Tumor Purity as an Underlying Key Factor in Glioma Free

This article is featured in Highlights of This Issue, p. 5997

Glioma is the most common and lethal type of tumor in the central nervous system (CNS). Traditional histopathology classifies gliomas chiefly based on glioma cell features, including microscopic similarities with putative origin cells and presumed levels of differentiation. Emerging molecular advances integrate the genotype of glioma cells, including IDH mutation and the 1p/19q codeletion, into the 2016 CNS World Health Organization (WHO) classification (1). Even though increasing studies uncover mechanisms affecting the malignant properties of glioma cells and identify reasonable therapeutic targets (2, 3), the current clinical prediction and treatment outcome are still not satisfactory.

At present, we are deeply aware that glioma tissues contain abundant glioma-associated nontumor cells within their microenvironment. These nontumor cells play important roles in glioma development and are represented by stromal and immune cells (4, 5). Studies reveal that stromal cells are deeply involved in glioma proliferation, invasion, and angiogenesis (6–9). Furthermore, increasing evidence indicates that infiltrating immune cells have diverse roles in glioma biology (10, 11). Indeed, even nontumor cells dilute the purity of glioma cells. They mutually cooperate to maintain a subtle homeostasis for gliomagenesis, malignancy progression, and treatment resistance (12). However, current classification systems apply little attention to the nontumor fractions within glioma tissue, and there is limited knowledge on the characteristics of glioma cells under various purities.

A proper analytical model is necessary for exploring this issue. Recently, Yoshihara and colleagues have developed a universal algorithm based on transcriptomic expression data to infer tumor purity—quantifying tumor cell content in the tissue—and estimate infiltration of stromal and immune cells (13). Here, transcriptomic data of 2,249 gliomas and 29 normal brain tissues were collected for analysis. Tumor purity was calculated using the ESTIMATE algorithm (13). The role of glioma purity was systematically investigated in clinical, genomic, and biological conditions, which may provide novel insights into the glioma microenvironment and aid in the development of robust classification and predictive systems.

This study collected 2,249 gliomas and 29 normal brain tissues from four databases: CGGA (China), TCGA (the United States), REMBRANDT (the United States), and GSE16011 (the Netherlands). There are two available cohorts of gliomas in the TCGA database. Accordingly, five glioma cohorts were involved in this study: CGGA, TCGA-RNAseq, TCGA-microarray, REMBRANDT, and GSE16011. Based on the transcriptomic profiling method, we grouped CGGA and TCGA-RNAseq cohorts into an RNAseq set for discovery. Whereas, cases from TCGA-microarray, REMBRANDT, and GSE16011 cohorts were classified into a microarray set for validation.

The CGGA cohort consisted of 325 cases whose clinical and molecular information was obtained from the CGGA database (http://www.cgga.org.cn). Each case enrolled was treated by members of the CGGA group. Tumor tissue samples were collected at the time of surgery after informed consent. Neuropathologists established the diagnosis and ensured the quality of the tissue for molecular testing (14). Overall survival (OS) was calculated from the date of diagnosis until death or the end of follow-up. The point of death is defined by death certification, which could be obtained from local hospitals and police stations. Another 1,953 glioma cases were included from the TCGA (https://gdc.cancer.gov/), REMBRANDT (https://caintegrator.nci.nih.gov/rembrandt/), and GSE16011 (http://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE16011) databases. Information on these patients is available from corresponding data portal. Patient characteristics are described in Supplementary Tables S1 and S2.

Collection of tumor tissue and clinicopathologic information was undertaken with informed consent. Study protocols were approved by the ethics committees of the participating institutions. The study was conducted in accordance with the Declaration of Helsinki.

The libraries were sequenced on the Illumina HiSeq 2000 platform by the 101-bp pair-end sequencing strategy. Base calling (Illumina pipeline CASAVA v1.8.2) was used to convert original image data into sequence data. The sequence data were further processed to define improper reads according to standard quality control criteria. All reads that fit any of the following parameters were removed:

• (1) The reads that aligned to adaptors or primers with no more than two mismatches.

• (2) The reads with more than 10% unknown bases (N bases).

• (3) The reads with more than 50% of low-quality bases (quality value ≤ 5) in one read.

We downloaded the Hg19 RefSeq (RNA sequences, GRCh37) from the UCSC Genome Browser (http://genome.ucsc.edu). The gene expression was estimated using the reads per kilobase transcriptome per million reads (RPKM) method (15). The RPKM method could remove the influence of different gene lengths and sequencing discrepancies from the expression calculation. Therefore, we could directly use the calculated gene expression data to compare the differences in gene expression among samples.

Genomic DNA was extracted from frozen tissues using a QIAamp DNA Mini kit (Qiagen) according to the manufacturer's protocol. DNA concentration and quality were measured using a Nano-Drop ND-1000 spectrophotometer. Pyrosequencing of mutations in IDH1 or IDH2 was supported by Gene Tech and performed on a Pyro-Mark Q96 ID System (Qiagen). PCR amplification was performed using the primers 5′-GCT TGT GAG TGG ATG GGT AAA AC-3′, 5′-Biotin- TTG CCA ACA TGA CTT ACT TGA TC-3′ for IDH1 and 5′-ATC CTG GGG GGG ACT GTC TT-3′, 5′-Biotin-CTC TCC ACC CTG GCC TAC CT-3′ for IDH2. Pyrosequencing was performed using the primers 5′-TGG ATG GGT AAA ACC T-3′ for IDH1 and 5′-AGC CCA TCA CCA TTG-3′ for IDH2 (16).

Mutational hotspots in the TERT core promoter were obtained [nucleotide numbers 1 295 228 (C228T) and 1 295 250 (C250T)] from the human reference sequence (GrCh37 February 2009; http://genome.ucsc.edu/). Nested PCR was used to amplify the mutational hotspots. The first PCR primer sequences were 5′-GTC CTG CCC CTT CAC CTT-3′ (forward) and 5′-GCA CCT CGC GGT AGT GG-3′ (reverse), which yielded a 273-bp product. Further nested PCR was conducted using primers (Primer set 2) 5′-CCG TCC TGC CCC TTC ACC-3′ (forward) and 5′-GGG CCG CGG AAA GGA AG-3′ (reverse), which yielded a 128-bp product. We purified products from the second PCR using Illustra ExoProStar (GE Healthcare). The purified products were subjected to direct sequencing on an ABI 3100 PRISM DNA sequencer with a Big-Dye Terminator cycle sequencing kit (Applied Biosystems). The quality of all PCR products was checked before sequencing using electrophoresis on 2% agarose gels (17).

Bisulfite modification of DNA was performed using an EpiTect kit (Qiagen). The MGMT promoter region was amplified using primers as follows: 5′-GTT TYG GAT ATG TTG GGA TA-3′ and 5′-biotin-ACC CAA ACA CTC ACC AAA TC-3′. Gene Tech performed pyrosequencing analysis. We averaged the methylation values across seven CpG loci within the MGMT promoter. We defined methylation as samples with an average methylation value of >10%.

We reviewed 12 glioma tissues from the CGGA cohort to infer glioma purity through the visual evaluation of hematoxylin and eosin–stained slides. Results of morphologic purity were independently confirmed by two neuropathologists.

Paraffin-embedded tissues of 43 cases from the CGGA cohort were collected. Because there is no specific antibody to distinguish macrophage and microglia in the CNS, we selected antibody against Iba-1 (Abcam; ab5076, 1:1,000) to detect macrophage/microglia (5). Anti-neutrophil elastase (Abcam; ab68672, 1:1,000) was used to stain neutrophil. Immunohistochemistry of paraffin sections was performed as previously described (18). Briefly, the sections were incubated with primary antibody overnight at 4°C, and then incubated with respective secondary antibody (ZSGB, 1:200) at room temperature for 1 hour. After washing with PBS buffer, the sections were stained with DAB for 5 minutes, rinsed in water, and counterstained with hematoxylin. Quantitative evaluation was performed by examining each section using at least five different high-power fields (X40 objective and X10 eyepiece) with the most abundant stained cells. Stained cells were manually counted 3 times for each photograph independently by two investigators. The number of stained cells was determined by the average method.

Stromal and immune scores were calculated using the ESTIMATE R package (https://sourceforge.net/projects/estimateproject/), and glioma purity was estimated according to the formula described in Yoshihara and colleagues (13). Based on the expression data, cases were classified into relevant subtypes using the TCGA and EM/PM (EM is the abbreviation for EGFR module; PM is the abbreviation for PDGFRA module) classification schemes (19, 20). We used GISTIC2.0 to assess copy-number alterations (CNA) associated with glioma purity. Locus with GISTIC value more than 1 or less than −1 was defined as an amplification or deletion, respectively. Knowledge on gene interactions and pathways was obtained from the KEGG website (http://www.kegg.jp/; ref. 21). The principal components analysis (PCA) package of R was used to profile transcriptomic patterns attributing to glioma purity. Significance analysis of microarrays (SAM) was used to identify differential genes based on the threshold of fold change more than 3 and FDR less than 0.05. Spearman correlation analysis was used to evaluate the association between gene expression and glioma purity. Gene sets were submitted to the DAVID website (http://david.abcc.ncifcrf.gov/home.jsp) to perform gene ontology (GO) analysis and to retrieve relevant biological implications (22). The biological phenotype was further verified using gene set enrichment analysis (GSEA; ref. 23). Associations between immune cells and glioma purity were analyzed using the gene set variation analysis (GSVA) package of R (24). The gene list of immune cells was summarized by Gabriela and colleagues (25) and Oleg and colleagues (26).

Associations between continuous variables were tested using Spearman correlation analysis. Differences in variables between groups were evaluated by the Student t test, one-way ANOVA, or Yate's continuity corrected χ² tests. The Kaplan–Meier survival curve was depicted to estimate survival distributions. The log-rank test was used to assess statistical significance between groups. Univariate cox and further backward stepwise multivariate cox regression analyses were performed to identify independent prognostic factors. A nomogram was formulated based on the results of multivariate cox regression analysis (27). The predictive validity was assessed by comparing nomogram-predicted with observed survival probability. Bootstraps with 1,000 resamples were applied to these activities. The performance of the nomogram was evaluated using the concordance index (C-index). A larger C-index indicated a more accurate prognostic prediction. Patients with missing information were excluded from corresponding analysis. All statistical analyses were conducted using R 3.3.0, SPSS software, and GraphPad Prism 6.0 software. A two-sided P value less than 0.05 was regarded as significant.

A total of 2,249 gliomas and 29 normal brain tissues were included for analysis (Supplementary Tables S1 and S2). Four glioma cell lines (U87, U251, H4, and LN229 cell lines) were analyzed as references. We assessed glioma purity, as well as stromal and immune scores using the ESTIMATE algorithm. Morphology evaluation of glioma purity was conducted for 12 cases from the CGGA cohort (Supplementary Fig. S1A). Consistent with previous reports, glioma purity based on the ESTIMATE algorithm exhibited moderate correlation with morphology evaluation (Supplementary Fig. S1B and S1C; refs. 13, 28, 29). A high degree of purity was observed in glioma cell lines (median purity: 99.44%; Supplementary Table S3), verifying the validity of the ESTIMATE algorithm. There is a wide range of purity among the cases, 696 (30.95%) of the cases were highly purified (>90%), whereas 213 (9.47%) had poor purity (<60%; Supplementary Fig. S2B). In accordance with a previous report (13), we found a highly positive correlation between stromal and immune scores in glioma (R = 0.9021; Supplementary Fig. S2A).

Gliomas from the RNAseq and microarray sets were arranged in order of increasing purity (Fig. 1). The purity distribution was evaluated among groups and in association with other factors. We found a negative correlation between glioma purity and age at diagnosis (Fig. 1). Low purity gliomas were more likely to belong to mesenchymal and EM subtype. On the contrary, high purity gliomas were enriched in proneural and PM subtype (Supplementary Fig. S3A and S3E). With regard to genomic alterations, the IDH mutation and 1p/19q codeletion indicated higher glioma purity (Fig. 1; Supplementary Fig. S3). Even the TERT promoter mutation conferred lower purity in LGGs, and its influences were abolished in glioblastoma (GBM; Supplementary Fig. S3D). In addition, we also examined the correlation between purity and other parameters in stratified histopathology and molecular classification (Supplementary Table S4).

According to the WHO grading system and histogenetic features, gliomas were classified into corresponding entities. Stromal and immune scores significantly increased along with malignancy progression, whereas glioma purity decreased in higher grades (Fig. 2; Supplementary Table S3). With respect to histopathologic classification, oligodendroglioma had the lowest level of stromal and immune scores, but the highest purity. Contrarily, GBM was accompanied with high stromal and immune score, but low purity (Fig. 2; Supplementary Table S3).

Recent studies have incorporated important genetic alterations, such as IDH mutation, 1p/19q codeletion, and TERT promoter mutation, into glioma classification (30–32). Here, we divided gliomas into five molecular entities including lower-grade glioma (LGG)-Oligo (IDH-mutant LGG with TERT promoter mutation or 1p/19q codeletion), LGG-Astro (IDH-mutant LGG without TERT promoter mutation or 1p/19q codeletion), early-GBM (LGG with wild-type IDH), GBM-IDHmut (GBM with mutant IDH), and GBM-IDHwt (GBM with wild-type IDH). We found survival time was significantly reduced in an order of LGG-Oligo, LGG-Astro, Early-GBM, GBM-IDHmut, and GBM-IDHwt (Supplementary Fig. S4, P < 0.0001). In accordance with survival variation, stromal and immune scores were highest in GBM-IDHwt but lowest in LGG-Oligo (Fig. 2A–D; Supplementary Table S3). Glioma purity was highest in LGG-Oligo and lowest in GBM-IDHwt (Fig. 2E and F; Supplementary Table S3).

In addition, in RNAseq set, recurrent glioma or secondary GBM showed similar tumor purity as primary glioma (Supplementary Fig. S5A). When we focused on GBM, limited differences in tumor purity were observed among primary, recurrent, and secondary GBM (Supplementary Fig. S5B and S5C). These results indicated that recurrent glioma and secondary GBM had similar purity as primary cases.

We used the dichotomization on median value to separate cases for depicting the survival curves. Parallel analyses were conducted in both RNAseq and microarray sets. Cases with low purity generally had shorter OS than those with higher purity (Fig. 3A and B). Stratified analyses revealed that low purity conferred reduced OS in most WHO grades except grade II in the RNAseq set (Fig. 3C–H). Survival curves were also depicted in different subtypes (Supplementary Fig. S6), recurrent gliomas, and secondary GBMs (Supplementary Fig. S7). Even though some of the tests failed to achieve statistical significance, low purity still implied a trend suggestive of inferior outcome in most subtypes and recurrent gliomas (Supplementary Figs. S6 and S7). In addition, similar analyses were applied for stromal (Supplementary Figs. S8 and S9) and immune scores (Supplementary Figs. S10 and S11), indicating that they were unfavorable indicators for prognosis.

We next conducted cox regression analyses to assess the prognostic independence of glioma purity among other factors. Glioma purity independently correlated with a better outcome in both RNAseq and microarray sets (Table 1). Among all cases, multivariate cox regression also confirmed that glioma purity along with age, Karnofsky performance score (KPS), radiation, chemotherapy, and 1p/19q status were independent prognostic indicators in glioma (Table 1). Afterward, these independent factors were integrated to develop a prognostic nomogram (Supplementary Fig. S12A). The calibration plot for the probability of survival at 1, 2, and 3 years showed an optimal agreement between the prediction by nomogram and actual observation (Supplementary Fig. S12B). We compared the predictive power for OS between nomograms and its constituting factors. The C-index for OS prediction was 0.736 (95% confidence interval, 0.656–0.816) by the nomogram, which was significantly higher than the C-indices of the constituting factors (Supplementary Fig. S12C). These results indicated that integrating glioma purity into a predictive model significantly improved the predictive validity.

Subsequently, the prognostic significance of known indicators was evaluated in gliomas with various purities. We sought to group all cases based on the median purity, but the sample size became limited owing to missing information and different proportion of cases among cohorts. Accordingly, we conducted the analyses in cases from the TCGA and CGGA databases, respectively. Cases were divided into low and high purity groups based on their median purity. Among traditional clinical and molecular factors, only KPS was verified as an independent prognostic factor for low purity gliomas in both TCGA and CGGA databases (Supplementary Table S5). In the high purity group, only histopathologic classification was identified with statistical significance. However, the improper HR, which was discrepant from the histopathology reality, undermined its validity (Supplementary Table S6). Accordingly, no factor could be deemed as an independent prognostic indicator for high purity gliomas in the present study. These results indicate the necessity of considering glioma purity for better prognostication.

To uncover the molecular mechanisms influencing tumor cell percentage within glioma, TCGA cases with available mutation and CNA information were analyzed for that purpose. We found a negative correlation between somatic mutations and glioma purity (R = −0.2434, P < 0.0001; Supplementary Fig. S13A). Based on the increasing purity, cases were respectively divided into two, three, or four subgroups. Parallel analyses were conducted in each kind of grouping to enhance the credibility of our findings. Here, we showed that the quartering condition was representative. First, we compared the frequency of mutations between cases with low and high purity. More somatic mutations were detected in cases with low purity (1st vs. 2nd half, 9,180 vs. 7,441 mutations; 1st vs. 3rd tertile, 6,532 vs. 4,790 mutations; 1st vs. 4th quarter, 4,872 vs. 3,518 mutations). Frequent mutations in IDH1, CIC, FUBP1, NOTCH1, as well as the 1p/19q codeletion were significantly enriched in cases with a higher purity (Fig. 4A; Supplementary Fig. S14A and S14B; Supplementary Table S7). On the other hand, mutations in PTEN, EGFR, and NF1 occurred more frequently in gliomas with a lower purity (Fig. 4A; Supplementary Fig. S14A and S14B; Supplementary Table S7). We also observed that a significantly different frequency of mutations in TTN, PCLO, MUC17, ZBTB20, NIPBL, FRAS1, and SPTA1 was attributed to various purity conditions, whereas existing studies have barely explored their roles in glioma.

Genomic mutations were further explored on the basis of KEGG pathways. Consistent with above findings, low purity cases carried more mutations in the 32 representative pathways (Supplementary Table S8). Calcium, FoxO, and cAMP pathways were the most significant differential pathways which had higher mutation rates among cases with low purity (Fig. 4B; Supplementary Fig. S14C and S14D; Supplementary Table S8). Meanwhile, a lower purity also indicated more mutations in genes involving focal adhesion, cell interaction, and recognized pathways in glioma, such as PI3K-AKT, Ras, and mTOR pathways (Fig. 4B; Supplementary Fig. S14C and S14D; Supplementary Table S8).

Subsequently, CNA data were investigated showing distinct chromosomal alteration patterns between gliomas with low and high purity. A negative correlation between CNAs and purity was found (R = −0.2939, P < 0.0001; Supplementary Fig. S13B). Consistently, more CNAs were detected in low purity gliomas (1st vs. 2nd half, 81,515 vs. 52,961 CNAs; 1st vs. 3rd tertile, 54,944 vs. 31,362 CNAs; 1st vs. 4th quarter, 38,258 vs. 18,875 CNAs). The incidence of the 1p/19q codeletion, which is a genomic hallmark in oligodendroglioma, increased along with the increasing glioma cell percentage (Fig. 4C). Whereas, Chr 7 amplification paired with Chr 10 loss, a representative event in GBM (19), were enriched in low purity cases (Fig. 4C). A total of 808 genes were identified with a different frequency of amplification attributing to various glioma purities. Meanwhile, another 44 genes were differentially deleted in different purity situations (Supplementary Dataset S1). All these differential alterations occurred more frequently in the lower purity subgroups. Among cases with low purity, the frequently deleted genomic regions were 9p21.3 encompassing the CDKN2A/CDKN2B locus (47.25% in the 1st quarter; P = 2.2E–17) as well as 10q23.3 encompassing the PTEN locus (19.51% in the 1st quarter; P = 0.0005; Fig. 4D; Supplementary Fig. S14E and S14F). On the other hand, the most commonly amplified region associated with low purity was 7p11.2 encompassing the EGFR (45.05% in the 1st quarter, P = 9.1E–24; Fig. 4D; Supplementary Fig. S14E and S14F). CDK4 (12q14.1; 12.64% in the 1st quarter; P = 0.0043) and PDGFRA (4q12; 13.19% in the 1st quarter; P = 0.0005) amplifications were also significantly frequent in low purity gliomas (Fig. 4D; Supplementary Fig. S14E and S14F).

Gene expression data were analyzed to explore the biological phenotypes associated with glioma purity. The TCGA-RNAseq cohort was selected as discovery set, and other cohorts were used for validation. We carried out PCA to explore the transcriptomic features associated with glioma purity. Here, we observed a tight association between whole transcriptome expression profile and glioma purity (Fig. 5A; Supplementary Fig. S15A–S15D), implying distinct general biological phenotypes attributing to various tumor cell percentage. Afterward, we sought to find out the biological features which are significantly enhanced in gliomas with low purity. We conducted spearman correlation analysis to identify genes associated with glioma purity. Genes whose expression were most negatively correlated with purity (R < −0.8) were submitted for GO analysis, showing that most of the top enriched biological implications were immune-relevant (Fig. 5B). Meanwhile, we divided cases into high and low purity groups based on the median purity. SAM was used to identify differential genes between cases with high and low purity. Genes that differentially upregulated in low purity cases were significantly enriched with terms related to immune response (Fig. 5C). Further, GSEA confirmed that cases with various glioma purity exhibited distinct immune status and low purity ones suffered a strengthened immune phenotype compared with high purity ones (Fig. 5D).

As lower purity confers intensive local immune status, we sought to investigate which type of immune cell forms the nontumor fraction and contributes to the enhanced immune response. We employed GSVA for estimating signatures of multiple immune cells and correlated the signatures with glioma purity (Fig. 5E; Supplementary Fig. S15E–S15H). According to the association with glioma purity, immune cells could be clustered into two subgroups. A threshold of absolute correlation coefficient more than 0.3 was used for filtration across cohorts (Supplementary Table S9). Macrophages, neutrophils, and microglia exhibited consistent negative correlation (R < −0.3) with glioma purity. Whereas, central memory T cells, follicular helper T cells, CD8 T cells, natural killer (NK) cells, eosinophils, and T helper 1 cells were consistently positively correlated (R > 0.3) with glioma purity. A mutual relationship clustered these immune cells into two subgroups (Fig. 5F; Supplementary Fig. S15I–S15L), which were in accordance with their correlation to glioma purity. We further collected 43 tissues from the CGGA cohort. IHC analysis confirmed that numbers of macrophage/microglia and neutrophil were enriched in low purity cases (Fig. 5G and H) and diversely correlated with glioma purity (Supplementary Fig. S16). Therefore, macrophages, microglia, and neutrophils could be a cluster of immune cells that contributed to the nontumor fraction within low purity gliomas. Considering macrophages, neutrophils, and microglia were highly enriched in gliomas with fewer tumor cells, we attempted to investigate their prognostic implications. Survival analyses showed macrophages and neutrophils, but not microglia, were significantly correlated with glioma prognosis (Fig. 5I–K). Similar analyses in other cohorts confirmed that enrichment of macrophages and neutrophils indicated reduced survival time (Supplementary Fig. S17).

Tumor tissues have a diverse mixture of tumor and nontumor cells within their microenvironments. Tumor purity is routinely determined by pathologists through visual evaluation, which could be affected by the sensitivity of histopathology, interobserver bias, and variability in accuracy (33). Other than routine pathology-based estimation, genomic advances introduce many computational methods to determine tumor purity based on multiform genomic data, such as somatic mutation (34), somatic CNA (29), and DNA methylation (35), which elicit objective and highly concordant results (28). Owing to the compatibility of ESTIMATE algorithm in both microarray- and RNAseq-based transcriptome profiles, we reviewed 2,249 glioma cases and systematically investigated the role of glioma purity in present study. Glioma purity was highly associated with major clinical and genomic features, strengthening that glioma purity is an intrinsic characteristic of the tumor cell in developing a suitable microenvironment (28).

We found cases with low purity glioma cells exhibited few favorable implications but were more likely to be ascribed in malignant entities and have reduced survival time. On the one side, glioma cells with limited proliferative and invasive properties tend to grow slowly and form a solid bulk with less nontumor cell infiltration. On the other side, tumor cells have been reported to be capable of dominating a microenvironment (11), giving rise to the hypothesis that malignant gliomas recruit abundant surrounding cells and subjugate them to compose a protective shield. Therefore, low tumor purity and correlated cellular heterogeneity are responsible for glioma's aggressive phenotype and poor prognosis. Purity of glioma cells may add a new dimension in estimating malignant identities and explain why most therapeutic strategies aimed purely against glioma cells do not have an ideal outcome.

Glioma purity was identified as a robust indicator for prognosis, whereas most prognostic indicators were no longer significantly effective even among gliomas with high purity. Tight associations between glioma purity and genomic indicators implied that the prognostic and predictive value of genomic indicators may be partially derived from their roles in monitoring a balance between tumor and nontumor cells (36). These results highlight the pivotal role of nontumor cells in determining prognosis and question the use of traditional indicators for clinical prediction, especially in individual cases. Here, we have developed a prognostic nomogram comprehensively taking glioma purity and other robust indicators into consideration, which showed superior validity in predicting survival time. Therefore, integrating glioma purity into a prognostic system would be helpful for precise prediction.

Stromal and immune cells constitute the major nontumor fraction within a glioma microenvironment. Moreover, there was a tight association between stromal and immune scores. Distinct local immune status was the primary differential phenotype that resulted from various glioma purities, which was significantly enhanced in cases with low purity. This verifies our previous report, which found that an intensive local immune response is unfavorable for glioma prognosis (37). Importantly, not all immune cells were enriched in tissues with low glioma purity, summarizing two subgroups of immune cells based on their relationship with glioma purity. This highlights the inability of antitumor immune cells in infiltrating into the established protective shield around the glioma cell. These antitumor cells could be represented by CD8 T cells and NK cells, whose enrichment could significantly prolong patients' survival (38, 39).

We should also note the ability of glioma cells in selectively recruiting immune cells to establish a proper microenvironment. We found that macrophages, microglia, and neutrophils were enriched in low purity cases, which could be verified using histologic methods. These cells were of myeloid lineage and associated with cellular immunosuppression in glioma (40). In particular, glioma-associated macrophages and microglia have been reported to form as many as 30% to 50% of the cells in gliomas (5). Different polarization statuses of macrophages and microglia have been defined based on stimulation in vitro, including the classical M1 phenotype and alternative M2 phenotype. Their enrichment generally contributes to the glioma growth, invasion, and increasing histopathologic grade (41). Previous studies deemed macrophage and microglia as unfavorable indicators for prognosis (41–43). Whereas, only macrophage was identified with prognostic significance in the present study. Unlike previous results which were mainly derived from analysis on a single marker, we evaluated the enrichment of macrophage and microglia by conducting a GSVA algorithm that was based on the expression profile of a characteristic gene set and generated a signature from a more comprehensive perspective. Although there are abundant studies that focus on the role of macrophages and microglia, further analysis is still needed to confirm our results and explore the different roles of polarization in glioma.

However, low purity glioma contains abundant neutrophil infiltration, which is positively associated with grade progression, shorter survival, and treatment resistance (44–46). A previous study reported complex interaction among neutrophils, macrophages, and microglia. This interaction is critical in maintaining glioma's local immunosuppressive phenotype (47). Neutrophils also inhibit the cytolytic activity of NK cells and CD8 T cells (45). Therefore, blocking neutrophil infiltration could be suggested for combined treatment strategy. Considering the tight association between the neutrophil-to-lymphocyte ratio in peripheral blood and local neutrophil infiltration, the neutrophil-to-lymphocyte ratio may be an invasive indicator for evaluating the treatment response (45).

Specific genomic alterations have been established as early event initiating gliomagenesis (32). Here, we observed distinct genomic patterns attributing to various glioma purities. Despite there being doubt that low purity may reduce the sensitivity of somatic mutation and CNA testing (48), we still detected negative correlation between glioma purity and events of genomic alteration, indicating an instable genomic status in low purity cases. Therefore, genomic alterations and heterogeneity may have substantial roles in editing the glioma microenvironment. The low purity induced an intensive immune phenotype that could further aggravate the genomic instability (49), thus creating a positive feedback that exacerbates the poor prognosis and treatment resistance (50).

The main advantage of our research was the use of large sample-sized glioma cohorts and systematic analysis in multidimensional conditions. Our findings highlight the important role of tumor purity and local immune cells in glioma biology and clinical management. However, we still need to clarify the extent and type of immune responses that can promote glioma progression. It would be interesting to depict cellular components within a microenvironment, interpret their clinical implications, and explore the immunoregulatory role of glioma cells.

Taken together, glioma purity exerted a considerable effect on clinical, genomic, and biological conditions. Assessing the proportion of different cell types may help in elucidating the complex role of glioma microenvironment and provide new insights into its clinical management. An ideal classification system should focus not only on the properties of tumor cells but also on nontumor cells.

No potential conflicts of interest were disclosed.

Conception and design: W. Cheng, G. Li, S. Han, T. Jiang, A. Wu

Development of methodology: C. Zhang, W. Cheng, A. Wu

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): C. Zhang, W. Cheng, X. Ren, X. Liu, G. Li, T. Jiang, A. Wu

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): C. Zhang, W. Cheng, X. Ren, Z. Wang, G. Li, S. Han, A. Wu

Writing, review, and/or revision of the manuscript: C. Zhang, W. Cheng, G. Li, T. Jiang, A. Wu

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): C. Zhang, W. Cheng, X. Ren, G. Li, T. Jiang

Study supervision: T. Jiang, A. Wu

The authors conducting this work represent the Chinese Glioma Cooperative Group (CGCG). We thank Lin Fu and Qingchang Li at the Department of Pathology, the First Affiliated Hospital of China Medical University, for assistance in morphology analysis.

This work was supported by grants from the National Natural Science Foundation of China [Grant Numbers: 81172409 (A. Wu), 81472360 (A. Wu), and 81402045 (S. Han)], the Science and Technology Department of Liaoning Province [Grant Number: 2011225034 (A. Wu)], the National Key Research and Development Plan [Grant Number: 2016YFC0902500 (T. Jiang)], and the Capital Medical Development Research Fund [Grant Number: 2016-1-1072 (T. Jiang)].

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.