This study included 44 patients who underwent surgery for primary and recurrent glioma by MRI imaging at the same hospital from 2005 to 2014 (Table S2 in Supplementary Material). The average period between diagnosis and surgeries was 5.4 days (range 2–10 days). The inclusion criteria included the following: greater than 18 years of age at the time of primary diagnosis, a histopathological diagnosis of diffuse astrocytomas (DA) or GBM, and the availability of formalin-fixed, paraffin-embedded (FFPE) tumor tissue blocks. The patients were separated into three groups: (1) the DA–DA (DA, diffuse astrocytoma) group (DAs that recur as DAs, 15 pairs); (2) the GBM–GBM group (GBMs that recur as GBMs, 15 pairs); and (3) the DA–GBM group (DAs that recur as GBMs, 14 pairs). None of the 44 patients received treatments before their primary surgery. Subsequently, patients received standard radiotherapy (60 Gy), while some received concomitant–adjuvant chemotherapy (temozolomide/other). The recurrence-free survival (RFS) was defined as the time from the date of the primary surgery to the first MRI-confirmed tumor recurrence. All recurrent patients underwent recurrent surgery. OS was defined as the interval between primary surgery and death or last follow-up. Survival after recurrence (SR) was defined from the recurrent surgery to the date of death or the last follow-up. The mean follow-up period was 1,355.1 days (range 289–3,520 days), during which 36 patients died, and 8 patients remained alive. No patients were lost to follow-up. Tumor size was calculated based on enhanced MR or CT images (on the layer with the maximum amount of tumor) as follows: long diameter (centimeter) × wide diameter (centimeter) × thickness (centimeter) × 0.5 (25). All specimens were selected after histopathological review to ensure that we would have sufficient tissue for analysis. Treatment information was obtained from the time of primary surgery, including any adjuvant treatment. The latest updated OS information was obtained on February 13, 2016. For patients who were still alive, the date of the last follow-up was used in the data analysis. All subjects gave written informed consent in accordance with the Declaration of Helsinki and research protocols were reviewed and approved by the Committee on Human Research at the Harbin Medical University, China.

To generate high-quality RNA from paired FFPE samples, a Sensation Plus™ FFPE Amplification and WT Labeling Kit was used to extract and amplify total RNA that was derived from whole tumor FFPE samples. The total RNA obtained from each sample was quantified using a NanoDrop ND-1000, and we selected three of the highest-quality paired samples from each group. The samples were used for labeling and submitted for hybridization array scanning using an Affymetrix Gene Chip^® Human Transcriptome Array 2.0. The microarray was processed, and the data were analyzed by Beijing Biolancet Technology, Co., Ltd. Nine pairs of samples were processed by Affymetrix HTA2.0 kit. After sequencing and QC, raw microarray data (CEL) were analyzed by Affymetrix Expression Console 1.4 software. HTA-2_0 library and annotation files were downloaded from Expression Console’s build-in database. Raw data were further normalized by the RMA-algorithm (Table S3 in Supplementary Material). Differentially expressed genes between samples were identified using fold-change filtering to identify distinct gene expression profiles between samples. For gene enrichment analysis, paired recurrent and primary gliomas samples’ microarray data were normalized by the log2 algorithm of genes mean value. Then, Gene Set Enrichment Analysis (GSEA) (26, 27) was applied to analyze the normalized data by groups. Enrichment Score, Normalized Enrichment Score, False Discovery Rate, and Nominal p-Value were reported by GSEA. Interpreting GSEA Results of the above four key statistics can be found in http://software.broadinstitute.org/gsea/doc/GSEAUserGuideFrame.html.

A systematic neuro-pathological review was performed based on the 2007 World Health Organization classification guidelines for CNS tumors (28). Paired samples of slides were reviewed to determine their histological classifications. Immunohistochemical staining was performed for CD3 (clone LN10, dilution 1:200, Quanhui, China), CD4 (clone UMAB64, dilution 1:50; ZSGB-BIO, China), CD8 (clone SP16, dilution 1:100; ZSGB-BIO, China), Foxp3 (clone mAbcam 450, dilution 1:50; Abcam), and CD34 (clone EP88, dilution 1:150; ZSGB-BIO, China). The secondary antibodies came as an immunohistochemical kit (KIT-5930, Maxim, China) and were incubated for 40 min at room temperature. Each staining batch included a negative control that was processed without the primary antibody, a biological negative control that consisted of normal brain tissues (the donor had died of myocardial infarction), and a positive control (tonsil tissues).

The IHC analyses were performed using a quantitative approach under a light microscope (Leica SP2, Leica Optical Co. Ltd., Germany). Subpopulations of TILs expressing CD3, CD4, Foxp3, or CD8 were divided into intratumoral fractions, which contained no vessels in each high powered field (HPF) (40× objective and 10× eyepiece), and perivascular fraction, which contained more than one vessel in each HPF (29, 30). The mean counts of CD34+ vascular circles were obtained from 10 consecutive HPF by three pathologists blinded to outcome data. Similarly, for TIL frequencies, the results were determined from a mean of 10 consecutive HPFs of typical tumor regions. The average scores counted by three experienced pathologists blinded to the clinical background were recorded as the final result. When there was a large difference in scores between observers, the score was re-evaluated to reach an agreement. Results from CD3+, CD4+, FoxP3+, CD8+ cell, and CD34+ vascular circle counts in each patient were used in the statistical analysis.

The following antibodies were used for two-color immunofluorescence: rabbit anti-CD34 and mouse anti-CD3 monoclonal antibodies (as described above). The primary antibodies were incubated overnight at 4°C. The sections were then incubated with goat anti-mouse fluorescein-isothiocyanate-conjugated secondary antibodies and goat anti-rabbit rhodamine-conjugated secondary antibodies for 2 h at room temperature. The nuclei were stained with DAPI. Finally, the sections were visualized using a confocal microscope (as described above).

The raw data were tabulated using Microsoft Excel (Microsoft Corp.) and analyzed using nonparametric tests and matched-paired-tests (Wilcoxon tests or Mann–Whitney U tests). Chi-square tests were used to identify differences in chemotherapy and sex. Spearman’s rank order correlation analyses were performed to detect significant associations with positive marker expression. A Kaplan–Meier survival analysis was used to evaluate differences in RFS, SR, and OS. To adjust for potential confounders, Cox proportional hazards models were used to evaluate hazard ratios (HRs) for recurrence or death according to the number of identified TIL subpopulations and clinical features. All statistical analyses were performed using SPSS 13.0 (SPSS Inc. Chicago, IL, USA) and GraphPad Prism 6.0 (GraphPad Software Inc., La Jolla, CA, USA). All tests used to determine the level of significance were two-sided. A (two-tailed) p-value threshold of 0.05 was considered to indicate statistical significance (*p < 0.05, **p < 0.01, ***p < 0.001).

To learn more about overall genetic changes before and after recurrence across the three recurrent groups (DA–DA, DA–GBM, and GBM-GBM), three pairs of FFPE glioma samples from each group were analyzed by Microarray. A distinct pattern of gene expression was observed before and after each recurrence (Figure S1A in Supplementary Material). Among these altered genes upon recurrence, VEGF-A was found increased by 10.2-fold after tumor recurrence in the DA–GBM recurrent group (Table S1 in Supplementary Material). Distinct clinical outcomes were also revealed in the three groups; the longest RFS was the DA to GBM recurrence (Table S2 and Figure S1B in Supplementary Material). Based on these results, blood vessels were measured using a vascular endothelial cell marker, CD34. Tumor-infiltrating T cells, such as CD3+, CD4+, and CD8+ T cells were determined by immunohistochemistry (Figure 1A; Figure S2A–C in Supplementary Material). Compared to other two groups, the primary tumors in the GBM-GBM group showed relatively higher on the average density of CD34+ circles and T cells, while the DA–GBM group displayed a wider-range of T cell infiltrations (Figures 1B–E). Similar kinetics of CD34+ circles and infiltrating T cells were observed when these parameters were compared before and after tumor recurrence/progression; compared to recurrent GBM, significantly increased CD34+ circles and T cell subsets were found in secondary GBM (Figures 1F–I).

To test which infiltrating T cell subset associated most closely with tumor angiogenesis, we next did a Spearman’s rank correlation analysis between the density of tumor vessels and T cell infiltrating subsets. CD4+ T cells correlated strongly with CD34+ circles in both primary and recurrent tumors across all three groups; notwithstanding primary tumors of DA–GBM, there was no association with CD8+ T cells (Figures 2A–C).

We found that CD4 transcript was predominantly expressed in primary GBM (>29-fold higher than CD8 alpha and beta chains), based on the RNA-seq analysis solely culled from TCGA database (TCGA Research Network: http://cancergenome.nih.gov/, Figure 3A). No similar results were observed for some few other cancer types, such as kidney renal clear cell, testicular germ cell tumor using TCGA Pan-Cancer datasets (30 different cancer types) (Figure 3B). To confirm our observation that infiltrating CD4+ T cells may be associated with tumor angiogenesis, microarray data of paired bevacizumab resistant tumors from previously published reports were analyzed and paired chemo-naïve tumors were used as a comparison control (31, 32). Only signals of activated CD4+ T cells, not other cell populations, were significantly increased in tumors after a recurrence of bevacizumab-treated patients (Figure 3C); no such trend found in the chemo-naïve recurrent patient tumors (Figure 3D).

T cell tumor trafficking and tumor angiogenesis have been shown to be linked biological phenomena (3). We found both factors had an upwards trend upon recurrence; we next separated tumor-infiltrating T cells into perivascular and intratumoral CD3+, CD4+, CD8+, and Foxp3+ T cell subsets (Figure S2A–C in Supplementary Material). We demonstrated that there were more perivascular than intratumoral TILs in primary and recurrent tumors. In primary tumors, only one-third of the TILs were located in the intratumoral space, and two-thirds of the TILs were in the perivascular niche; these findings reproduce previously published reports (30). When comparing the cell numbers before and after recurrence, the perivascular T cells were enriched after tumor progression across all three groups; the intratumoral T cells, however, were only increased in recurrent tumors of the DA–GBM group (Figures 4A–F).

The association between tumor-infiltrating CD4+ T cells and tumor angiogenesis led us to further test the association between angiogenesis and a subset of immune inhibitory CD4+ T cells or regulatory T cells (Tregs). We found that perivascular Foxp3+ (Figures 5A–B upper panel), and CD4+ T cells (Figures 5A–B, lower panel) in the primary tumors correlated with tumor blood vessels. No similar trend was found for the CD8+ T cells in both geographic locations (data not shown). We sought to confirm the observation by using gene enrichment for Treg signature genes using paired samples. The results indicate that tumor recurrence significantly increases CD4 (similar to the results from bevacizumab resistant tumors) and Tregs signals in these tumors, whereas other cell subsets, i.e., NK, Th17, show contrasting results (Figure 5C and data not shown). To evaluate the correlation between differential geographic T cell subsets and clinical outcomes (RFS, OS), Kaplan–Meier survival curves were plotted using these tested parameters against RFS and OS. We found that only perivascular CD4+ T cells and Tregs were associated with shorter RFS (p = 0.007 for CD34, p = 0.01 for CD4+ T cells and p = 0.001 for Foxp3+ T cells). No association was observed for these factors with respect to OS (Figures 5D–E). Additionally, there was no association found for perivascular or intratumoral CD8+ T cells with respect to RFS and OS (Figure S3 in Supplementary Material). Next, to assess if key T cell subsets independently predict glioma progression, Cox multivariate regression analyses were performed (by including factors, such as age, sex, chemo/radiotherapy, CD34+ circles and perivascular/intratumoral infiltrating T cell subsets). The results concluded that only perivascular Foxp3+ T cells were found to be an independent predictor of shortened RFS (HR = 4.199, 95% CI 1.522–11.584, p = 0.006) when all the patients were included.