Introduction

Glioblastoma (GBM) represents the most frequent and devastating brain malignancy. Despite standard therapeutic modalities, the outcome of patients remains dismal [1,2,3]. Isocitrate dehydrogenase (IDH) gene mutation plays a fundamental role in the carbohydrate metabolism, the tumor microenvironment (TME), and was involved in compromising the anti-tumor immune response [4, 5]. Approximately 90% of GBM is IDH wild-type, which represents the most lethal subtype of glioma [5, 6]. As there lacks effective treatment modalities, it is in urgent need to accurately predict the prognosis of IDH wild-type GBM for individual tailored therapeutic regimens and improvement of clinical management.

The outcome of IDH wild-type GBM can be heterogeneous. IDH wild-type GBM is classified as proneural, classical, and mesenchymal subtypes based on the transcriptome, with the latter having the worst prognosis [4, 7]. Recently, four tumor cell states have been identified using the advanced single-cell RNA sequencing, which was implicated in treatment resistance and prognosis [8]. In addition, O-6-methylguanine-DNA methyltransferase (MGMT) promoter methylation, telomerase reverse transcriptase (TERT) promoter mutations, and N6-methyladenosine-mediated RNA modification were found to be associated with patients’ outcome [9,10,11,12,13]. Nevertheless, the clinical application of these novel biomarkers for prognostic prediction remains limited. Recently, several studies have endeavored to develop prognostic models for IDH wild-type GBM based on multi-omics data [14,15,16,17,18,19,20]. Notably, gene–gene (G × G) interactions have profound biological implications. For instance, the prognostic value of HIF1A in non-small cell lung cancer is altered with EGLN2 expression [21], and gene signature associated with T cell dysfunction can be screened by assessing G × G interactions [22]. In addition, G × G interactions can serve as the basis for constructing prognostic models [23, 24].

Compelling evidence illuminates that stem-like tumor cell, which is at the interface of neural and glioma biology, is essential in tumor progression and treatment resistance [25,26,27]. One step further, neural progenitor cell-like (NPC-like) tumor cells characterized by CDK4 has been defined in GBM [8]. These cells retain the potential to differentiate into other cells and have an enhanced invasive capacity upon neuronal activity induced calcium signals [28, 29]. Despite being an ideal therapeutic target, the inherent cytotoxic reagents resistance of NPC-like tumor cells prompts a deeper understanding of the cancer biology of these cells.

In this study, we developed a robust prognostic model for IDH wild-type GBM through incorporating a novel parameter, G × G interaction [23, 24], to effectively identify patients at high mortality risk. In addition, comprehensive bioinformatics analysis pinpointed a subset of NPC-like cells as an essential player in driving unfavorable outcome.

Materials and methods

Data collection and sample pre-procession

The expression profile and corresponding demographic of 800 IDH wild-type GBM from 3 programs were first enrolled for the construction of COVPRIG, including TCGA RNA-seq (RNAseq, n = 143) [30], CGGA693 (RNAseq, n = 190) [31], TCGA microarray (microarray, n = 372) and GSE16011 (microarray, n = 95) [32]. Any sample missing survival information or simultaneously missing age, IDH gene mutation, and MGMT promoter (MGMTp) methylation status was excluded. The RNA-seq profiles were fragments per kilobase of exon model per million mapped fragments (FPKM) normalized, and both RNA-seq and microarray data sets were log-transformed. Three 10 × Genomics-based (GSE131928, GSE138794, GSE139448) and a Smartseq2-based (GSE131928) GBM single-cell expression profiles were employed for exploring cellular dynamics and interactions [8]. To expand the application of COVPRIG, all GBM samples (IDH wild-type and IDH mutant) included in the TCGA RNA-seq, TCGA microarray, CGGA693, GSE16011, GSE4271 [33], and GSE7696 [34] with complete survival information were included (nsample = 956) for assessing the prognostic significance of the model. All-cause death was defined as the outcome, which was curated by each program. Patients were followed for a period of at least 2 years’ post-surgery or until death. Demographics of samples were curated in Additional file 7: Tables S1 and S2. Gene signatures, including NABA core matrisome, NABA extracellular matrix (ECM) affiliated, NABA matrisome associated, Reactome ECM organization, and HALLMARK interferon alpha response was collected from the MSigDB [35] (https://www.gsea-msigdb.org/gsea/msigdb).

Gene main effects, G × G interactions selection and COVPRIG construction

Gene main effects selection

A total of 9594 common mRNAs from TCGA RNA-seq, TCGA microarray, CGGA693, and GSE16011 cohorts were included. Individual genes were enrolled into the Cox-ph model (Model 1) to determine their independent prognostic significance. Given the potential differences between RNA-seq and microarray data [36], gene main effects of significant coefficients in the TCGA RNA-seq and TCGA microarray were intersected. Then, candidates were further screened using multivariate Cox regression, with age being the covariate.

$$\mathrm{Model} \,1{:}\, h(t)={h}_{0}(t)exp({beta}_{gene i} \times {gene}_{i} +{beta}_{age} \times age)$$

G × G interactions selection

The prognostic significance of G × G interactions was determined based on Model 2. To ensure the potential significance while incorporating an appropriate number of variables, we calculated the median absolute deviation (MAD) was calculated for each gene. The top 1200 highly variable genes in TCGA RNA-seq, TCGA microarray, CGGA693, and GSE16011 were intersected, resulting in 434 common genes, which were randomly paired into 93,961 G × G interactions. Likewise, G × G interactions of prognostic significance in both the TCGA RNA-seq and microarray were intersected.

$$\mathrm{Model}\, 2{:}\, h(t)={h}_{0}(t)exp({beta}_{gene i} \times {gene}_{i} +{beta}_{gene j}\times {gene}_{j}+{beta}_{i j}\times {gene}_{i}\times {gene}_{j} +{beta}_{age} \times age)$$

To determine the optimal number of G × G interactions to be included, we traversed the model constructed from 1 to 14 random G × G interactions and calculated the p-value. When satisfactory p-values can be obtained on average, the optimal model was more likely to be constructed by including the corresponding number of G × G interactions.

Model construction

The GG score was calculated using the optimal G × G interaction combination and incorporated in the Cox-ph model with prevalent clinicopathological features (age, gender, MRMTp methylation status) to determine independent prognostic significance. COVPRIG was constructed using age and GG score, with coefficients determined based on TCGA RNA-seq cohort for RNA-seq datasets and TCGA microarray for microarray datasets. The prognostic model was constructed and validated following the TRIPOD principle.

Model evaluation and systemic literature review

The Cox regression analysis was performed to assess the discriminative ability of COVPRIG. IDH wild-type GBM samples were divided into three groups in ascending order of COVPRIG score, and all GBM samples were divided into four groups, with group 1 being the reference. The predictive ability of the model was evaluated using the receiver operating characteristic curve (ROC) and the corresponding AUC. Decision curves were employed to assess the net clinical benefit to patients.

To compare with other prognostic models, we searched two major databases (PubMed and Web of Science) using the following search strings: “((IDH wildtype) OR (IDHwt) OR (IDH wild-type)) AND ((glioblastoma) OR (gbm)) AND ((progn*) OR (survival)) AND ((predict*) OR (auc) OR (area under the curve) OR (receiver operator characteristic curve) OR (c-index) OR (c statistic) OR (roc) OR (calibration))”. All studies of the development and validation of IDH wild-type GBM prognostic models were included. Publications were restricted to being written in English and published before Aug 30, 2022. The exclusion criteria were set as 1) not histological GBM, 2) without a declaration of IDH wild-type, or with a mixture of IDH wild-type and mutation samples, 3) models also included WHO grade II and III gliomas, 4) models without overall survival as the outcome, 5) without evaluation of predictive efficacy, 6) conference abstracts, commentaries, editorials, or letters. Systematic review was conducted following the PRISMA principle.

Bioinformatics

R packages ‘edgeR’ and ‘limma’ were employed to calculate the differentially expressed genes (DEGs) [37, 38]. Gene ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analysis was performed using the online tool Metascape (https://metascape.org) [39]. The abundance of immune infiltration was deconvoluted using CIBERSORT algorithm [40]. Tumor Immune Dysfunction and Exclusion (TIDE) was employed to [35] estimate T cell function and potential sample responsiveness to immune checkpoint inhibitors (ICI) [22].

Single-cell transcriptome profiles of primary GBM were processed using the R package 'Seurat' (v4.3.0) [41, 42]. In short, genes that were expressed in less than three cells and cells that did not express over 300 genes were excluded. Expression matrices underwent independent quality control prior to integration. Batch effects were corrected by canonical correlation analysis (CCA) and mutual nearest neighbors (MNN) for the three 10 × single-cell transcriptome expression profiles [43]. Potential doublets were identified using the R package ‘DoubletFinder’ at a criteria of credible 4% [44]. Previously identified conserved tumor cell marker genes were employed to define cell identity [8]. Gene set activity was assessed using ‘AUCell’ [45]. The correlation of cell identity and COVPRIG risk-group was analyzed using ‘Scissor’ [46]. Cell–cell communication analysis was performed based on the R package ‘CellChat’ [47]. R package ‘SCENIC’ and Cytoscape software were employed for inferring and constructing the transcriptional network [45, 48].

Statistics

All statistics were based on R software (v4.2.1). Continuous variables were summarized using mean and standard deviation, and categorized variables were described by frequency and proportion. Gene main effects and G × G interactions of prognostic significance were screened using the Cox-ph model. Kaplan–Meier (K–M) analysis and log-rank test were performed to exhibit survival differences. ROC curves and corresponding AUC, C-index, and decision curves were used to assess the predictive validity of the model. Integrated discriminative improvement was calculated for comparation of models. Wilcoxon test was employed to compare non-normally distributed continuous variables. Fisher’s exact test was performed to compare the composition ratio. A two-sided p-value <  = 0.05 was considered statistically significant when no additional declaration was made.

Results

Screening of G × G interactions and construction of COVPRIG

The workflow of this study was exhibited (Fig. 1). The only gene main effect of independent prognostic significance was KIAA1671 (multi-Cox p = 0.007), which was eliminated by p-value correction (adjust p = 0.08) (Additional file 7: Tables S3 and S4). In distilling for G × G interactions, 434 highly variable genes yielded 93,961 G × G interactions. Univariate cox regression analysis yielded 13 G × G interactions of independent prognostic significance by taking the intersection of TCGA RNAseq and microarray (Additional file 7: Table S5). Given that the prognostic performance of individual G × G interaction was not superior to that of single gene main effect, the number of G × G interactions to be included was determined. We tried combinations from a single G × G interaction to a maximum of 14 G × G interaction and assessed the performance of the model. As a result, the log-rank test p-value of the model approached 0 when over 5 G × G interactions were included simultaneously (Additional file 1: Fig. S1). Therefore, a combination of 6 G × G interactions seemed to be sufficient. 1716 combinations consisting of any 6 of the 13 G × G interactions were traversed to find the optimal one. By ranking the uniCox p-values ascendingly and taking intersections in the TCGA RNA-seq and microarray, combination No. 599 was the most statistically significant, which included RIT2 × OAS1, HOXA5 × MLLT11, SLC1A2 × FAM189A2, LOXL1 × NCAPG, C21orf62 × GOLGA8A, and UBE2S × NRXN1. Based on the coefficients derived from TCGA RNA-seq and microarray respectively, these G × G interactions were assembled into the GG score using Eq. 1 (for RNA-seq) and 2 (for microarray). K–M and univariate Cox analysis disclosed that an increased GG score was suggestive of an unfavorable outcome, and ROC analysis depicted a good predictive performance (Additional file 2: Fig. S2A, B). Given the universal association between age and the prognosis of GBM (Additional file 2: Fig. S2C), the COVPRIG was constructed based on Eqs. 3 and 4 using coefficients derived from the TCGA RNA-seq and microarray, respectively.

Fig. 1
figure 1

Overview of the workflow

Full size image
$$GG Score=0.518\times RIT2+0.170 \times OAS1-0.135\times RIT2\times OAS1-0.935\times HOXA5-0.322\times MLLT11+0.178\times HOXA5\times MLLT11-0.318\times SLC1A2-0.623\times FAM189A2+0.163\times SLC1A2\times FAM189A2+0.846\times LOXL1+0.593\times NCAPG-0.253\times LOXL1\times NCAPG-0.479\times C21orf62-0.570\times GOLGA8A+0.166\times C21orf62\times GOLGA8A-0.399\times UBE2S-1.032\times NRXN1+0.254\times UBE2S\times NRXN1$$
(1)
$$GG Score=0.435\times RIT2+0.394 \times OAS1-0.056\times RIT2\times OAS1-0.734\times HOXA5-0.472\times MLLT11+0.074\times HOXA5\times MLLT11-0.483\times SLC1A2-0.558\times FAM189A2+0.076\times SLC1A2\times FAM189A2+0.480\times LOXL1+0.313\times NCAPG-0.066\times LOXL1\times NCAPG-0.479\times C21orf62-0.541\times GOLGA8A+0.064\times C21orf62\times GOLGA8A-0.682\times UBE2S-0.717\times NRXN1+0.104\times UBE2S\times NRXN1$$
(2)
$${COVPRIG}_{RNA-seq}=1.101\times GG score+0.042\times Age$$
(3)
$${COVPRIG}_{Microarray}= 0.535\times GG score+0.025\times Age$$
(4)

The prognostic significance of COVPRIG

To demonstrate the prognostic significance of COVPRIG, samples were split into high- and low-risk groups by the median value. K–M analysis found that an increased COVPRIG score was suggestive of a significantly decreased overall survival (maximum C-index = 0.642 in the TCGA RNAseq cohort) (Fig. 2A, B). The ROC curves depicted that the prediction performance of COVPRIG was satisfactory in the TCGA RNA-seq cohort, with a minimum AUC of 6-month survival over 0.7 and a maximum of 18-month being 0.79, while humbly in CGGA693 (Fig. 2C, D). For GSE16011 and TCGA microarray, the highest AUC value exceeded 0.75 (GSE16011, 15-month survival). As for discriminative ability, samples were split into three equal groups with group 1 being the reference. The hazard ratio (HR) showed a dose–response association with groups (Fig. 2E). For example, the HRgroup3vs.1 was 3.14 in the TCGA RNA-seq cohort, higher than HRgroup2vs.1 (1.63). Decision curves depicted that COVPRIG offered more net benefit than the base model containing age, gender, IDH gene mutation, and MGMTp methylation status at 9- and 15-month survival, especially for the RNA-seq datasets (Fig. 2F, G). In sum, nomograms were constructed based on RNA-seq and microarray datasets for individualized prognostic prediction (Fig. 2H).

Fig. 2
figure 2

Construction and evaluation of COVPRIG. K–M analysis of the A TCGA RNAseq (train) and CGGA693 (external validation) and B TCGA microarray (train) and GSE16011 (external validation). C ROC curves based on the TCGA RNAseq cohort. D AUCs of the 4 cohorts. E The discriminative ability of COVPRIG. Group 1 was set as the reference. Decision curves based on the F TCGA RNAseq and G TCGA microarray cohorts. H The Nomogram for RNAseq and microarray-based cohorts

Full size image

As IDH wild-type GBM accounts for the majority of GBM samples, the performance of COVPRIG in the entire GBM (including IDH mutant) was also tested. A total of 956 samples were collected from six cohorts. Univariate Cox analysis found that an increased COVPRIG score predicted unfavorable outcome (Additional file 2: Fig. S2D). In accordance with findings in IDH wild-type samples, the overall survival decreased significantly with increased quantile groups (Additional file 2: Fig. S2E, H). Further, ROC analysis showed satisfactory predictive performance of COVPRIG. The AUC values at 12-month or later were above 0.7 in GSE16011, TCGA microarray, and GSE7696 (Additional file 2: Fig. S2F, G). Therefore, COVPRIG was also applicable to GBM with unknown IDH mutation status.

Comparison of COVPRIG with existing models

To compare with previous studies, a systemic literature review was conducted. A total of 258 records were collected from the Pubmed and Web of Science. 69 duplicate and 54 irrelevant records were excluded by evaluating titles and abstracts. Through full-text assessment, editorial and conference abstracts (n = 3), a mixture with LGG or other central nervous system (CNS) tumors (n = 53), a mixture with IDH mutant GBM (n = 22), no prognostic model constructed (n = 17), without evaluation of the predictive ability (n = 26), and not overall survival as the outcome variable (n = 6) were excluded, resulting in 7 eligible studies [14,15,16,17,18,19,4H, I). These findings can be further validated by an independent Smartseq2-based single-cell RNAseq dataset (Additional file 6: Fig. S6A–C). Interestingly, Scissor algorithm found that NPC-like.1 and to a lesser extent, OPC-like.3 were most correlated with the COVPRIG high-risk group, while the opposite was true for a subset of MES-like.2 and AC-like.1/2/5 cells (Fig. 4D), in accordance with the AUCell. Together, these results highlighting the role of NPC-like.1 tumor cells in IDH-wild type GBM.

Fig. 4
figure 4

Association between NPC-like tumor cell with poor outcome. A Cell types and B marker genes of each type of cell. C The activity COVPRIG_up (DEGs with log2FC > 0.35 and p.val < 0.05, ngene = 73) and COVPRIG_down (log2FC > 0.35, p.val < 0.05, ngene = 94) in each type of cell. D Scissor algorithm assessed the correlation between COVPRIG-based risk groups and specific cell types. pos: positively correlated, neg: negatively correlated

Full size image

Two subsets of NPC-like tumor cells were identified. NPC-like.1 uniquely expressed METTL1, ATP23, and CYP27B1, and had different transcriptional regulation network from NPC-like.2 (Fig. 5A, B, Additional file 6: Fig. S6D, E). High-confidence target genes of the NPC-like.1 transcriptional factors were primarily enriched in the E2F targets (Fig. 5C), which is associated the cell cycle controlling. NPC-like cells are implicated in tumor invasion [26], and in this scenario, it was mainly monocytes/macrophages and some MES-like cells that involved in the modification of ECM. Therefore, NPC-like.1 may be involved in ECM modification by regulating the function of these cells. Cellchat found that NPC-like.1 interact with MES-like cells and monocytes/macrophages mainly in a PTN, MIF, APP and CD99-dependent manner (Fig. 5D, Additional file 6: Fig. S6F), with MIF and APP playing a role in ECM remodelling [50, 51]. On the other hand, PTPRZ1 played an essential role in receiving signaling from outside and therefore may be a potential target for regulating the function of NPC-like.1.

Fig. 5
figure 5

Features of NPC-like.1. A Genes uniquely expressed by NPC-like.1 cells and their expression density. B Transcriptional regulatory network and their high confidence target genes of NPC-like.1. The size of the node is proportional to the percentage of cells expressing the gene and the color is proportional to the average expression. C Enrichment analysis of high confidence target genes. D Signal export from NPC-like.1 cell to MES-like cells and monocytes/macrophages and important molecules. E Signal input to NPC-like.1 cell from other cells and important molecules

Full size image

Discussion

Multiple mechanisms dominate the heterogeneous prognosis of IDH wild-type GBM patients. Due to the lack of extremely effective therapy for GBM, it is worthwhile for clinicians to estimate the mortality risk and decide on available therapies to maximize the clinical benefit for individuals. Obstacles include the limited data available for study, and potential geographic and demographic disparities [52]. Herein, we constructed a prognostic model using G × G interactions that greatly increased the number of common candidates of prognostic significance between small sample datasets. The model was made robust through a multi-step screening process and separate determination of coefficients from RNA-seq and microarray data, as confirmed by systematic performance comparison. Collectively, this study provided a reference for accurately determining the prognosis of IDH wild-type GBM.

In total, six gene pairs were used to construct COVPRIG, including RIT2 and OAS1, HOXA5 and MLLT11, SLC1A2 and FAM189A2, LOXL1 and NCAPG, C21orf62 and GOLGA8A, and UBE2S and NRXN1. The biological implications of some of these genes in GBM have been explored. OAS1 is an interferon-inducible gene that encodes a protein involved in the synthesis of 2',5'-oligoadenylates and the innate immune response. OAS1 may affect the prognosis of GBM in a TRIM5-dependent manner [53]. Increased expression of HOXA5, encoding a transcription factor named homeobox genes, was associated with the tumorigenic potential of glioma stem cell [54, 55]. In addition, ubiquitin-conjugating enzyme E2S encoded by UBE2S is associated with the activity of PI3K-Akt pathway in GBM and may serve as a therapeutic target [56]. This suggests that the candidate genes derived from our screening for G × G interactions may have biological significance in IDH wild-type GBM.

The essential role of G × G interactions in prognosis was inspired by Zhang et al. [21]. They found that the expression of EGLN2 was negatively correlated with HIF1A in non-small cell lung cancer. Interestingly, HIF1A shifted from an independent risk factor to a protective factor as EGLN2 expression increased. Likewise, we found that the prognostic significance of LOXL1 and C21orf62 was associated with the expression of NCAPG and GOLGA8A (data not shown). The mathematical representation of G × G interactions include the z-score in the Cox-ph model, and the product of normalized gene expressions [22, 23]. Zhang et al. and Chen et al. constructed effective prognostic models using the product of gene expressions as a metric for G × G interactions [23, 24]. However, these studies included 613 and 505 patients as train sets, and the total sample size was around 1000, which provided a guarantee for develo** robust models. Currently, individual GBM cohorts rarely reach such a magnitude, in which poses challenges in identifying genes with shared prognostic value. Notably, G × G interaction greatly increased the number of candidates for model construction, thus, to some extent compensating for screening candidates available for more cohorts.

The performance of COVPRIG versus previously constructed models was evaluated. Some models seemed to outperform COVPRIG in terms of AUC and the C-index. However, it is difficult to avoid statistical overfitting without external validation, therefore limiting the application of some these models. In the models externally validated, COVPRIG had a 12-month AUC of 0.74, which was comparable with models 5 and 6. The highest AUC of COVPRIG occurred at 18-month (0.79), second to models 4 and 6, but still satisfactory. Notably, COVPRIG showed a 6–16% improvement relative to the externally validated and best performing model 4 after calibration to the same conditions, indicating that it should be the optimal prognostic model available.

The molecular underpinnings of COVPRIG were addressed. The dysregulated genes were mainly enriched in ECM-related signaling pathways, suggesting that COVPRIG high-risk group was characterized by ECM remodeling. Further, AUCell identified an association between ECM-associated pathways with monocytes/macrophages and some MES-like cells, in line with a previous study [57]. Interestingly, both AUCell and Scissor identified that genes up-regulated in the COVPRIG high-risk group were primarily enriched in NPC-like.1, indicating that the unfavorable outcome of the COVPRIG high-risk group was driven by NPC-like.1, and the tumor invasion associated with NPC-like cells was dependent on its interaction with cells including macrophages and MES-like cells. Further, we addressed the gene expression, transcriptional regulation, and essential cell communication molecules of NPC-like.1, which may provide a preliminary basis for subsequent studies and therapies targeting these cells.

There were several limitations. We should acknowledge that COVPRIG were not internally validated, which may lead to an under-fitting of the model. In the initial stages of exploration, tenfold cross-validation was performed instead of dividing samples into train and validation sets. Despite being able to achieve an AUC approaching 0.9 and having a satisfactory C-index in one training cohort, the excellent performance never passing on to others. This may be related to the intrinsic differences in different GBM cohorts and exploring such differences is one of the directions for future update of COVPRIG. In addition, as COVPRIG was constructed primarily with Caucasians and Asians, it should be cautiously applicated to other ethnic populations.

Conclusions

This study provided a current robust prognostic tool for GBM which was applicable to both microarray and RNA sequencing data. In addition, this study highlighted the role of a class of neuronal progenitor cells in driving poor prognosis in GBM. These results may provide basis for future research.