years Years Years Number at risk Number at risk Number at risk 62 45 31 24 13 10 35 29 17 14 4 2 16 8 3 1 0 0 62 44 23 11 8 5 35 26 13 9 3 2 17 5 0 0 0 0 35 30 20 13 6 5 30 26 11 6 2 1 21132000 35 22 12 9 4 2 30 24 15 11 8 8 21123000 Fig. 5 Positive associations between Notch-Hedgehog+ CSCs and tumour hypoxia in glioma and clear cell renal cell carcinoma. a Scatter plots demonstrate significant positive correlations between 13-gene and hypoxia scores. Patients were stratified into four categories based on median 13- gene and hypoxia scores. Density plots depict the distribution of 13-gene and hypoxia scores at the y-andx-axes, respectively. b Kaplan–Meier analyses were performed on the four patient categories to ascertain the combined relationship of the signature and tumour hypoxia on overall survival. Contribution of hypoxia on Notch-Hedgehog+ CSCs were also determined in histological subtypes of glioma (astrocytoma, oligoastrocytoma and glioblastoma multiforme). c Table inset demonstrates univariate Cox proportional hazards analyses of the relationship between tumour hypoxia and the signature in glioma and clear cell renal cell carcinoma. CI: confidence interval. Significant P-values are highlighted in bold Aberrations in Notch-Hedgehog signalling reveal cancer stem cells. . . WH Chang and AG Lai 676 a

Glioma High signature and high Treg scores High signature and low Treg scores Low signature and high Treg scores Low signature and low Treg scores 9 c

8 Hazard Ratio (95% CI) P-value Glioma High signature and high Treg scores vs. low signature and low Treg scores 4.921 (3.277–7.391) 1.60E–14

Signature score 7 = 0.43 High signature and low Treg scores vs. low signature and low Treg scores 1.671 (0.989–2.822) 0.055 p < 2.2e−16 Low signature and high Treg scores vs. low signature and low Treg scores 2.377 (1.459–3.872) 0.000503

Astrocytoma 45678 High signature and high Treg scores vs. low signature and low Treg scores 2.721 (1.398–5.298) 3.20E–03 Treg score High signature and low Treg scores vs. low signature and low Treg scores 1.288 (0.561–2.960) 0.55 Low signature and high Treg scores vs. low signature and low Treg scores 1.457 (0.597–3.554) 0.41

Oligoastrocytoma High signature and high Treg scores vs. low signature and low Treg scores 5.431 (1.522–19.374) 9.10E–03 Renal clear cell High signature and low Treg scores vs. low signature and low Treg scores 2.470 (0.576–10.603) 0.22 Low signature and high Treg scores vs. low signature and low Treg scores 1.032 (0.222–4.804) 0.97

Glioblastoma 8 High signature and high Treg scores vs. low signature and low Treg scores 3.065 (1.362–6.900) 0.0068 High signature and low Treg scores vs. low signature and low Treg scores 1.955 (0.824–4.639) 0.13 Low signature and high Treg scores vs. low signature and low Treg scores 1.274 (0.570–2.846) 0.55 7 Renal clear cell High signature and high Treg scores vs. low signature and low Treg scores 2.968 (1.922–4.582) 9.20E–07 High signature and low Treg scores vs. low signature and low Treg scores 1.649 (0.997–2.728) 0.051 Signature score = 0.31 6 Low signature and high Treg scores vs. low signature and low Treg scores 2.132 (1.342–3.385) 0.0013 p < 7.6e–13

56789 Treg score

b Glioma Renal clear cell 1.00 1.00

0.75 0.75 Low signature and low Treg scores 0.50 0.50 High signature and high Treg scores High signature and low Treg scores 0.25 0.25 Low signature and high Treg scores

Survival probability p < 0.0001 p < 0.0001

0.00 0.00 012345 012345 Years Years Number at risk Number at risk 196 155 103 70 42 29 147 125 107 92 75 55 198 129 63 36 20 13 149 121 93 72 51 32 98 83 43 27 14 11 114 91 76 63 45 27 98 67 42 27 15 12 114 98 79 62 50 38

Astrocytoma Oligoastrocytoma Glioblastoma 1.00 1.00 1.00

0.75 0.75 0.75

0.50 0.50 0.50 p = 0.021

0.25 0.25 0.25

Survival probability p = 0.012 p = 0.0037

0.00 0.00 0.00 012345 012345 012345 Years Years Years Number at risk Number at risk Number at risk 65 49 31 23 12 8 38 31 17 14 5 3 15 9 3 0 0 0 65 46 27 13 8 5 38 28 13 8 4 2 22 11 1 0 0 0 32 28 16 11 6 5 27 24 11 7 1 1 16 7 1 0 0 0 32 18 12 10 5 4 27 22 15 11 7 7 22 11 3 1 0 0 Aberrations in Notch-Hedgehog signalling reveal cancer stem cells. . . WH Chang and AG Lai 677 Fig. 6 Positive associations between Notch-Hedgehog+ CSCs and immunosuppression in glioma and clear cell renal cell carcinoma. a Scatter plots demonstrate significant positive correlations between 13-gene and Treg scores. Patients were stratified into four categories based on median 13-gene and Treg scores. Density plots depict the distribution of 13-gene and Treg scores at the y- and x-axes respectively. b Kaplan–Meier analyses were performed on the four patient categories to ascertain the combined relationship of the signature and Treg- mediated immunosuppression on overall survival. Contribution of Tregs on Notch-Hedgehog+ CSCs were also determined in histological subtypes of glioma (astrocytoma, oligoastrocytoma and glioblastoma multiforme). c Table inset demonstrates univariate Cox proportional hazards analyses of the relationship between Tregs and the signature in glioma and clear cell renal cell carcinoma. CI: confidence interval. Significant P-values are highlighted in bold

10. Dakubo, G. D., Mazerolle, C. J. & Wallace, V. A. Expression of Notch and Wnt gene signature may serve as a staging point for exploring pathway components and activation of Notch signaling in medulloblastomas combinatorial treatments that simultaneously target CSCs, from heterozygous patched mice. J. Neurooncol. 79, 221–227 (2006). hypoxia and tumour immunity. 11. Katoh, M. Networking of WNT, FGF, Notch, BMP, and Hedgehog signaling pathways during carcinogenesis. Stem Cell Rev. 3,30–38 (2007). 12. Steg, A. D., Katre, A. A., Goodman, B. W., Han, H.-D., Nick, A. M., Stone, R. L. et al. AUTHOR CONTRIBUTIONS Targeting the notch ligand JAGGED1 in both tumor cells and stroma in ovarian cancer. Clin. Cancer Res. 17, 5674–5685 (2011). clincanres–0432. A.G.L. designed the study and supervised the research. W.H.C. and A.G.L. analysed the 13. Domingo-Domenech, J., Vidal, S. J., Rodriguez-Bravo, V., Castillo-Martin, M., Quinn, data, interpreted the results and wrote the initial manuscript draft. A.G.L. revised the S. A., Rodriguez-Barrueco, R. et al. Suppression of acquired docetaxel resistance in manuscript draft. All authors approved the final submitted manuscript. prostate cancer through depletion of notch-and hedgehog-dependent tumor- initiating cells. Cancer Cell. 22, 373–388 (2012). 14. Schreck, K. C., Taylor, P., Marchionni, L., Gopalakrishnan, V., Bar, E. E., Gaiano, N. ADDITIONAL INFORMATION et al. The Notch target Hes1 directly modulates Gli1 expression and Hedgehog Supplementary information is available for this paper at https://doi.org/10.1038/ signaling: a potential mechanism of therapeutic resistance. Clin. Cancer Res. 16, s41416-019-0572-9. 6060–6070 (2010). 15. Weinstein, J. N., Collisson, E. A., Mills, G. B., Shaw, K. R. M., Ozenberger, B. A., Ellrott, Competing interests: The authors declare no competing interests. K. et al. The cancer genome atlas pan-cancer analysis project. Nat. Genet. 45, 1113 (2013). Ethics approval and consent to participate: No ethics approval or consent to 16. Mermel, C. H., Schumacher, S. E., Hill, B., Meyerson, M. L., Beroukhim, R. & Getz, G. participate is required as this study is conducted using publicly available datasets GISTIC2. 0 facilitates sensitive and confident localization of the targets of focal from The Cancer Genome Atlas. somatic copy-number alteration in human cancers. Genome Biol. 12, R41 (2011). 17. Buffa, F. M., Harris, A. L., West, C. M. & Miller, C. J. Large meta-analysis of multiple cancers reveals a common, compact and highly prognostic hypoxia metagene. Funding: None Br. J. Cancer [Internet]. 102, 428–435 (2010). 18. Zheng, C., Zheng, L., Yoo, J. K., Guo, H., Zhang, Y., Guo, X. et al. Landscape of Consent to publish: None. infiltrating T cells in liver cancer revealed by single-cell sequencing. Cell 169, 1342–1356.e16 (2017). Data availability: The datasets supporting the conclusions of this article are included 19. De Simone, M., Arrigoni, A., Rossetti, G., Gruarin, P., Ranzani, V., Politano, C. et al. within the article and its Supplementary files. Transcriptional Landscape of Human Tissue Lymphocytes Unveils Uniqueness of Tumor-Infiltrating T Regulatory Cells. Immunity 45, 1135–1147 (2016). 20. Plitas, G., Konopacki, C., Wu, K., Bos, P. D., Morrow, M., Putintseva, E. V. et al. Note: This work is published under the standard license to publish agreement. After Regulatory T cells exhibit distinct features in human breast cancer. Immunity. 45, 12 months the work will become freely available and the license terms will switch to 1122–1134 (2016). a Creative Commons Attribution 4.0 International (CC BY 4.0). 21. Tirosh, I., Izar, B., Prakadan, S. M., Wadsworth, M. H., Treacy, D., Trombetta, J. J. et al. Dissecting the multicellular ecosystem of metastatic melanoma by single- Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims cell RNA-seq. Science 352, 189–196 (2016). in published maps and institutional affiliations. 22. Chang, W. H., Forde, D. & Lai, A. G. A novel signature derived from immunor- egulatory and hypoxia genes predicts prognosis in liver and five other cancers. J. Transl. Med. 17(Jan), 14 (2019). 23. Chang W. H., Forde D., Lai A. G. Dual prognostic role for 2-oxoglutarate oxyge- REFERENCES nases in ten diverse cancer types: Implications for cell cycle regulation and cell 1. Maugeri-Saccà, M., Bartucci, M. & De Maria, R. DNA damage repair pathways in adhesion maintenance. Cancer Commun. 39, 23 (2019). cancer stem cells. Mol. Cancer Ther. 11, 1627–1636 (2012). 24. Chang, W. H. & Lai, A. G. Timing gone awry: distinct tumour suppressive and 2. Chang, W. H. & Lai, A. G. Transcriptional landscape of DNA repair genes underpins oncogenic roles of the circadian clock and crosstalk with hypoxia signalling in a pan-cancer prognostic signature associated with cell cycle dysregulation and diverse malignancies. J. Transl. Med. 17(Apr), 132 (2019). tumor hypoxia. DNA Repair (Amst) [Internet]. 78(Jun), 142–153 (2019). 25. Tabas-Madrid, D., Nogales-Cadenas, R. & Pascual-Montano, A. GeneCodis3: a non- 3. Schatton, T. & Frank, M. H. Antitumor immunity and cancer stem cells. Ann. N Y redundant and modular enrichment analysis tool for functional genomics. Nucleic Acad. Sci. 1176, 154–169 (2009). Acids Res. 40(W1), W478–W483 (2012). 4. Clevers, H. The cancer stem cell: premises, promises and challenges. Nat. Med. 17, 26. Kuleshov, M. V., Jones, M. R., Rouillard, A. D., Fernandez, N. F., Duan, Q., Wang, Z. 313 (2011). et al. Enrichr: a comprehensive gene set enrichment analysis web server 2016 5. Takebe, N., Harris, P. J., Warren, R. Q. & Ivy, S. P. Targeting cancer stem cells by update. Nucleic Acids Res. 44(W1), W90–W97 (2016). inhibiting Wnt, Notch, and Hedgehog pathways. Nat. Rev. Clin. Oncol. 8, 97 (2011). 27. Haider, S., McIntyre, A., van Stiphout, R. G. P. M., Winchester, L. M., Wigfield, S., 6. Ellisen, L. W., Bird, J., West, D. C., Soreng, A. L., Reynolds, T. C., Smith, S. D. et al. Harris, A. L. et al. Genomic alterations underlie a pan-cancer metabolic shift TAN-1, the human homolog of the Drosophila notch gene, is broken by chro- associated with tumour hypoxia. Genome Biol. 17,1–17 (2016). mosomal translocations in T lymphoblastic neoplasms. Cell. 66, 649–661 (1991). 28. Chang W. H., Lai A. G. An immunoevasive strategy through clinically-relevant 7. Lobry, C., Oh, P. & Aifantis, I. Oncogenic and tumor suppressor functions of Notch pan-cancer genomic and transcriptomic alterations of JAK-STAT signaling com- in cancer: it’s NOTCH what you think. J. Exp. Med. 208, 1931–1935 (2011). ponents. bioRxiv [Internet]. 2019; Available from: https://www.biorxiv.org/ 8. Briscoe, J. & Thérond, P. P. The mechanisms of Hedgehog signalling and its roles content/early/2019/03/14/576645. in development and disease. Nat. Rev. Mol. Cell Biol. 14, 416 (2013). 29. Chang, W. H. & Lai, A. G. The pan-cancer mutational landscape of the 9. Solecki, D. J., Liu, X., Tomoda, T., Fang, Y. & Hatten, M. E. Activated PPAR pathway reveals universal patterns of dysregulated metabolism and Notch2 signaling inhibits differentiation of cerebellar granule neuron precursors interactions with tumor immunity and hypoxia. Ann. N Y Acad. Sci. 1448(Jul), by maintaining proliferation. Neuron. 31, 557–568 (2001). 65–82 (2019). Aberrations in Notch-Hedgehog signalling reveal cancer stem cells. . . WH Chang and AG Lai 678 30. Bao, B., Wang, Z., Ali, S., Kong, D., Banerjee, S., Ahmad, A. et al. Over-expression of 49. Sjölund, J., Johansson, M., Manna, S., Norin, C., Pietras, A., Beckman, S. et al. FoxM1 leads to epithelial–mesenchymal transition and cancer stem cell pheno- Suppression of renal cell carcinoma growth by inhibition of Notch signaling type in pancreatic cancer cells. J Cell Biochem. 112, 2296–2306 (2011). in vitro and in vivo. J Clin Invest. 118, 217–228 (2008). 31. Zingg, D., Debbache, J., Schaefer, S. M., Tuncer, E., Frommel, S. C., Cheng, P. et al. 50. Yeh, T.-S., Wu, C.-W., Hsu, K.-W., Liao, W.-J., Yang, M.-C., Li, A. F.-Y. et al. The The epigenetic modifier EZH2 controls melanoma growth and metastasis activated Notch1 signal pathway is associated with gastric cancer progression through silencing of distinct tumour suppressors. Nat. Commun. 6, 6051 (2015). through cyclooxygenase-2. Cancer Res. 69, 5039–5048 (2009). 32. Li, H., Cai, Q., Wu, H., Vathipadiekal, V., Dobbin, Z. C., Li, T. et al. SUZ12 promotes 51. Chang, W. H. & Lai, A. G. Pan-cancer genomic amplifications underlie a Wnt human epithelial ovarian cancer by suppressing apoptosis via silencing HRK. Mol hyperactivation phenotype associated with stem cell-like features leading to Cancer Res. 10, 1462–1472 (2012). poor prognosis. Transl Res. 208,47–62 (2019). 33. Hovinga, K. E., Shimizu, F., Wang, R., Panagiotakos, G., Van Der Heijden, M., 52. Palagani, V., Bozko, P., El Khatib, M., Belahmer, H., Giese, N., Sipos, B. et al. Moayedpardazi, H. et al. Inhibition of notch signaling in glioblastoma targets Combined inhibition of Notch and JAK/STAT is superior to monotherapies and cancer stem cells via an endothelial cell intermediate. Stem Cells. 28, 1019–1029 impairs pancreatic cancer progression. Carcinogenesis. 35, 859–866 (2013). (2010). 53. Javelaud, D., Alexaki, V. I., Dennler, S., Mohammad, K. S., Guise, T. A. & Mauviel, A. 34. Bussolati, B., Dekel, B., Azzarone, B. & Camussi, G. Human renal cancer stem cells. TGF-β/SMAD/GLI2 signaling axis in cancer progression and metastasis. Cancer Cancer Lett. 338, 141–146 (2013). Res. 71, 5606–5610 (2011). 35. Suvà, M.-L., Riggi, N., Janiszewska, M., Radovanovic, I., Provero, P., Stehle, J.-C. 54. Kim, E., Kim, M., Woo, D.-H., Shin, Y., Shin, J., Chang, N. et al. Phosphorylation of et al. EZH2 is essential for glioblastoma cancer stem cell maintenance. Cancer Res. EZH2 activates STAT3 signaling via STAT3 methylation and promotes tumor- 69, 9211–9218 (2009). 8–5472. igenicity of glioblastoma stem-like cells. Cancer Cell. 23, 839–852 (2013). 36. Sauvageau, M. & Sauvageau, G. Polycomb group proteins: multi-faceted reg- 55. Sasaki, M., Ikeda, H., Itatsu, K., Yamaguchi, J., Sawada, S., Minato, H. et al. The ulators of somatic stem cells and cancer. Cell Stem Cell. 7, 299–313 (2010). overexpression of polycomb group proteins Bmi1 and EZH2 is associated with 37. Johnson, R., Teh, C. H., Kunarso, G., Wong, K. Y., Srinivasan, G., Cooper, M. L. et al. the progression and aggressive biological behavior of hepatocellular carcinoma. REST regulates distinct transcriptional networks in embryonic and neural stem Lab. Investig. 88, 873 (2008). cells. PLoS Biol. 6, e256 (2008). 56. Cheng, A. S. L., Lau, S. S., Chen, Y., Kondo, Y., Li, M. S., Feng, H. et al. EZH2- 38. Keith, B. & Simon, M. C. Hypoxia-inducible factors, stem cells, and cancer. Cell. mediated concordant repression of Wnt antagonists promotes 129, 465–472 (2007). β-catenin–dependent hepatocarcinogenesis. Cancer Res. 71, 4028–4039 (2011). 39. Gustafsson, M. V., Zheng, X., Pereira, T., Gradin, K., Jin, S., Lundkvist, J. et al. 57. Wagener, N., Macher-Goeppinger, S., Pritsch, M., Hüsing, J., Hoppe-Seyler, K., Hypoxia requires notch signaling to maintain the undifferentiated cell state. Dev Schirmacher, P. et al. Enhancer of zeste homolog 2 (EZH2) expression is an Cell. 9, 617–628 (2005). independent prognostic factor in renal cell carcinoma. BMC Cancer. 10, 524 40. Li, Z., Bao, S., Wu, Q., Wang, H., Eyler, C., Sathornsumetee, S. et al. Hypoxia- (2010). inducible factors regulate tumorigenic capacity of glioma stem cells. Cancer Cell. 58. Zhang, J., Wu, W., Xu, S., Zhang, J., Zhang, J., Yu, Q. et al. MicroRNA-105 inhibits 15, 501–513 (2009). human glioma cell malignancy by directly targeting SUZ12. Tumor Biol. 39, 41. Bao, S., Wu, Q., Sathornsumetee, S., Hao, Y., Li, Z., Hjelmeland, A. B. et al. Stem 1010428317705766 (2017). cell–like glioma cells promote tumor angiogenesis through vascular endothelial 59. Pan, Q., Li, Q., Liu, S., Ning, N., Zhang, X., Xu, Y. et al. Concise review: targeting growth factor. Cancer Res. 66, 7843–7848 (2006). cancer stem cells using immunologic approaches. Stem Cells. 33, 2085–2092 42. Kawasaki, B. T. & Farrar, W. L. Cancer stem cells, CD200 and immunoevasion. (2015). Trends Immunol. 29, 464–468 (2008). 60. Wu, A., Wiesner, S., Xiao, J., Ericson, K., Chen, W., Hall, W. A. et al. Expression of 43. Ledaki, I., McIntyre, A., Wigfield, S., Buffa, F., McGowan, S., Baban, D. et al. Carbonic MHC I and NK ligands on human CD133+ glioma cells: possible targets of anhydrase IX induction defines a heterogeneous cancer cell response to hypoxia immunotherapy. J Neurooncol. 83, 121–131 (2007). and mediates stem cell-like properties and sensitivity to HDAC inhibition. 61. Wang, B., Wang, Q., Wang, Z., Jiang, J., Yu, S.-C., Ping, Y.-F. et al. Metastatic Oncotarget 6, 19413–19427 (2015). consequences of immune escape from NK cell cytotoxicity by human breast 44. Clambey, E. T., McNamee, E. N., Westrich, J. A., Glover, L. E., Campbell, E. L., cancer stem cells. Cancer Res. 74, 5746–5757 (2014). Jedlicka, P. et al. Hypoxia-inducible factor-1 alpha–dependent induction of FoxP3 62. Soeda, A., Park, M., Lee, D., Mintz, A., Androutsellis-Theotokis, A., McKay, R. D. et al. drives regulatory T-cell abundance and function during inflammatory hypoxia of Hypoxia promotes expansion of the CD133-positive glioma stem cells through the mucosa. Proc. Natl. Acad. Sci. USA 109, E2784–E2793 (2012). activation of HIF-1α. Oncogene. 28, 3949 (2009). 45. Doedens, A. L., Stockmann, C., Rubinstein, M. P., Liao, D., Zhang, N., 63. Heddleston, J. M., Li, Z., McLendon, R. E., Hjelmeland, A. B. & Rich, J. N. The DeNardo, D. G. et al. Macrophage expression of hypoxia-inducible factor-1α hypoxic microenvironment maintains glioblastoma stem cells and promotes suppresses T-cell function and promotes tumor progression. Cancer Res. 70, reprogramming towards a cancer stem cell phenotype. Cell cycle. 8, 3274–3284 7465–7475 (2010). (2009). 46. Noman, M. Z., Hasmim, M., Messai, Y., Terry, S., Kieda, C., Janji, B. et al. Hypoxia: a 64. Azzi, S., Bruno, S., Giron-Michel, J., Clay, D., Devocelle, A., Croce, M. et al. Differ- key player in antitumor immune response. A Review in the Theme: Cellular entiation therapy: targeting human renal cancer stem cells with interleukin 15. J. Responses to Hypoxia. Am. J. Physiol. Cell. Physiol. 309, C569–C579 (2015). Natl. Cancer Inst. 103, 1884–1898 (2011). 47. Huang, S., He, J., Zhang, X., Bian, Y., Yang, L., Xie, G. et al. Activation of the 65. Mathieu, J., Zhang, Z., Zhou, W., Wang, A. J., Heddleston, J. M., Pinna, C. M. A. et al. hedgehog pathway in human hepatocellular carcinomas. Carcinogenesis. 27, HIF induces human embryonic stem cell markers in cancer cells. Cancer Res. 71, 1334–1340 (2006). 4640–4652 (2011). 48. Villanueva, A., Alsinet, C., Yanger, K., Hoshida, Y., Zong, Y., Toffanin, S. et al. Notch 66. Topalian, S. L., Drake, C. G. & Pardoll, D. M. Immune checkpoint blockade: a signaling is activated in human hepatocellular carcinoma and induces tumor common denominator approach to cancer therapy. Cancer Cell. 27, 450–461 formation in mice. Gastroenterology. 143, 1660–1669 (2012). (2015).