Archival FFPE (formalin fixed paraffin embedded) tumor tissue sections were obtained from patients diagnosed with primary gliomas(n=505) when undergone tumor resection between 2001 and 2017 at the Department of Neurosurgery, Huashan Hospital, Fudan University. Three FFPE tissue microarrays (TMA) sections including 276 samples were prepared using 1.5 mm tissue cores, and the other 229 specimens were overlapped as described previously.4 The tumor area for each core in TMA was confirmed by the neuropathologist using HE staining. Detailed clinicopathological characteristics are available in online supplementary table S1 ( online supplementary additional file 2: table S1 ).

FFPE glioma tumor tissue sections were deparaffinized in xylene and rehydrated through a graded series of ethanol concentrations; the sections were blocked with 3% H₂O₂ for 20 min at room temperature. Sections for B7-H4 and CD3 were retrieved by Tris-Ethylene Diamine Tetraacetie Acid (EDTA, pH=9.0) and sodium citrate buffer (pH=6.0) respectively, and sections for PD-L1 and CD163 were retrieved by lycium hydrochloric acid (pH=6.0). B7-H4 antibody (Ab), PD-L1 Ab and CD3 Ab were diluted at the same dilution of 1:100, while CD163 Ab was diluted at 1:50. After incubated overnight at 4℃, primary Abs were washed out and followed by VECTASTAIN ABC kit (PK-6200) for 30 min. DAB substrate kit (SK4100) was used to develop DAB chromogen and Hematoxylin staining solution was used to stain nuclei. The PD-L1, CD3 and CD163 Abs have been validated in gliomas by previous studies (online supplementary additional file 2: table S2). For B7-H4 Ab, GL261 cells transfected with human B7-H4 gene were prepared as FFPE cell blocks (online supplementary additional file 4: supplementary materials and methods, additional file 1: figure S1); IHC was performed with or without primary Ab as positive and negative controls, respectively. FFPE cell blocks derived from GL261 cells transfected with an empty vector were used for negative controls. B7-H4-positive GL261 cells showed brownish stain while cells in negative controls were stained with hematoxylin showing blue nuclei (online supplementary additional file 1: figure S1). Moreover, the staining in normal control brain tissues (non-tumor tissues) with primary Abs (online supplementary additional file 1: figure S1) and in the tumor tissue sections without primary Abs were conducted as negative controls for each targeted protein.

B7-H4 and PD-L1 specific IHC staining was semiquantitatively scored based on the intensity and percentage as previously described.4 11 Briefly, the cells with partial or complete IHC staining at any intensity are considered as positive cells. The positive reaction of B7-H4/PD-L1 was scored into four grades according to the intensity of the staining: 0, 1, 2 and 3. The percentages of B7-H4/PD-L1–positive cells were also scored into four categories: 0 (0%), 1 (1%–33%), 2 (34%–66%) and 3 (67%–100%). The sum of the intensity and percentage scores was used as the final IHC score from 0 to 6. The staining pattern of the tissues was defined as follows according to the final IHC score: low expression, 0–2; high expression, 3–6. Extent of TIL presence in TMA slide cores was assessed by semiquantitative measurement of the density of CD3 positive (CD3⁺) cells, and was scored as previously published9 17: 1, no (less than four TILs in the entire specimen) or sporadic (more than four but no accumulation) CD3⁺ cells; 2, moderate numbers of CD3⁺ cells (single areas with accumulation of TILs); 3, abundant occurrence of CD3⁺ cells (TILs throughout the tumor section); and 4, highly abundant occurrence of CD3⁺ cells (high frequency of TILs throughout the entire tumor section). The IHC of CD163 was assessed as the percentage of positive cells related to the total cell number.

The IHC staining was analyzed by two observers blinded to clinical characteristics. A consensus for the final scores of PD-L1, B7-H4 and CD3 were reached for each sample, and the mean number of positive proportions from two observers was calculated for CD163. Representative photographs of each category for staining intensity and percentage of PD-L1/B7-H4 were shown in online supplementary figure S2 ( online supplementary additional file 1: figure S2 ). Fleiss’s kappa for the categorical values (high expression and low expression for PD-L1/B7-H4) and ranked numerical values (1, 2, 3 and 4 for CD3⁺) were calculated for the interobserver agreement between two observers. The correlation analysis based on Spearman’s testing for continuous numerical values (positive proportion for CD163⁺) was performed to eliminate a potential interobserver variation of manual quantification. Concordant agreements were observed for the interobserver assessment in each protein molecule assessed ( online supplementary additional file 2: table S3 and S4 ).

Two published glioma transcriptome datasets were downloaded from The Cancer Genome Atlas (TCGA) and Chinese Glioma Genome Atlas (CGGA).18 19 RNA sequencing (RNA-seq) expression data were further normalized using upper quartile FPKM (FPKM-UQ) as described in TCGA website. The RNA-seq data were log2 transformed before analysis between several groups. Patient demographics were obtained from each data portal. Differentially expressed genes (DEGs) between different groups were identified using DESeq2 package20 with the criteria of |fold change|>2 and the false discovery rate <0.05).

In order to identify PD-L1/B7-H4 classifier genes (genes that are coexpressed with PD-L1 or B7-H4) in the training dataset (TCGA), we used biweight midcorrelation (bicor) to evaluate ‘similarity’ between gene expression profiles, which is thought to be a good measurement for gene coexpression module analysis.21 First, bicor for each pair of genes was calculated, and the 95th percentile of the bicor values was defined as the cut-off, which was 0.37027. Genes whose patterns are highly correlated to PD-L1 or B7-H4 were further selected if positively correlated with each other (the bicor values are larger than the cut-off). Finally, 62 PD-L1 classifier genes and 35 B7-H4 classifier genes were determined ( online supplementary additional file 2: table S5 ).

We performed non-negative matrix factorization (NMF) algorithm with the R package NMF22 to define glioma clusters.23 The following steps were used to determine the optimal factorization rank (the number of clusters) in each data set: (1) Random expression data were generated by permutating columns (patients) in the original dataset; (2) We performed 100 runs of NMF and select the best fitted model for both the original and the corresponding random data sets with rank from 2 to 7, and obtained the residual sum of squares (RSS) for each rank; (3) The smallest rank was selected if RSS decrease is less than that in random data as described previously.24 With the optimal rank k, the partitioning around medoids algorithm was used to perform clustering on the coefficient matrix of each run of the corresponding NMF model where the number of clusters is k. The consensus matrix of clustering was then generated and the final clusters of samples were determined by cutting the hierarchical clustering of the consensus matrix into k groups.

The R package ESTIMATE (Estimation of STromal and Immune cells in MAlignant Tumour tissues using Expression data) was applied to predict tumor purity based on the estimation of stromal and immune scores in tumor tissues using gene expression data.25 Samples from TCGA having genomic sequencing data were labeled with ABSOLUTE purity instead of ESTIMATE purity and were extracted from previous paper.26 The gene set variation analysis (GSVA) package of R was used to calculate infiltrating scores for 22 immune cell subtypes and microglia.27 In order to rule out possible bias from different algorithms, we calculated absolute immune cell fractions for 22 immune cell populations with CIBERSORT (Cell-type Identification By Estimating Relative Subsets Of RNA Transcripts),28 followed by being corrected using tumor purity per sample.29 Gene ontology (GO) term enrichment analysis was carried out with R language using clusterProfiler packages.30 Two gene sets extracted from the Molecular Signatures Database (immune system process, M13664; immune response M19817) were used to profile immune phenotype between modules using principal components analysis (PCA).31 Ontology gene sets were obtained from MSigDB32 and the results of gene set enrichment analysis (GSEA) were visualized by cluster Profiler package.30

Spearman’s rank correlation coefficient was used for correlation analysis. χ² test or Fisher’s exact test was used for categorical variables and differences in continuous variables between groups were evaluated by t-test or one-way analysis of variance. All tests were two-sided performed by SPSS software or R language and p<0.05 was considered statistically significant labeled with * for p<0.05, ** for p<0.01, *** for p<0.001, and **** for p<0.0001.

Although the majority of newly diagnosed gliomas patients were PD-L1-positive (61%, IHC score >0) in the IHC cohort (n=505), only 23.4% of them were defined as high expression of PD-L1 (IHC score >2). For GBM patients in this cohort, about 73.5% samples were PD-L1-positive while 29.1% were determined as high expression. For lower grade gliomas (LGG, grades II and III), we found that 49.4% of patients expressed PD-L1 while 18.5% expressed it at high levels. The relationship between PD-L1 expression and clinicopathological variables were investigated as shown in online supplementary table S6 ( online supplementary additional file 2: table S6 ). Tumor location, histology, WHO grade and IDH gene status had a significant correlation with PD-L1 expression. The percentages of B7-H4-positive and B7-H4 high-expression samples were 54.1% and 19.6% in the IHC cohort, 53.8% and 14.1% in GBM patients, and 54.2% and 24.4% in LGG patients, respectively. The B7-H4 expression was significantly associated with gender, histology, WHO grade, TERT gene status and chromosome 1 p/19q (Chr1p/19q) status ( online supplementary additional file 2: table S6 ).

In the cohort assessed for IHC (n=505), 61% of glioma patients had low expression of both PD-L1 and B7-H4 (double-low subgroup), 20% had high expression of PD-L1 but low expression of B7-H4 (high-PD-L1 subgroup) and 17% had high expression of B7-H4 but low expression of PD-L1 (high-B7-H4 subgroup). Interestingly, only 2% of the patients had high expression of both PD-L1 and B7-H4 (double-high subgroup, figure 1A). Significant associations were detected in histology, WHO grade, IDH status, TERT gene status and Chr1p/19q status among PD-L1/B7 H4 subgroups ( online supplementary additional file 1: figure S3 and additional file 2: table S7 ). Representatives of these subgroups with mutually exclusive expression of PD-L1 and B7-H4 are shown in figure 1B. Moreover, H-scores of PD-L1 and B7-H4 in TMA cores ( online supplementary additional file 4: online supplementary materials and methods ) also showed limited coexpression of both biomarkers ( online supplementary additional file 1: figure S4 ).

In order to avoid possible bias by tumor tissue sampling in the TMA cores, we analyzed the mRNA expression levels of PD-L1and B7-H4 based on two published RNA-seq datasets obtained from TCGA and CGGA. As shown in figure 2, a similarly restricted coexpression pattern was observed in TCGA cohort ( figure 2A, n=666) and CGGA cohort ( figure 2B, n=305) with significant negative correlation between PD-L1 and B7-H4. Moreover, we found most other checkpoint immune molecules that have been verified as therapeutic targets in clinical or preclinical trials were tightly associated with PD-L1 while demonstrating mutually exclusive expression with B7-H4 (figure 2).

Identification of gene coexpression modules around key signaling pathways may enable molecular classification of gliomas. We, therefore, decided to generate gene coexpression modules by classifier genes that are strongly associated with PD-L1 and B7-H4 using transcriptomics datasets. The classifier genes were extracted from TCGA training set (n=666), identifying genes with similar expression patterns with PD-L1 (62 genes) and B7-H4 (35 genes), analyzes of which could help decipher key signaling pathways related to these molecules ( online supplementary additional file 2: table S5 ). GO analysis for classifier genes revealed that they were significantly enriched in many biological processes related to immune responses and intercellular communication ( online supplementary additional file 2: table S8 ).

According to the expression profiles of predicted PD-L1/B7-H4 classifier genes, each grade of gliomas from TCGA could be classified into three new subtypes: high-PD-L1 (high expression of PD-L1 but low expression of B7-H4 classifier genes), high-B7-H4 (high expression of B7-H4 but low expression of PD-L1 classifier genes) and double-low modules (low expression of both PD-L1 and B7-H4 classifier genes) ( figure 3A and online supplementary additional file 1: figures S5a and S6a ). Here, we confirmed that high-PD-L1 module gliomas highly expressed PD-L1 with lower levels of B7-H4 while High-B7-H4 module gliomas highly expressed B7-H4 with lower levels of PD-L1 ( online supplementary additional file 1: figures S5b, S6b and S7a ). Strikingly, we found there was no double-High module (high PD-L1 and high B7-H4 classifier genes expression) which supported our previous finding of a restricted coexpression pattern between PD-L1 and B7-H4. We then used the transcriptome dataset from CGGA to validate such classification criteria ( figure 3B and online supplementary additional file 1: figures S7-S9). The number of possible glioma subtypes was determined based on appropriate rank number with stable cophenetic correlation coefficients, which is consistent with the above observations (online supplementary additional file 1: figures S10-S12). In addition, most of the clinicopathological parameters exhibited statistically significant differences between different subtypes for each grade, such as age, histology, TCGA subtype and molecular classifications ( online supplementary additional file 2: tables S9-S14 ). Further investigation of the relationship between new subtypes and known TCGA subtypes revealed that almost all of the mesenchymal subtype belonged to the high-PD-L1 subtype, whereas the high-B7-H4 subtype mainly were contained in the proneural subtype.

Considering that many PD-L1/B7-H4 classifier genes are related to tumor–immune system interactions, it is possible that modules defined by PD-L1/B7-H4 signatures may be distinct in terms of the tumor immune microenvironment. Indeed, the immune and stromal score calculated by ESTIMATE in TCGA set revealed that gliomas in the High-PD-L1 module had more immune cell infiltrations whereas a significantly smaller immune infiltrating score was obtained for High-B7-H4 module ( figure 4A and online supplementary additional file 1: figures S13a and S14a ). We also found that high-B7-H4 module had higher tumor purity has been shown to contribute to poor immune cell infiltration ( figure 4B and online supplementary additional file 1: figures S13b and S14b ). The GSVA score represented the extent of 23 immune cells infiltrations suggested smaller fractions of TILs and tumor-associated macrophages (TAMs) in the high-B7-H4 module ( figure 4C and online supplementary additional file 1: figures S13c and S14c ). Further evaluation of absolute fractions of different immune cell types verified that high-PD-L1 module tumors had more TILs and TAMs than tumors of the high-B7-H4 module ( online supplementary additional file 1: figures S15a, S16a and S17a ). The data from CGGA validation dataset also showed that high-B7-H4 module gliomas were deficient of TILs in the TME ( online supplementary additional file 1: figures S15b, S16b, S17b and S18-S20 ). Correlation analysis revealed that B7-H4 expression level negatively correlates with TILs and TAMs presence ( online supplementary additional file 1: figures S21-S25 ). We performed IHC staining of CD3 and CD163 in TMA cores to determine the presence of TILs and TAMs in PD-L1/B7-H4 subgroups, respectively, and found that gliomas in the high-B7-H4 subgroup gliomas appeared as ‘super-cold’ tumors with less TILs and TAMs compared with high-PD-L1 subgroup ( figure 5 and online supplementary additional file 1: figure S26 ).

Considering such distinct levels of immune cell infiltration among PD-L1/B7-H4 modules, we explored the differences between High-PD-L1 module and High-B7-H4 module gliomas using immune-related gene sets and whole genome expression data. The PCA analysis showed remarkable differences between High-PD-L1 and High-B7-H4 modules in immune phenotype and whole genome expression signature for each grade gliomas ( online supplementary additional file 1: figures S27-S29 ). Next, we detailed these findings with GSEA analysis between high-PD-L1 and high-B7-H4 modules. Here, we found that many immune response pathways were enriched in high-PD-L1 module gliomas, such as T-cell receptor signaling pathway, regulation of antigen processing and presentation, macrophage differentiation status and toll-like receptor signaling pathway ( online supplementary additional file 1: figure S30 ). DEGs were obtained within each glioma grade from two datasets, and DEGs shared between high-PD-L1 and high-B7-H4 modules were identified ( online supplementary additional file 1: figure S31 and additional file 3: table S15 ).