Cancer stemness, intratumoral heterogeneity, andimmune response across cancersAlex Mirandaa,1, Phineas T. Hamiltona,1, Allen W. Zhangb,c,d, Swetansu Pattnaike, Etienne Bechtf, Artur Mezheyeuskig,Jarle Bruunh, Patrick Mickeg, Aurélien de Reyniesi, and Brad H. Nelsona,j,k,2

aDeeley Research Centre, BC Cancer, Victoria, BC V8R 6V5, Canada; bDepartment of Molecular Oncology, BC Cancer, Vancouver, BC V5Z 4E6, Canada;cCentre for Molecular Medicine and Therapeutics, BC Children’s Hospital Research Institute, Vancouver, BC V5Z 4H4, Canada; dGraduate BioinformaticsTraining Program, University of British Columbia, Vancouver, BC V6T 1Z3, Canada; eThe Kinghorn Cancer Centre and Cancer Division, Garvan Institute ofMedical Research, Darlinghurst, NSW 2010, Australia; fSingapore Immunology Network, Agency for Science, Technology and Research, 138648 Singapore;gDepartment of Immunology, Genetics, and Pathology, Uppsala University, 751 85 Uppsala, Sweden; hDepartment of Molecular Oncology, Institute forCancer Research, The Norwegian Radium Hospital, Oslo University Hospital, 0379 Oslo, Norway; iProgramme Cartes d’Identité des Tumeurs, Ligue NationaleContre le Cancer, 75013 Paris, France; jDepartment of Biochemistry and Microbiology, University of Victoria, Victoria, BC V8P 3E6, Canada; and kDepartmentof Medical Genetics, University of British Columbia, Vancouver, BC V6T 1Z3, Canada

Edited by Douglas Hanahan, Swiss Federal Institute of Technology Lausanne, Lausanne, Switzerland, and approved March 22, 2019 (received for reviewOctober 22, 2018)

Regulatory programs that control the function of stem cells areactive in cancer and confer properties that promote progressionand therapy resistance. However, the impact of a stem cell-liketumor phenotype (“stemness”) on the immunological properties ofcancer has not been systematically explored. Using gene-expres-sion–based metrics, we evaluated the association of stemness withimmune cell infiltration and genomic, transcriptomic, and clinicalparameters across 21 solid cancers. We found pervasive negativeassociations between cancer stemness and anticancer immunity.This occurred despite high stemness cancers exhibiting increasedmutation load, cancer-testis antigen expression, and intratumoralheterogeneity. Stemness was also strongly associated with cell-intrinsic suppression of endogenous retroviruses and type I IFNsignaling, and increased expression of multiple therapeutically ac-cessible immunosuppressive pathways. Thus, stemness is not onlya fundamental process in cancer progression but may provide amechanistic link between antigenicity, intratumoral heterogene-ity, and immune suppression across cancers.

cancer stemness | antitumor immunity | intratumoral heterogeneity

Tumor infiltration by T cells has been associated with improvedclinical outcomes in a broad range of tumor types. Despite this,

a large proportion of solid cancers appears nonpermissive to lym-phocyte infiltration or nonimmunogenic (immunologically “cold”),and thus protected from cytolytic attack by lymphocytes, such asCD8+ T cells (1). The advent of immunotherapies, such as immunecheckpoint inhibitors that rely on preexisting antitumor immuneresponses, has made an improved understanding of the mechanismsunderlying the cold tumor phenotype essential.Mounting evidence suggests that tumor cells can exhibit stem

cell-like properties, ranging from characteristic gene-expressionprofiles to experimentally validated long-term self-renewal andrepopulation capacities. The cancer stem cell (CSC) hypothesisposits that a subpopulation of tumor cells is capable of self-renewal and is responsible for the long-term maintenance oftumors (2). This hypothesis provides compelling explanations forclinical observations, such as therapeutic resistance, tumor dor-mancy, and metastasis (3). CSCs have been identified in a varietyof human tumors, as assayed by their ability to initiate tumorgrowth in immunocompromised mice (4, 5). However, consid-erable controversy remains as to how best to define CSCs andthe extent to which different tumor types exhibit a hierarchicalorganization. These controversies notwithstanding, there is in-creasing evidence that stem cell-associated molecular features,often referred to as “stemness,” are biologically important incancer (6). It is unclear whether the stemness phenotype reflectsthe presence of bona fide CSCs in tumors or simply the cooptingof stem cell-associated programs by non-CSC tumor cells (or

both). Whatever the underlying mechanism may be, stemness hasemerged as an important phenomenon due to its strong associ-ation with poor outcomes in a wide variety of cancers (7, 8).Moreover, stemness appears to be a convergent phenotype incancer evolution (9, 10), suggesting it is a fundamentally im-portant property of malignancy.The evolution of transformed cells in the tumor microenvi-

ronment is shaped by diverse selective pressures, including thehost immune response. Experimental work has shown that em-bryonic, mesenchymal, and induced pluripotent stem cellspossess immune modulatory properties, while resistance to immune-mediated destruction has also recently been shown to be an intrin-sic property of quiescent adult tissue stem cells (11) andCSCs (12). Similarly, immune selection has been shown to drivetumor evolution toward a stemness phenotype that inhibits cy-totoxic T cell responses (13). Moreover, a recent analysis of TheCancer Genome Atlas (TCGA) revealed negative associationsbetween stemness and some metrics of tumor leukocyte in-filtration (14). Finally, CSCs have been proposed as a driver ofintratumoral heterogeneity (6, 9). Consistent with this, we (15)

Significance

The exclusion of immune cells from the tumor microenvironmenthas been associated with poor prognosis in the majority of can-cers. We report that when considering 21 solid cancer types, im-mune cell exclusion is widely associated with the presence of astem cell-like phenotype in tumors (“stemness”). Stemness posi-tively correlates with higher intratumoral heterogeneity, possiblyby protecting antigenic clones from elimination by the immunesystem. The activation of a stemness program appears to limitantitumor immune responses via tumor cell-intrinsic silencing ofendogenous retrovirus expression, repression of type I interferonsignaling, and up-regulation of immunosuppressive checkpoints.Our work suggests that targeting the stemness phenotype incancer will promote T cell infiltration and render tumors moreresponsive to immune control.

Author contributions: A. Miranda, P.T.H., and B.H.N. designed research; A. Miranda andP.T.H. performed research; E.B., A. Mezheyeuski, J.B., P.M., and A.d.R. contributed newreagents/analytic tools; A. Miranda, P.T.H., A.W.Z., S.P., and B.H.N. analyzed data; B.H.N.conducted study supervision; and A. Miranda, P.T.H., and B.H.N. wrote the paper.

The authors declare no conflict of interest.

This article is a PNAS Direct Submission.

Published under the PNAS license.1A. Miranda. and P.T.H. contributed equally to this work.2To whom correspondence should be addressed. Email: bnelson@bccrc.ca.

This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.1073/pnas.1818210116/-/DCSupplemental.

Published online April 17, 2019.

9020–9029 | PNAS | April 30, 2019 | vol. 116 | no. 18 www.pnas.org/cgi/doi/10.1073/pnas.1818210116

Dow

nloa

ded

by g

uest

on

June

19,

202

0

and others (16) have reported negative associations betweenimmune cell infiltration and intratumoral heterogeneity.Motivated by these observations, we hypothesized that the

stemness phenotype of cancer cells may confer immunosup-pressive properties on tumors, resulting in immunologically coldmicroenvironments that both foster and maintain intratumoralheterogeneity. To address this, we performed an integratedanalysis of stemness, immune response, and intratumoral het-erogeneity across cancers. We recover pervasive negative asso-ciations between antitumor immunity and stemness, and strongpositive associations between stemness and intratumoral het-erogeneity. We further find that cancer cell lines with highstemness have cell-intrinsic immunosuppressive features, sug-gesting that immunologically cold microenvironments can arisedue to the presence of high-stemness cancer cells. We proposethat cancer stemness provides a link between tumor antigenicity,intratumoral heterogeneity, immune suppression, and theresulting evolutionary trajectories in human cancer.

ResultsDerivation and Comparison of Stemness Signatures. Recent studieshave provided evidence that cancer stemness can be representedby core gene-expression programs across diverse cancer types (8,10, 14, 17, 18). Building on this prior work, we inferred tumorstemness from cancer transcriptomes using single-sample geneset enrichment analysis (ssGSEA) with a modified version of agene set developed by Palmer et al. (17) to measure the level ofplasticity and differentiation of mesenchymal stem cells, plurip-otent stem cells, terminally differentiated tissues, and humantumors across >3,200 microarray samples. Intriguingly, the au-thors of this gene set identified a cluster of “immune” genes thatnegatively loaded the principal components they used to inferstemness; however, they did not further explore this relationship.To adopt this gene set for use in ssGSEA and avoid biasing ouranalysis toward recovering negative associations between stemnessand immunity, we omitted this immune gene cluster from our sig-nature. We also omitted cell proliferation markers to avoid recov-ering a signature of proliferation rather than stemness (19).We validated the performance of the resulting 109 gene signa-

tures (Dataset S1A) on diverse datasets, finding that it recapitulatedthe expected degree of stemness in both malignant and nonmalig-nant cell populations (SI Appendix, Fig. S1). We further validatedthis signature using the stem cell-based validation dataset used byMalta et al. (14) in their recent pan-cancer stemness analysis(GSE30652), which revealed similarly high classification accuracy forour signature compared with theirs [multiclass area under the curve(AUC) 0.92 vs. 0.91, respectively] (SI Appendix, Fig. S2 A and B).We then analyzed RNA sequencing data from 8,290 primary cancersrepresenting 21 solid cancer types (from TCGA) and found oursignature showed good concordance with that of Malta et al. andtwo other recently published signatures (Spearman’s ρ = 0.43, 0.74,and 0.66, respectively) (SI Appendix, Fig. S2 C–F).

Stemness Varies Across Cancers and Predicts Patient Survival. Usingour signature, we found that stemness varied strongly acrossTCGA samples, with cancer type explaining 54% of the variation(ANOVA; adjusted R2) (Fig. 1A). Consistent with prior reportsof stemness being a negative prognostic factor (7, 8), we found astrong negative relationship between median stemness and medianoverall survival across cancer types (Fig. 1B) (ρ = −0.60; P = 0.004;n = 21 cancers). Within cancer types, Cox regression likewiseshowed stemness to be significantly negatively prognostic for overallsurvival in the majority of cancers (Cox proportional hazards; P <0.05), and significantly positively prognostic for none (Fig. 1C),underscoring the relevance of this signature both within and acrosscancers. We also noted a significant decrease in the magnitude ofthe hazard associated with stemness within cancers as medianstemness increased (ρ = −0.59; P < 0.01) (Fig. 1C), pointing to a

potential threshold effect with a saturating hazard in cancers withhigher average stemness. For reference, we compared the prog-nostic association of our ssGSEA-based stemness with the mRNAsisignature of Malta et al. (14) derived using one-class logistic re-gression (OCLR), for which positive associations in some cancerswere reported. Using pan-cancer Cox regressions stratified bycancer, we found ssGSEA-based stemness to be substantially morepredictive of survival in this modeling framework [log hazard ra-tio = 0.23 ± 0.03 (coefficient ± SE); P < 10−15 vs. not significant formRNA stemness index (mRNAsi)], demonstrating this signatureuncovers the negative outcomes associated with high-stemnesscancers expected from previous reports (7, 8). This relationshipheld when controlling for tumor purity, demonstrating that a re-lationship between tumor purity and patient survival is not a con-founding factor (and see below).

Stemness Negatively Associates with Immune Cell Infiltration AcrossSolid Cancers. To evaluate the relationship between stemness andantitumor immunity, we generated signatures of predicted immunecell infiltration for each patient sample using xCell, an ssGSEA-based tool that infers cellular content in the tumor microenviron-ment (20). CD8+ T cells, which have a well-established association

Fig. 1. Stemness and survival across cancers. (A) Stemness score varieswidely across 21 solid cancers from TCGA. Each point represents an indi-vidual case, and cancer types are ordered by median stemness score (z-scoredssGSEA). (B) Median survival decreases with increasing median stemness (P =0.004). Gray points represent cancers in which median overall survival timeswere not evaluable. (C) Stemness associates with poor outcome withincancers. Log hazard ratio (±95% CI) for the association of stemness withoverall survival is shown. Hazard decreases with increasing average stemnessof cancers (P = 0.008). Cox models control for patient age and tumor purity.Cancer acronyms are used as defined by TCGA (https://portal.gdc.cancer.gov).

Miranda et al. PNAS | April 30, 2019 | vol. 116 | no. 18 | 9021

MED

ICALSC

IENCE

S

Dow

nloa

ded

by g

uest

on

June

19,

202

0

with favorable prognosis in a majority of solid cancers (21), showeda clear negative association with stemness for most cancers (Fig.2A). We also considered other cell types important for antitumorimmunity, including NK cells and B cells, and again observed re-current negative associations with stemness (Fig. 2A). Additionalcell types, such as CD4+ T cells, Tregs, and neutrophils, showedmore variable associations with stemness, indicating this relation-ship does not apply to all infiltrating immune cell populations in allcancers (Fig. 2A).To generate a robust single score for antitumor immunity

(hereafter referred to as “immune signature”), we aggregatedxCell predictions for CD8+ T cell, NK cell, and B cell infiltration,reasoning that these cells represent important anticancer effec-tor cells across diverse cancers (22, 23). Indeed, this immunesignature was significantly associated with increased survival inthe majority of cancers (SI Appendix, Fig. S3A; overall survivalcurves for all patients stratified by median stemness and immunesignature are shown in SI Appendix, Fig. S3B). A notable exception

was kidney cancer (i.e., KIRP), where a negative association wasobserved, in accord with prior reports for this malignancy (24).Consistent with our analyses using single immune cell-type scores(Fig. 2A), the immune signature showed a negative association withstemness within nearly all cancers (Fig. 2B). This negative associ-ation was also recovered using other published stemness gene sets(18, 19, 25) (Dataset S1B), most clearly with stemness scoresreflecting NANOG, SOX2, and MYC signaling, and to a lesserextent with those reflecting embryonic stem cell programs (SIAppendix, Fig. S4). Furthermore, when we used CIBERSORTinfiltration scores (26) in place of our immune signature, wefound that increased stemness was associated with strong polari-zation of infiltrating leukocyte populations toward a macrophage-dominated, CD8+ T cell-depleted composition for most cancers(linear model controlling for cancer type; P < 10−15 for bothcell fractions).Whereas the preceding analyses were performed within cancer

types, we also evaluated the relationship between stemness and

Fig. 2. Stemness negatively associates with immune cell signatures. (A) Circos plot showing the association between stemness score and the presence of8 inferred immune cell types across 21 cancer types (colored bars in outer ring). The color and height of the inner bars represent the Spearman correlation ρvalues for each cell type and cancer type. (B) Volcano plot showing the association between stemness score and immune signature (sum of z-scored signaturesof CD8+ T cells, NK cells, and B cells) for each cancer. The dashed line indicates Padj = 0.05. (C) Association between stemness score and immune signature inthe different molecular subtypes of endometrial (UCEC) and breast (BRCA) cancer, and within primary and metastatic melanoma (SKCM) samples (Padj < 10−7).Each point represents one case. Colors indicate the different molecular subtypes of UCEC and BRCA, or sample types for SKCM. CN, copy number; MSI,microsatellite instable; POLE, polymerase epsilon.

9022 | www.pnas.org/cgi/doi/10.1073/pnas.1818210116 Miranda et al.

Dow

nloa

ded

by g

uest

on

June

19,

202

0

immune signature across cancer types. Unexpectedly, we foundno significant association between median immune signature andmedian stemness score across cancers (P = 0.82). This suggeststhat factors other than stemness control the differences in im-mune cell infiltration across cancer types, while the associationwith stemness applies within individual cancer types. It alsodemonstrates that our stemness metric is not recovering negativeassociations with immunity simply due to the lower tumor puritythat is inextricably associated with the presence of infiltratingimmune cells.We next examined whether the negative association between

immune signature and stemness was influenced by tumor subtypeor stage. Here, we focused on breast and endometrial cancers,which have well-characterized subtypes with strong prognosticassociations, and melanoma, for which both primary and meta-static samples are available within TCGA. In breast cancer,stemness varied markedly across subtypes (ANOVA; adjusted R2 =0.26), with the basal subtype having the highest stemness, asexpected (19), and the luminal-A subtype the lowest (Fig. 2C). Inendometrial cancer, the highest stemness score was observed inhigh-copy number (CN) alteration (CN-high) and polymerase-epsilon mutant tumors, and the lowest stemness score was seenin low-CN alteration tumors (CN-low) [Tukey honest significantdifference (HSD); P = 0.003; CN-high vs. CN-low tumors]. Fi-nally, we observed a substantially higher stemness score inmetastatic compared with primary melanoma lesions (Fig. 2C).In all these cancers, we observed recurrent negative associationsbetween stemness and immune signature which remained sig-nificant when controlling for cancer subtype and tumor purity(see below; linear models; P < 10−7).To investigate in an unbiased manner whether processes apart

from antitumor immunity negatively correlate with stemness, weconducted differential expression tests to identify gene-expressionpatterns associated with the lowest versus highest stemness quintiles(<20th versus >80th percentiles) for each analyzed cancer type.Even with this unbiased approach, nearly all of the pathways re-currently enriched in low-stemness samples within a cancer wereimmune-related (SI Appendix, Fig. S5). Recognizing that the pres-ence of nonmalignant cells can confound expression analyses ofbulk-sequenced tumor samples by diluting tumor-specific expres-sion signatures, we performed additional analyses to control forsuch effects. First, using recently published estimates of purityacross TCGA (27), we found that the association between tumorpurity and stemness was negligible and failed to reach significancein a pan-cancer linear model controlling for cancer type (P = 0.23).Second, we refit the differential expression models described aboveto control for tumor purity and repeated the pathway enrichmentanalyses, still finding that immune-associated pathways wereenriched in low-stemness tumors (SI Appendix, Fig. S6).To evaluate the relative contributions of malignant versus

nonmalignant cells (e.g., stromal cells) to the stemness score, weapplied our stemness metric to a single-cell RNA sequencing(scRNA-seq) dataset of lung cancer (28), which comprises acomprehensive inventory of different cell types from the tumormicroenvironment (>52,000 cells from five patients). The aver-age stemness score was much higher in cancer cells than anyother cell type, including fibroblasts, myeloid cells, or lympho-cytes (SI Appendix, Fig. S7) (Tukey HSD; P < 10−15), stronglysupporting the notion that the stemness signature largely ema-nates from cancer cells rather than stromal or other cell types inthe tumor microenvironment.

Stemness Associates with Immunologically Cold Cancers Measuredvia Immunohistochemistry. To confirm the negative associationbetween stemness and lymphocyte infiltration, we turned to threepatient cohorts with matched immunohistochemistry (IHC)-basedT cell infiltration scores and gene-expression data suitable forcomputing a stemness score. Using a cohort of 33 colorectal cancer

patients (29), we found a strong negative association betweenstemness and total infiltrating CD3+ T cells (ρ = −0.63; P < 0.001)(Fig. 3A). Furthermore, in this cohort the xCell-based immunesignature was strongly correlated with infiltrating CD3+ cells (ρ =0.69; P < 0.001), supporting its fidelity for measuring immune-cellinfiltration. With this validation in hand, we compared the xCell-based immune signature and stemness for the total patient cohortwith available microarray data (n = 585), and again observed a clearnegative correlation (ρ = −0.22; P < 10−7).We next evaluated this relationship in a cohort of 35 lung cancer

patients with matched RNA-sequencing and IHC-based quantita-tion of immune cell infiltrates (30). For consistency with the aboveanalysis, we calculated the infiltration of T cells by summing pre-viously calculated CD4+ and CD8+ cell fractions. Although weobserved a negative association between the stemness score and thepercent of infiltrating T cells, this did not reach statistical signifi-cance (ρ = −0.32; P = 0.07) (Fig. 3B). Nonetheless, the additionalRNA-seq data in this cohort revealed a clear negative associationbetween the stemness and immune signature at the transcriptionallevel (n = 199 tumor samples; ρ = −0.19; P = 0.007).Finally, we evaluated a small cohort of high-grade serous

ovarian cancer (HGSC) cases (15) for which matched IHC-basedT cell counts and microarray-based gene-expression data wereavailable for multiple tumor sites within each patient (n =44 samples from 12 patients). Consistent with the above findings,we found a negative association between stemness and totalCD3+ T cells (negative binomial mixed effects model; P =0.0028) (Fig. 3C). In addition, we previously subjected samplesfrom this cohort to Getis-Ord Gi* “hotspot” analysis to quantifyimmune cell engagement with tumor cells (15). Intriguingly, allhotspot metrics showed a clear negative association with stem-ness (mixed-effects models; P < 0.01; 44 samples from 12 pa-tients) (Fig. 3D, FCI shown) (31), suggesting that stemnessnegatively influences lymphocyte engagement with tumor cells.

Stemness Associates with Intratumoral Heterogeneity. Stemness hasbeen proposed to foster tumor clone diversity, with the replicativepotential of CSCs enabling greater tumor heterogeneity (6, 9). Ourhypothesis suggests that stemness could additionally promoteintratumoral heterogeneity by inhibiting immune selection againstnew cancer clones. These predictions have not, to our knowledge,been systematically tested within or across cancers. Therefore, wecompared stemness and intratumoral heterogeneity using data fromtwo recent TCGA studies (26, 32). Using data from the first study(32), we found a dramatic positive correlation between medianstemness and median number of clones in a cancer (n = 935 pa-tients across 11 cancers; ρ = 0.75; P = 0.008) (Fig. 4A). Further-more, we found a positive association between stemness and tumorclone count in a linear model controlling for cancer site and tumorpurity, indicating that these associations are discernible both acrossand within cancers (P = 0.0002) (Fig. 4B). Using pan-cancer pre-dictions of intratumoral heterogeneity from a second, largerTCGA-based study (26), we recovered a similarly strong association(ρ = 0.64; P = 0.002; n = 6,791 samples across 21 cancers) (Fig. 4C),which was again significant across all samples when controlling forcancer site and tumor purity in a linear model (P < 10−15) (Fig. 4D).

Potential Mechanisms and Consequences of Stemness-AssociatedImmunosuppression. We evaluated the association betweenstemness and several known genetic and environmental factorsthat affect antitumor immunity. Antitumor immunity involvesT cell recognition of neo-antigens arising from somatic muta-tions (33). Therefore, we examined the association betweenstemness, immune signature, and nonsynonymous mutation load,analyzing TCGA samples with available mutation calls (n =6,682). While median mutation load correlated with medianimmune signature across cancers (ρ = 0.49; P = 0.02, n = 21),there was generally little correlation within cancers, as has been

Miranda et al. PNAS | April 30, 2019 | vol. 116 | no. 18 | 9023

MED

ICALSC

IENCE

S

Dow

nloa

ded

by g

uest

on

June

19,

202

0

reported in other TCGA-based analyses (34) (SI Appendix, Fig.S8A). Across cancers, median mutation load showed a positiveassociation with median stemness (ρ = 0.52; P = 0.015, n = 21) (Fig.5A). Similarly, within cancers, we generally found positive associa-tions between mutation load and stemness (Fig. 5B). We recoveredqualitatively similar but slightly stronger associations betweenstemness and neoantigen load computed with NetMHCpan in theabove pan-cancer analysis (26) (SI Appendix, Fig. S8 B and C).Cancer-testis (CT) antigens are tumor antigens that normally

are expressed in gametogenic tissue but become aberrantlyexpressed in a broad range of malignancies, often leading to animmune response (35). Using a set of 201 CT genes curated bythe CTdatabase (Dataset S2) (36), we generated an ssGSEA CTantigen score and found strong positive associations with stem-ness within cancers (Fig. 5C). Accordingly, there was a generallynegative association between CT antigen score and the immunesignature, which was significant in 8 of 21 cancers (Padj < 0.05)(SI Appendix, Fig. S9). Thus, like neoantigens, CT antigens showa positive association with stemness and a negative associationwith immune signature.Normal stem cells have been shown to suppress endogenous

retrovirus (ERV) expression, presumably to prevent insertionalmutagenesis in long-lived stem cell lineages (37). Conversely,ERV expression can be activated in cancer cells (38, 39), where itcan potentially elicit antitumor immune responses by activatingviral defense mechanisms and the type I IFN response (40) or by

yielding immunogenic foreign epitopes (41). Despite these pos-sibilities, associations between immunity and ERV expressionwere inconsistently observed in a recent pan-cancer analysis (42).To better understand this relationship, we investigated interac-tions between stemness, immune signature, and ERV expression.Because of the repetitive nature of ERVs, we used ERV-specificread-mappings (42) to evaluate ERV expression in 4,252 TCGAsamples that overlapped with our stemness and immune signatureanalysis. Using redundancy analysis (a constrained extension ofprincipal components analysis), we found that multivariate ERVexpression was not clearly associated with the immune signature,consistent with prior reports (42); members of the ERVK familywere an exception, showing moderate positive associations withimmune signature (Fig. 5D). In contrast, ERV expression showed apervasive negative association with stemness (P < 0.001) (Fig. 5D),consistent with the notion that suppression of ERV expression is afeature of the stem cell phenotype (43). Thus, both immune sig-nature and ERV expression are negatively associated with stem-ness, but these appear to be largely orthogonal relationships.

Tumor Cell Intrinsic Mechanisms of Stemness-Mediated Immuno-suppression. To address whether the negative association be-tween stemness and immune signature is attributable to cancercell-intrinsic processes, we calculated stemness scores for 1,048cancer cell lines using gene-expression data from the Cancer CellLine Encyclopedia (CCLE) (44), as well as the cancer cell fractionfrom the scRNA-seq study of lung cancer, mentioned above (28).

Fig. 3. Relationship between stemness and immune cell infiltrates in different cancer cohorts scored via IHC. (A) Stemness score negatively associates withtumor-infiltrating CD3+ T cells in colorectal cancer (P < 0.01; n = 33; data from ref. 29). Each point represents one patient sample. (B) Stemness negativelyassociates with tumor-infiltrating T cells (in this case the sum of CD4+ and CD8+ cells) in lung cancer (P = 0.07; n = 35; data from ref. 30). Colors denote adenoversus squamous cell lung cancers. (C) Stemness is significantly associated with tumor-infiltrating CD3+ T cells in a multisite dataset of high-grade ovariancancer (P = 0.028; n = 44 samples from 12 patients; data from ref. 15). Colors represent individual patients. (D) By hotspot analysis, the fraction of tissue areaoccupied by colocalizing tumor and immune cells (FCI) is negatively associated with stemness in a multisite dataset from high grade ovarian cancer (P < 0.01;44 samples from 12 patients; data from ref. 15). Colors represent individual patients.

9024 | www.pnas.org/cgi/doi/10.1073/pnas.1818210116 Miranda et al.

Dow

nloa

ded

by g

uest

on

June

19,

202

0

We first assessed associations between stemness scores and theexpression of 11 mapped ERVs in CCLE transcriptomes and foundthat three of three ERVs with nonnegligible expression levelsacross cell lines were significantly negatively associated with stem-ness, two of which remained significant when controlling for tissueof origin (linear models; Padj < 0.05) (SI Appendix, Fig. S10). Wenext generated an ssGSEA score for type I IFN signaling (reactomeIFN α/β pathway gene set) and generally observed negative corre-lations with stemness within cancers in the TCGA dataset (Fig. 6A).When evaluating this association in the CCLE dataset, we observeda clear negative relationship between the IFN signature andstemness (ρ = −0.22. P < 10−10) (SI Appendix, Fig. S11A), whichremained significant when controlling for tissue of origin of the cellline (linear model; P < 0.001) and when omitting cell lines derivedfrom hematopoietic lineages (linear model; P = 0.02).Analysis of the aforementioned lung cancer scRNA-seq

dataset (28) also revealed a striking negative association be-tween cancer cell-intrinsic stemness and IFN α/β signaling in fourof five patients (Fig. 6B). To further test this association innonneoplastic lineages, we took advantage of the stem cell gene-expression dataset previously used to validate stemness signa-tures (GSE30652), which likewise showed a striking negativeassociation between stemness and type I IFN signaling (ρ = −0.81;P < 10−15) (SI Appendix, Fig. S11B).Finally, to examine other cell-intrinsic mechanisms of immu-

nosuppression, we analyzed a curated list of immunosuppressivegenes previously reported to be expressed in human cancer cells.Using both the CCLE and pan-cancer TCGA datasets, the ex-pression of each of these genes was assessed in relation to ourstemness signature (Fig. 6 C and D). This revealed positive as-sociations between stemness and a number of immunosuppres-

sive genes, including CD276 [B7-H3, shown to inhibit T cellactivation and autoimmunity (45)], PVR [CD155, a member ofthe B7/CD28 superfamily, shown to exhibit potent inhibitoryaction in different subsets of immune cells (46)], and TGFB1 [akey player in the induction of immunological tolerance (47)].Thus, the stemness phenotype is associated with expression ofseveral gene products that could potentially serve as targets forimmune modulation.

DiscussionAlthough cancer stemness, antitumor immunity, and intra-tumoral heterogeneity have all emerged as important featuresof cancer in recent years, their covariation across cancers hasnot been systematically investigated. Here we report thatstemness is associated with suppressed immune response,higher intratumoral heterogeneity, and dramatically worseoutcome for the majority of cancers. Although correlativeanalyses such as ours do not reveal causality, we propose thatthe stemness phenotype found in cancer cells, similar to that innormal stem cells, involves the expression of immunosuppres-sive factors that engender the formation of immune-privilegedmicroenvironments in which tumor clone diversification canoccur. The resulting heterogeneity may provide a substrate forthe selection of treatment-resistant clones, resulting in inferiorclinical outcomes. Thus, our findings implicate stemness as a sharedtherapeutic target to achieve the dual objectives of constrainingtumor evolution and enhancing antitumor immunity.A recent pan-cancer analysis reported inconsistent relation-

ships between cancer stemness and immunity, recovering nega-tive relationships between stemness and tumor-infiltratinglymphocytes for some cancers and positive relationships for

Fig. 4. Stemness associates with intratumoral heterogeneity within and across cancers. (A) Median stemness and median clonality [inferred by Andor et al.(32)] are strongly correlated across cancers (n = 11; P = 0.008). (B) Stemness score and clonality [inferred by Andor et al. (32)] are correlated across patientswhile controlling for cancer type (n = 935; P = 0.0002). Colored points represent different tumor sites. (C) Median stemness score and median intratumoralheterogeneity score [inferred by Thorsson et al. (26)] are strongly correlated across cancers (n = 20; P = 0.002). (D) Stemness score and intratumoral het-erogeneity [inferred by Thorsson et al. (26)] are correlated across patients while controlling for cancer type (n = 6,791; P < 10−15). Colored points representdifferent tumor sites. Spearman ρ values are shown.

Miranda et al. PNAS | April 30, 2019 | vol. 116 | no. 18 | 9025

MED

ICALSC

IENCE

S

Dow

nloa

ded

by g

uest

on

June

19,

202

0

others (14). While this work provided a valuable perspective onstemness across cancers, our results differ in that we recovermuch stronger and more pervasive negative relationships withimmune infiltration and survival. In contrast to our ssGSEAapproach, the OCLR approach used above was validated on acohort that lacked any malignant samples. Moreover, when wereproduced their analysis, we found many components of theirstemness score (i.e., positive OCLR model weights) were im-munologically relevant genes; for example, among the 50 mostpositive gene weights were IDO1, LCK, KLRG2, PSMB9 (acomponent of the immunoproteasome), and multiple TNF-receptors. With the OCLR approach, higher expression ofsuch immune genes contributes positively to the stemness score,precluding an unbiased assessment of the relationship betweenstemness and tumor immunity. As mentioned, it is also possiblethat an immune cell signature in the tumor microenvironmentcould negatively correlate with transcriptional signatures fromcancer cells simply through dilution of cancer-specific tran-scripts by infiltrating immune cells. We took numerous precautionsto ensure this was not driving our results. These included con-trolling for tumor purity in linear models throughout the analyses,

showing a nonsignificant relationship between stemness andtumor purity, and identifying a negative relationship betweenstemness and cell-intrinsic IFN signaling in individual lungcancer cells (by scRNA-seq; Fig. 6B) and in cancer cell line tran-scriptomes (SI Appendix, Fig. S11A).A contribution of cancer stemness to intratumoral heteroge-

neity has been postulated for some time (6, 9), but direct evi-dence has been lacking. We recovered a dramatic positiveassociation between stemness and multiple metrics of intra-tumoral heterogeneity across cancers (Fig. 4), which is especiallynoteworthy given that these metrics were derived from differentdata types (i.e., mRNA vs. DNA). Given recent work from ourgroup and others linking increased intratumoral heterogeneitywith decreased immune cell infiltration (15, 16), one couldspeculate that stemness might contribute to intratumoral het-erogeneity by both increasing the replicative capacities of indi-vidual tumor clones and by shielding antigenic clones fromelimination by the immune system.We found generally positive associations between stemness

and mutation load within cancers (Fig. 5B), and clear evidence ofthis across cancers (Fig. 5A), consistent with studies demonstrating

Fig. 5. Mutation load, CT antigen expression and ERV associations with stemness. (A) Median stemness and median mutation load are positively correlatedacross cancers (n = 21; P = 0.015). Mutation load is represented as log-transformed nonsynonymous mutations per base (log10 ns mutations per base pair). (B)Volcano plot reveals stemness score and mutation load correlate within some cancers (upper right quadrant). The x axis represents Spearman correlation ρvalues, and the y axis represents −log10-adjusted P values (Padj). Dashed red line indicates the significance threshold, Padj value = 0.05. (C) Stemness score andCT antigen expression (ssGSEA of CT antigen gene set) positively correlate in most cancers. Bar plots show the Spearman ρ values for each cancer type, andasterisks denote Padj < 0.05. (D) Redundancy analysis triplot reveals stemness negatively associates with multivariate ERV expression (P < 0.001; 33 ERVsevaluated in 4,252 samples, analysis conditioned by cancer type).

9026 | www.pnas.org/cgi/doi/10.1073/pnas.1818210116 Miranda et al.

Dow

nloa

ded

by g

uest

on

June

19,

202

0

accumulation of mutations in normal adult stem cells (48). Wealso found strong positive associations between stemness and CTantigen expression (Fig. 5C), which is consistent with reports ofCT antigen expression in mesenchymal (49), embryonic stemcells (50), and CSCs (51). It is also consistent with prior studies(42) and the present work finding nonsignificant or negativecorrelations between immune infiltration and CT antigen ex-pression (Fig. 5D). Thus, the negative association between stemnessand immune response is not readily attributable to low neoantigenor CT antigen load, strongly implicating the involvement of othermechanisms.ERVs, which constitute ∼8% of the human genome, are

known to be suppressed in pluripotent and embryonic stemcells (43) yet activated in human cancer (38, 39), leading us toask which behavior would predominate in high stemness

cancers. We found a strong negative association betweenstemness and ERV expression in both TCGA (Fig. 5D) andcancer cell line data (SI Appendix, Fig. S10), indicating thatthe stemness phenotype in human cancer retains this propertyof normal stem cells.We identified a negative correlation between stemness and

type I IFN signaling in the TCGA (Fig. 6A), CCLE (SI Ap-pendix, Fig. S11A), and scRNA-seq datasets (Fig. 6B).Whether ERV suppression underlies the low intrinsic IFNsignaling in high stemness cancer cells awaits experimentalinvestigation. In support of our results, an attenuated innateimmune response is a major characteristic of embryonic stemcells (52), and we find clear confirmation of this in nonneo-plastic stem cell transcriptomes (SI Appendix, Fig. S11B).Collectively, these data suggest that activation of a stemness

Fig. 6. Cell-intrinsic stemness score associates with decreased type I IFN signaling and increased expression of CD276 and PVR. (A) Volcano plot reveals thatstemness score and type I IFN signaling (reactome IFN α/β pathway ssGSEA) are negatively correlated in most cancers (upper left quadrant). The x axisrepresents Spearman correlation ρ values, and the y axis represents −log10 adjusted P values (Padj). Dashed red line indicates the significance threshold,Padj = 0.05. (B) Stemness score is negatively associated with type I IFN signaling (reactome IFN α/β pathway ssGSEA) in tumor cells from four of five lungcancer patients, based on scRNA-seq data from ref. 28. Each point represents a single tumor cell. (C and D) Heatmaps showing Spearman correlations forstemness and select immunosuppressive genes based on data from the CCLE (C) and TCGA (D). Spearman correlations were calculated within tissuesrepresented in the CCLE by more than 10 cell lines. Genes are ranked according to the final column (“overall”), which represents the correlation across allsamples, irrespective of cell line or tumor type. Red-blue intensities reflect the correlation ρ values. Asterisks denote Benjamini–Hochberg-corrected sig-nificant associations (Padj < 0.05).

Miranda et al. PNAS | April 30, 2019 | vol. 116 | no. 18 | 9027

MED

ICALSC

IENCE

S

Dow

nloa

ded

by g

uest

on

June

19,

202

0

program in tumors could limit antitumor immune responses bysilencing ERVs and repressing type I IFN signaling in a cell-intrinsic manner (53).Using CCLE data, we found a clear association between

stemness and the expression of several immunosuppressivegenes, including CD276, PVR, and TGFB1 (Fig. 6C). These as-sociations are especially intriguing given that CCLE-based geneexpression profiles are independent of any ongoing influence ofthe immune system. CD276, a B7 family ligand, is now beingclinically targeted due to expression on both cancer cells andtumor-infiltrating blood vessels (54). Intriguingly, CD276 iscoexpressed with CD133, a marker that distinguishes cell pop-ulations enriched for CSCs in colorectal cancer (55). PVR is akey ligand in an emerging checkpoint pathway involving TIGIT,an inhibitory receptor expressed on T cells and other immunecells (46). Although no association between stem cells and PVRhas been described so far in mammalian stem cells, expression ofPVR can be activated by sonic hedgehog signaling (56), apathway essential for self-renewal and cell fate determination innormal and CSCs (57). TGFB1 and other TGFB family membershave well-documented roles in development (58) and CSC pro-liferation and maintenance (59, 60).Recent work has demonstrated that tumor-intrinsic onco-

genic signaling pathways have immune suppressive properties(61), but this too can be understood in the context of stemness.For example, molecular pathways involving WNT/B-catenin,MYC, PTEN, and LKB1 have been implicated in the in-hibition of antitumor immunity (61), yet they also play impor-tant roles in stem cell maintenance (62–65). Thus, stemnessmay provide a unifying framework for understanding how var-ious oncogenic signaling pathways engender an immunosup-pressive tumor microenvironment.Although it remains unclear if stemness metrics derived from

bulk tumor samples represent rare populations of bona fideCSCs or a wholesale shift of the cancer cell population toward ahigher stemness phenotype, our findings provide rationale fortherapeutic targeting of the stemness phenotype itself. In par-ticular, if stemness plays a causative role in the formation of coldtumor microenvironments, it may prove beneficial to targetspecific molecules or pathways that are inherent to the stemnessphenotype, such as the aforementioned immunosuppressivemolecules. High-stemness cancers might also be rendered moresensitive to immune control by administering drugs that inducecell differentiation to irreversibly disrupt the stemness phenotype(66). This might bring the additional benefit of constrainingfurther tumor evolution, creating the conditions for durableclinical responses.

Materials and MethodsData Acquisition and Processing. All analyses were conducted with R soft-ware newer than version 3.4.2. We extracted clinical parameters andmolecular subtypes for TCGA data from the pan-cancer curated clinicaldata of Liu et al. (67). Mutation data were downloaded from Firebrowse(www.firebrowse.org) as the number of nonsynonymous mutations perbase. Tumor purity (27), mRNAsi (14), intratumoral heterogeneity (26, 32),and CIBERSORT and neoantigen scores (26) were accessed by extractingthe relevant data for overlapping samples from the respective supplementalmaterials.

For correlations between stemness scores and IHC-based immune cellcounts, immune cell infiltration data were provided by the authors of therespective studies (29). We obtained matching expression data (micro-array or RNA-sequencing) from the Gene Expression Omnibus (GEO:GSE39582, GSE81089), or from the authors (15, 29, 30). Other expressiondata for validation was obtained from the GEO (GSE30652, GSE15192,GSE31257, GSE76009).

We accessed RNA sequencing as upper quartile-normalized fragments perkilobase of transcript per million mapped reads using the TCGAbiolinks R/Bioconductor package (68), for each cancer of interest, and expression datawere merged across cancers. For genes with multiple annotated transcripts,we selected the transcript with the highest expression to represent the gene,

then filtered the expression set to include only primary samples (except formelanoma, for which we included metastases), removed patients withduplicate samples, and removed any patients without a consensus purityscore in Aran et al. (27) to enable purity corrections in analyses. Formicroarray datasets, we converted probe IDs to human gene symbolsusing biomaRt (69), and retained the probe with the highest expressionfor each gene, as above. We accessed the scRNA-seq data and tSNE em-beddings in Lambrechts et al. (28) from EBI (E-MTAB-6149), while cell typeannotations and anonymized patient codes were provided by the au-thors. Where appropriate (e.g., for linear modeling), expression datawere log2(x + 1) -transformed.

We calculated stemness and other ssGSEA signatures using the GSVApackage in R (gene sets in Dataset S1) (70) without normalization, andsubsequently scaled values as z-scores within datasets of interest. For ssGSEAcalculations on scRNA-seq, we first omitted genes below the median of av-erage expression across samples; this left representation for 61 of 109 of thegenes in the stemness signature. xCell enrichment scores were calculated inR, using the rawEnrichment analysis (20) function, which omits scaling scoresto [0, 1] and correction for correlations among related cell types (20), as wesought to avoid introducing nonlinearities from these steps into analysis. Togenerate the immune signature, we summed z-scored signatures of celltypes of interest (CD8+ T cells, NK cells, B cells). Because z-scoring of ssGSEAscores was done within each dataset, these scores should not be directlycompared across datasets.

Quantification and Statistical Analysis. We used nonparametric Spearman’scorrelation to assess pairwise associations between variables of interestwithin cancers, or for median values across cancers, adjusting for multipletests using the Benjamini–Hochberg method, where appropriate. Foranalyses across multiple cancer types or subtypes, we used linear models,controlling for site or subtype as fixed main effects, and inspecting modelresiduals to ensure model assumptions were reasonable. For analysescontrolling for purity, purity (as consensus purity estimate from ref. 27) wasincluded as a covariate (main effect) in linear models.

We conducted survival analyses using Cox proportional hazards models,calculating 95% CIs on log hazard ratios. We tested model assumptionsusing cox.zph (71). Where there were significant violations of model as-sumptions (P < 0.05), we inspected model Schoenfield residuals. We foundthat for the few instances in which model assumptions were violated, thiswas attributable to higher than predicted survival for some long-termsurvivors. For analyses across all patients and cancer types, we stratifiedCox models by cancer type. For IHC data with multiple samples taken fromthe same patient, we modeled patient as a random effect in linear mixed-effects models implemented in lme4 in R (72) and assessed the signifi-cance of fixed effects of interest using likelihood-ratio tests on nestedmodels.

Pathway enrichment analysis was conducted using ReactomePA (73) aftertesting for differential expression between quintiles using limma (74). Forenrichment analysis, we selected the top 1,000 significantly down-regulatedgenes in high stemness cancers based on the moderated t statistic; in cancerswhere <1,000 genes were significantly down-regulated (Padj < 0.05), weselected all significantly down-regulated genes for downstream analysis.Significance of enrichment was evaluated at Padj < 0.05, and recurrentlyenriched pathways were defined as those that were significantly enriched inthe greatest number of cancers. In limma analyses that included purity as acovariate, purity was log-transformed for consistency with the trans-formation of expression values.

To assess ERV expression, we first variance-filtered mapped ERVs to selectthose above the median interquartile range of expression using the gene-filter R package (75). We conducted partial redundancy analysis using thevegan package (76), implementing the default distance metric (Euclidean).We conditioned the analysis by cancer type to control for cancer type-specific effects, and tested the significance of multivariate associations us-ing permutation tests (n = 1,000).

ACKNOWLEDGMENTS. We thank Dr. John Dick (University of Toronto) andDr. Julian Lum (BC Cancer) for critical review of the manuscript, and Dr. PeterWatson (BC Cancer) for helpful discussions. We are grateful to The CancerGenome Atlas and Cancer Cell Line Encyclopedia for access to data thatenabled this study. Funding was provided by the BC Cancer Foundation (toB.H.N.), Canadian Cancer Society (to B.H.N.), Canada’s Networks and Cen-tres of Excellence (to B.H.N.), Canadian Institutes of Health Research (toB.H.N.), Terry Fox Research Institute (to B.H.N.), Cancer Research Society(to B.H.N.), Vanier Canada Graduate Scholarship (to A.W.Z.), and Cana-dian Institutes of Health Research Postdoctoral fellowships (to P.T.H. andA. Miranda).

9028 | www.pnas.org/cgi/doi/10.1073/pnas.1818210116 Miranda et al.

Dow

nloa

ded

by g

uest

on

June

19,

202

0

1. Gajewski TF (2015) The next hurdle in cancer immunotherapy: Overcoming the non-T-cell-inflamed tumor microenvironment. Semin Oncol 42:663–671.

2. Batlle E, Clevers H (2017) Cancer stem cells revisited. Nat Med 23:1124–1134.3. Nassar D, Blanpain C (2016) Cancer stem cells: Basic concepts and therapeutic impli-

cations. Annu Rev Pathol 11:47–76.4. O’Brien CA, Pollett A, Gallinger S, Dick JE (2007) A human colon cancer cell capable of

initiating tumour growth in immunodeficient mice. Nature 445:106–110.5. Al-Hajj M, Wicha MS, Benito-Hernandez A, Morrison SJ, Clarke MF (2003) Prospective

identification of tumorigenic breast cancer cells. Proc Natl Acad Sci USA 100:3983–3988.

6. Kreso A, Dick JE (2014) Evolution of the cancer stem cell model. Cell Stem Cell 14:275–291.

7. Ng SW, et al. (2016) A 17-gene stemness score for rapid determination of risk in acuteleukaemia. Nature 540:433–437.

8. Smith BA, et al. (2018) A human adult stem cell signature marks aggressive variantsacross epithelial cancers. Cell Rep 24:3353–3366.e5.

9. Greaves M (2013) Cancer stem cells as ‘units of selection’. Evol Appl 6:102–108.10. Chen H, He X (2016) The convergent cancer evolution toward a single cellular des-

tination. Mol Biol Evol 33:4–12.11. Agudo J, et al. (2018) Quiescent tissue stem cells evade immune surveillance.

Immunity 48:271–285.e5.12. Bruttel VS, Wischhusen J (2014) Cancer stem cell immunology: Key to understanding

tumorigenesis and tumor immune escape? Front Immunol 5:360.13. Noh KH, et al. (2012) Cancer vaccination drives Nanog-dependent evolution of tumor

cells toward an immune-resistant and stem-like phenotype. Cancer Res 72:1717–1727.14. Malta TM, et al. (2018) Machine learning identifies stemness features associated with

oncogenic dedifferentiation. Cell 173:338–354.e15.15. Zhang AW, et al. (2018) Interfaces of malignant and immunologic clonal dynamics in

ovarian cancer. Cell 173:1755–1769.16. Safonov A, et al. (2017) Immune gene expression is associated with genomic aber-

rations in breast cancer. Cancer Res 77:3317–3324.17. Palmer NP, Schmid PR, Berger B, Kohane IS (2012) A gene expression profile of stem

cell pluripotentiality and differentiation is conserved across diverse solid and hema-topoietic cancers. Genome Biol 13:R71.

18. Shats I, et al. (2011) Using a stem cell-based signature to guide therapeutic selectionin cancer. Cancer Res 71:1772–1780.

19. Ben-Porath I, et al. (2008) An embryonic stem cell-like gene expression signature inpoorly differentiated aggressive human tumors. Nat Genet 40:499–507.

20. Aran D, Hu Z, Butte AJ (2017) xCell: Digitally portraying the tissue cellular hetero-geneity landscape. Genome Biol 18:220.

21. Gooden MJ, de Bock GH, Leffers N, Daemen T, Nijman HW (2011) The prognosticinfluence of tumour-infiltrating lymphocytes in cancer: A systematic review withmeta-analysis. Br J Cancer 105:93–103.

22. Melero I, Rouzaut A, Motz GT, Coukos G (2014) T-cell and NK-cell infiltration intosolid tumors: A key limiting factor for efficacious cancer immunotherapy. CancerDiscov 4:522–526.

23. Sarvaria A, Madrigal JA, Saudemont A (2017) B cell regulation in cancer and anti-tumor immunity. Cell Mol Immunol 14:662–674.

24. Jochems C, Schlom J (2011) Tumor-infiltrating immune cells and prognosis: The po-tential link between conventional cancer therapy and immunity. Exp Biol Med(Maywood) 236:567–579.

25. Bhattacharya B, et al. (2004) Gene expression in human embryonic stem cell lines:Unique molecular signature. Blood 103:2956–2964.

26. Thorsson V, et al.; Cancer Genome Atlas Research Network (2018) The immunelandscape of cancer. Immunity 48:812–830.e14.

27. Aran D, Sirota M, Butte AJ (2015) Systematic pan-cancer analysis of tumour purity.Nat Commun 6:8971.

28. Lambrechts D, et al. (2018) Phenotype molding of stromal cells in the lung tumormicroenvironment. Nat Med 24:1277–1289.

29. Becht E, et al. (2016) Estimating the population abundance of tissue-infiltrating im-mune and stromal cell populations using gene expression. Genome Biol 17:218.

30. Mezheyeuski A, et al. (2018) Multispectral imaging for quantitative andcompartment-specific immune infiltrates reveals distinct immune profiles that classifylung cancer patients. J Pathol 244:421–431.

31. Nawaz S, Heindl A, Koelble K, Yuan Y (2015) Beyond immune density: Critical role ofspatial heterogeneity in estrogen receptor-negative breast cancer. Mod Pathol 28:766–777.

32. Andor N, et al. (2016) Pan-cancer analysis of the extent and consequences of intra-tumor heterogeneity. Nat Med 22:105–113.

33. Matsushita H, et al. (2012) Cancer exome analysis reveals a T-cell-dependent mech-anism of cancer immunoediting. Nature 482:400–404.

34. Senbabaoglu Y, et al. (2016) Tumor immune microenvironment characterization inclear cell renal cell carcinoma identifies prognostic and immunotherapeutically rele-vant messenger RNA signatures. Genome Biol 17:231.

35. Simpson AJ, Caballero OL, Jungbluth A, Chen YT, Old LJ (2005) Cancer/testis antigens,gametogenesis and cancer. Nat Rev Cancer 5:615–625.

36. Almeida LG, et al. (2009) CTdatabase: A knowledge-base of high-throughput andcurated data on cancer-testis antigens. Nucleic Acids Res 37:D816–D819.

37. Gaudet F, et al. (2004) Dnmt1 expression in pre- and postimplantation embryogenesisand the maintenance of IAP silencing. Mol Cell Biol 24:1640–1648.

38. Strissel PL, et al. (2012) Reactivation of codogenic endogenous retroviral (ERV) en-velope genes in human endometrial carcinoma and prestages: Emergence of newmolecular targets. Oncotarget 3:1204–1219.

39. Schmitt K, Reichrath J, Roesch A, Meese E, Mayer J (2013) Transcriptional profiling ofhuman endogenous retrovirus group HERV-K(HML-2) loci in melanoma. Genome BiolEvol 5:307–328.

40. Chiappinelli KB, et al. (2015) Inhibiting DNA methylation causes an interferon re-sponse in cancer via dsRNA including endogenous retroviruses. Cell 162:974–986.

41. Boller K, Janssen O, Schuldes H, Tönjes RR, Kurth R (1997) Characterization of theantibody response specific for the human endogenous retrovirus HTDV/HERV-K.J Virol 71:4581–4588.

42. Rooney MS, Shukla SA, Wu CJ, Getz G, Hacohen N (2015) Molecular and geneticproperties of tumors associated with local immune cytolytic activity. Cell 160:48–61.

43. Yang BX, et al. (2015) Systematic identification of factors for provirus silencing inembryonic stem cells. Cell 163:230–245.

44. Barretina J, et al. (2012) The Cancer Cell Line Encyclopedia enables predictive mod-elling of anticancer drug sensitivity. Nature 483:603–607.

45. Lee YH, et al. (2017) Inhibition of the B7-H3 immune checkpoint limits tumor growthby enhancing cytotoxic lymphocyte function. Cell Res 27:1034–1045.

46. Mahnke K, Enk AH (2016) TIGIT-CD155 interactions in melanoma: A novel co-inhibitory pathway with potential for clinical intervention. J Invest Dermatol 136:9–11.

47. Johnston CJ, Smyth DJ, Dresser DW, Maizels RM (2016) TGF-β in tolerance, develop-ment and regulation of immunity. Cell Immunol 299:14–22.

48. Blokzijl F, et al. (2016) Tissue-specific mutation accumulation in human adult stemcells during life. Nature 538:260–264.

49. Saldanha-Araujo F, et al. (2010) Cancer/testis antigen expression on mesenchymalstem cells isolated from different tissues. Anticancer Res 30:5023–5027.

50. Lifantseva N, et al. (2011) Expression patterns of cancer-testis antigens in humanembryonic stem cells and their cell derivatives indicate lineage tracks. Stem Cells Int2011:795239.

51. Yamada R, et al. (2013) Preferential expression of cancer/testis genes in cancer stem-like cells: Proposal of a novel sub-category, cancer/testis/stem gene. Tissue Antigens81:428–434.

52. Guo YL, et al. (2015) Attenuated innate immunity in embryonic stem cells and itsimplications in developmental biology and regenerative medicine. Stem Cells 33:3165–3173.

53. Fuertes MB, Woo SR, Burnett B, Fu YX, Gajewski TF (2013) Type I interferon responseand innate immune sensing of cancer. Trends Immunol 34:67–73.

54. Seaman S, et al. (2017) Eradication of tumors through simultaneous ablation ofCD276/B7-H3-positive tumor cells and tumor vasculature. Cancer cell 31:501–515.e8.

55. Bin Z, et al. (2014) Overexpression of B7-H3 in CD133+ colorectal cancer cells is as-sociated with cancer progression and survival in human patients. J Surg Res 188:396–403.

56. Solecki DJ, Gromeier M, Mueller S, Bernhardt G, Wimmer E (2002) Expression of thehuman poliovirus receptor/CD155 gene is activated by sonic hedgehog. J Biol Chem277:25697–25702.

57. Cochrane CR, Szczepny A, Watkins DN, Cain JE (2015) Hedgehog signaling in themaintenance of cancer stem cells. Cancers (Basel) 7:1554–1585.

58. Watabe T, Miyazono K (2009) Roles of TGF-beta family signaling in stem cell renewaland differentiation. Cell Res 19:103–115.

59. Peñuelas S, et al. (2009) TGF-beta increases glioma-initiating cell self-renewal throughthe induction of LIF in human glioblastoma. Cancer Cell 15:315–327.

60. Ikushima H, et al. (2009) Autocrine TGF-beta signaling maintains tumorigenicity ofglioma-initiating cells through Sry-related HMG-box factors. Cell Stem Cell 5:504–514.

61. Spranger S, Gajewski TF (2018) Impact of oncogenic pathways on evasion of anti-tumour immune responses. Nat Rev Cancer 18:139–147.

62. Nusse R (2008) Wnt signaling and stem cell control. Cell Res 18:523–527.63. Murphy MJ, Wilson A, Trumpp A (2005) More than just proliferation: Myc function in

stem cells. Trends Cell Biol 15:128–137.64. Hill R, Wu H (2009) PTEN, stem cells, and cancer stem cells. J Biol Chem 284:

11755–11759.65. Gurumurthy S, et al. (2010) The Lkb1 metabolic sensor maintains haematopoietic

stem cell survival. Nature 468:659–663.66. de Thé H (2018) Differentiation therapy revisited. Nat Rev Cancer 18:117–127.67. Liu J, et al. (2018) An integrated TCGA pan-cancer clinical data resource to drive high-

quality survival outcome analytics. Cell 173:400–416.e11.68. Colaprico A, et al. (2016) TCGAbiolinks: An R/bioconductor package for integrative

analysis of TCGA data. Nucleic Acids Res 44:e71.69. Durinck S, Spellman PT, Birney E, Huber W (2009) Mapping identifiers for the in-

tegration of genomic datasets with the R/bioconductor package biomaRt. Nat Protoc4:1184–1191.

70. Hänzelmann S, Castelo R, Guinney J (2013) GSVA: Gene set variation analysis formicroarray and RNA-seq data. BMC Bioinformatics 14:7.

71. Therneau TM (2018) survival: Survival Analysis. R Package Version 2.38. Available athttps://CRAN.R-project.org/package=survival. Accessed on December 20, 2018.

72. Bates D, Mächler M, Bolker BM, Walker SC (2015) Fitting linear mixed-effects modelsusing lme4. J Stat Softw 67:1–48.

73. Yu G, He QY (2016) ReactomePA: An R/bioconductor package for reactome pathwayanalysis and visualization. Mol Biosyst 12:477–479.

74. Ritchie ME, et al. (2015) limma powers differential expression analyses for RNA-sequencing and microarray studies. Nucleic Acids Res 43:e47.

75. Gentleman R, Carey V, Huber W, Hahne F (2018) genefilter: Methods for FilteringGenes from High-Throughput Experiments. R Package Version 1.64.0. Available atbioconductor.org/packages/genefilter. Accessed on December 20, 2018.

76. Oksanen J, et al. (2015) vegan: Community Ecology Package. R Package Version 2.5-4.Available at https://CRAN.R-project.org/package=vegan. Accessed on December 20,2018.

Miranda et al. PNAS | April 30, 2019 | vol. 116 | no. 18 | 9029

MED

ICALSC

IENCE

S

Dow

nloa

ded

by g

uest

on

June

19,

202

0